Training system for infield training of a vision-based object detector

• US 10,002,430 B1
• Filed: 04/28/2016
• Issued: 06/19/2018
• Est. Priority Date: 06/04/2013
• Status: Active Grant

• Current Assignee: HRL Laboratories LLC (The Boeing Co.)
• Original Assignee: HRL Laboratories LLC (The Boeing Co.)
• Inventors: Mundhenk, Terrell N., Flores, A. Arturo, Neely, III, Howard, Daily, Michael J.
• Primary Examiner(s): Gilliard, Delomia L
• Application Number:
  US15/141,416
• Time in Patent Office: 782 Days
• Field of Search: None
• CPC Class Codes: G06F 18/2193 based on specific statistical tests
• CPC Class Codes: G06F 18/2453 non-linear, e.g. polynomial classifier
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• CPC Class Codes: G06T 2207/20081 Training; Learning
• CPC Class Codes: G06T 7/73 using feature-based methods
• CPC Class Codes: G06V 10/462 Salient features, e.g. scale invariant feature transforms [SIFT]
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G06V 10/7796 based on specific statistical tests
• CPC Class Codes: G06V 20/52 Surveillance or monitoring of activities, e.g. for recognising suspicious objects recognising microscopic objects G06V20/69

Abstract

Described is a training system for training a vision-based object detector. The system is configured to run an object detector on an image of a cleared scene to detect objects in the cleared scene. The object detector includes a support vector machine (SVM) or similar classifier with a feature model to generate an SVM score for object features and a spatial bias threshold to generate augmented object scores. The system designated detected objects in the cleared scene as false detections and, based on that, updates at least one of the feature model and spatial bias threshold to designate the false detections as background. The updated feature model or updated spatial bias threshold are then stored for use in object detection.

Claims

1. The vision-based object detector, the system comprising:

one or more processors and a memory, the memory having executable instructions encoded thereon, such that upon execution of the instructions, the one or more processors perform operations of;

running an object detector on an image of a cleared scene to detect objects in the cleared scene, the object detector having a detection classifier with a feature model to generate a detection score for object features and a spatial bias threshold to generate augmented object scores;

designating detected objects in the cleared scene as false detections; and

updating at least one of the feature model and spatial bias threshold to designate the false detections as background; and

storing the updated feature model or updated spatial bias threshold in the vision based object detector.

2. The vision-based object detector as set forth in claim 1, wherein updating the spatial bias threshold further comprises operations of:

loading an old bias file into the object detector;

combining the old bias file with a new bias file by concatenating the old bias file and the new bias file to generate a learned bias file; and

storing the learned bias file.

3. The vision-based object detector as set forth in claim 2, wherein updating the feature model further comprises operations of:

obtaining feature descriptors from all detected objects designated as false detections;

running the feature descriptors through the detection classifier in a randomized order with particular training parameters to generate detection scores;

augmenting weights of the feature model in the detection classifier until all detection scores provided are above a predetermined threshold, indicating no false detections; and

saving augmented weights in the detection classifier as the updated feature model.

4. The vision-based object detector as set forth in claim 3, wherein each detection score is associated with a score rectangle in an image of a scene and wherein the learned bias file includes bias rectangles that inhibit false detections of objects within the bias rectangle.

5. The vision-based object detector as set forth in claim 4, further comprising operations of:

determining, for all detection classification scores above a predetermined threshold, if a corresponding score rectangle overlaps with a bias rectangle;

if a score rectangle does not overlap with a bias rectangle, then obtaining a next bias rectangle and score rectangle and determining overlap;

if a score rectangle does overlap with a bias rectangle, then subtracting the bias rectangle from the score rectangle; and

generating an augmented object score, the augmented object score reflecting a likelihood that a target object is in the scene.

6. The vision-based object detector as set forth in claim 2, wherein each detection score is associated with a score rectangle in an image of a scene and wherein the learned bias file includes bias rectangles that inhibit false detections of objects within the bias rectangle;

and

further comprising operations of;

determining, for all detection classification scores above a predetermined threshold, if a corresponding score rectangle overlaps with a bias rectangle;

if a score rectangle does not overlap with a bias rectangle, then obtaining a next bias rectangle and score rectangle and determining overlap;

if a score rectangle does overlap with a bias rectangle, then subtracting the bias rectangle from the score rectangle; and

generating an augmented object score, the augmented object score reflecting a likelihood that a target object is in the scene.

7. The vision-based object detector as set forth in claim 1, wherein the detection classifier is support vector machine classifier.

8. The vision-based object detector as set forth in claim 1, wherein updating the feature model further comprises operations of:

obtaining feature descriptors from all detected objects designated as false detections;

running the feature descriptors through the detection classifier in a randomized order with particular training parameters to generate detection scores;

augmenting weights of the feature model in the detection classifier until all detection scores provided are above a predetermined threshold, indicating no false detections; and

saving augmented weights in the detection classifier as the updated feature model.

25. The vision-based object detector as set forth in claim 1, wherein running an object detector further comprises:

using a feature model, generating a detection score for object features; and

using a spatial bias threshold, generating augmented object scores.

9. A method for training a vision-based object detector, the method comprising an act of:

causing one or more processors to execute instructions encoded on a non-transitory computer-readable medium, such that upon execution, the one or more processors perform operations of;

running an object detector on an image of a cleared scene to detect objects in the cleared scene, the object detector having a detection classifier with a feature model to generate a detection score for object features and a spatial bias threshold to generate augmented object scores;

designating detected objects in the cleared scene as false detections; and

updating at least one of the feature model and spatial bias threshold to designate the false detections as background; and

storing the updated feature model or updated spatial bias threshold in the vision based object detector.

10. The method as set forth in claim 9, wherein updating the spatial bias threshold further comprises operations of:

loading an old bias file into the object detector;

combining the old bias file with a new bias file by concatenating the old bias file and the new bias file to generate a learned bias file; and

storing the learned bias file.

11. The method as set forth in claim 10, wherein updating the feature model further comprises operations of:

obtaining feature descriptors from all detected objects designated as false detections;

running the feature descriptors through the detection classifier in a randomized order with particular training parameters to generate detection scores;

augmenting weights of the feature model in the detection classifier until all detection scores provided are above a predetermined threshold, indicating no false detections; and

saving augmented weights in the detection classifier as the updated feature model.

12. The method as set forth in claim 11, wherein each detection score is associated with a score rectangle in an image of a scene and wherein the learned bias file includes bias rectangles that inhibit false detections of objects within the bias rectangle.

13. The method as set forth in claim 12, further comprising operations of:

determining, for all detection classification scores above a predetermined threshold, if a corresponding score rectangle overlaps with a bias rectangle;

if a score rectangle does not overlap with a bias rectangle, then obtaining a next bias rectangle and score rectangle and determining overlap;

if a score rectangle does overlap with a bias rectangle, then subtracting the bias rectangle from the score rectangle; and

generating an augmented object score, the augmented object score reflecting a likelihood that a target object is in the scene.

14. The method as set forth in claim 10, wherein each detection score is associated with a score rectangle in an image of a scene and wherein the learned bias file includes bias rectangles that inhibit false detections of objects within the bias rectangle;

and

further comprising operations of;

determining, for all detection classification scores above a predetermined threshold, if a corresponding score rectangle overlaps with a bias rectangle;

if a score rectangle does not overlap with a bias rectangle, then obtaining a next bias rectangle and score rectangle and determining overlap;

if a score rectangle does overlap with a bias rectangle, then subtracting the bias rectangle from the score rectangle; and

generating an augmented object score, the augmented object score reflecting a likelihood that a target object is in the scene.

15. The method as set forth in claim 9, wherein the detection classifier is support vector machine classifier.

16. The method as set forth in claim 9, wherein updating the feature model further comprises operations of:

obtaining feature descriptors from all detected objects designated as false detections;

running the feature descriptors through the detection classifier in a randomized order with particular training parameters to generate detection scores;

augmenting weights of the feature model in the detection classifier until all detection scores provided are above a predetermined threshold, indicating no false detections; and

saving augmented weights in the detection classifier as the updated feature model.

17. A computer program product for training a vision-based object detector, the computer program product comprising:

a non-transitory computer-readable medium having executable instructions encoded thereon, such that upon execution of the instructions by one or more processors, the one or more processors perform operations of;

running an object detector on an image of a cleared scene to detect objects in the cleared scene, the object detector having a detection classifier with a feature model to generate a detection score for object features and a spatial bias threshold to generate augmented object scores;

designating detected objects in the cleared scene as false detections; and

updating at least one of the feature model and spatial bias threshold to designate the false detections as background; and

storing the updated feature model or updated spatial bias threshold in the vision based object detector.

18. The computer program product as set forth in claim 17, wherein updating the spatial bias threshold further comprises operations of:

loading an old bias file into the object detector;

combining the old bias file with a new bias file by concatenating the old bias file and the new bias file to generate a learned bias file; and

storing the learned bias file.

19. The computer program product as set forth in claim 18, wherein updating the feature model further comprises operations of:

obtaining feature descriptors from all detected objects designated as false detections;

running the feature descriptors through the detection classifier in a randomized order with particular training parameters to generate detection scores;

augmenting weights of the feature model in the detection classifier until all detection scores provided are above a predetermined threshold, indicating no false detections; and

saving augmented weights in the detection classifier as the updated feature model.

20. The computer program product as set forth in claim 19, wherein each detection score is associated with a score rectangle in an image of a scene and wherein the learned bias file includes bias rectangles that inhibit false detections of objects within the bias rectangle.

21. The computer program product as set forth in claim 20, further comprising operations of:

determining, for all detection classification scores above a predetermined threshold, if a corresponding score rectangle overlaps with a bias rectangle;

if a score rectangle does not overlap with a bias rectangle, then obtaining a next bias rectangle and score rectangle and determining overlap;

if a score rectangle does overlap with a bias rectangle, then subtracting the bias rectangle from the score rectangle; and

generating an augmented object score, the augmented object score reflecting a likelihood that a target object is in the scene.

22. The computer program product as set forth in claim 18, wherein each detection score is associated with a score rectangle in an image of a scene and wherein the learned bias file includes bias rectangles that inhibit false detections of objects within the bias rectangle;

and

further comprising operations of;

determining, for all detection classification scores above a predetermined threshold, if a corresponding score rectangle overlaps with a bias rectangle;

if a score rectangle does not overlap with a bias rectangle, then obtaining a next bias rectangle and score rectangle and determining overlap;

if a score rectangle does overlap with a bias rectangle, then subtracting the bias rectangle from the score rectangle; and

generating an augmented object score, the augmented object score reflecting a likelihood that a target object is in the scene.

23. The computer program product as set forth in claim 17, wherein the detection classifier is support vector machine classifier.

24. The computer program product as set forth in claim 17, wherein updating the feature model further comprises operations of:

obtaining feature descriptors from all detected objects designated as false detections;

running the feature descriptors through the detection classifier in a randomized order with particular training parameters to generate detection scores;

augmenting weights of the feature model in the detection classifier until all detection scores provided are above a predetermined threshold, indicating no false detections; and

saving augmented weights in the detection classifier as the updated feature model.