26
Automatic Car Detection Using Statistic-based Boosting Cascade Weilong Yang, Wei Song, Zhigang Qiao, Michael Fang 1

Automatic Car Detection

Embed Size (px)

DESCRIPTION

Automatic Car Detection. Using Statistic-based Boosting Cascade. Weilong Yang, Wei Song, Zhigang Qiao, Michael Fang. System Overview. Haar -like feature generation Integral image AdaBoost (feature selection) Limitations Statistic-based boosting ( StatBoost ) Cascade structure - PowerPoint PPT Presentation

Citation preview

Page 1: Automatic Car Detection

1

Automatic Car Detection

Using Statistic-based Boosting Cascade

Weilong Yang, Wei Song, Zhigang Qiao, Michael Fang

Page 2: Automatic Car Detection

2

Haar-like feature generation◦ Integral image

AdaBoost (feature selection)◦ Limitations

Statistic-based boosting (StatBoost) Cascade structure Performance evaluation

System Overview

Page 3: Automatic Car Detection

3

Features:

Training Images Set:

500 negative samples (w/o cars)

500 positive samples (with cars)

Training Image size: 40X100 pixels

imresize(X, 0.35), resize ratio, 0.35

Feature Templates:

Rectangle features were used.

Total 9 types of features were used

Integral graph was used to accelerate the calculation of rectangle features.

Page 4: Automatic Car Detection

4

Features Templates:

Edge Features: (1) (2)

Line Features: (3) (4)

Diagonal Line Features:

(5)

Corner Features: (6) (7) (8) (9)

Page 5: Automatic Car Detection

Integral Image

B D

A C

4

3

2

1

(x ,y )

yyxx

yxiyxii','

)','(),(

The sum within D can be computed as: 4 + 1 − (2 + 3).

5

Page 6: Automatic Car Detection

6

Features in 14X35 windows

The feature templates can be laid at any position with any possible size in the 14X35 windows.

One possible position and size of some type of feature define a feature. All the possible features form the total feature set of the 14X35 windows.

Total 177660 features

Page 7: Automatic Car Detection

7

The objective is to learn a sequence of best weak classifiers and the best combining weights

Each weak classifier is constructed based on one feature (1D).

Key Concepts of AdaBoost

Page 8: Automatic Car Detection

8

AdaBoost (illustrated)

Weak classifier 3

Final classifier: linear combination of the weak classifiers

WeightsIncreased

Weak Classifier 2

Weak Classifier 1

Page 9: Automatic Car Detection

9

Formal Procedure of AdaBoost

Page 10: Automatic Car Detection

10

It takes WEEKS to train!

Computational Issue

Page 11: Automatic Car Detection

11

The Bottleneck

Factor

Description Common value

N number of examples

10,000

M number of weak classifiers in total

4,000 - 6,000

T number of Haar-like features

40,000

A view of a face detector training algorithm

for weak classifier m from 1 to M:…update weights – O(N)for feature t from 1 to T:

compute N feature values – O(N)sort N feature values – O(N log N)train feature classifier – O(N)

select best feature classifier – O(T)…

Training of the weak classifier

Figure courtesy to Minh-Tri Pham, 2007

Page 12: Automatic Car Detection

12

Train the weak classifiers using statistics Assume the feature values of each class are

of normal distribution

Fast-StatBoost (1)

Non-car

Car

Optimalthreshold

Featurevalue

Figure courtesy to Minh-Tri Pham, 2007

Page 13: Automatic Car Detection

13

Fast-StatBoost (2)

A view of our face detector training algorithm

for weak classifier m from 1 to M:…update weights – O(N)Extract statistics of integral image – O(Nd2)for feature t from 1 to T:

project statistics into 1D – O(1)train feature classifier – O(1)

select best feature classifier – O(T)…

Constant Time

Figure courtesy to Minh-Tri Pham, 2007

Page 14: Automatic Car Detection

14

Motivation◦ Increase detection performance◦ Reduce computation time

Key insight◦ Use simpler classifiers to reject the majority of

sub-windows◦ Use more complex ones to achieve low false

positive rates on the rest of them

The Cascade

Page 15: Automatic Car Detection

15

Overall form: degenerate decision tree Reflects the fact: within any single image an

overwhelming majority of sub-windows are negative [1]

The Cascade (illustrated)

All Sub-windows

1 2 3

Rejected Sub-windows

Further Processing

T T T

FF F

Page 16: Automatic Car Detection

16

False positive rate

Detection rate

A Cascade of Classifiers

Page 17: Automatic Car Detection

17

Three-layer cascade Each layer has 4, 11, and 73 weak

classifiers, respectively

Our Cascaded Car Detector

Page 18: Automatic Car Detection

18

Training Set (1)

Number of car images: 500

(Courtesy to UIUC Image Database for Car Detection)

Page 19: Automatic Car Detection

19

Training Set (2)

Number of non-car images: 500

(Courtesy to UIUC Image Database for Car Detection)

Page 20: Automatic Car Detection

20

First Three Features

1st Feature 2nd Feature

3rd Feature

Page 21: Automatic Car Detection

21

Distribution of the First Three Features

Page 22: Automatic Car Detection

22

Car Detection (Uniform Scale)

Page 23: Automatic Car Detection

23

Car Detection (Different Scales)

Page 24: Automatic Car Detection

24

False Positives

Page 25: Automatic Car Detection

25

The ROC curveTest results on 170 Images:

Page 26: Automatic Car Detection

26

References

1. P. Viola and M. Jones, “Rapid Object Detection Using a Boosted Cascade of Simple Features,” In Proc. CVPR, 2001.

2. M. Pham, T. Cham, “Fast training and selection of Haar features using statistics in boosting-based face detection.” In Proc. ICCV 2007