free space target signatures and analysis of discriminants ... · discriminants for human targets...

Free space target signatures and analysis of discriminants for human targets for the 3-D L-band through-wall experimental SAR system Chan. B. Sévigny, P. DiFilipo, D.J.

The information contained herein is proprietary to Her Majesty and is provided to the recipient on the understanding that it will be used for information and evaluation purposes only. Any commercial use including use for manufacture is prohibited. .

Defence R&D Canada – Ottawa Technical Memorandum

DRDC Ottawa TM 2012-182 November 2012

Free space target signatures and analysis of discriminants for human targets for the 3-D L-band through-wall experimental SAR system

Chan, B. Sévigny, P. DiFilippo, D.J.

The information contained herein is proprietary to Her Majesty and is provided to the recipient on the understanding that it will be used for information and evaluation purposes only. Any commercial use including use for manufacture is prohibited.

Defence R&D Canada – Ottawa Technical Memorandum DRDC Ottawa TM 2012-182 November 2012

Principal Author

B. Chan

Approved by

A. Damini

Acting Head/Radar Sensing and Exploitation Section

Approved for release by

C. MacMillan

Head/Document Review Panel

© Her Majesty the Queen in Right of Canada, as represented by the Minister of National Defence, 2012

© Sa Majesté la Reine (en droit du Canada), telle que représentée par le ministre de la Défense nationale, 2012

Original signed by

Original signed by

Original signed by

DRDC Ottawa TM 2012-182 i

Abstract ……..

DRDC Ottawa has been investigating 3-D through wall synthetic aperture radar (SAR) imaging from an experimental L-band through-wall SAR prototype. Tools and algorithms for 3-D visualization are being developed to exploit the resulting imagery. The through-wall technology and data exploitation algorithms and tools have the capability to enhance situational awareness for military forces operating in an urban environment. Current work involves analyzing signatures of human targets behind a wall and understanding the clutter and multipath signals in a room of interest. In this report, a comprehensive study of the characteristics of free-space target signatures is presented using 3-D SAR data. The aim of this investigation is to gain a better appreciation of the signatures of targets when placed behind different wall materials while identifying potential discriminants for classification of human targets. An analysis of different target signatures is provided. Six features are investigated as potential discriminants and five of them are identified as good candidates.

Résumé ….....

L’imagerie 3-D par radar à synthèse d’ouverture (SAR) à travers les murs est étudiée à RDDC Ottawa pour un banc d’essai d’un radar expérimental dans la bande L. Des outils et algorithmes pour la visualisation en 3-D sont en développement pour l’exploitation de l’imagerie qui en résulte. La technologie à travers les murs et les algorithmes et outils d’exploitation des données permettront d’améliorer considérablement la perception de la situation par les forces militaires dans l’espace de combat en milieu urbain. Les travaux en cours incluent l’analyse de la signature d’une cible humaine derrière une structure de mur et la compréhension du fouillis d’échos et des signaux dûs à la propagation par trajets multiples, visibles à l’intérieur d’une pièce d’intérêt. Dans ce rapport, une étude compréhensive des caractéristiques des signatures des cibles en espace libre est présentée utilisant des données SAR en 3-D. L’objectif de cette investigation est d’amasser une meilleure appréciation des signatures de cibles lorsqu’elles sont placées derrière différents matériaux de mur et d’identifier des discriminants potentiels pour la classification des cibles humaines. Une analyse des signatures de différentes cibles est fournie. Six traits ont été investigués comme discriminants potentiels et cinq d’entre eux ont été identifiés comme de bons candidats.

ii DRDC Ottawa TM 2012-182

This page intentionally left blank.

DRDC Ottawa TM 2012-182 iii

Executive summary


Brigitte Chan; Pascale Sévigny; David J. DiFilippo; DRDC Ottawa TM 2012-182; Defence R&D Canada – Ottawa; November 2012.

Introduction or background: DRDC Ottawa has been investigating 3-D through wall synthetic aperture radar (SAR) imaging from an experimental L-band through-wall SAR prototype. Tools and algorithms for 3-D visualization are being developed to exploit the resulting imagery. The through-wall technology and data exploitation algorithms and tools have the capability to enhance situational awareness for military forces operating in an urban environment. Current work involves analyzing signatures of human targets behind a wall and understanding the clutter and multipath signals in a room of interest. A comprehensive study of the characteristics of free-space target signatures has been conducted using 3-D SAR data. The aim of this investigation was to gain a better appreciation of signatures of targets placed behind different wall material while identifying potential discriminants for the human target for future classification.

Results: Signatures of four target types were analyzed and different characteristics of each target were presented. Six different features were analyzed for their utility in discriminating the human targets from all others in free space. The elevation coordinate at the maximum intensity value of the target, the value of the maximum intensity of the target, the 3dB width in azimuth of the no squint case, the number of range resolution cells of 3-D volumes, and the correlation between target pairs of the maximum intensity profiles all show significant potential. None of these features could be used alone to discriminate the human target from all others described in this study. A combination of at least two different features would be able to achieve this.

Significance: A capability of the through-wall SAR project desired by our military client is the ability to detect human targets behind different wall structures. The free-space study reported in this technical memorandum provides a greater appreciation for target signatures and helps to identify potential features that could be used for discrimination. This preliminary work will form the basis for analysis of target signatures behind various wall structures.

Future plans: This study will continue to search for more features and will be enlarged to include more targets in different orientations as well as more exemplars of the human target in various positions in order to create a robust strategy for human target detection and classification. Efforts will be made to choose the most effective classifier for human target discrimination. Different approaches of a one-class classifier for discrimination of the human target from all others, including density estimation, boundary methods, and reconstruction methods, will be investigated. The study will simultaneously progress to the analysis of target signatures behind various wall structures, where clutter, multipath, and wall signatures will be considered. More features will be investigated with the purpose of determining which ones would be best suited for the one class classifier.

iv DRDC Ottawa TM 2012-182

Sommaire .....


Brigitte Chan; Pascale Sévigny; David J. DiFilippo ; DRDC Ottawa TM 2012-182 R & D pour la défense Canada – Ottawa; novembre 2012.

Introduction ou contexte : L’imagerie 3-D par radar à synthèse d’ouverture (SAR) à travers les murs est étudiée à RDDC Ottawa pour un banc d’essai d’un radar expérimental dans la bande L. Des outils et algorithmes pour la visualisation en 3-D sont en développement pour l’exploitation de l’imagerie qui en résulte. La technologie à travers les murs et les algorithmes et outils d’exploitation des données permettront d’améliorer considérablement la perception de la situation par les forces militaires dans l’espace de combat en milieu urbain. Les travaux en cours incluent l’analyse de la signature d’une cible humaine derrière une structure de mur et la compréhension du fouillis d’échos et des signaux dûs à la propagation par trajets multiples, visibles à l’intérieur d’une pièce d’intérêt. Une étude compréhensive des caractéristiques des signatures des cibles en espace libre a été menée utilisant des données SAR en 3-D. L’objectif de cette investigation était d’amasser une meilleure appréciation des signatures de cibles lorsqu’elles sont placées derrière différents matériaux de mur et d’identifier des discriminants potentiels pour la classification des cibles humaines.

Résultats : Les signatures de quatre différentes cibles ont été analysées et différentes caractéristiques de chaque cible sont ont été présentées. Six traits différents ont été investigués pour leurs utilités de discriminer les cibles humaines des autres cibles en espace libre. La coordonnées en élévation à la valeur d’intensité maximum de la cible, la valeur d’intensité maximum de la cible, le ‘3dB width’ en azimut pour le cas du ‘no squint’, le nombre de cellules en distance des volumes en 3-D, et la corrélation entres des paires de cibles du profile maximum en intensité démontrent tous un potentiel significatif. Aucun des traits ne pouvait être utilisé seul pour discriminer la cible humaine des autres cibles décrites dans cette étude. Il faut une combinaison d’au moins deux différents traits pour accomplir cet objectif.

Importance : Une capacité du projet SAR à travers les murs désirée par notre client militaire est l’habilité de détecter des cibles humaines derrière différentes structures de mur. L’étude en espace libre reportée dans ce mémorandum technique a permit d’amasser une meilleure appréciation des signatures de cibles et d’identifier des discriminants potentiels pour la classification des cibles humaines. Ce travail préliminaire formera la base pour l’analyse des signatures des cibles derrière différentes structures de mur.

Perspectives : Cette étude continuera d’investiguer d’autres traits et sera agrandi pour inclure plus de cibles en différentes orientations en plus de plusieurs exemplaires des cibles humaines en positions variées pour créer une stratégie robuste pour la détection et la classification des cibles humaines. Des efforts viseront à choisir le classificateur le plus efficace pour la discrimination des cibles humaines. Différentes approches d’un classificateur d’une classe pour la discrimination de la cible humaine de toutes autres, incluant les méthodes d’estimation de densité, et les méthodes de limites et de reconstruction, seront investiguées. Simultanément, l’étude

DRDC Ottawa TM 2012-182 v

progressera à l’analyse des signatures des cibles derrière des structures de mur variées, où les fouillis d’échos, les signaux dûs à la propagation par trajets multiples, et les signatures des murs seront considérées. D’autres traits seront investigués pour déterminer lesquels seront les meilleurs pour le classificateur d’une classe.

vi DRDC Ottawa TM 2012-182


DRDC Ottawa TM 2012-182 vii

Table of contents

Abstract …….. ................................................................................................................................. i Résumé …..... ................................................................................................................................... i Executive summary ....................................................................................................................... iii Sommaire ..... .................................................................................................................................. iv

Table of contents .......................................................................................................................... vii List of figures .............................................................................................................................. viii List of tables ................................................................................................................................... x

Acknowledgements ....................................................................................................................... xi 1 Introduction ............................................................................................................................... 1

2 Processing of 3-D SAR sigal data ............................................................................................. 4

2.1 Image frame coordinate system ..................................................................................... 4

2.2 Processing kernel for 3-D SAR image formation .......................................................... 5

2.3 Radar resolution ............................................................................................................. 7

2.4 Viewing of the SAR data ............................................................................................... 7

3 Target signatures ....................................................................................................................... 9

3.1 Human target signature .................................................................................................. 9

3.1.1 Human standing ................................................................................................. 10

3.1.2 Human arms stretched ....................................................................................... 16

3.1.3 Analysis of human target signatures .................................................................. 21

3.2 Chair target signature .................................................................................................. 21

3.3 Metallic plate target signature ..................................................................................... 27

3.4 Table target signature .................................................................................................. 31

3.5 Discussion of target signatures .................................................................................... 38

4 Target features ........................................................................................................................ 39

4.1 PMI elevation coordinate ............................................................................................ 39

4.2 PMI intensity ............................................................................................................... 40

4.3 3dB width .................................................................................................................... 42

4.4 3dB width profile ......................................................................................................... 44

4.5 Maximum intensity profile .......................................................................................... 51

4.6 Number of resolution cells for 3-D volumes ............................................................... 57

4.7 Summary and discussion of results ............................................................................. 63

5 Conclusions and future research efforts .................................................................................. 65

References ..... ............................................................................................................................... 68

List of symbols/abbreviations/acronyms/initialisms .................................................................... 70

Distribution list ............................................................................................................................. 71

viii DRDC Ottawa TM 2012-182

List of figures

Figure 1: 3-D through-wall SAR system. ........................................................................................ 1

Figure 2: Photographs of the test scenes: (a) Scene 2010-1; (b) Scene 2010-11; (c) close-up of standing human target in Scene 2010-1; (d) close-up of human arms stretched target in Scene 2010-11; (e) close-up of chair target in Scene 2010-11; (f) close-up of table target in Scene 2010-11; (g) Scene 2010-12; (h) Scene 2011-68; and (i) Scene 2011-70. .............................................................................................................. 3

Figure 3: Image and building image coordinate frames for cylindrical and Cartesian coordinate systems. ....................................................................................................... 5

Figure 4: Through-wall SAR data as 2-D slice representations intersecting at a point of interest (POI) where a human target is located. ............................................................ 8

Figure 5: Human subject with arms at his side. ............................................................................. 10

Figure 6: Three 2-D slices at the PMI intersection for standing human at varying elevation for the three cases. ............................................................................................................ 11

Figure 7: Top view 2-D slices of human target for no squint at varying elevation. ...................... 12

Figure 8: Top view 2-D slices of human target for leading squint at varying elevation. .............. 13

Figure 9: Top view 2-D slices of human target for trailing squint at varying elevation. .............. 14

Figure 10: Side view 2-D slice of human target for no squint. ...................................................... 15

Figure 11: Human subject with arms stretched. ............................................................................ 16

Figure 12: Three 2-D slices at the PMI intersection for human arms stretched at varying elevation for the three cases. ....................................................................................... 17

Figure 13: Top view 2-D slices of human target with arms stretched for no squint at varying elevation. ..................................................................................................................... 18

Figure 14: Top view 2-D slices of human target with arms stretched for leading squint at varying elevation. ........................................................................................................ 19

Figure 15: Top view 2-D slices of human target with arms stretched for trailing squint at varying elevation. ........................................................................................................ 20

Figure 16: Chair. ............................................................................................................................ 21

Figure 17: Chair dimensions. ........................................................................................................ 22

Figure 18: Top view 2-D slices of chair target for no squint at varying elevation. ....................... 24

Figure 19: Top view 2-D slices of chair target for leading squint at varying elevation. ............... 24

Figure 20: Top view 2-D slices of chair target for trailing squint at varying elevation. ............... 25

Figure 21: Metallic plate. .............................................................................................................. 27

Figure 22: Top view 2-D slices of metallic plate for no squint at varying elevation. ................... 28

Figure 23: Top view 2-D slices of metallic plate target for leading squint at varying elevation. . 29

DRDC Ottawa TM 2012-182 ix

Figure 24: Top view 2-D slices of metallic plate target for trailing squint at varying elevation. .. 30

Figure 25: Table. ........................................................................................................................... 32

Figure 26: Table dimension measurements. .................................................................................. 33

Figure 27: Top view 2-D slices of table target for no squint at varying elevation. ....................... 34

Figure 28: Top view 2-D slices of table target no leading squint at varying elevation. ................ 35

Figure 29: Top view 2-D slices of table target for trailing squint at varying elevation. ................ 36

Figure 30: Box plots of PMI elevation coordinate for human, table, chair, and metallic plate targets. ......................................................................................................................... 40

Figure 31: Box plots of PMI intensities of targets for no squint case. .......................................... 41

Figure 32: Box plots of PMI intensities of targets for leading squint. .......................................... 41

Figure 33: Box plots of PMI intensities of targets for trailing squint. ........................................... 42

Figure 34: Box plots of 3dB width in azimuth of human (H), table (T), chair (C), and metallic plate (P) targets for no squint (NS), leading squint (LS), and trailing squint (TS): (a) original plot, and (b) close-up of plot without TNS and PNS in the figure. .......... 43

Figure 35: Box plots of 3dB width in range of human (H), table (T), chair (C), and metallic plate (P) targets for no squint (NS), leading squint (LS), and trailing squint (TS). .... 44

Figure 36: Azimuth 3dB widths at varying range positions for all targets. ................................... 47

Figure 37: Correlation matrix of azimuth 3dB widths at varying range positions for all targets. . 48

Figure 38: Range 3dB widths at varying azimuth positions for all targets.................................... 49

Figure 39: Correlation matrix of range 3dB widths at varying azimuth positions for all targets. . 50

Figure 40: Maximum intensity in azimuth at varying range positions for all targets. ................... 53

Figure 41: Correlation matrix of maximum intensity in azimuth at varying range positions for all targets. .................................................................................................................... 54

Figure 42: Maximum intensity in range at varying azimuth positions for all targets. ................... 55

Figure 43: Correlation matrix of maximum intensity in range at varying azimuth positions for all targets. .................................................................................................................... 57

Figure 44: Human target volumes: (a) 3dB of human standing, (b) 3dB of human arms stretched, (c) 6dB of human standing, and (d) 6dB of human arms stretched. ........... 58

Figure 45: Segmented 3dB volumes of all targets. ........................................................................ 59

Figure 46: Human standing 3-D volume at 12dB threshold. ......................................................... 60

Figure 47: Human arms stretched 3-D volume at 24dB threshold. ............................................... 61

Figure 48: Chair 3-D volume at 15dB threshold. .......................................................................... 61

Figure 49: Table 3-D volume at 14dB threshold. .......................................................................... 62

Figure 50: Metallic plate 3-D volume at 18dB threshold. ............................................................. 62

x DRDC Ottawa TM 2012-182

List of tables

Table 1: June 2010 and October 2011 trial targets. ......................................................................... 2

Table 2: Parameters from processing parameter file. ...................................................................... 6

Table 3: Parameter configurations for analysis of target signatures. ............................................... 9

Table 4: Measured dimensions of metallic chair features. ............................................................ 26

Table 5: Target symbols ................................................................................................................ 45

Table 6: Number of resolution cells for 3-D volumes at 24dB threshold...................................... 63

DRDC Ottawa TM 2012-182 xi

Acknowledgements

We are grateful to Janice Lang, official DRDC Ottawa photographer, for all photographs in this technical memorandum.

xii DRDC Ottawa TM 2012-182


DRDC Ottawa TM 2012-182 1

1 Introduction

Development of a prototype 3-D through-wall synthetic aperture radar (SAR) system is currently underway. The intent is to map out building wall layouts and to detect targets of interest behind walls such as humans, arms caches, and furniture such as chairs and tables. This situational awareness capability can be invaluable to the military working in an urban environment.

An experimental high resolution L-band through-wall SAR radar testbed has been developed by DRDC Ottawa [1]. The side-looking radar is truck-mounted with data collected as the vehicle is driven past the front of a building of interest. The system operates over a 1.86 GHz bandwidth between 0.85 and 2.71 GHz. It has two transmit and eight receive antennas, all of which consist of compact printed bowtie antenna elements [2]. The transmit antennas are vertically spaced 1.2 m apart and have a toggling capability that increases the resolution in elevation [3] by a factor of two for this configuration. The receive antennas are organized in a vertical array with a spacing of 15cm. The radar control hardware and data acquisition system are located in the vehicle. The radar system is shown in Figure 1. It is complemented by a light detection and ranging system (LIDAR) and a positioning system, both truck-mounted as well, as seen in this figure. The LIDAR is a Riegl LMS Q240i 2-D laser scanner [4] which provides high resolution images of the exterior and part of the interior of a building. It adds another layer of information to help with the interpretation of the SAR data and could potentially be used in future algorithm development for wall compensation. The positioning system is an Applanix POS-LV 420 [5] used for motion compensation as well as to provide an accurate point of reference. It consists of a positioning computer system, two global positioning system (GPS) antennas, an inertial measurement unit (IMU) and a distance measurement indicator (DMI).

A small portable LIDAR sensor, the Velodyne HDL-32E [6], is used inside a building to create a 3-D map of the interior of a room. It provides ground truth for analyzing the detection capabilities of the SAR data.

Figure 1: 3-D through-wall SAR system.

Riegl LIDAR and motion compensation unit

transmit antennas

receive antenna

array

2 DRDC Ottawa TM 2012-182

The SAR and LIDAR data are collected while the truck moves along the SAR path parallel to the building of interest. It is typically located approximately 10 meters from the front wall if the road structure allows. The azimuth direction is parallel to the SAR path while the range direction is perpendicular to it. Offline, following acquisition, the raw SAR data are pre-processed for proper formatting and motion compensation, pulse-compressed in range, and then a backprojection algorithm is applied to form the SAR image [6]. A comprehensive study is currently underway to examine the characteristics of free-space target signatures using 3-D SAR data from this system. The aim is to gain a better appreciation of the signatures of targets which can later be placed behind different wall materials and to identify features for discrimination of the human target from other targets. 3-D free space data acquired during two sets of trials were used to study these characteristics. Three scenes with four different types of targets (standing human, human with arms stretched, table, and chair) were examined from the June 2010 Trials (Scenes 2010-1, 11, and 12). Two scenes with two different types of targets (human with arms stretched and metallic plate) were used from the October 2011 Trials (Scenes 2011-68 and 70). A summary of all scenes and targets is provided in Table 1. All range values listed in the target description column are given with reference to the truck-mounted radar system at closest approach. Photographs of the imaged scenes and targets are presented in Figure 2.

Table 1: June 2010 and October 2011 trial targets.

Scene Identifier

Scene Filename Target Description and Location with respect to the testbed

2010-1 TDfreeMedSP01 Human at 10m in range (H1) 2010-11 TDfreeTCPH04 Table at 10m in range

Chair at 10m in range Human with arms out at 10m in range (H6)

2010-12 TDfreeHGroup04 4 humans in a rectangular figuration spaced 1.5m apart in range, 3.1m apart in azimuth at the front and 0.8m apart in azimuth at the back at 10m and 12m in range (H2, H3, H4, H5)

2011-68 ts_setupb_02 Human arms out at 10m in range (H7) 2011-70 ts_setupc_01 Human arms out at 5m in range (H8)

Metal plate at 5m in range Descriptions of the image frame coordinate system, the processing kernel used for 3-D SAR image formation, and the SAR data viewing platform are provided in Section 2.0. In Section 3.0, signatures of the human, chair, table, and metallic plate are presented. In Section 4.0, discriminants for human target detection and future feature classification schemes are identified and an analysis of the utility of the target signature discriminants is given. Conclusions and future research are presented in Section 5.0.


(a) (b)

(c) (d) (e) (f)

(g) (h)

(i)

Figure 2: Photographs of the test scenes: (a) Scene 2010-1; (b) Scene 2010-11; (c) close-up of standing human target in Scene 2010-1; (d) close-up of human arms stretched target in

Scene 2010-11; (e) close-up of chair target in Scene 2010-11; (f) close-up of table target in Scene 2010-11; (g) Scene 2010-12; (h) Scene 2011-68; and (i) Scene 2011-70.


2 Processing of 3-D SAR sigal data

In this section, the processing of the 3-D SAR signal data is described. The image frame coordinate system is presented in Section 2.1. A summary of the processing kernel is provided in Section 2.2. A summary of the expected radar resolution is given in Section 2.3. In Section 2.4, the process used to manually detect target locations is described.

2.1 Image frame coordinate system The five coordinate frames used for calculations in the SAR image formation algorithm were presented in [7]. The image frame used which has been updated since that time is described here. The image frame is the operator-defined coordinate system that establishes the grid points at which the image intensities are to be determined. For example, in a scene that contains a building and targets in free space, the operator can define which part of the signal data containing only one target in free space is to be processed and viewed. This avoids processing parts of the scene not needed to conduct an investigation and thus decreases the computing time. In the processing kernel, the operator can define the image frame as either a cylindrical or a Cartesian coordinate system. Typically, when the operator chooses a Cartesian coordinate system (𝑋𝑐𝑎𝑟𝑡, 𝑌𝑐𝑎𝑟𝑡, 𝑍𝑐𝑎𝑟𝑡), the origin of the image frame is fixed at the location of the IMU reference point at the beginning of the SAR path. When defining the image grid points, the operator is required to specify the start and end points of the desired processed data in azimuth, range, and elevation as values in meters along with the grid spacing (desired distance between grid points in meters). Due to the orientation of the coordinate frame axes, azimuth values are positive, range values are negative, and elevation values could either be positive or negative depending on whether the desired grid points are above or below the IMU. When the operator chooses a cylindrical coordinate system (𝑋𝑐𝑦𝑙, 𝑌𝑐𝑦𝑙, 𝑍𝑐𝑦𝑙), the origin of the image frame is translated from the Cartesian coordinate system origin, along the y-axis to the point where the middle of the receive array is located and along the z-axis to the centre of the array. When defining the grid points, the operator is required to specify the start and end points in azimuth and range as positive and negative values in meters, respectively, in the same manner as for the Cartesian coordinate system. The elevation start and end grid points, as well as the grid spacing in elevation, are specified as sin𝜃 values (unitless), where 𝜃 is defined as the elevation angle above or below the horizontal plane, or the direction of the boresight. During the October 2011 trials, the origin of the image frame was redefined to be located at the top right corner of a building (𝑋𝑏𝑢𝑖𝑙, 𝑌𝑏𝑢𝑖𝑙, 𝑍𝑏𝑢𝑖𝑙) to ensure that the same origin could be used for all scenes for that trial. This new coordinate frame is called the “building image frame”. This was carried out for processing of the data of and near three of the buildings, namely Troop Shelter, Brouillette, and Grenade Range. The targets used from the October 2011 trails in this report were located next to the Troop Shelter building. Use of this coordinate frame changes the definition of grid points by the operator from that used previously. In the case of the Cartesian coordinate system, the operator is required to specify the start and end points as well as the grid spacing values in the same manner as previously defined while taking into account the new


location of the origin and the orientation of the axes. In the case of the cylindrical coordinate system, care must be taken when specifying grid point values since they are now obtained from two different coordinate frames. The azimuth grid points are the same as those defined in the building image frame while the range and elevation grid points are the same as those defined in the original image frame for a cylindrical coordinate system (𝑋𝑏𝑢𝑖𝑙, 𝑌𝑐𝑦𝑙, 𝑍𝑐𝑦𝑙 ). Image andbuilding image coordinate frames are presented in Figure 3.

Figure 3: Image and building image coordinate frames for cylindrical and Cartesian coordinate systems.

2.2 Processing kernel for 3-D SAR image formation The processing kernel to form the 3-D SAR image from the backscattered data was created in MATLAB. It is divided into three main steps. In the first step, pre-processing of the raw data is carried out; in the second step, Fast Fourier Transform (FFT) based pulse compression in range is performed; and in the third step, SAR processing of the pulse compressed data using backprojection is used to form the 3-D SAR image. The processing kernel including its input parameters is documented in detail in [7] and [8].

In the second step, the operator can currently specify one of four weighting windows to be used prior to the FFT. They include uniform, Hamming, Hanning, and Chebychev windows; however, the Hamming window has since been discarded as a useful option due to extra undesirable sidelobes introduced by the discontinuity created after zero-padding. Hanning windowing is currently recommended for pulse compression. In the third step, there are many parameters that can currently be specified by the operator. These are provided in Table 2.

IMU

SAR path

truck at start of SAR path

antenna array

θ

image point

boresight

building of interest


Table 2: Parameters from processing parameter file.

Processing Parameters Possible inputs Channels to process 1 through 8

Remove toggle option 0 to allow toggling 1 to remove toggle

cylindrical coordinate system Image grid origin Image grid end

Image grid spacing

[Azimuth (m) Range (m) sin𝜃] [Azimuth (m) Range (m) sin𝜃]

[∆Azimuth (m) ∆Range (m) ∆sin𝜃] Cartesian coordinate system

Image grid origin Image grid end

Image grid spacing

[Azimuth (m) Range (m) range∙sin𝜃(m)] [Azimuth (m) Range (m) range∙sin𝜃(m)]

[∆Azimuth (m) ∆Range (m) ∆range∙sin𝜃(m)] Image grid title User defined descriptive title Image grid type 1 for cylindrical coordinate system

2 for Cartesian coordinate system Process flag PC 0 for uniform window during pulse compression

1 for Hamming window during pulse compression 2 for Chebychev window during pulse compression

3 for Hanning window during pulse compression Process flag el 0 for uniform window during coherent integration

1 for Hamming window during coherent integration Process flag az 0 for uniform window during coherent integration

1 for Hamming window during coherent integration Process flag azBeamwidth [Beamwidth angle 1 Beamwidth angle 2] (rad)

[-25° +25°] for a 50° azimuth beamwidth [-60° -10°] for a leading squinted SAR geometry [+10° +60°] for a trailing squinted SAR geometry

The backprojection algorithm is used to carry out azimuth and elevation compression by performing time domain coherent integration of the signals from all of the receive channels for all transmit elements as well as all transmit locations in the synthetic aperture. The intensity is obtained at each specified image point. Hamming windowing is currently recommended to weight the input for both azimuth and elevation compression. Elevation compression is accomplished by coherent summation over the receive channels and transmit elements. Azimuth compression is accomplished by adding the contributions from the different transmit locations used to form the synthetic aperture. Limiting the processed beamwidth is a way to reduce processing time as it reduces the size of the summation during azimuth compression. The backprojection algorithm includes the capability to compensate for wall effects although this implementation has yet to be tested. The outputs of the SAR processing routine include 3-D SAR image data, processing parameters used during the processing, a 2-D top view slice of the 3-D SAR image, and a SAR processing log file.


2.3 Radar resolution Radar resolution is conventionally defined as the 3dB width of the point spread function (PSF). The expected radar resolution of the SAR image was calculated based on the through-wall SAR system parameters. The formulas for elevation, range, and azimuth resolution were derived in [8] for free-space propagation of the radar signal. The through-wall SAR system parameters required for the calculations include the radar bandwidth of 1.86GHz, the centre frequency of 1.78GHz, the element spacing of 0.15m, and the number of receive elements for the antenna of 8 (the value effectively doubles to 16 for the toggled case).

Based on the results of the analysis conducted by varying radar parameters and weighting windows, all SAR data were processed using the toggling option, with Hamming windowing used for the azimuth and elevation compressions, Hanning windowing used during pulse compression, and an azimuth beamwidth of 50°. The formulas to calculate resolutions for free-space propagation of the radar signal were then used for the through-wall 3-D SAR radar system. The resolutions are expected to be 0.12m in range, 0.11m in azimuth, and 5.2𝜊 in elevation.

2.4 Viewing of the SAR data When attempting to detect targets of interest, the 3-D SAR data are typically displayed in three 2-D slices that intersect at a pixel of interest (POI): one slice displays range vs azimuth (top view), another slice displays elevation vs azimuth (front view), and the last slice displays elevation vs range (side view). An example is provided in Figure 4 where the white lines show the locations of the slices which intersect at the POI. Along the azimuth axis, values are displayed in reverse to emulate the SAR path taken by the vehicle during data collection. Along the elevation axis, the value of 0.0m is at ground level. Values below ground level appearing in the image represent the bottom part of the “curvature” created as a result of the point spread function.

An image enhancement scaling factor directly related to the POI intensity value is applied to the linearly scaled power intensity data. Low power intensities of the data below a specified threshold are mapped to the first color of the colormap while high power intensities of the data above another specified threshold are mapped to the last color in the colormap. This has the effect of expanding the mid-range “color resolution”. Threshold values were chosen to ensure that sufficient contrast exists between the target signal and the clutter. The lower threshold was set at 10 percent of the intensity at the POI while the upper threshold was set at 110 percent of the intensity at the POI.

Since the scaling is based on the values within the displayed part of the data, the choice of pixel of interest can greatly affect the image displayed. If the pixel is on-target, that is, it is at the location where a target has the largest backscatter, it is highly visible and all surrounding clutter and noise is minimal. When the pixel of interest is off-target, the clutter and noise dominate the whole image due to saturation. Viewing 2-D slices at different POIs provides a mechanism to manually detect the locations of targets of interest. When the POI is on-target at the highest intensity value of that target, the POI at that value becomes the pixel at maximum intensity (PMI).


Figure 4: Through-wall SAR data as 2-D slice representations intersecting at a point of interest

(POI) where a human target is located.


3 Target signatures

Transmit and receive antenna elements of the SAR through-wall radar have a very broad beamwidth in azimuth up to 90°. The radar operator can take advantage of the wide beamwidth and use a squinted SAR geometry which makes it possible to eliminate or minimize specular signal returns parallel to the SAR path, such as those due to a front wall while at the same time emphasizing features perpendicular to the SAR path, such as side walls. This is carried out by processing of the signal data to effectively illuminate at an angle so that some of the microwaves incident on an object which is oriented parallel to the track of the vehicle are scattered away from the radar return path. Studying signatures of targets at varying squint angles can be used to investigate if they differ significantly with and without squint.

For each scene presented in Table 1 in Section 1, the SAR data were processed using three different parameter configurations: no squint, leading squint, and trailing squint. A summary of these configurations is provided in Table 3.

Table 3: Parameter configurations for analysis of target signatures.

No squint Leading squint Trailing squint Beamwidth -25 to +25 degrees +10 to +60 degrees +60 to -10 degrees

Weighting Azimuth compression Hamming Hamming Hamming Elevation compression Hamming Hamming Hamming

Pulse compression Hanning uniform uniform

Using the Cartesian coordinate system, the grid spacing was specified at 3cm in both azimuth and range and 10cm in elevation. For each scene, all targets were manually detected and the corresponding PMIs were identified. A close look at signatures of the targets is presented in this section: a human target in Section 3.1, a chair in Section 3.2, a metallic plate in Section 3.3, and a table in Section 3.4. A discussion of target signatures is provided in Section 3.5.

3.1 Human target signature In this section, the signature of a human target in free-space is examined for two body positions. In the first case, the human is standing with arms at his side; in the second the human is standing with both arms stretched parallel to the ground. Several human targets were studied; however, results for only one are discussed here as this was the only example available where the same human target was studied in both positions. The results for the standing position are presented in Section 3.1.1 and for the arms stretched position in Section 3.1.2. An analysis of the results is given in Section 3.1.3.

According to simulation studies of the signatures of the human body in free space conducted by the United States Army Research Laboratory [9], the main contribution to the radar return comes from the human torso. That research shows that the relative position of the limbs does not modify the average return of the radar cross-section by a significant amount. When investigating the human signature in SAR imagery in 2-D, they demonstrate that the brightest spot corresponds to the front torso of the human. They conclude that arms and shoulders contribute to bright spots on


the sides of the torso, that the legs contribute to a “late-time” return caused by the effect of double bouncing of the incident wave from one leg to the other before transmission to the receive antenna, and there is significant backscatter appearing to come from behind the main return corresponding to the torso. In the next two subsections, verification of the results of those simulations using 3-D SAR images of human targets in free space is conducted, with the added advantage of examining how the backscatter varies with elevation.

3.1.1 Human standing The human subject shown in Figure 5 is 1.7m tall, of average height and build. Three 2-D slices that intersect at the PMI for this human target are shown in Figure 6 for all three cases (no squint, leading squint, and trailing squint). In each case, one slice displays range vs azimuth (top view), another slice displays elevation vs azimuth (front view), and the last slice displays elevation vs range (side view). The data were normalized to ensure scaling remained the same for all cases. Although it might be intuitive to consider that the head represents the strongest point return, that is not the case since the PMI elevation coordinate is located at the subject’s lower sternum.

Figure 5: Human subject with arms at his side.

The PMI elevation coordinates corresponding to the subject’s lower sternum were found to be 1.2m for the no squint and leading squint and 1.0m for trailing squint, slightly lower by 20cm. The location of the PMI azimuth and range coordinates changed very little between the three cases – with a maximum variation of 20cm. This is above the azimuth and range resolutions but can be accounted for by considering the human physique and the geometry of the squint angle. In this case, squint SAR has the same effect as pointing the radar 35° from boresight in azimuth. Consequently, more of the side of the human is illuminated than the front. Since it is believed that the radar backscatter from a human is from the skin [9], it follows that the PMI is simply shifted from the front of the lower sternum to the side of the lower ribs, by at most 20cm in azimuth and 10cm in range. 3-D SAR data are displayed as progressive top view 2-D slices at varying elevation in Figure 7, Figure 8, and Figure 9 for no squint, leading squint, and trailing squint, respectively, for elevations between -0.30m and 2.20m, at steps of 50cm (half the elevation resolution at a distance of 10m from the radar).


No squint Leading squint

Trailing squint

Figure 6: Three 2-D slices at the PMI intersection for standing human at varying elevation for the three cases.


-0.30m 0.2m

0.70m 1.2m

1.70m 2.2m

Figure 7: Top view 2-D slices of human target for no squint at varying elevation.


-0.30m 0.2m

0.70m 1.2m

1.70m 2.2m

Figure 8: Top view 2-D slices of human target for leading squint at varying elevation.


-0.30m 0.2m

0.70m 1.2m

1.70m 2.2m

Figure 9: Top view 2-D slices of human target for trailing squint at varying elevation.


The strongest returns (0dB) for the no squint case occur at elevations from 0.7m to 1.7m. This 1m variation in elevation could be expected as this is the elevation resolution at a range of 10m, the location of the human target. At these elevations, point scatterers located where the shoulder features are expected are present and 10dB lower than the main return from the torso. A widening of the main feature of 30cm in azimuth is observed between elevations of 0.7m and 2.2m compared to the main feature observed at elevations of 0.2m and below. A significant return behind that from the main return of the torso, as was reported in the simulations of the human signature in [9], is visible between elevations of 0.7m and 2.2m. Significant returns are also observed at the level of the ground (0.0m) where the ground and feet/legs intersect to form a corner. Returns from ground clutter at -30dB to -40dB are prominent at lower elevations and are no longer visible at an elevation of 0.7m and higher. The side view 2-D slice of the 3-D SAR data in Figure 10 makes it possible to view the difference in elevation between the ground and leg interaction centered at an elevation of 0.0m and the torso return centered at 1.2m.

Figure 10: Side view 2-D slice of human target for no squint.

In the case of leading squint, the human signature is now viewed at 35° from the normal to the front of the target. The strongest return is visible between 0.2m and 1.7m in elevation. It is no longer rounded but rather elongated in the azimuth-range plane with a length of 55cm. It appears as though the returns from the human torso and the strong returns from the shoulders are combined into one feature at a peak normalized intensity of 0dB. The bright features seen behind the main torso appear to be 10dB lower than the main peak and more elongated than what was seen with no squint.

In the case of trailing squint, the human signature appears at -35° from the normal to the front of the target. The strongest return appears between 0.2m and 1.7m in elevation. The same observations as for the leading squinted SAR geometry can be made here, with the strongest return elongated in the azimuth-range plane with a length of 50cm.

In both the leading and trailing squint SAR cases, a somewhat curved “smearing effect” occurs for the target’s bright features. This is currently unexplained and needs to be investigated further.


3.1.2 Human arms stretched A picture of the same human subject with arms stretched is shown in Figure 11. The distance between finger tips was measured to be 1.5m at a height of 1.4m from the ground. Three 2-D slices that intersect at the PMI for the human target are shown in Figure 12 for each of the three cases (no squint, leading squint, and trailing squint) showing the top view, front view, and side view. The data were normalized to ensure scaling remained the same for the different cases.

Figure 11: Human subject with arms stretched.

The PMI elevation coordinates corresponding to a height slightly lower than the subject’s lower sternum were found to be 0.8m for no squint, 0.9m for leading squint, and 1.0m for trailing squint. These values are within 20cm of one another. The PMI elevation coordinate is lower than that found for the human standing position, by 40cm for the no squint case, 30cm for the leading squint case, and the same for the trailing squint case. The location of the PMI in azimuth and range varies by 30cm in azimuth and 10cm in range. This can once again be accounted for by considering the human physique and the geometry of the squint angle. 3-D SAR data are displayed as progressive top view 2-D slices in Figure 13, Figure 14, and Figure 15 for no squint, leading squint, and trailing squint, respectively. In the images for the no squint case, the torso and the extended arms of the human are readily discernible. The strongest point returns (0dB) occur at elevations between 0.4m and 1.4m, where a 1m variation in elevation could be expected due to the resolution in elevation. At elevation levels between 0.4m and 1.4m, point scatterers formed by the extended arms are evident, appearing 10dB to 20dB lower in intensity with a width of 2.0m, 50cm wider than expected. For some unexplained reason, there seems to be a strong return to the left of the human. It gives the impression that the left arm is more dominant in the image than the right arm. This rather large feature seems to dominate at all levels of elevation indicating that it does not form part of the arms of the human. In fact, it might be masking the left arm and thus could also explain the wider than expected “wing span” of the human arms. It is possible that the watch being worn on that arm could cause this strong signal return but more investigation would be required to identify the source of this feature. Significant returns are once again seen behind the main torso and at the ground level where the ground and feet/legs form a corner.


No squint Leading squint

Trailing squint

Figure 12: Three 2-D slices at the PMI intersection for human arms stretched at varying elevation for the three cases.


-0.10m 0.4m

0.90m 1.4m

1.90m 2.4m

Figure 13: Top view 2-D slices of human target with arms stretched for no squint at varying elevation.


-0.10m 0.4m

0.90m 1.4m

1.90m 2.4m

Figure 14: Top view 2-D slices of human target with arms stretched for leading squint at varying elevation.


-0.10m 0.4m

0.90m 1.4m

1.90m 2.4m

Figure 15: Top view 2-D slices of human target with arms stretched for trailing squint at varying elevation.


In the case of leading squint, the human signature is now viewed at 35° from the normal to the target. Between elevations of -0.1m and 1.9m, one strong point return of 0dB is observed to the right of the human. It appears to be due to the same unexplained reason as for the no squint images of Figure 13. Between elevations of 0.4m and 1.9m, a second strong return of 0dB is observed which has more of a rounded shape in the azimuth-range plane than expected. A third strong return 10db lower than the main feature appears between elevations of 0.9m and 1.9m. It is interesting to note that the two extra point sources appear further back than the torso as opposed to next to it suggesting that those returns do not come from part of the arms. In fact, the arms are not expected to be at as high a relative intensity in a squinted SAR geometry as in the no squint case due to the scattering of the radar signal away from the radar return path in azimuth. They do appear at the correct location but are 20dB to 30dB lower than the strong return from the torso.

For the trailing squint case, the human signature is now viewed at -35° from the normal to the target. Only one strong return is observed between elevations 0.4m and 1.9m. Returns from the extended arms are again 20dB to 30dB lower than the strong return from the torso.

3.1.3 Analysis of human target signatures The difference in the human signature between the standing position and the arms stretched position is easily seen in the progressive 2-D slice views. The arms appear at an intensity 10dB below that at the PMI elevation coordinate for imaging with no squint whereas they appear at nearly 20 to 30dB below for imaging with both leading and trailing squint, de-emphasizing the arm returns. The difference in the human signature between no squint and squint is also seen in the progressive 2-D slice views. The signature for the human standing position for no squint is very similar to the signature for leading and trailing squint, with only a slight shift in the PMI azimuth coordinate. The signature of the human arms stretched position for no squint is different than the signature for leading and trailing squint, where returns of the extended arms are 10dB to 20dB lower for leading and trailing squint than for no squint.

3.2 Chair target signature The signature of a chair in free-space is presented in this sub-section. Only one type of chair, a picture of which is shown in Figure 16, was studied. Chair dimensions are presented in Figure 17. The chair is composed of three materials: metal, cloth, and plastic. The chair legs which support the armrest and back, as well as the cushion support brackets underneath the chair, are metallic. The armrests are hard plastic. The seat cushion and chair back consist of a cloth material covering foam and corkboard. Only the metallic pieces are expected to have strong SAR returns. Dimensions of interest for this target are summarized in Table 4.

Figure 16: Chair.

side

front back

bottom

met

hard

cushio


3-D SAR data were displayed as progressive top view 2-D slices in Figure 18, Figure 19, and Figure 20 for no squint, leading squint, and trailing squint, respectively. The data were normalized to ensure scaling remained the same for the different cases. The dimensions of interest as measured from the SAR image are summarized in Table 4.

-0.50m 0.00m

0.50m 1.00m

1.50m 2.00m

Figure 18: Top view 2-D slices of chair target for leading squint at varying elevation


-0.50m 0.00m

0.50m 1.00m

1.50m 2.00m

Figure 19: Top view 2-D slices of chair target for leading squint at varying elevation.


-0.50m 0.00m

0.50m 1.00m

1.50m 2.00m

Figure 20: Top view 2-D slices of chair target for trailing squint at varying elevation.


Table 4: Measured dimensions of metallic chair features.

Dimensions of interest

Physical measurement

(m)

SAR image measurement

no squint (m)

SAR image measurement

leading squint (m)

SAR image measurement trailing squint

(m) Length of front metal legs

0.54 -0.30 and -0.40 -0.30 and -0.3 -0.40 and -0.40

Length of back metal legs

0.60 -0.20 and -0.20 -0.20 and -0.40 -0.40 and -0.50

Height of cushion support

0.35 0.0 N/A N/A

Distance between both front legs

0.58 0.51 0.54 0.51

Distance between both back legs

0.41 0.39 0.45 0.48

Distance between front and back legs

0.49 0.45 0.39 and 0.42 0.39 and 0.45

Five strong point returns are identifiable in the SAR images for the no squint case. Four of the strong returns correspond to each of the four metal legs of the chair, as expected. The PMI elevation coordinates for each chair leg are -0.3m, -0.2m, -0.2m, and -0.4m, suggesting that the strongest return corresponds to the 90° angle junction created between the ground and each leg. There is very strong agreement between physical measurements and SAR image measurements for each leg, varying by 7cm, well within the resolution in both azimuth and range. The fifth return is the strongest and appears slightly higher in elevation than the other four returns in the SAR image. It is centrally located between the front and the back legs in the azimuth direction but is positioned at the same range as the back legs. When looking at the metal frame supporting the cushion, it becomes clear that a wide angled metallic corner is formed at the junction between each half of the bracket. This corner is what is creating the fifth large return in the SAR data. It is approximately 30cm higher in range than its physical location because the microwave signal bounces between the two sides of the bracket, resulting in a delay before returning to the radar. In the case of leading squint, only four strong target returns are observed, corresponding to each of the four metal chair legs. The fifth return seen in the no squint case does not appear in the image due to the scattering of the microwave signal away from the radar. As expected, the strong returns of the chair signature appear at an angle 35° from the normal to the target. The dimensions of interest as measured from the SAR image are summarized in Table 4. The PMI elevation coordinate for each leg are -0.3m, -0.4m, -0.2m, and -0.3 m, suggesting that the strongest return corresponds to the 90° angle junction between the ground and each leg. Again, there is very strong agreement between physical measurements and SAR image measurements for each leg, with distances between them varying by 10cm, which is within the resolution in both azimuth and range. Due to the squinted SAR geometry, the four strong point returns have been shifted to the right in azimuth by 20cm. In the case of trailing squint, only three of the strong returns, representing three of the four metal chair legs, are observed. The back left leg does not appears as a strong return as for the no squint


case but it can still be seen. The maximum intensity for this chair leg return was 10dB lower than for the other legs; unfortunately, that is also the intensity level observed for the somewhat curved “smearing effect” observed for leading and trailing squint. The results of both the leading and trailing squint should be similar, showing all four strong returns for the chair legs. It is unexplained at this time why the fourth leg did not appear brighter. The fifth return seen in the no squint case is again absent. As expected, the strong returns of the chair signature appear at an angle -35° from the normal to the target. The dimensions of interest were measured in the SAR image and are summarized in Table 8. The PMI elevation coordinate for each leg are -0.4m, -0.5m, -0.4m, and -0.4 m, suggesting that the strongest return corresponds to the 90° angle junction created between the ground and each leg. Again, there is very strong agreement between physical dimensions and SAR image measurements for each leg, with distances varying by 10cm, which is within the resolution in both azimuth and range. Due to the squinted SAR geometry, the three strong point returns have been shifted to the left in azimuth by 15cm.

3.3 Metallic plate target signature The signature of a flat metallic plate in free-space, shown in Figure 21, is presented in this sub-section. The plate was measured to be 1.8m wide, 1.2m high, and 3cm thick. It is supported by plastic tubes at each end.

Figure 21: Metallic plate.

3-D SAR data are displayed as progressive top view 2-D slices in Figure 22, Figure 23, and Figure 24 for no squint, leading squint, and trailing squint, respectively. The data were normalized to ensure scaling remained the same for the different cases.


-0.50m 0.00m

0.50m 1.00m

1.50m 2.00m

Figure 22: Top view 2-D slices of metallic plate for no squint at varying elevation.


-0.50m 0.00m

0.50m 1.00m

1.50m 2.00m

Figure 23: Top view 2-D slices of metallic plate target for leading squint at varying elevation.


-0.50m 0.00m

0.50m 1.00m

1.50m 2.00m

Figure 24: Top view 2-D slices of metallic plate target for trailing squint at varying elevation.


Two strong elongated returns are identifiable in the SAR images for the no squint case, centered at the same range at two distinct elevations. The PMI elevation coordinate for the strongest return is 1.0m, suggesting that this return corresponds to the top edge of the metallic plate. The other strong return signal occurs at a lower elevation, suggesting that it comes from the corner formed by the 90° junction between the ground and the metallic plate. The width of the return is measured to be 2.0m in the azimuth direction. This is in close agreement with the physical length of the plate, differing by 20cm from that value. Secondary returns appear 10dB lower in intensity all around the main strong return signal. At lower elevations, the secondary returns appear in front of the main return. At higher elevations, the secondary return equally surrounds the front and back of the main return. For still higher elevations, the secondary return appears behind the main return. This could be due to multipath. As the elevation increases, the return from this interaction takes longer to reach the radar; hence making this feature appear at a higher range. There are also secondary returns on each end of the metallic plate, suggesting interaction of the microwaves with the side edges of the metallic plate.

For the leading squint case, the strong elongated return in azimuth seen for the no squint case does not appear. What remains is a strong return at a location 20cm to the right of the right edge of the metallic plate. This strong return appears for elevations between 0.0m and 1.0m, suggesting that it comes mainly from the right edge of the plate. Another return corresponding to the left edge of the metallic plate can be seen at an intensity that is 20dB lower.

For the trailing squint case, the strong elongated return in azimuth seen for the no squint case does not appear. What remains are two strong returns. They are located 45cm to the left of the left edge of the metallic plate. This stronger return appears for elevations between 0.5m and 1.5m, suggesting that it comes mainly from the left top corner edge of the plate. The other return appears at -0.5m, suggesting a strong interaction of the microwaves with the ground and the bottom corner of the metallic plate. Another return at the right edge of the metallic plate appears at the same elevation but is 20dB lower than the stronger return, suggesting that it comes mainly from the edge of the right top corner of the plate.

The results for the leading and trailing squint cases were expected to be similar; however, only one strong return appeared at the right edge of the metallic plate for the leading squint case whereas two strong returns appeared at the left edge of the metallic plate for the trailing squint case. More work is necessary to understand why this unexpected result occurred. This will be part of future studies.

3.4 Table target signature The signature of a table in free-space is presented in this sub-section. Only one type of table, a picture of which is shown in Figure 25, is studied. The dimensions of the table are presented in Figure 26. The table is composed of two materials: wood and metal. The table legs, crossbar plate, and under table frame are metallic while the tabletop is wood. Only the metallic pieces are expected to have strong returns in the SAR imagery. Dimensions of interest include the length of the crossbar (1.4m) and the height of the table legs (0.70m).

3-D SAR data are displayed as progressive top view 2-D slices in Figure 27, Figure 28, and Figure 29 for no squint, leading squint, and trailing squint, respectively. The data were normalized to ensure scaling remained the same for the different cases.


Figure 25: Table.


Figure 26: Table dimension measurements.

metal legs

metal frame below tabletop

metal crossbar

wood

67cm

42cm 18cm

7cm 72cm

80cm

146cm

7cm

15cm

48cm

160cm

7cm

3cm 140cm

2cm

3cm

11cm

10cm

45cm

6cm

7cm

3cm

154cm

140cm

18cm

42cm

24cm 6cm

7cm

7cm

80cm

side view

bottom view

front view


-0.50m 0.00m

0.50m 1.00m

1.50m 2.00m

Figure 27: Top view 2-D slices of table target for no squint at varying elevation.


-0.50m 0.00m

0.50m 1.00m

1.50m 2.00m

Figure 28: Top view 2-D slices of table target no leading squint at varying elevation.


-0.50m 0.00m

0.50m 1.00m

1.50m 2.00m

Figure 29: Top view 2-D slices of table target for trailing squint at varying elevation.


As expected, the signature of the table is more complex than that of the metallic plate. Two strong elongated uniform returns are identifiable for the no squint case, with a difference in intensity of 7dB between them. There is a slight yet noticeable angle to this signature of approximately -10° from the normal to the SAR path, implying that one of the table legs was placed slightly closer to the radar than the other. The two returns are separated in range by 50cm and appear at elevations between -0.5m and 1.0m. The range distance is in close agreement with the physical distance between the front support frame under the tabletop and the crossbar, well within the range resolution. This suggests that the strong return at higher range corresponds to the crossbar and the strong return at lower range corresponds to the front of the support frame under the tabletop. The length of the strong uniform return at higher range is approximately 1.5m, which is also in close agreement with the physical length of the crossbar. The length of the strong return at the smaller range is approximately 1.7m, 15cm longer than the physical length of the support frame.

The PMI elevation coordinate of both returns is 0.5m, at the height of the bottom of the crossbar and 20cm down from the height of the support frame under the tabletop. This suggests that the strong return corresponds to the bottom edge of the crossbar where a corner is formed and the secondary return corresponds to the support frame under the tabletop. Because the returns are also present at lower elevations, part of the contribution to the returns is created by the corners formed between the table legs and feet. The same secondary returns that appear around the main strong return signal of the metallic plate are also present in the table signature. This is caused by the crossbar which also acts as a flat plate. The signal is 10dB lower all around the strong return at higher range.

For leading squint, the strong elongated returns seen for the no squint case do not appear. What remains are three strong returns at the same range as the strongest return for the no squint case at the location of the crossbar. Two returns, one at each side of the table, appear between -0.5m and 0.5m in elevation, suggesting that they are due to the corner created by the table foot and the ground and the corner between the table leg and foot. The third strong return, seen on the left of the image, occurs between 0.0m and 1.0m, suggesting that it is due to the corner formed between the metal crossbar and the metal leg. The strong return on the right of the image appears shifted by as much as 66cm to the right from the azimuth location of the right-most edge of the table leg in the no squint case. The strong returns on the left of the image appear shifted by 12cm to the right from the azimuth location of the left-most edge of the table leg compared to the no squint case.

For trailing squint, the strong elongated returns seen for the no squint case do not appear. What remains are two strong returns at the same range as the strongest return for the no squint case at the location of the crossbar. Three strong returns were expected, as is observed for the leading squint case. It is possible that the orientation of the table approximately -10° from the normal to the SAR path contributed to this difference. The two returns, one at each side of the table, appear between -0.5m and 1.5m in elevation, suggesting that they are due to the corner created by the table foot and the ground and the corner between the table leg and foot. The strong return on the left of the image appears shifted by as much as 52cm to the left from the azimuth location of the left-most edge of the table leg in the no squint case. The strong return on the right of the image does not appear shifted from the azimuth location to the right-most edge of the table leg compared to the no squint case.


3.5 Discussion of target signatures Radar signatures in 3-D SAR imagery were studied for five different targets, namely, a standing human, a human with arms stretched, a chair, a metallic plate, and a table. In general, there was very close agreement between the measured physical dimensions of the targets and the values measured from the SAR imagery for the strong returns. There was very close agreement between simulations and the strong SAR image returns produced from the human target. The strong returns in the SAR image from the other targets are mostly due to returns from metallic components and interactions between the metallic components and the ground. The effects of a squinted SAR geometry on the target signatures were investigated by comparing these results to those for the no squint case. Returns from wide objects in the azimuth direction (parallel to the SAR path), such as those for the human arms, the chair cushion support bracket, the table crossbar and tabletop frame, and the top and bottom edges of the metallic plate, were significantly smaller than for no squint or non-existent. Features perpendicular to the SAR path that would otherwise not be seen in the no squint case became visible and dominant in the SAR images, including the sides of the human torso, the sides of the table legs and the side edges of the metallic plate. There was more similarity between the signatures of the human target for no squint and both leading and trailing squint compared to the differences observed for the other targets. This phenomenon could potentially serve as a feature used in discriminating the human target from all others.

Viewing of the SAR data as 2-D slices provides a qualitative means of discriminating between different target signatures. A more useful approach to discrimination would be to quantify these differences. In the next section, a look at different quantitative features as potential discriminants is investigated.


4 Target features

The differences in signatures for a number of targets were presented in Section 3. To the human eye, the differences between the signatures for the human, table, chair, and metallic plate are readily visible in the SAR images. In this section, potential quantitative features are investigated with the goal of finding discriminants for eventual detection and classification of human targets behind walls in the 3-D SAR images. For this analysis, all human targets listed in Table 1 are included. These include humans standing or with arms stretched who may be positioned behind other humans.

As observed in the previous section, some targets, like the human target and the metallic plate, give rise to only one strong return while others, such as the chair and table, give rise to several strong returns or peaks. For targets having several strong returns, a decision must be made whether to evaluate each strong return separately, or to attempt to analyze them as a group when considering features. Both approaches have been investigated to see if discriminants can be identified.

Six different features are investigated: the PMI elevation coordinate feature in Section 4.1, the PMI intensity feature in Section 4.2, the 3dB width features in range and azimuth of the PMI profile in Section 4.3, the 3dB widths at varying range and azimuth in Section 4.4, the PMI intensity feature at varying range and azimuth in Section 4.5, and the number of resolution cells for 3-D volumes in Section 4.6. A summary and discussion of the results are presented in Section 4.7.

4.1 PMI elevation coordinate It was seen in Section 3 that the elevation of the PMI can differ between targets. Some targets produce very strong returns at the contact with the ground while for others, such returns are at various elevations above the ground. This feature is investigated as a potential discriminating factor for the human target.

Box plots of the PMI elevation coordinate data for each target were created and can be seen in Figure 30. A box plot is a graphical representation of a distribution. It displays differences between populations without making any assumptions of the underlying statistical distribution. For each box plot, the central mark in red represents the median of the data. The edges of the blue box represent the 25th and 75th percentiles at the bottom and at the top, respectively. Whiskers extend to the most extreme data points not considered as outliers. Where outliers exist, they are plotted individually as red stars. In order for a PMI elevation coordinate feature to be a good candidate for human target discrimination, no overlap between the human box plot and all others is desired.

All data, including those for no squint, leading squint, and trailing squint, are grouped together since the variation between the locations of the PMI elevation coordinate values is minimal. Results for all human targets whether standing or with arms stretched are used for the first box plot. Each individual strong return from the chair and table targets is included in the creation of the second and third box plots. Results for the metallic plate are used for the last box plot.


Figure 30: Box plots of PMI elevation coordinate for human, table, chair, and metallic plate

targets.

There is clearly a difference between PMI elevation coordinate of the human target and the table and chair targets. Whereas this value is 1.0m above ground on average for the human target, the values for the table and chair are much closer to the level of the ground. There is no overlap in any of the elevation values between the human target and the chair and table targets. The metallic plate has the same median location of the PMI elevation coordinate as the human target and cannot be discriminated from the human target with this feature. The PMI elevation coordinate feature is a great candidate for discriminating the human target from other targets; however, it cannot be used to discriminate between them all. Other discriminants are needed to fully and robustly discriminate the human target from all others. Furthermore, the orientation and position of the human target is an important consideration. Close agreement between the locations of the PMI elevation coordinate is observed for both the standing human and the human with arms stretched. What still needs to be investigated is a human target in various other poses, such as seated and kneeling. The PMI elevation coordinate is sure to be lower in those circumstances which would make this feature ineffective for human target discrimination for a human in positions closer to the ground.

4.2 PMI intensity The peak intensity value at the PMI is considered as a feature for the no squint, leading squint, and trailing squint cases. Box plots of the peak intensity value at the PMI for each target are presented in Figure 31, Figure 32, and Figure 33 for no squint, leading squint, and trailing squint, respectively.


Figure 31: Box plots of PMI intensities of targets for no squint case.

Figure 32: Box plots of PMI intensities of targets for leading squint.


Figure 33: Box plots of PMI intensities of targets for trailing squint.

As seen in Figure 31 to Figure 33, there is a discernible difference in peak intensity values between the target types although some overlap exists. The median PMI intensity values of the chair and table are all below the median PMI intensity values of the human target in all instances. For the no squint and the trailing squint cases, the chair can be completely discriminated from the human target as no overlap exists. Although the median PMI intensity of the table is lower than that of the human, overlap still occurs, especially for the no squint case. The overlap is very minimal for the leading and trailing squint cases. The PMI intensity values for the metallic plate are closest to the PMI intensity values for the human target although on the higher end when considering the no squint case. More exemplars of each target type are required to determine the true statistical nature of the data before discrimination can be achieved with this feature. The PMI intensity feature is a good candidate for discriminating the human target from other targets; however, again, it cannot be used to discriminate fully. Other discriminants are needed to fully and robustly discriminate the human target from all others.

4.3 3dB width The 3dB width in each of the range and azimuth directions of each target is considered as a feature. Unlike the 3dB width of a point target, where it is defined as radar resolution, the 3dB width of the different targets under consideration is a measure of the extent of the strong return. When investigating target signatures in Section 3, the widths of the strong returns from the human target and the individual peak returns from the chair were relatively small. The strong returns from the table and metallic plate were present over a larger number of azimuth cells than the other targets, measuring between 1m and 2m. The 3dB width feature provides a means to quantify this characteristic and should help in discriminating the metallic plate from the human target, which cannot be done with the features discussed in Section 4.1 and 4.2.


Box plots of the 3dB width in azimuth for each target were created for the no squint, leading squint, and trailing squint cases. The 3dB width in azimuth is shown in Figure 34 while the 3dB width in range is shown in Figure 35. The lettering of the labels on the x-axis has the format XYY, where X is the first letter of the target type (H for human, T for table, C for chair, and P for metallic plate) and YY provides the initials for the squint geometry used (NS for no squint, LS for leading squint, and TS for trailing squint).

(a)

(b)

Figure 34: Box plots of 3dB width in azimuth of human (H), table (T), chair (C), and metallic plate (P) targets for no squint (NS), leading squint (LS), and trailing squint (TS): (a) original

plot, and (b) close-up of plot without TNS and PNS in the figure.


Figure 35: Box plots of 3dB width in range of human (H), table (T), chair (C), and metallic plate

(P) targets for no squint (NS), leading squint (LS), and trailing squint (TS).

Two features stand out in Figure 34(a). The values for the table and metallic plate targets for the no squint condition are well above those of all other targets by a factor varying from 10 to almost 30 in the azimuth direction. A closer look at the box plots of Figure 34(a), without these two high 3dB width values can be seen in Figure 34(b). There appears to be a high level of overlap between all other targets along the azimuth dimension, where the median varies very little (between 0.10dB and 0.18dB). There appears to be significant overlap between all targets for the range 3dB widths in Figure 35. As seen for the azimuth 3dB width, the median varies very little again (between 0.11dB and 0.16dB). Even though the values for the metallic plate appear to be above that for the human, the difference is not great enough to confidently use the 3dB width in range for discrimination. It can still remain a consideration if other objects with large 3dB widths along the range direction are considered as targets in the future. Using the average 3dB width in the azimuth direction for the no squint case would be a practical choice to consider in future human target discrimination applications even though it also cannot be used to fully discriminate the human target from all other targets. The average 3dB width in the range direction could still remain a consideration as a way to discriminate the human target from objects having large 3dB widths in that direction.

4.4 3dB width profile Variations in the 3dB width in azimuth and range in the range and azimuth directions, respectively, for increasing distances from the PMI have been examined. These two dimensions were chosen due to their higher spatial resolutions.


To create the profile of azimuth 3dB widths in the range direction, the 3dB width in azimuth at the location of the PMI is found. The 3dB width in azimuth at one range grid point further from the location of the PMI is then determined. This process of finding the 3dB width at the next range grid point continues until a 3dB width profile in the range direction is created. The maximum intensity often occurs at an azimuth coordinate that is different from the PMI azimuth coordinate. Care is taken to ensure that the maximum, regardless of its azimuth coordinate, is used to find the 3dB width. To create the 3dB width profile in the azimuth direction, the same process as that described for the 3dB width profile in the range direction is used to find the 3dB widths at increasing azimuth distances from the location of the PMI. In both cases, the elevation remains constant throughout at the value for the PMI.

For this study, the distance from the PMI was arbitrarily chosen to vary from 30cm in front of to 30cm behind the PMI in the range direction and similarly from 30cm to the left to 30cm to the right of the PMI in the azimuth direction. The assumption is that at increasing distances from the PMI, the 3dB width becomes larger or exhibits a certain characteristic that can make it possible to discriminate one target from all others.

The azimuth 3dB widths at varying range locations were determined for all targets including all humans (H1 to H5 representing standing human and H6 to H8 representing human arms stretched), each bright spot for each end of the table (T1 to T4), each bright spot of the chair (C1 to C5), and the metallic plate (P) for the no squint case. The target symbols are presented in Table 5. The azimuth 3dB widths at varying ranges are shown in Figure 36.

Table 5: Target symbols

Symbol Target Type NOTES H1 Human standing in Scene 2010-1 H2 Human standing front right in Scene 2010-12 H3 and H4 were placed behind H1 and

H5; therefore these signatures were not technically in free space. Regardless, data were still used in the analysis.

H3 Human standing back right in Scene 2010-12 H4 Human standing back left in Scene 2010-12 H5 Human standing front left in Scene 2010-12 H6 Human arms stretched in Scene 2010-11 H7 Human arms stretched in Scene 2011-68 H8 Human arms stretched in Scene 2011-70 T1 Table peak front right in Scene 2010-11 Although not easily seen in the imagery,

two peaks exist at each end of the two strong table returns.

T2 Table peak back right in Scene 2010-11 T3 Table peak back left in Scene 2010-11 T4 Table peak front left in Scene 2010-11 C1 Chair leg front right in Scene 2010-11 C2 Chair leg back right in Scene 2010-11 C3 Chair seat support bracket in Scene 2010-11 C4 Chair leg back left in Scene 2010-11 C5 Chair leg front left in Scene 2010-11 P Metallic plate in Scene 2011-70

The 3dB width profiles in the range direction for the human targets appear to follow a similar pattern in that a peak exists in all cases at a range of +0.1m from the PMI. This corresponds to the arm and shoulder features. Other peaks appear at different range locations.


H1 H2 H3

H4 H5 H6

H7 H8 T1

T2 T3 T4

C1 C2 C3


C4 C5 P

Figure 36: Azimuth 3dB widths at varying range positions for all targets.

A pattern of distinctive peaks similar to that seen for the human targets is observed for the strong returns from the chair legs. When considering the five profiles for the chair as a group, a pattern is discernible. The two outer figures (C1 and C5) and the two adjacent figures (C2 and C4) are very similar to one another, although an extra peak appears in the profile of C4 that is not in that of C2.

A large difference is observed between the 3dB width profiles of the table and metal plate and those of the other targets. This is due to the relatively constant elongated shape of the return in the azimuth direction for both targets. In the case of the table target, the variation in the 3dB widths over the 0.6m extent in range is much larger than for other targets, varying between 10cm and 70cm, whereas it only varies between 4cm and 14cm for the human target. The 3dB width of the metallic plate has a much wider variation, between 40cm and almost 380cm. When the peaks for the table are considered as a group, a pattern is also observed. T1 and T3 as well as T2 and T4 appear to be nearly identical mirror images of one another. This symmetry is due to the symmetry of the table.

To quantify the difference between the azimuth 3dB width variations in the range direction, a correlation matrix was obtained for all pairs of the curves shown in Figure 36. A color visual representation of the correlation matrix is given in Figure 37. For this feature to be a good candidate for discrimination of the human target from all others, a strong correlation of 0.8 and above is desired between each pair of the profiles for the human, and a weak correlation of less than 0.6 is desired between each pair for the human and all others. There is certainly more correlation observed between each pair of curves for the humans than between each pair for the humans and all others, particularly for the human arms stretched position; however, the correlation value is not as high as desired. As a result, this correlation feature might not prove to be a very strong candidate for discrimination of a human target from all others.

The range 3dB widths at varying azimuth locations were determined for all targets including all humans (H1 to H5 representing standing human and H6 to H8 representing human arms stretched), each bright spot for each end of the table (T1 to T4), each bright spot of the chair (C1 to C5), and the metallic plate (P) for the no squint case. The range 3dB widths at varying azimuths are shown in Figure 38.


Figure 37: Correlation matrix of azimuth 3dB widths at varying range positions for all targets.

H1 H2 H3

H4 H5 H6


H7 H8 T1

T2 T3 T4

C1 C2 C3

C4 C5 P

Figure 38: Range 3dB widths at varying azimuth positions for all targets.

The 3dB width profiles in the azimuth direction for the human targets do not seem to follow a similar pattern to those in the range direction. Peaks are appearing at random locations.

When considering the five signals for the chair as a group, a pattern is again discernible. The two outer figures (C1 and C5) and the two adjacent figures in (C2 and C4) are nearly identical mirror images of one another. This symmetry is due to the symmetry of the chair.

A smaller variation is observed between the range 3dB width profiles of the table and metal plate curves compared to those for the other targets. The 3dB widths of the table peaks T1 and T4


either increase or decrease with increasing distance from the PMI. Furthermore, the variation in the 3dB widths over the 0.6m extent in azimuth is very small, between 3cm and 5cm, whereas it varied between 3cm and 9cm for the human target. The 3dB width of the metallic plate has an even smaller variation, between 4.1cm and 4.4cm. A correlation matrix was obtained between all pairs of the curves shown in Figure 38. A color visual representation of the correlation matrix is presented in Figure 39. Examination shows that there is very little correlation between any of the human targets and between any human target and all others. This is expected since peaks appear at random azimuth values in most of the curves. This indicates that this feature does not appear to be a good candidate for discrimination between human targets and all others. The azimuth and range 3dB width profiles in the range and azimuth directions, respectively, do not seem effective for discriminating the human target from other targets. Although some correlation exists for the azimuth 3dB width profiles between any of the human targets and between any human target and all others, particularly for the human arms stretched position, the correlation value is not as high as desired. For the range 3dB width profiles, very little correlation exists. Consequently, using this feature for complete discrimination between the human targets and all others is not recommended.

Figure 39: Correlation matrix of range 3dB widths at varying azimuth positions for all targets.


4.5 Maximum intensity profile Variations in the maximum intensity in azimuth and range in the range and azimuth directions, respectively, for increasing distances from the PMI have been examined. These two dimensions were chosen due to their higher spatial resolution.

To create the profile of maximum intensities in the range direction, the maximum intensity at the location of the PMI is found. The maximum intensity in azimuth at one range grid point further from the location of the PMI is then determined. This process of finding the maximum intensity at the next range grid point continues until a maximum intensity profile in the range direction is created. To create the maximum intensity profile in the azimuth direction, the same process as that described for the maximum intensity profile in the range direction is used to find the maximum intensities at increasing azimuth distances from the location of the PMI. The elevation remains constant throughout at the value for the PMI.

As for the 3dB width profiles described in Section 4.4, the distance from the PMI is arbitrarily chosen to vary from 30cm in front of to 30cm behind the PMI in the range direction and similarly from 30cm to the left to 30cm to the right of the PMI in the azimuth direction. The assumption is that at increasing distances from the PMI, the maximum intensity decreases at different rates or in different patterns for each of the targets which can make it possible to discriminate one target from all others.

The maximum intensity profiles in the range direction for all targets are shown in Figure 40.

The curves for the human targets appear to follow a similar pattern of decreasing with increasing distance from the PMI. The left side of the profile exhibits a gradual decrease to between 5 and 10dB whereas the right side exhibits a smaller decrease to only between 20 and 30dB. The smaller decrease might be due to the effect of the secondary feature present at the back of the human torso. An exception to this pattern occurs for the human target H4 where the slight initial decrease on the right side of the profile is followed by an increase in maximum intensities.

A similar pattern to that for the human targets is also discernible for the profiles for the individual table peaks T2 and T3. The left side of the profile exhibits a gradual decrease to 18dB and 19dB, respectively, although at a lower value than was observed for the human targets, whereas the right side exhibits a slightly smaller decrease to 23dB and 21dB, respectively, similar to that observed for the human targets.

The profiles for the individual table peaks T1 and T4 exhibit a similar pattern to that of the curves for the human target H4 where the slight decrease on the right side is followed by an increase in maximum intensities. The left side of the profile exhibits a gradual decrease to 7dB, similar to that observed for H4 whereas the right side exhibits a drop followed by a rise to 20dB, although at a lower value than for H4. When all four curves for the table are considered as a group, a pattern is discernible. The two outer curves (T1 and T4) are due to the maximum intensities surrounding the first strong return of the table legs with the ground. The two inner curves (T2 and T3) are due to the maximum intensities surrounding the second strong return from the metallic crossbar of the table leg support. The two strong returns for each end of the table exhibit the same pattern.


H1 H2 H3

H4 H5 H6

H7 H8 T1

T2 T3 T4

C1 C2 C3


C4 C5 P

Figure 40: Maximum intensity in azimuth at varying range positions for all targets.

A similar pattern to that seen in the profiles for the human targets is discernible in the curves for the individual chair peaks C1 and C5. The left side of the profiles exhibit a gradual decrease to 3dB and 5dB, respectively, whereas the right side show a slightly smaller decrease to 19dB and 20dB, respectively, followed by an increase to 26dB and 25dB, respectively. A completely different pattern is observed for the individual chair peaks C2, C3, and C4. The left side of the profiles exhibit a gradual decrease to between 18dB and 20dB followed by an increase to between 24dB and 28dB. The right side of the profiles show a gradual decrease to between 16dB and 18dB. If the five curves for the chair are considered as a group, a pattern is discernible. The two outer curves (C1 and C5), due to the maximum intensities surrounding the two front chair legs, and the three inner curves (C2, C3, and C4), due to the two back chair legs and the cushion support brackets, are nearly identical.

The profile for the metallic plate exhibits a similar pattern to that for the human target H4 where the decrease on the right side of the profile is followed by an increase in maximum intensities. The left side of the profile exhibits a gradual decrease to 10dB whereas on the right side there is a drop to 17dB followed by a rise to 34dB.

As was done in Section 4.4, the correlation matrix was created and is shown in Figure 41. Very high correlation is observed between each pair of curves for the human targets. Very high correlation is also seen between the human target curves and those for the peaks of the table and the metallic plate. This is not desirable for discrimination of the human target from the others. There is, however, very low correlation between the human target and the three peaks of the chair C2, C3, and C4, which is desirable. Considering the high level of similarity between each curve except for the three chair peaks C2, C3, and C4, this correlation matrix result is expected. Although the human target cannot be discriminated from the table and metallic plate using this metric, it can strongly be discriminated from the chair, making this feature a strong candidate for future discrimination of human targets.

The maximum intensity profiles in the azimuth direction for all targets are shown in Figure 42.

The curves for the human targets appear to follow a similar pattern of decreasing with increasing distance from the PMI. Both sides of the profile exhibit a gradual decrease to between 13dB and 28dB with some cases, namely, H2, H7, and H8, exhibiting decreases followed by an increase to between 30dB and 38dB. They are more symmetrical about the PMI than was observed for the maximum intensity profiles in the range direction. This symmetry is not surprising since the


Figure 41: Correlation matrix of maximum intensity in azimuth at varying range positions for all

targets.

H1 H2 H3

H4 H5 H6


H7 H8 T1

T2 T3 T4

C1 C2 C3

C4 C5 P

Figure 42: Maximum intensity in range at varying azimuth positions for all targets.

human signature is also symmetrical in the azimuth direction. Although the curves for the human targets are similar, there are still some variations observed between them.

The maximum intensity profiles for the table are completely different from those of the human targets. Profiles for table peaks T1 and T2 and for table peaks T3 and T4 are very similar to each other. The left side of the profiles for table peaks T1 and T2 exhibits a gradual decrease to 18dB and 26dB, respectively. The right side of the profiles has a constant value of 32dB and 38dB, respectively. The left side of the profiles for table peaks T3 and T4 is constant at approximately


38dB and 31dB, respectively. The right side of the profiles exhibits a gradual decrease to 32dB and 26dB, respectively. Some similarities exist between the profiles for the chair peaks C1 and C5 and the human targets, although the former exhibit less symmetry about the location of the PMI in the gradual decrease of the curve with increasing distance from the PMI. The left side of the profile for chair peak C1 exhibits a gradual decrease to 22dB followed by a local maximum at 27dB. The right side shows a gradual decrease to 12dB. The left side of the profile for chair peak C5 exhibits a gradual decrease to 10dB. The right side shows a gradual decrease to 21dB followed by a local maximum at 25dB. Some similarities also exist between the waveform for the chair peak C3 and the human targets, although both gradual decreases seen for C3 are followed by a local maximum on both sides. These smaller peaks are due to the two back chair legs. The left side of the profile for chair peak C3 exhibits a gradual decrease to 21dB followed by a local maximum peak at 27dB. The right side shows a gradual decrease to 26dB followed by a local maximum at 28dB. The curves for the chair peaks C2 and C4 exhibit very different patterns than those for C1 and C5. The left side and the right side of the profiles for chair peaks C2 and C4, respectively, have a local maximum that is greater than the maximum intensity at the PMI by 2dB. This is because the signature due to the cushion support bracket seen in C3 also appears in these images at a higher maximum intensity than the chair peaks due to the back legs. When considering the five profiles for the chair as a group, a pattern is discernible. The two outer curves (C1 and C5), representing the maximum intensities surrounding the two front chair legs, and the two adjacent curves (C2 and C4), representing the two back chair legs, are nearly identical mirror images. This symmetry is due to the symmetry of the chair. The profile for the metallic plate exhibits a similar pattern to that for the human targets although variations in the values are less than 1dB compared to those for the human target of approximately 30dB. The left side of the profile exhibits a slight decrease from 41.8dB to 41.4dB with the right side having a gradual decrease to 41.5dB. This is expected since the intensity values for the metallic plate are essentially constant in the azimuth direction over a large area. The correlation matrix was created and is shown in Figure 59. High correlation is observed between each pair of curves for the human targets. This is desirable. High correlation is also observed between the human target and metallic plate, and between the human targets and the chair peaks C1 and C5. This is not desirable; however, low correlation exists between the human targets and all four table peaks, and between the human targets and three chair peaks C2, C3, and C4, which is again desirable. Although the human target cannot be discriminated from the metallic plate using this metric, it can strongly be discriminated from the chair and table, making this feature a strong candidate for future discrimination of human targets from other targets.


Figure 43: Correlation matrix of maximum intensity in range at varying azimuth positions for all targets.

4.6 Number of resolution cells for 3-D volumes

So far, effects on resolution have been investigated on 2-D profiles of the 3-D SAR data. Another perspective can be gained from viewing 3-D volumes of the SAR data. The PMI can be used as a seed point for a region growing algorithm. There are different statistical methods that can be selected to extract the segmented data. To begin, the choice was made to extract all intensities that fall within the 3dB and the 6dB intensities of the main peak intensity of a human target in the standing position and in the arms stretched position. The same targets that were used in Section 3 are used here. The segmented 3-D volumes produced are shown in Figure 44. Since the region growing algorithm is only set to extract values within 3dB and 6dB of the intensity of the PMI, the extended arms were not expected to appear in the 3-D volumes. This is indeed the case. The difference between human standing and human arms stretched is the slightly larger azimuth and range extent for the latter. The only difference between the 3dB and the 6dB volumes for this target is that a larger number of cubes are used to define the volume. The “blob-like” shape remains relatively the same – no other features become visible.

Viewing the SAR data as 3-D volumes was then investigated for all targets listed in Table 1. As described above, the PMI of each target was used as a seed point for a region growing algorithm. All pixels with intensities within 3dB of the peak value were extracted. The segmented 3-D volumes produced from the region growing algorithm are shown in Figure 45.


(a) (b)

(c) (d)

Figure 44: Human target volumes: (a) 3dB of human standing, (b) 3dB of human arms stretched, (c) 6dB of human standing, and (d) 6dB of human arms stretched.


H1 H2 H3 H4

H5 H6 H7 H8

C1 C2 C3 C4

C5 T1 T2

T3 T4 P

Figure 45: Segmented 3dB volumes of all targets.


Examination of Figure 45 shows that there does not seem to be much difference between the 3-D volumes of the human targets and those of the individual chair peaks C1 through C5. Quantitatively, the number of cubes occupying the space appears to be very similar in all directions for all of these. The similarity in 3dB 3-D volumes for the table and the metallic plate is significant, since there is a much larger number of cubes in the azimuth direction for both cases compared to the human targets and chair peaks. The number of cubes present in the azimuth direction could potentially be used as a feature to discriminate the human target from the table and the metallic plate. Increasing the dB threshold used was investigated next to see at what value other features of the target signatures, if any, become apparent in the 3-D volumes. For the human target H1 (standing human), the front of the torso and the arms and shoulders are already prominent for a 3dB threshold. The signature feature of the legs behind the torso first appears at a threshold of 12dB. The 3-D volume of the human standing is shown in Figure 46 where both features can be seen.

Figure 46: Human standing 3-D volume at 12dB threshold.

For the human target H6 (arms stretched), the same signature as that observed for the human standing target is seen up to a threshold of 20dB. At this value, the strong return seen on the right of the human appears. The strong return of the arm on the left of the human appears at a threshold of 24dB. The 3-D volume of the human target H6 at this threshold is shown in Figure 47 where the torso and the arms stretched features are readily visible.


Figure 47: Human arms stretched 3-D volume at 24dB threshold.

For the chair target, the four legs and bracket features become distinctively observable at a threshold of 15dB. The 3-D volume of the chair at this threshold is shown in Figure 48. The back legs and the bracket are visible at the same range but the return from the bracket is greater. The front legs are visible at a smaller range.

Figure 48: Chair 3-D volume at 15dB threshold.

For the table target, the two elongated strong returns in the azimuth direction begin to appear at a threshold of 14dB. The 3-D volume of the table at this threshold is shown in Figure 49. In this


view, the front feature due to the support from under the tabletop at a smaller range and the interaction of the legs with the ground at a lower elevation can be seen. The feature at higher range and due to the bottom edge of the crossbar, the corner junction between the crossbar and the table legs, and the corner junction between the table legs and the table foot, at a higher elevation and a higher range are also visible.

Figure 49: Table 3-D volume at 14dB threshold.

For the metallic plate, the two elongated strong returns in the azimuth direction begin to appear at a threshold of 18dB. The 3-D volume of the metallic plate at this threshold is shown in Figure 50. The front feature is due to the interaction of the microwaves between the bottom edge of the plate and the ground. The feature at a higher range and elevation is due to the top edge of the plate.

Figure 50: Metallic plate 3-D volume at 18dB threshold.

A good metric for the 3-D volume analysis is the number of resolution cells the cubes occupy for each target. For consistency in comparison between each target, the same threshold must be used. A threshold of 24dB was chosen as it was the lowest threshold value necessary in order that all


target features be visible. The number of resolution cells in azimuth, range, and elevation are presented in Table 6 for all five types of targets.

Table 6: Number of resolution cells for 3-D volumes at 24dB threshold.

Number of azimuth resolution cells

Number of range resolution cells

Number of elevation resolution cells

Human standing 9 11 6 Human arms stretched 23 12 6

Chair 28 21 6 Table 13 13 6 Plate 28 35 3

At a threshold of 24dB, there is enough difference in the numbers of azimuth resolution cells (varying from 14 to 19) to be able to strongly discriminate the human standing target from the human arms stretched, chair, and metallic plate targets. There is also enough difference in the number of range resolution cells (between 9 and 24) to be able to discriminate the human standing and human arms stretched targets from the chair and metallic plate. The difference in the numbers of resolution cells in the azimuth and range directions between the table and the human standing target is only 2 to 4 cells. The number of resolution cells in elevation is similar for all cases. The number of resolution cells in azimuth feature does not appear to be a good candidate because of the high value of azimuth resolution cells for the human arms stretched target compared to that for the human standing target, making discrimination of all human targets from the others impossible for this case. Perhaps the unexplained strong return to the left of the human that dominates in the azimuth direction is skewing the results. More investigation is required to determine the true number of azimuth resolution cells for the human arms stretched target.

It is concluded that the number of resolution cells in range feature would be a good candidate for the discrimination of the human target from the other targets but more investigation is required to conclude whether the number of resolution cells in azimuth feature would be a good candidate or not.

4.7 Summary and discussion of results Six different approaches were introduced and analyzed for their capability to discriminate human targets from others in free space. The PMI elevation coordinate, PMI intensity, 3dB width in azimuth of the no squint case, and the number of range resolution cells for 3-D volumes all showed significant potential as features that could be used to discriminate the human target from all others. The correlation between the 3dB width profiles in the range direction of both human targets and all others showed limited potential and the correlation between the 3dB width profiles in the azimuth direction of both human target and all others showed no potential for discriminating the human target from all others. The correlations between the two 3dB width profiles in the range direction of the human target and three peaks of the chair (C2, C3, and C4) show significant potential for discriminating the human target from the chair. The correlation between the pair of maximum intensity profiles in the azimuth direction of the human target and all four peaks of the table (T1, T2, T3, and T4) as well as three peaks of the chair (C1, C2, and C3) show great potential for discriminating the human target from the chair and the table. The


number of resolution cells in range shows great potential for discrimination of both human targets from the chair and the metallic plate. The number of resolution cells in azimuth shows weak potential for discrimination of both human targets from the chair and metallic plate but might still be a consideration once the source of the unexplained strong return on the left of the human with arms stretched target is found and negated.

Results presented in this section are preliminary. A larger set of different target types, especially those typically found in buildings of interest, is necessary. As well, the limited number of exemplars of each of the targets used in this study is not sufficient to make conclusive remarks. Use of a larger set of exemplars of the same type of target, in the same position and at different orientations and positions, would give more confidence in the values obtained for the different features. This would be a more robust analysis and would be a prudent approach to eliminating anomalies observed in the data. For example, if multiple subjects and objects of a given type could be used, a more accurate box plot can be generated, perhaps with smaller variations between the lowest and highest values in the data. This would provide a more realistic metric to investigate if this feature could be used to discriminate the human target from the table.

Although each approach showed potential for discriminating the human target from other targets, not one feature could be used to fully discriminate between all targets. Combining several features could perform this task. For example, the PMI elevation coordinate and the 3dB width features could provide a means to fully discriminate the human target from all others presented in this study.


5 Conclusions and future research efforts

Tools and algorithms for 3-D visualization are being developed for the analysis of signatures of human targets behind a wall and to develop an understanding of the clutter and multipath signals in a room of interest. In this report, a comprehensive study of the characteristics of free-space target signatures was presented using 3-D SAR data. The aim of this investigation was to gain a better appreciation of the signatures of targets placed behind different wall materials while identifying potential discriminants for classification of human targets. An analysis of different target signatures was provided. Six features were investigated as potential discriminants and five of them were identified as good candidates.

In general, there was very close agreement between the measured physical dimensions of the targets and those obtained from the strong returns in the SAR imagery. There was very close agreement between simulations and the strong SAR image returns for the human targets. Most of the sources of the strong returns seen in the SAR images were explained, taking into consideration the location of the PMI, the measurements between different returns, and the location of the strong returns with respect to different physical features of each target. There are still some strong returns that need to be investigated to confirm, verify, or discover their sources.

The effects of imaging in a squinted SAR geometry on target signatures were investigated. Strong returns from wide objects along the azimuth direction parallel to the SAR path were either minimized or eliminated such as those from the human arms, the chair cushion support bracket, the table crossbar and tabletop frame, and the top and bottom edges of the metallic plate. Features perpendicular to the SAR path that are not seen in the no squint case are visible and even dominant in the squinted SAR images, including the sides of the human torso, the sides of the table legs, and the side edges of the metallic plate. There is much about the squint SAR that is not yet understood, such as the “smearing effect” of the bright target features, the curvature of some of these bright features, and strong returns that appeared more like point target returns in the no squint cases but appeared elongated in the range-azimuth planes. Another phenomenon that is not well understood is the shifting of the bright features in azimuth and range. This phenomenon can be explained in the case of the human target but requires more understanding for wider objects.

Simulations are a great tool to investigate radar returns of different targets, as was seen in the analysis of the human target signatures. Simulations allow certain features to be added or removed to investigate their effects on the SAR image. They also make it possible to study the effects of different beamwidths or different squinted SAR geometries on target signatures. Unfortunately, simulation capabilities are not currently available to the through-wall SAR team at DRDC. Research in this area is limited to literature searches and review of the scientific literature.

There are different ways to view the 3-D SAR data. In this report, three different ways to observe them were used. Three 2-D slices that intersect at the PMI of a target gave an idea of how the target signature looks from a top view, front view, and side view. Data displayed as progressive top view 2-D slices provided a way to view, measure, and observe changes to the target signature from a top view perspective as the elevation was varied. 3-D volumes of segmented target signatures provided a way to look for features that fall within a certain threshold from the peak


value. Each viewing method provides a different perspective and aids in understanding how the target signatures are formed, the sources of the strong scattering, and what features can be used to discriminate one target from another. The one view that was not exploited to its fullest in the work described was the 3-D volume view. The implementation of the region growing algorithm was simplified to act as a thresholding tool. There are many more complex and sophisticated statistical methods available to segment the data. This is one area that is currently being investigated with the through-wall team and promises to impact our ability to discriminate between targets in a positive way.

All of the discriminants measured or calculated in this report have relied heavily on the PMI and the PMI intensity. The PMI is not only important for the extraction of the discriminants, but also for forming the seed to many available segmentation algorithms. Currently, the PMI is extracted manually after inspection of each target in each scene for each different squint SAR geometry for each SAR parameter and weighting window investigated. This is a tedious and onerous task. An automatic algorithm to extract these values would be invaluable and a real time saver.

Five different discrimination approaches were identified as showing great potential of features that could be used to discriminate the human target from all others: PMI elevation coordinate, PMI intensity, 3dB width in azimuth of the no squint case, and the number of resolution cells in range. Based on the free-space scenario and the different targets in this study, no single feature could be used to fully discriminate the human targets from all others. A combination of at least two different features is required to achieve this.

This study will be continued to search for more features and will be enlarged to include more targets in different orientations as well as more exemplars of the human target in various positions in order to create a robust strategy for detection and classification of human targets. The research will be expanded to include an investigation of how the signatures of targets vary as they are moved behind different types of walls. Furthermore, a thorough understanding of the clutter and multipath present behind walls are required if the signature of the human target is to be extracted and discriminated from these noise sources. Eventually, efforts will be made to choose the most effective classifier for the discrimination of human targets from all others. Different approaches for a one-class classifier such as density estimation, boundary methods, and reconstruction methods, will be investigated [10].


References .....

[1] Sévigny, P., DiFilippo, D.J., Laneve, T., Chan, B., Fournier, J., Roy, S., Ricard, B., and Maheux, J. (2010). Concept of operation and preliminary experimental results for the DRDC through-wall SAR system. Proc. of SPIE, 7669, 766907-766907-11.

[2] Raut, S. And Petosa, A. (2009). A compact printed bowtie antenna for ultra-wideband applications. Proc. European Microwave Conf. EUMC 2009, Rome, Italy, pp. 081-084.

[3] Hoctor, R.T. and Kassam, S.A. (1990). The unifying role of the coarray in aperture synthesis for coherent and incoherent imaging. Proc IEEE, 78(4), 735-752.

[4] (2007). LMS Q240i Airborne Laser Scanner Technical Documentation and User’s Instructions. Technical Report. Riegl Laser Measurement Systems.

[5] (2000). POS LV V3 Installation and Operation Guide. (Technical Report PUBS-MAN-000048). Applanix Corporation.

[6] (2011). HDL-32E high definition LiDAR sensor user’s manual and programming guide. (Technical Report, 63-9113 Rev B). Velodyne Lidar Inc.

[7] Sévigny, P. and DiFilippo, D.J. (2010). Software design and functional description of the L-band through-wall experimental radar testbed processor. (DRDC Ottawa TM 2010-204). Defence R&D Canada – Ottawa.

[8] DiFilippo, D.J. (2008). A signal processing kernel for 3-D through-wall SAR imaging. (DRDC Ottawa TM 2008-335). Defence R&D Canada – Ottawa.

[9] Dogaru, T., Nguyen, L., and Le, C. (2007). Computer models of the human body signature for sensing through the wall radar applications. (ARL-TR-4290). Army Research Lab.

[10] Khan, S.S., and Madden, M.G. (2009). A survey of recent trends in one class classification. Proc. Of the 20th Irish Conference on Artificial Intelligence and Cognitive Science, Dublin, pp. 181-190.


List of symbols/abbreviations/acronyms/initialisms

C Chair target

DMI Distance Measurement Indicator

DRDC Defence Research and Development Canada

GPS Global Positioning System

H Human target

IMU Inertial Measurement Unit

LIDAR Light Detection and Ranging System

LS Leading Squint

NS No Squint

P Metallic plate target

PMI Pixel at Maximum Intensity

POI Pixel of Interest

SAR Synthetic Aperture Radar

T Table target

TS Trailing Squint

DOCUMENT CONTROL DATA (Security classification of title, body of abstract and indexing annotation must be entered when the overall document is classified)

1. ORIGINATOR (The name and address of the organization preparing the document.Organizations for whom the document was prepared, e.g. Centre sponsoring a contractor's report, or tasking agency, are entered in section 8.)

Defence R&D Canada – Ottawa 3701 Carling Avenue Ottawa, Ontario K1A 0Z4

2. SECURITY CLASSIFICATION (Overall security classification of the document including special warning terms if applicable.)

UNCLASSIFIED (NON-CONTROLLED GOODS) DMC A REVIEW: GCEC JUNE 2010

3. TITLE (The complete document title as indicated on the title page. Its classification should be indicated by the appropriate abbreviation (S, C or U) in parentheses after the title.)

Free space target signatures and analysis of discriminants for human targets for the 3-D L-band through-wall experimental SAR system 4. AUTHORS (last name, followed by initials – ranks, titles, etc. not to be used)

Chan, B.; Sévigny, P.; DiFilippo, D.J.

5. DATE OF PUBLICATION(Month and year of publication of document.)

November 2012

6a. NO. OF PAGES (Total containing information, including Annexes, Appendices, etc.)

88

6b. NO. OF REFS(Total cited in document.)

10

7. DESCRIPTIVE NOTES (The category of the document, e.g. technical report, technical note or memorandum. If appropriate, enter the type of report, e.g. interim, progress, summary, annual or final. Give the inclusive dates when a specific reporting period is covered.)

Technical Memorandum 8. SPONSORING ACTIVITY (The name of the department project office or laboratory sponsoring the research and development – include address.)

Defence R&D Canada – Ottawa 3701 Carling Avenue Ottawa, Ontario K1A 0Z4

9a. PROJECT OR GRANT NO. (If appropriate, the applicable research and development project or grant number under which the document was written. Please specify whether project or grant.)

9b. CONTRACT NO. (If appropriate, the applicable number under which the document was written.)

10a. ORIGINATOR'S DOCUMENT NUMBER (The official document number by which the document is identified by the originating activity. This number must be unique to this document.)

DRDC Ottawa TM 2012-182

10b. OTHER DOCUMENT NO(s). (Any other numbers which may be assigned this document either by the originator or by the sponsor.)

11. DOCUMENT AVAILABILITY (Any limitations on further dissemination of the document, other than those imposed by security classification.)

Unlimited 12. DOCUMENT ANNOUNCEMENT (Any limitation to the bibliographic announcement of this document. This will normally correspond to theDocument Availability (11). However, where further distribution (beyond the audience specified in (11) is possible, a wider announcement audience may be selected.))

Unlimited

13. ABSTRACT (A brief and factual summary of the document. It may also appear elsewhere in the body of the document itself. It is highly desirable that the abstract of classified documents be unclassified. Each paragraph of the abstract shall begin with an indication of the security classification of the information in the paragraph (unless the document itself is unclassified) represented as (S), (C), (R), or (U). It is not necessary to include here abstracts in both official languages unless the text is bilingual.)

DRDC Ottawa has been investigating 3-D through wall synthetic aperture radar (SAR) imaging from an experimental L-band through-wall SAR prototype. Tools and algorithms for 3-D visualization are being developed to exploit the resulting imagery. The through-wall technology and data exploitation algorithms and tools have the capability to enhance situational awareness for military forces operating in an urban environment. Current work involves analyzing signatures of human targets behind a wall and understanding the clutter and multipath signals in a room of interest. In this report, a comprehensive study of the characteristics of free-space target signatures is presented using 3-D SAR data. The aim of this investigation is to gain a better appreciation of the signatures of targets when placed behind different wall materials while identifying potential discriminants for classification of human targets. An analysis of different target signatures is provided. Six features are investigated as potential discriminants and five of them are identified as good candidates.

L’imagerie 3-D par radar à synthèse d’ouverture (SAR) à travers les murs est étudiée à RDDC Ottawa pour un banc d’essai d’un radar expérimental dans la bande L. Des outils et algorithmes pour la visualisation en 3-D sont en développement pour l’exploitation de l’imagerie qui en résulte. La technologie à travers les murs et les algorithmes et outils d’exploitation des données permettront d’améliorer considérablement la perception de la situation par les forces militaires dans l’espace de combat en milieu urbain. Les travaux en cours incluent l’analyse de la signature d’une cible humaine derrière une structure de mur et la compréhension du fouillis d’échos et des signaux dûs à la propagation par trajets multiples, visibles à l’intérieur d’une pièce d’intérêt. Dans ce rapport, une étude compréhensive des caractéristiques des signatures des cibles en espace libre est présentée utilisant des données SAR en 3-D. L’objectif de cette investigation est d’amasser une meilleure appréciation des signatures de cibles lorsqu’elles sont placées derrière différents matériaux de mur et d’identifier des discriminants potentiels pour la classification des cibles humaines. Une analyse des signatures de différentes cibles est fournie. Six traits ont été investigués comme discriminants potentiels et cinq d’entre eux ont été identifiés comme de bons candidats.

14. KEYWORDS, DESCRIPTORS or IDENTIFIERS (Technically meaningful terms or short phrases that characterize a document and could be helpful in cataloguing the document. They should be selected so that no security classification is required. Identifiers, such as equipment model designation, trade name, military project code name, geographic location may also be included. If possible keywords should be selected from a published thesaurus, e.g. Thesaurus of Engineering and Scientific Terms (TEST) and that thesaurus identified. If it is not possible to select indexing terms which are Unclassified, the classification of each should be indicated as with the title.)

through-wall, human target signature, discriminants, L-band, SAR, detection, classification

free space target signatures and analysis of discriminants ... · discriminants for human targets...

Documents