jiu xu, axel beaugendre and satoshi goto computer sciences and convergence information technology...

45
REAL-TIME HUMAN TRACKING BY DETECTION BASED ON HOG AND PARTICLE FILTER Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time Human Tracking by Detection based on HOG and Particle Filter

Upload: emory-porter

Post on 29-Jan-2016

218 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

1

REAL-TIME HUMAN TRACKING BY

DETECTION BASED ON HOG AND

PARTICLE FILTER

Jiu XU, Axel BEAUGENDRE and Satoshi GOTOComputer Sciences and Convergence Information

Technology (ICCIT), 2011 6th International Conference on

Page 2: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

2

Outline Introduction Proposed Method

Human DetectionMoving object feature, Hierarchical search algorithm

Human TrackingColor-EdgeTexture histogram, Occlusion Handling

Strategy

Experimental Results Conclusion

Page 3: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

3

Introduction The real-time foreground objects

tracking and detection is the most critical and fundamental step in video surveillance systems.

Different from vehicles, it is much more difficult to locate and track the human body out of the background.

Page 4: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

4

Introduction

Pedestrian detection by Histogram of Oriented Gradient (HOG) put forward by Dalal [1] [2].

The main idea of this method is to use gradient direction histogram of small pieces to describe the image.

[1] N. Dalal, "Finding People in Images and Videos," PhD thesis, Institut National Polytechnique de Grenoble, 2006.[2] N. Dalal and B. Triggs, "Histogram of oriented gradient for human detection," in CVPR, 2005.

Detection

Page 5: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

5

Introduction Feature extraction by adjusting the

parameters, the HOG feature can effectively describe the body’s shape information.

Moreover, it also has the ability of invariance when the small local area occurs dithering and rotation.

Page 6: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

6

Introduction However, the computational complexity

of feature extraction makes the method very slow, and difficult to meet the needs of the practical application of the system.

Page 7: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

7

Introduction

Some common methods perform tracking by pre-initialized trackers based on Kalman filter [6] or particle filter [7].

Tracking

[6] Gutman, P., Velger, M. “Tracking Targets Using Adaptive Kalman Filtering”, IEEE Transactions on Aerospace and Electronic Systems Vol. 26, No. 5: pp. 691-699 1990.[7] B.Ristic, “Beyond the Kalman Filter: Particle Filters for Tracking Applications”. Arthech House, 2004.

Page 8: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

8

Introduction Particle filter is based on the Bayes

principle and is a sequential Monte-Carlo simulation method indicated by probability density of particles.

Page 9: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

9

Introduction one of the intractable problems is that

the target is usually occluded by other objects frequently and instantaneously.

To select powerful features to avoid hijacking problems when tracking similar objects.

If several humans walk together as a group, we cannot separate them individually.

Page 10: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

10

Introduction The target of this paper is to build a

powerful tracking system for real-time surveillance system.

Combining human detection and tracking and do tracking by detection in order to achieve high accuracy and low time consumption together with occlusion solutions and group segmentations.

Page 11: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

11

Outline Introduction Proposed Method

Human DetectionMoving object feature, Hierarchical search algorithm

Human TrackingColor-EdgeTexture histogram, Occlusion Handling

Strategy

Experimental Results Conclusion

Page 12: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

12

Proposed Method

a Tracking by Detection System using human detection together with human tracking

Page 13: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

13

Human Detection Histogram of Oriented Gradient (HOG)

is quite a popular method of detecting people in static image.

Page 14: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

14

Human Detection The image is first divided into blocks

while these blocks overlap with each other. Each block contains four cells.

Page 15: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

15

Human Detection For each pixel I(x,y), the

orientation Θ(x,y) and the magnitude m (x, y) of the gradient

Page 16: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

16

Human Detection A histogram is calculated for each cell,

and the volume of each bin is the sum of magnitude of the pixels whose orientations are in the corresponding angle interval.

Page 17: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

17

Human Detection However, HOG can’t be used into real-

time system due to its high time consumption.

Improvements Moving object feature Hierarchical search algorithm

Page 18: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

18

Human Detection Moving object feature Kim [10] presented a new kind of non-

parametric algorithm for background subtraction.

For each pixel, it builds a codebook consisting of one or more codewords.

[10] K. Kim, T. H. Chalidabhonse, D. Harwood, and L. Davis. “Real-time foreground-background segmentation using codebook model”. Elsevier Real-Time Imaging, vol. 11, no.3, 167–256, June 2005.

Page 19: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

19

Human Detection In our previous work [13], a block-based

codebook method has been proposed to make it more adoptable for human tracking.

The block feature heighten the pixel-based method to block level, thus we could take advantage of the relationship between neighboring pixels.

[13] Jiu Xu, Ning Jiang, Satoshi Goto, “Block-based Codebook Model with Oriented-Gradient Feature for Real-time Foreground Detection”, IEEE 13th International Workshop on Multimedia Signal Processing(MMSP), 2011.

Page 20: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

20

Human Detection

morphological operation(opening, closing)

Page 21: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

21

Human Detection Hierarchical search algorithm The expected target might appear at

any position and the size would also keep changing

Page 22: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

22

Human Detection From our experiments, we find that the

curve of the number of detected pedestrians under various scaling satisfied the curve of normal distribution probability density function.

We could estimate the minimum interval [a, b].

Image scale levels only from a to b should be detected, reduce the searching levels and the amount of computation as well.

Page 23: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

23

Human Tracking

Page 24: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

24

Human Tracking Particle tracker initialization

Page 25: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

25

Transition model p()- how objects move between frames. Observation model p()- specifies the likelihood of an object being in a specific state. Initial state p()- describes initial distribution

Human Tracking

Page 26: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

26

Human Tracking Color-EdgeTexture histogram We proposed a Color-EdgeTexture

Histogram to generate the weight for the observation models.

We choose the HSV color. The brightness variations while in HSV color space we could better separate the brightness with others.HSV = Hue, Saturation, Value

Page 27: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

27

Human Tracking Moreover, we add the edge local binary

pattern to describe the shape texture of the moving objects.

Local binary pattern (LBP) [14] is an effective texture description operator, which could be used to measure and extract texture information from the local neighborhood in a gray image.

[14] T. Ojala, M. Pietikainen, and D. Harwood, “A Comparative Study of Texture Measures with Classification Based on Feature Distributions”, Pattern Recognition, vol. 29, pp. 51-59. 1996

Page 28: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

28

Human Tracking Local binary pattern (LBP)Consider a pixel (, )

Page 29: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

Human Tracking LBP value together with H-S color

information improve the performance Two problems :1. time consumption. Since LBP is a pixel-

wise coding, if we calculate all the LBP value inside the whole regions of predicted position of the particles, the calculation is really quite huge, thus it will extremely decrease the real-time ability.

Page 30: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

Human Tracking2. the LBP value for background part of

the regions within the trackers are also calculated, it will greatly reduce the tracking rate when the size of the object is changing, and the portion of the background becomes larger and larger since the weight of the background is increased.

Page 31: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

Human Tracking To solve these problems, we use a

concept of edge LBP and the edge LBP only focuses on the edge points of the foreground objects.

canny edge detector

Page 32: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

Human Tracking we use a kind of H-S-

ForegroundEdgeLBP histogram in which the size is 8x8x8 for each component.

H-S-LBP histogram

Page 33: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

Human Tracking Particles are weighted according to the

similarity between the target histogram distribution q(u) and the histogram distributions p(u) given by particles.

Page 34: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

Human Tracking the weight of the i-th particle is defined

as

color histogram of the target at time k histogram of the i-th particle

Page 35: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

35

Human Tracking Occlusion Handling Strategy In traditional particle filter tracking, if

the object meets some partial or total occlusions, the observation model will turn to the occluder and will not track the previous objects any longer.

Page 36: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

36

Human Tracking After color-edgetextue histogram, we

define a threshold to this weight.

① If the tracker moves out of the margin of the frame

② If the tracker is still inside the frame and the max weight is great than the threshold

③ If the tracker is still inside the frame and the max weight is less than the threshold

Delete

Update

Keep increase the number of the particles, searching range

Page 37: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

37

Outline Introduction Proposed Method

Human DetectionMoving object feature, Hierarchical search algorithm

Human TrackingColor-EdgeTexture histogram, Occlusion Handling

Strategy

Experimental Results Conclusion

Page 38: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

38

Experimental Results

Other method

Page 39: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

39

Experimental Results

Other method

Page 40: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

40

Experimental Results

Other method

Page 41: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

41

Experimental Results

Separate group

Page 42: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

42

Experimental Results

The time consumption of our method is much less than [8][9][11].

(400 frames long)

Page 43: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

43

[8] L.M.Fuentes and S.A.Velastin, “People tracking in surveillance applications”, Image and Vision Computing, pp.1165-1171, 2006

[9] Tao Yang, Quan Pan, Jing Li and Li, S.Z. “Real-time Multiple Objects Tracking with Occlusion Handling in Dynamic Scene” In CVPR, Vol.1. pp. 970-975, 2005

[11] R. Hess and A. Fern, “Discriminatively Trained Particle Filters for Complex Multi-Object Tracking”. In CVPR, 2009.

Experimental Results

Page 44: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

44

Uses a combination strategy of HOG human detection method and particle filter human tracking algorithm in surveillance system.

Original HOG feature is not be so efficient in real-time system

In order to enhance the performance in color-based particle tracking

The proposed method has a good robustness in all kinds of situations together with low time consumption.

Conclusion

Page 45: Jiu XU, Axel BEAUGENDRE and Satoshi GOTO Computer Sciences and Convergence Information Technology (ICCIT), 2011 6th International Conference on 1 Real-time

Real-time Human Tracking by Detection based on HOG and Particle Filter

45

Thanks for listening !!