vision-based multiple vehicle detection and tracking for driver assistant system

32
Qing Ming [email protected] May. 11. 2012 Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

Upload: lavada

Post on 18-Mar-2016

41 views

Category:

Documents


3 download

DESCRIPTION

Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System. Qing Ming [email protected] May. 11. 2012. Vision Based Driver Assistant System. Problem setting. Main Goal : Multiple vehicle detection and tracking. Challenge work:. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

Qing [email protected]

May. 11. 2012

Vision-Based Multiple Vehicle Detection and Tracking for Driver

Assistant System

Page 2: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

2Intelligent Systems

Lab.

Vision Based Driver Assistant System

Page 3: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

3Intelligent Systems

Lab.

Problem setting

Main Goal: Multiple vehicle detection and tracking

Challenge work:-Different environment(illumination, distance, background…)-Tracking windows scale dynamic adjustment -Vehicle partial occlusion-Vehicle temporary missing

Page 4: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

4Intelligent Systems

Lab.

yes

System architecture

Frame 1

Vehicle detection unit

Detectednew vehicle?

yes

Frame 2

Vehicle detection unit

Detectednew vehicle?

yes

Frame n

Vehicle detection unit

Detectednew vehicle?

Particle filter 1

Internal storage

Particle filter 2

Particle filter m

Vehicle Tracking Unit Return to Vehicle

detection unit

Page 5: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

5Intelligent Systems

Lab.

Offline process

Vehicle Detection Algorithm

Image Sequence

Gabor FeatureExtraction

BP Neural NetworkTraining

BP Neural NetworkClassifier

Test Image

Vehicle Candidate Detection

Vehicle Candidate Verification

Detected Vehicle

Page 6: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

6Intelligent Systems

Lab.

Color segmentation

Morphological operation

Original image

Vehicle candidate detection

+

Threshold obtained by tail light image statistical value

Page 7: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

7Intelligent Systems

Lab.

Vehicle candidate generation

hc2

1 2min maxc cw w w 1 2c ch h d

wc1c2: width between vehicle light pair.

hc1, hc2: height of C1 and C2

d : a constant which depends on image size

hc1

wc1c2

Page 8: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

8Intelligent Systems

Lab.

Gabor feature2 2

2 212 ( )1(x,y, , ) exp exp

2x y

x y

j x y

x y

G

8 orientations5 scale

Tai Sing LEE, “Image Representation using 2D Gabor Wavelets,”IEEE Transactions on Parrern Analysis and Machine Intelligence,Vol 18,No.10, pp959-971,October 1996

Page 9: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

9Intelligent Systems

Lab.

Training

Back propagationNeural network

Non-vehicle database

vehicle database

-1

1

Stuart Russell and Peter Norvig, “Artificial Intelligence A Modern Approach”. p. 578. 1969

Page 10: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

10Intelligent Systems

Lab.

training

(No)

(Yes)

(Yes)BPNNclassifier

Gabor feature set

Vehicle candidate verification

Page 11: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

11Intelligent Systems

Lab.

Vehicle Tracking

Frame t

Frame t+1

Detected vehicle

Histogram generation

Colorhistogram

Particle generation

Particle selection

Similaritycomputation

TrackingWindow

estimation

Histogram updating

Updatedhistogram …

Page 12: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

12Intelligent Systems

Lab.

Target Vehicle Representation

Frame t

Color space representation

Detected vehicle

Histogram representation

Split into uniform Histogram bins

Page 13: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

13Intelligent Systems

Lab.

Color PDF

Particle Generation

Frame t+1 Particle selection Each Particle is

consider as one pixel

• Randomly generate particles at the position of tracking window in previous frame

Page 14: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

14Intelligent Systems

Lab.

Similarity Computing

……

•Bhattacharyya coefficient•Mean state of the particle set

Each selected Particle is consider as one region

Dorin Comaniciu, Visvanathan Ramesh, P eterMeer, "Real-Time Tracking of Non-Rigid Objects using Mean Shift, " IEEE Conference on Computer Vision and Pattern Recognition, June 13 -15, Hilton Head, SC, USA, 2000

Page 15: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

15Intelligent Systems

Lab.

Partial occlusion and temporary missing

Frame 1 Frame 45 Frame 25

Frame 120 Frame 80 Frame 148

Target vehicle Partial occlusion Temporary missing

Temporary missing Partial occlusion Target vehicle re-tracking

Page 16: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

16Intelligent Systems

Lab.

Temporary missing

Frame 80

Particles are generated nearby the covering vehicle bounding box

effective particles are searched

Frame 125

When enough effective particles are searched, the missing vehicle start Tracking again

Page 17: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

17Intelligent Systems

Lab.

Color Histogram Updating

Target vehicle color histogram changing under different condition

(ex: different distance, different illumination…)

Page 18: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

18Intelligent Systems

Lab.

Color Histogram Updating

Frame 1 Frame 55 Frame 62 Frame 115

Frame 1 Frame 55 Frame 62 Frame 115

Tracking without color histogram updating

Tracking with color histogram updating

Page 19: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

19Intelligent Systems

Lab.

Detection result

High way Urban Road Campus

Page 20: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

20Intelligent Systems

Lab.

Multiple vehicle detection result

Image originHigh way Urban road Campus

Total number of vehicles 56 42 36

Number of vehicle Correct detection 43 34 30

Number of vehicle fail detection 13 8 6

Detect rate (%) 76.8 80.1 83.3

Page 21: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

21Intelligent Systems

Lab.

Tracking result

Page 22: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

22Intelligent Systems

Lab.

Tracking result

Horizontal trajectory Vertical trajectory

Page 23: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

23Intelligent Systems

Lab.

Tracking result

Trajectory in image plane Tracking error

Page 24: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

24Intelligent Systems

Lab.

Partial occlusion and temporary missing result

Page 25: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

25Intelligent Systems

Lab.

Multiple vehicle tracking result

Video origin

High way Urban Road Campus

Total number of frames 5254 952 1326

Image Size 640×480 1920×1080 640×480

Tracking window size (in pixel) 40×32-130×104 60×48-320×256 32×25-130×104

AVG. of tracking trajectory error (in pixels) 10 23 7

Total number of moving vehicle 12 5 6

Number of miss tracking vehicle 3 1 0

Page 26: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

26Intelligent Systems

Lab.

Conclusion

Detected multiple vehicles in different environment (different light condition, different size vehicle, different speed) Tracked partial occluded vehicle Re-tracked temporary missing vehicle Tracking windows dynamically adapt according to target vehicle scale changing Color histogram self-updating

Advantage

Disadvantage Only color model based multiple vehicle tracking is not suitable for same color Vehicle occlusion problem

Future works -Camera stabilization for smooth trajectory generation-Combine with odometry information to predict dangerous situations

Page 27: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

27Intelligent Systems

Lab.

Publications Qing Ming, Kang-Hyun Jo, “Vehicle Detection Using Tail Light Segmentation,” International Forum on Strategic Technology, August 22-24, Harbin, 2011.Ming Qing and Kang-Hyun Jo, “Vehicle Detection and Scale-adaptive Tracking Using Tail Light Segmentation,” proc. of image and vision computing New Zealand, pp. 115-119, 2011.Ming Qing and Kang-Hyun Jo, “A novel particle filter implementation for Multiple-Vehicle Detection and Tracking System using Tail Light Segmentation”, International Journal of Control, Automation, and Systems (Reviewing)Ming Qing and Kang-Hyun Jo, “Vision Based Multiple Vehicle Detection and Tracking Using Tail Light Segmentation”, IECON Montreal (reviewing), October 25-28, 2012 Mecatronics 2012, Paris, Date line: 30 May,2012

Page 28: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

28Intelligent Systems

Lab.

Page 29: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

29Intelligent Systems

Lab.

BPNNBack propagation nerual network:

Xi : input Zi :output t1: expect output wij : weights between input layer and hidden layer wjk : weights between hidden layer and output layer. The input goes through the neural network in order to obtain the forward propagation’s output. Compare with expect output, difference value backward propagation through network to update weights

Page 30: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

30Intelligent Systems

Lab.

HSV color model

H: HueS: SaturationV: Value

Page 31: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

31Intelligent Systems

Lab.

Particle filter

Particle filter:

PropagateParticle generation evaluation …

p(xt|zt) p(xt+1|zt+1)p(xt+1|zt)

Zt+1

Zt: observation

P(xt|Zt): particle state under current observation

P(xt+1|Zt): particle state prediction

Propagate

p(xt+2|zt+1)

Page 32: Vision-Based Multiple Vehicle Detection and Tracking for Driver Assistant System

32Intelligent Systems

Lab.

Bhattacharyya coefficient

Bhattacharyya coefficient is an approximate measurement of the amount of overlap between two statistical samples. This coefficient can be used to describe the similarity of two discrete and normalized distributions.

1

[ , ]m

u uu

p q p q