System and method for detection of multi-view/multi-pose objects

• US 8,391,592 B2
• Filed: 06/20/2011
• Issued: 03/05/2013
• Est. Priority Date: 06/13/2006
• Status: Active Grant

• Current Assignee: SRI International, Inc.
• Original Assignee: SRI International, Inc.
• Inventors: Han, Feng, Shan, Ying, Sawhney, Harpreet Singh, Kumar, Rakesh
• Primary Examiner(s): Werner, Brian P
• Application Number:
  US13/134,885
• Publication Number:
  US 20120002869A1
• Time in Patent Office: 624 Days
• Field of Search: None
• US Class Current: 382/159
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Abstract

The present invention provides a computer implemented process for detecting multi-view multi-pose objects. The process comprises training of a classifier for each intra-class exemplar, training of a strong classifier and combining the individual exemplar-based classifiers with a single objective function. This function is optimized using the two nested AdaBoost loops. The first loop is the outer loop that selects discriminative candidate exemplars. The second loop, the inner loop selects the discriminative candidate features on the selected exemplars to compute all weak classifiers for a specific position such as a view/pose. Then all the computed weak classifiers are automatically combined into a final classifier (strong classifier) which is the object to be detected.

Claims

1. A computer implemented method of training an image classifier comprising:

segmenting a candidate exemplar into a number of grid cell image regions;

computing a gradient orientation histogram for at least one of the grid cell image regions of the candidate exemplar, where the candidate exemplar comprises at least one image of an object to be detected;

computing a gradient orientation histogram for a plurality of training samples;

calculating a distance value between the gradient orientation histogram for at least one of the grid cell image regions of the candidate exemplar and the gradient orientation histogram for the plurality of training samples; and

training the image classifier based on the distance value.

2. The method of claim 1 wherein the plurality of training samples include additional images of the at least one object to be detected.

3. The method of claim 2 further comprising randomly selecting a subset of positive samples from the plurality of training samples to create the candidate exemplar, wherein the positive training samples comprise images of an object to be detected using the trained image classifier.

4. The method of claim 1 wherein said distance value represents commonality of image features between the candidate exemplar and the plurality of training samples.

5. The method of claim 1 wherein the classifier is a weak classifier.

6. The method of claim 5 further comprising a plurality of weak classifiers and, after being trained, combining the plurality of weak classifiers to form a final classifier.

7. The method of claim 5 wherein the final classifier represents a multi-view, multi-pose object to be detected.

8. A computer implemented method of training and using a plurality of weak classifiers to form an image classifier comprising:

segmenting a candidate exemplar into a plurality of grid cell image regions, where the candidate exemplar comprises at least one image of an object to be detected;

computing a gradient orientation histogram for at least one of the grid cell image regions of the candidate exemplar;

computing a gradient orientation histogram for a plurality of training samples;

calculating a distance value between the gradient orientation histogram for the at least one of the grid cell image regions of the candidate exemplar and the gradient orientation histogram for the plurality of training samples; and

training the plurality weak classifiers based on the distance value.

9. The method of claim 8 wherein the plurality of training samples include additional images of the at least one object to be detected.

10. The method of claim 8 further comprising randomly selecting a subset of positive samples from the plurality of training samples to create the candidate exemplar, wherein the positive training samples comprise images of an object to be detected using the trained image classifier.

11. The method of claim 8 wherein the distance value represents commonality of image features between the candidate exemplar and the plurality of training samples.

12. The method of claim 8 further comprising:

computing error rates for each of the trained weak classifiers;

selecting the trained weak classifier with the lowest error rate as an individual classifier;

continuing to train the weak classifiers until all of the weak classifiers in the plurality of weak classifiers have been selected as an individual classifier;

combining the individual classifiers into a final classifier, wherein said final classifier is the object to be detected.

13. The method of claim 12 wherein the plurality of weak classifiers are generated based on a location of the corresponding weak classifier to recognize the object.

14. The method of claim 13 wherein the weak classifier is trained as:

ε

′

_b=Σ

_iw′

<sub>τ

,i</sub>|g(I;

s_b^cλ

_b^c)−

y_i^s|.

wherein the weak classifier g(I;

s_b^c, λ

_b^c ) is trained based on the distance value d_b from an image feature of the candidate exemplar s_b^c to the corresponding image feature of all the training samples in B_s.

15. The method of claim 8 further comprising, after being trained, combining the plurality of weak classifiers to form a final classifier.

16. The method of claim 15 wherein the final classifier represents a multi-view, multi-pose object to be detected.

17. Apparatus for training and using a plurality of weak classifiers to form an image classifier comprising:

a plurality of weak classifiers, where each weak classifier in the plurality of weak classifiers is trained using a first gradient orientation histogram representation of a candidate exemplar and a second gradient orientation histogram representing a plurality of training samples, where the candidate exemplar comprises positive samples of images of an object to be detected; and

a combiner, coupled to the plurality of weak classifiers, for linearly combining the plurality of weak classifiers to form a final classifier.

18. The apparatus of claim 17 wherein each weak classifier comprises an AdaBoost loop.

19. The apparatus of claim 17 wherein the final classifier represents a multi-view, multi-pose object to be detected.