Upload
hilary-hudson
View
220
Download
6
Tags:
Embed Size (px)
Citation preview
Pedestrian Detection and Localization
Members:Đặng Trương Khánh Linh 0612743
Bùi Huỳnh Lam Bửu 0612733
Advisor:
A.Professor Lê Hoài Bắc
UNIVERSITY OF SCIENCE
ADVANCED PROGRAM IN COMPUTER SCIENCE
Year 20111
Outline Introduction
Problem statement & application Challenges Existing approaches. Review HOG and SVM Motivation
Overview of methodology Learning phase Detection phase
Our contributions: Spatial selective approach Multi-level based approach
Fusion Algorithm - Mean Shift Conclusions Future work Reference
2
Problem statement
Build up a system which automatically detects and localizes pedestrians in static image.
Pedestrians: up-right and fully visible.
3
Applications
Automated Automobile Driver , or smart camera in general.
Build a software to categorize personal album images to proper catalogue.
Video tracking smart surveillance.Action recognition.
4
Challenges
5
• Huge variation in intra-class.
• Non-constraints illumination.
• Variable appearance and clothing.
• Complex background.
• Occlusions, different scales.
Existing approachesHaar wavelets + SVM: Papageorgiou &
Poggio, 2000; Mohan et al 2000Rectangular differential features + adaBoost:
Viola & Jones, 2001Model based methods: Felzenszwalb 2008
Local Binary Pattern: Wang 2009
Histogram of Oriented Gradients: Dalal and Trigg 2005
6
FelzenszwalbCVPR 2008
Wang ICCV 2009
Histogram of Oriented Gradients (HOG)
Base on the gradient of pixels.
7
Histogram of Oriented Gradients Review
8
9 orientation bins0 - 180° degrees
cell
9x4 feature vector per cell
normalizeFeature vector
f = […,…,…, ,…]
block
Histogram of Oriented Gradients Review
9
9 orientation bins0 - 180° degrees
cell
9x4 feature vector per cell
normalizeFeature vector
f = […,…,…, ,…]
block
Histogram of Oriented Gradients Review
10
9 orientation bins0 - 180° degrees
cell
9x4 feature vector per cell
normalizeFeature vector
f = […,…,…, ,…]
block
Histogram of Oriented Gradients Review
11
9 orientation bins0 - 180° degrees
cell
9x4 feature vector per cell
normalizeFeature vector
f = […,…,…, ,…]
block
SVM Review
12
1xw
1xw
1c
1c
1)( ii xwc
1)( subject to 2
1 minimize
2 ii xwcw
w
2
Motivation of choosing HOG
The blob structure based methods have fail to object detection problem.
Object detection methods via edge detection are unreliable.
Use the advantage of rigid shape of object.Has a good performance and low complexity.Disadvantages:
Very high dimensional feature vector. Lack of multi-scale shape of object.
13
Contributions
Re-implement HOG-based pedestrian detector.Spatial Selective Method.Multi-level Method.
14
Dataset
INRIA pedestrian dataset
Train:1208 positive windows1218 negative images
Test:566 positive windows453 negative images
15
Dataset
16
Positive images Negative images Positive windows Negative windows
Overview of methodology
Learning Phase: Input: positive windows and negative images. Output: binary classifier
Detection Phase: Input: arbitrary image. Output: bounding boxes containing
pedestrians.
17
Learning Phase
18
Detection Phase
19
Result of re-implementation
20
Examples
21
Contributions
IMPROVEMENTS
PERFORMANCE(Spatial Selective Approach)
ACCURACY(Multi-Level Approach)
22
Spatial Selective Approach
Less informative region
[A1,..,Z1] [A2,..,Z2] [A3,..,Z3] [A4,..,Z4]
[A1,..,Z1, A2,..,Z2, A3,..,Z3, A4,..,Z4]
Descriptor
23
Spatial Selective Approach
Examples:
24
Result Spatial Selective Method
25
Vector Length v.s Speed
Origin 1:0 1:1 1.5:0 1.5:10
500
1000
1500
2000
2500
3000
3500
4000
0
1
2
3
4
5
6
Vector lengthSpeed (in minutes)
A:B Deleted cell(s): Overlap cell(s)26
Examples
27
Original Re-implement
Multi-level Approach
Purpose: enhance the performance by getting more information about shape of object.
Inspired by pyramid model
28
Multi-level Approach
[A1,..,Z1] [A2,..,Z2] [A3,..,Z3]
[A1,..,Z1, A2,..,Z2, A3,..,Z3, A4,..,Z4]
HOG
29
Result(cont…)
30
Feature vector length v.s Time
31
3000 3500 4000 4500 5000 5500 6000 65004
5
6
7
8
9
10
11
12
Original
3 levels
5 levelsVector length v.s Time
Examples
32
Original Multi-level
Examples (cont)
33
Fusion Algorithm – Mean Shift
34
Region ofinterest
Center ofmass
Mean Shiftvector
Mean shift
Slide by Y. Ukrainitz & B. Sarel35
Region ofinterest
Center ofmass
Mean Shiftvector
Mean shift
Slide by Y. Ukrainitz & B. Sarel36
Region ofinterest
Center ofmass
Mean Shiftvector
Mean shift
Slide by Y. Ukrainitz & B. Sarel37
Region ofinterest
Center ofmass
Mean Shiftvector
Mean shift
Slide by Y. Ukrainitz & B. Sarel38
Region ofinterest
Center ofmass
Mean Shiftvector
Mean shift
Slide by Y. Ukrainitz & B. Sarel39
Region ofinterest
Center ofmass
Mean Shiftvector
Mean shift
Slide by Y. Ukrainitz & B. Sarel40
Region ofinterest
Center ofmass
Mean shift
Slide by Y. Ukrainitz & B. Sarel41
Mean shift clustering
• Cluster: all data points in the attraction basin of a mode
• Attraction basin: the region for which all trajectories lead to the same mode
Slide by Y. Ukrainitz & B. Sarel42
Non-maximum suppression
Using non-maximum suppression such as mean shift to find the modes.
43
Conclusions
Successfully re-implement HOG descriptor.Propose the Spatial Selective Approach which take
advantages of less informative center region of image window.
Multi-level has more information about shape of object.
It is general model, and can apply to any object.
44
Future work
Non-uniform grid of points.Combination of Spatial Selective and Multi-level
approach.
45
Non-uniform grid of points
46
Demonstration
47
References
N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in IEEE Conference on Computer Vision and Pattern Recognition, 2005.
Subhransu Maji et al. Classification using Intersection Kernel Support Vector Machines is Efficient. IEEE Computer Vision and Pattern Recognition 2008
C. Harris and M. Stephens. A combined corner and edge detector. In Alvey Vision Conference, pages 147–151, 1988.
D. G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91–110, 2004.
48
Scan image at all positions and scales
Object/Non-object
classifier
49
Miss rate = 1 – recall
50
True FalseTrue tp fnFalse fp tn
Overview of methodology
51