dprg chapters 1 – 7: overview · • lidar data includes ground/terrain and...
TRANSCRIPT
Ayman F. Habib1
DPRG
Laser Scanning
Chapters 1 – 7: Overview• Photogrammetric mapping: introduction, applications, and
tools• GNSS/INS-assisted photogrammetric and LiDAR
mapping• LiDAR mapping: principles, applications, mathematical
model, and error sources and their impact.• QA/QC of LiDAR mapping• Registration of Laser scanning data• Point cloud characterization, segmentation, and QC
• This chapter will be focusing on LiDAR-based orthophoto and Digital Terrain Model (DTM) generation.
Ayman F. Habib2
DPRG
Laser Scanning
OCCLUSION-BASED PROCEDURE FOR TRUE ORTHOPHOTOGENERATION AND LIDAR DATA CLASSIFICATION
Chapter 8
Ayman F. Habib3
DPRG
Laser Scanning
Overview• Introduction• Orthophoto generation
– Literature review– Procedure
• LiDAR data classification– Literature review– Procedure– Experimental results
• Concluding remarks
Ayman F. Habib4
DPRG
Laser Scanning
True Orthophoto Generation
Ayman F. Habib5
DPRG
Laser Scanning
Image and Map characteristics
object
Image plane
Image
map
Relief displacement
No relief displacement
Non-uniform scale
Uniform scaleOrthogonal projection
Perspective projection
An orthophoto is a digital image which has the same characteristics of a map.
Ayman F. Habib6
DPRG
Laser Scanning
Perspective Image
Ayman F. Habib7
DPRG
Laser Scanning
Orthophoto
Ayman F. Habib8
DPRG
Laser Scanning
Orthophoto Generation: Prerequisites• Digital image:
– Wide range of operational photogrammetric systems• Interior Orientation Parameters (IOPs) of the used
camera:– Camera calibration procedure
• Exterior Orientation Parameters (EOPs) of that image: – Image georeferencing techniques
• Digital Surface Model (DSM) or Digital Terrain Model (DTM)– LiDAR, imagery, Radar, …
Ayman F. Habib9
DPRG
Laser Scanning
Digital Image
PC
(x, y)
Backward Projection (EOP & IOP)
Datum
Terrain
g(resampling)
G(X, Y) = g (x, y)
Z(X, Y)
Interpolation
(X, Y)
Differential Orthophoto Generation
Ayman F. Habib10
DPRG
Laser Scanning
Digital Surface Model
perspective center
imagery
AB C
D
Orthophoto
Differential Orthophoto Generation
ghost image/double-mapped area
Ayman F. Habib11
DPRG
Laser Scanning
Differential Orthophoto Generation
Generated OrthophotoOriginal Imagery
Double-mapped areas
Ayman F. Habib12
DPRG
Laser Scanning
Digital Image
PC
Indirect (backward) transformation
Orthophoto Generation & Visibility Analysis
Ayman F. Habib13
DPRG
Laser Scanning
perspective center
imagery
0 01 1 120
AB C
D
a b c
DC P.C
longer
Invisible point
Digital Surface Model
Orthophoto
Z-Buffer Method
True Orthophoto Process – Existing Method
Ayman F. Habib14
DPRG
Laser Scanning
True Orthophoto Process – Existing Method
Generated True OrthophotoOriginal Imagery
Z-Buffer Method
Ayman F. Habib15
DPRG
Laser Scanning
True Orthophoto Process – Existing Method
Generated True Orthophoto
Z-Buffer Method
Ayman F. Habib16
DPRG
Laser Scanning
True Orthophoto Generation
perspective center
A
BInvisible point
DE
C
5 ° visible
12° visible
15° invisible
12 °
15°
14°
A
B
C
D
E 20° 15° visible
max angle
visible/ hiddenpoint angle comparison
0° visible5° >
>
>
<=
>
Nadir point
Digital Surface Model
Orthophoto
Angle-based Method
Ayman F. Habib17
DPRG
Laser Scanning
00
ii
Radial Sweep for the Angle-Based Method
True Orthophoto Generation
Angle-based Method
Ayman F. Habib18
DPRG
Laser Scanning
True Orthophoto Gen.: Adaptive Radial Sweep
DSM partitioning for the adaptive radial sweep method
DSM
column
row
nadir pointsection 1
section 2
section 3
1
23
321
Ayman F. Habib19
DPRG
Laser Scanning
Conceptual procedural flow of the spiral sweep method
DSM
column
row
target point
nadir point
True Orthophoto Gen.: Spiral Sweep
Ayman F. Habib20
DPRG
Laser Scanning
Comparative Analysis
Z-buffer method
Angle-based (spiral sweep) method
Differential rectification
Angle-based (adaptive radial sweep) method
Ayman F. Habib21
DPRG
Laser Scanning
Original Image
Ayman F. Habib22
DPRG
Laser Scanning
LiDAR Surface Model
Elevation Data Intensity Data
Ayman F. Habib23
DPRG
Laser Scanning
Orthophoto with Ghost Images
Ayman F. Habib24
DPRG
Laser Scanning
True Orthophoto without Ghost Images
Ayman F. Habib25
DPRG
Laser Scanning
True Orthophoto After Occlusion Filling
Ayman F. Habib26
DPRG
Laser Scanning
perspective center
A
BInvisible point
DE
C
5 ° visible
12° visible
15° invisible
12 °
15°
14°
15°
A
B
C
D
E 20° visible
max angle
visible/ hiddenpoint angle comparison
0° visible5° >
>
>
<=
>
Nadir point
Digital Surface Model
Orthophoto
Occlusion Extension
Angle-based Method
16°
Ayman F. Habib27
DPRG
Laser Scanning
True Orthophoto After Occlusion Filling
Ayman F. Habib28
DPRG
Laser Scanning
True Orthophoto After Occlusion Extension
Ayman F. Habib29
DPRG
Laser Scanning
True Orthophoto After Boundary Enhancement
Ayman F. Habib30
DPRG
Laser Scanning
Orthophoto with Ghost Images
Ayman F. Habib31
DPRG
Laser Scanning
True Orthophoto without Ghost Images
Ayman F. Habib32
DPRG
Laser Scanning
True Orthophoto After Occlusion Filling
Ayman F. Habib33
DPRG
Laser Scanning
True Orthophoto After Occlusion Extension
Ayman F. Habib34
DPRG
Laser Scanning
True Orthophoto After Boundary Enhancement
Ayman F. Habib35
DPRG
Laser Scanning
Classification of LiDAR Data(Ground/Non-Ground Points)
Ayman F. Habib36
DPRG
Laser Scanning
LiDAR Classification: Introduction• LiDAR data includes ground/terrain and non-ground/off-
terrain points.– Knowledge of the terrain is useful for deriving contour lines,
road network planning, and flood monitoring.– Knowledge of the off-terrain points is useful for DBM detection,
DBM reconstruction, 3D city modeling, and 3D visualization. – Knowledge of terrain and off-terrain points is useful for change
detection applications.
Ayman F. Habib37
DPRG
Laser Scanning
LiDAR Classification: Introduction• Definition of ground/non-
ground (Sithole & Vosselman, 2003)– Ground: Topsoil or any thin
layering (asphalt, pavement, etc.) covering it
– Non-ground: Vegetation and artificial features
• How to distinguish ground points from non-ground points in LiDAR data?
Ground Profile
Non-Ground Profile
LiDAR Profile
Ayman F. Habib38
DPRG
Laser Scanning
LiDAR Classification: Literature• Categories (Sithole & Vosselman 2003):
– Slope-based– Block-minimum– Surface-based– Clustering/segmentation
Ayman F. Habib39
DPRG
Laser Scanning
LiDAR Classification: Literature Review• Modified Block Minimum (Wack and Wimmer, 2002)• Modified Slope-based Filter (Vosselman, 2000)• Morphological Filter (Zhang et al., 2003)• Active Contour (Elmqvist et al., 2001)• Progressive TIN Densification (Axelsson, 2000)• Robust Interpolation (Pfeifer et al., 2001)• Spline Interpolation (Brovelli et al., 2002)
Ayman F. Habib40
DPRG
Laser Scanning
LiDAR Classification: Concept• Assumption: Non-ground
objects produce occlusions in synthesized perspective views.
• Search for occlusions Non-ground objects can be detected as those causing occlusions.
Perspective Projection
Ayman F. Habib41
DPRG
Laser Scanning
LiDAR Classification: Processing Flow
Ayman F. Habib42
DPRG
Laser Scanning
LiDAR Classification: Methodology• LiDAR data is irregularly
distributed. • We start by interpolating the
LiDAR data.– The average point density is
used to estimate the optimum GSD for resampling.
– We use the nearest neighbor interpolation to avoid blurring the height discontinuities.
Point C
Point A
Point B
The jthColumn
The ith
Row
DSM( i , j ) = Height of Point B
Ayman F. Habib43
DPRG
Laser Scanning
LiDAR Classification: Methodology• If there is more than 1 point located in a given cell, we
pick the one with the lowest height and assign its height to that cell.
Point A
Point B
Point CThe jthColumn
The ith
Row
DSM( i , j ) = Height of Point C
Ayman F. Habib44
DPRG
Laser Scanning
LiDAR Classification: Methodology
• Occlusion Detection(Angle-based)
A
C
PC
Nadir Point
Off-Nadir Angle
A B C D E
B
D
E
DE E: Occlusion!!!
AB
Visible Point (B)
BC
CD
Visible Point (C)
Visible Point (D)
Last Visible Point (D)
First Occluded Point (E)
Ayman F. Habib45
DPRG
Laser Scanning
LiDAR Classification: Methodology
PC
Last Visible Point
First Occluded Point
D E
E
Nadir Point
D
C
• Detect the Points Causing Occlusion
C
B
EC
EB
C: Non-Ground
D: Non-GroundB: Ground
B
Non-Ground
Ayman F. Habib46
DPRG
Laser Scanning
LiDAR Classification: Methodology• How can we maximize our
ability to detect the majority of non-ground objects?– Manipulate the location &
number of synthesized projection center(s)
Ayman F. Habib47
DPRG
Laser Scanning
LiDAR Classification: Methodology
• Non-ground points detected from projection centers with different horizontal locations
Ayman F. Habib48
DPRG
Laser Scanning
• Non-ground points detected from projection centers with different vertical locations
LiDAR Classification: Methodology
Ayman F. Habib49
DPRG
Laser Scanning
LiDAR Classification: Methodology
• Two opposite projection centers will allow for the detection of a larger non-ground area
PC A
Detected Non-Ground Points From PC A
Detected Non-Ground Points From PC B
Combined Results
PC B
Ayman F. Habib50
DPRG
Laser Scanning
LiDAR Classification: Methodology
The eight neighbors of any given pixel are checked to see if they are occluded by that pixel or not.
Ayman F. Habib51
DPRG
Laser Scanning
A
PC1
For Pixel A
PC 2For Pixel A
PC 3For Pixel A
PC 4For Pixel A
PC 8For Pixel A
PC 7For Pixel A
PC 6For Pixel A
PC 5For Pixel A
dd
d
d
B
LiDAR Classification: Methodology
Ayman F. Habib52
DPRG
Laser Scanning
LiDAR Classification: Methodology
A
PC 1For Pixel B
PC 2For Pixel B
PC 3For Pixel B
PC 4For Pixel B
PC 8For Pixel B
PC 7For Pixel B
PC 6For Pixel B
PC 5For Pixel B
dd
d
d
B
Ayman F. Habib53
DPRG
Laser Scanning
The eight neighbors of any given pixel are checked to see if they are occluded by that pixel or not.
A
Perspective Center 1
For Pixel A
Perspective Center 2
For Pixel A
Perspective Center 3
For Pixel A
Perspective Center 4
For Pixel APerspective
Center 8For Pixel A
Perspective Center 7
For Pixel A
Perspective Center 6
For Pixel A
Perspective Center 5
For Pixel A
dd
d
d
45 Degree
B
LiDAR Classification: Methodology
Ayman F. Habib54
DPRG
Laser Scanning
LiDAR Classification: Results
Simulated Dataset DSM Identified Occluding Points (in white)
Simulated DatasetMisclassified ground points
Ayman F. Habib55
DPRG
Laser Scanning
LiDAR Classification: Methodology• Multiple projection centers at pre-specified locations will:
+ Improve our capability of detecting non-ground points• Useful when dealing with large and low buildings
– Enhance the noise and high-frequency components of the terrain• Will lead to false hypotheses regarding instances of non-ground points
• Solution: implement a statistical filter to refine the occlusion-based terrain/off-terrain classification procedure
Ayman F. Habib56
DPRG
Laser Scanning
LiDAR Classification: Methodology• Points producing occlusions (hypothesized off-terrain
point):– True non-ground points + false non-ground points
• Points not producing occlusions (hypothesized terrain point):– True ground points + false ground points
DSMIdentified Occluding Points (in white)
Less probable
Ayman F. Habib57
DPRG
Laser Scanning
LiDAR Classification: Filtering• We designed a statistical filter to remove the effects of
terrain roughness (e.g., noise in the LiDAR data and high frequency components of the surface – cliffs).
• The elevation “h” of the ground points can be assumed to be normally distributed with a mean “μ” and standard deviation “σ”.
Height
Freq
uenc
y
Ayman F. Habib58
DPRG
Laser Scanning
GroundThreshold : Threshold for modifying non-ground points
groundNonThreshold : Threshold for modifying ground points
OutlierThreshold : Threshold for detecting low outliers
Ground Non-Ground
OutlierThreshold
groundNonThreshold
GroundThreshold
Outliers
LiDAR Classification: Filtering• For each DSM cell, we define a local neighborhood that is
adaptively expanded until a pre-defined number of terrain points is located.– Derive a histogram of the terrain point elevations
Height
Freq
uenc
y
Ayman F. Habib59
DPRG
Laser Scanning
LiDAR Classification: Filtering
• Examples of outliers: multi-path errors, errors in the laser range finder
Ayman F. Habib60
DPRG
Laser Scanning
Ground Non-Ground Ground Non-Ground
GroundTARGETHeight
groundNonTARGETHeight
groundNonTARGETHeight
GroundTARGETHeight
GroundTARGETHeightKeep it
groundNonTARGETHeight
GroundTARGETHeight
OutlierTARGETHeight
Outlier
OutlierThreshold
HeightHeightFreq
uenc
yLiDAR Classification: Filtering
Ayman F. Habib61
DPRG
Laser Scanning
LiDAR Classification: Point Cloud Class.• If a cell is classified as
non-ground, all the LiDAR points in that cell are classified as non-ground points.
• If the cell is classified as a ground point, then– The lowest LiDAR point
in that cell is classified as ground.
– The LiDAR points that are at least 20 cm higher than the lowest LiDAR point are classified as non-ground points.
Point A
Point B
Point CThe i th Column
The j th Row
DSM( i , j ) = Height of Point C
Point C
(Ground)
20cmPoint B Ground
Point A Non-ground
Ayman F. Habib62
DPRG
Laser Scanning
LiDAR Classification: Results
DSM
Classification Results using filterClassification Results without filter
Simulated Dataset
Ayman F. Habib63
DPRG
Laser Scanning
LiDAR Classification: Results
Real Dataset (1 - Brazil)
Ayman F. Habib64
DPRG
Laser Scanning
LiDAR Classification: Results
Real Dataset (1 - Brazil)
Occluding points in white
Ayman F. Habib65
DPRG
Laser Scanning
LiDAR Classification: Results
Real Dataset (1 - Brazil)
After Statistical Filtering
Ayman F. Habib66
DPRG
Laser Scanning
LiDAR Classification: Results
DSM → Non-ground objects
Real Dataset (1 - Brazil)
Ayman F. Habib67
DPRG
Laser Scanning
LiDAR Classification: Results
• Using the LiDAR DSM and an orthophoto over the same area, we manually generated a ground truth for ground and non-ground points classification.
• Comparing our result with the ground truth, the number of misclassified points divided by the total number of points was found to be 4.7%.
Real Dataset (1 - Brazil)
Ayman F. Habib68
DPRG
Laser Scanning
LiDAR Classification: Results
Misclassified Points Misclassified Points displayed on DSM
Real Dataset (1 - Brazil)
Ayman F. Habib69
DPRG
Laser Scanning
LiDAR Classification: Results
Real Dataset (1 - Brazil)
Original DSM Derived DTM
Ayman F. Habib70
DPRG
Laser Scanning
LiDAR Classification: Results
DSM Occluding Points Non-ground Points
Discontinuous Terrain: Tunnels
Real Dataset (2 - Stuttgart)
Ayman F. Habib71
DPRG
Laser Scanning
DSM
Occluding Points
Non-ground Points
LiDAR Classification: Results
Real Dataset (2 - Stuttgart)
Ayman F. Habib72
DPRG
Laser Scanning
LiDAR Classification: Results
DSM Occluding Points Non-ground PointsReal Dataset (2 - Stuttgart)
Ayman F. Habib73
DPRG
Laser Scanning
LiDAR Classification: Results
• A ROI near the University of Calgary is selected as an experimental data.
• The Transit Train trail extends into a tunnel under the ground.
Real Dataset (3 - Calgary)
Ayman F. Habib74
DPRG
Laser Scanning
LiDAR Classification: Results
Non-ground points (TerraScan) Non-ground points (Occlusion-based)
Real Dataset (3 - Calgary)
Ayman F. Habib75
DPRG
Laser Scanning
LiDAR Classification: Results
TerraScan’s Result
Occlusion-Based Result
Real Dataset (3 - Calgary)
Ayman F. Habib76
DPRG
Laser Scanning
LiDAR Classification: Results
• Another ROI near the University station is selected as another experimental data.
• Complex contents – The Transit Train station,– Bridge,– Ramps, and– Trees.
Real Dataset (3 - Calgary)
Ayman F. Habib77
DPRG
Laser Scanning
Non-ground points (TerraScan) Non-ground points (Occlusion-Based Results)
LiDAR Classification: Results
Real Dataset (3 - Calgary)
Ayman F. Habib78
DPRG
Laser Scanning
LiDAR Classification: Results
TerraScan’s Result
Occlusion-Based Result
Real Dataset (3 - Calgary)
Ayman F. Habib79
DPRG
Laser Scanning
Original LiDAR Points over UofC
LiDAR Classification: Results
Real Dataset (4 - Calgary)
Ayman F. Habib80
DPRG
Laser Scanning
Aerial Photo over UofC
LiDAR Classification: Results
Real Dataset (4 - Calgary)
Ayman F. Habib81
DPRG
Laser Scanning
Ground/Non-Ground Points
Real Dataset (4 - Calgary)
LiDAR Classification: Results
Ayman F. Habib82
DPRG
Laser Scanning
LiDAR Classification: Conclusion• The achieved results proved the feasibility of the
suggested procedure.• Default parameters are sufficient for most cases.• The proposed procedure is capable of handling urban
areas with complex contents:– Tall buildings, low and nearby buildings, trees, bushes, fences,
bridges, ramps, cliffs, tunnels, etc.• Future work will focus on further testing of the proposed
methodology as well as improving its efficiency.• Also, the classified non-ground points will be further
classified into vegetation and man-made structures.– Building detection and change detection
Ayman F. Habib83
DPRG
Laser Scanning
Comments and Questions?