Method and apparatus for unsupervised learning of discriminative edge measures for vehicle matching between non-overlapping cameras

• US 7,650,030 B2
• Filed: 12/05/2005
• Issued: 01/19/2010
• Est. Priority Date: 12/03/2004
• Status: Expired due to Fees

• Current Assignee: SRI International, Inc.
• Original Assignee: Sarnoff Corporation (SRI International, Inc.)
• Inventors: Kumar, Rakesh, Sawhney, Harpreet, Shan, Ying
• Primary Examiner(s): Hung; Yubin
• Application Number:
  US11/295,143
• Publication Number:
  US 20060165277A1
• Time in Patent Office: 1,506 Days
• Field of Search: 382/103, 382/104, 382/155, 382/159, 382/160, 382/190, 382/224, 382/225, 382/228, 700/47
• US Class Current: 382/159
• CPC Class Codes: G06F 18/256 of results relating to different input data, e.g. multimodal recognition
• CPC Class Codes: G06T 7/33 using feature-based methods
• CPC Class Codes: G06V 10/255 Detecting or recognising potential candidate objects based on visual cues, e.g. shapes
• CPC Class Codes: G06V 10/811 the classifiers operating on different input data, e.g. multi-modal recognition
• CPC Class Codes: G06V 20/54 of traffic, e.g. cars on the road, trains or boats
• CPC Class Codes: G06V 2201/08 Detecting or categorising vehicles

Abstract

A method and apparatus for unsupervised learning of measures for matching objects between images from at least two non-overlapping cameras is disclosed The method includes collecting at least two pairs of feature maps, where the at least two pairs of feature maps are derived from features of objects captured in the images. The method further includes computing, as a function of at least two pairs of feature maps, at least one first and second match measures, wherein the first match measure is of a same class and the second match measure is of a different class.

Claims

1. A computer implemented method for unsupervised learning of measures for matching objects between images from at least two non-overlapping cameras, comprising:

using a processor to perform the steps of;

collecting at least two pairs of feature maps, wherein the at least two pairs of feature maps are derived from features of objects captured in the images from the at least two non-overlapping cameras;

computing, as a function of the at least two pairs of feature maps, at least first and second match measures, wherein the first match measure is of a same class and the second match measure is of a different class, and wherein objects in the same class are deemed to match; and

learning from the at least two pairs of feature maps outlier and inlier distributions, wherein determining the first and second match measures, comprises;

computing, as a function of the outlier and inlier distributions, a weight for each of the first and second match measures, and probabilities for the first and second match measures being of the same and different classes.

2. The method of claim 1, wherein computing at least first and second match measures:

determining a weight of at least one of the at least first and second match measures, wherein the weight is indicative of the same class or the different class.

3. The method of claim 1, further comprising:

using discriminative feature detection of objects captured in the images to derive the feature maps.

4. The method of claim 1, wherein computing at least first and second match measures comprises:

deriving at least one measurement vector from differences between oriented features of the objects captured in the images.

5. The method of claim 4, further comprising:

projecting the at least one measurement vector to a one-dimensional projection along a line defined by a weight for each of the first and second match measures so as to maximize a separation between the same class and the different class.

6. The method of claim 4, wherein the at least one measurement vector has at least 2 dimensions.

7. The method of claim 1, wherein computing the probabilities of the first and second match measures comprises:

fitting to a Gaussian distribution the same class and the different class.

8. The method of claim 1, wherein collecting, determining and computing are performed automatically on a periodic basis.

9. An apparatus for matching objects between images from at least two non-overlapping cameras, comprising:

means for collecting at least two pairs of feature maps, wherein the at least two pairs of feature maps are derived from features of objects captured in the images; and

means for computing, as a function of the at least two pairs of feature maps, at least first and second match measures, wherein the first match measure is of a same class and the second match measure is of a different class, and wherein objects in the same class are deemed to match; and

means for learning from the at least one two pairs of feature maps outlier and inlier distributions, wherein the means for computing the first and second match measures comprises;

means for computing, as a function of the outlier and inlier distributions, a weight for each of the first and second match measures, and probabilities for the first and second match measures being of the same and different classes.

10. The apparatus of claim 9, wherein the means for determining a same class and a different class comprises:

a means for determining a weight of the first and second match measures, wherein the weight is indicative of any of the same class and the different class.

11. The apparatus of claim 9, further comprising:

a means for deriving the feature map via discriminative feature detection of objects captured in the images.

12. The apparatus of claim 9, wherein the means for computing the first and second match measures comprises:

means for deriving at least one measurement vector from differences between oriented features of the objects captured in the images.

13. The apparatus of claim 12, wherein the means for computing the first and second match measures the further comprises:

projecting the at least one measurement vector to a one-dimensional projection along a line defined by a weight for each of the first and second match measures so as to maximize a separation between the same class and the different class.

14. The apparatus of claim 9, wherein means for computing the probabilities of the first and second match measure comprises:

means for fitting to a Gaussian distribution the same class and the different class.

15. A computer implemented method for unsupervised learning of measures for matching objects between images from at least two non-overlapping cameras, comprising:

using a processor to perform the steps of;

collecting at least two pairs of feature maps, wherein the at least two pairs of feature maps are derived from features of objects captured in the images from the at least two non-overlapping cameras;

learning from the at least two pairs of feature maps respective outlier and inlier distributions; and

computing, as a function of the at least two pairs of feature maps and the respective outlier and inlier distributions, at least first and second match measures, wherein the first match measure is of a same class and the second match measure is of a different class, and wherein objects in the same class are deemed to match.

16. The method of claim 15, wherein learning from the at least two pairs of feature maps respective outlier and inlier distributions comprises ranking the feature maps as a function of the features of the objects captured in the images.

17. The method of claim 16, wherein the images are unlabeled.

18. The method of claim 15, wherein computing at least first and second match measures comprises:

deriving at least one measurement vector from differences between oriented features of the objects captured in the images.