People detection in video and image data

• US 7,986,828 B2
• Filed: 10/10/2007
• Issued: 07/26/2011
• Est. Priority Date: 10/10/2007
• Status: Active Grant

• Current Assignee: Honeywell International Inc.
• Original Assignee: Honeywell International Inc.
• Inventors: Paturu, Chaitanya K., NC, Pramod, Rao, Supriya
• Primary Examiner(s): CHANG, JON CARLTON
• Application Number:
  US11/870,237
• Publication Number:
  US 20090097739A1
• Time in Patent Office: 1,385 Days
• Field of Search: 382/155, 382/159, 382/224, 382/226, 382/227, 382/103, 382/115, 382/118, 348/143, 348/152
• US Class Current: 382/159
• CPC Class Codes: G06F 18/254 of classification results, e.g. of results related to same input data
• CPC Class Codes: G06V 10/809 of classification results, e.g. where the classifiers operate on the same input data
• CPC Class Codes: G06V 40/10 Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands

Abstract

A process identifies a person in image data. The process first executes a training phase, and thereafter a detection phase. The training phase learns body parts using body part detectors, generates classifiers, and determines a spatial distribution and a set of probabilities. The execution phase applies the body part detector to an image, combines output of several body part detectors, and determines maxima of the combination of the output.

Claims

1. A process to identify a person in image data comprising:

executing a training phase, the training phase comprising;

using one or more body part detectors to learn one or more body parts;

generating classifiers for the one or more body parts;

determining a spatial distribution among the one or more of the body parts; and

determining a set of probabilities, the set of probabilities relating to one or more of a probable location and scale of the person in the image data given that a particular body part detector has fired, a probability that there is a person in the image data given that a particular body part detector has fired, and a probability that a particular body part detector can fire at a particular location and a particular scale; and

executing a detection phase, the detection phase comprising;

applying the one or more body part detectors to an image;

combining output of the one or more body part detectors; and

determining a maximum of the combination of output of the one or more body part detectors.

2. The process of claim 1, comprising:

modeling an accuracy of the one or more body part detectors; and

weighting the body part detectors as a function of the accuracy modeling.

3. The process of claim 2, comprising computing a ratio of true detections by the one or more body part detectors to the number of total detections by the one or more body part detectors to model the accuracy of the learned body parts.

4. The process of claim 1, comprising using a plurality of persons comprising different scales and different aspect ratios to learn the one or more body parts.

5. The process of claim 1, comprising voting for a centroid of the person and using a result of the voting to combine the output of the one or more body part detectors.

6. The process of claim 5, comprising using the spatial distribution in the voting.

7. The process of claim 1, wherein the body parts include one or more of a head, a leg, a shoulder, and a torso.

8. The process of claim 1, wherein the generating classifiers for the one or more body parts includes an Adaboost algorithm using Haar features.

9. The process of claim 1, comprising tuning the one or more body part detectors in the training phase to provide a high detection rate.

10. The process of claim 1, comprising verifying a spatial relationship between two or more body parts, thereby verifying the relative location of the two or more body parts.

11. A process to identify an object in image data comprising:

executing a training phase, the training phase comprising;

using one or more part detectors to learn one or more parts of the object;

generating classifiers for the one or more parts;

determining a spatial distribution among the one or more parts; and

determining a set of probabilities, the set of probabilities relating to one or more of a probable location and scale of the object in the image data given that a particular part detector has fired, a probability that there is an object in the image data given that a particular part detector has fired, and a probability that a particular part detector can fire at a particular location and a particular scale; and

executing a detection phase, the detection phase comprising;

applying the one or more part detectors to an image;

combining output of the one or more part detectors; and

determining a maximum of the combination of output of the one or more part detectors.

12. A non-transitory machine-readable medium comprising instructions, which when implemented by one or more processors perform the following operations:

execute a training phase, the training phase comprising;

use one or more body part detectors to learn one or more body parts;

generate classifiers for the one or more body parts;

determine a spatial distribution among the one or more of the body parts; and

determine a set of probabilities, the set of probabilities relating to one or more of a probable location and scale of the person in the image data given that a particular body part detector has fired, a probability that there is a person in the image data given that a particular body part detector has fired, and a probability that a particular body part detector can fire at a particular location and a particular scale; and

execute a detection phase, the detection phase comprising;

apply the one or more body part detectors to an image;

combine output of the one or more body part detectors; and

determine a maximum of the combination of output of the one or more body part detectors.

13. The non-transitory machine-readable medium of claim 12, comprising instructions to:

model an accuracy of the one or more body part detectors; and

weight the body part detectors as a function of the accuracy modeling.

14. The non-transitory machine-readable medium of claim 13, comprising instructions to compute a ratio of true detections by the one or more body part detectors to the number of total detections by the one or more body part detectors to model the accuracy of the learned body parts.

15. The non-transitory machine-readable medium of claim 12, comprising instructions to use a plurality of persons comprising different scales and different aspect ratios to learn the one or more body parts.

16. The non-transitory machine-readable medium of claim 12, comprising instructions to vote for a

centroid of the person and using a result of the voting to combine the output of the one or more body part detectors.

17. The non-transitory machine-readable medium of claim 16, comprising instructions to use the spatial distribution in the voting.

18. The non-transitory machine-readable medium of claim 12, wherein the body parts include one or more of a head, a leg, a shoulder, and a torso.

19. The non-transitory machine-readable medium of claim 12, wherein the generating classifiers for the one or more body parts includes an Adaboost algorithm using Haar features.

20. The non-transitory machine-readable medium of claim 12, comprising instructions to:

tune the one or more body part detectors in the training phase to provide a high detection rate; and

verify a spatial relationship between two or more body parts, thereby verifying the relative location of the two or more body parts.