PERFORMING OBJECT DETECTION OPERATIONS VIA A GRAPHICS PROCESSING UNIT

• US 20140270551A1
• Filed: 09/17/2013
• Published: 09/18/2014
• Est. Priority Date: 03/15/2013
• Status: Active Grant

• Current Assignee: NVIDIA Corporation
• Original Assignee: NVIDIA Corporation
• Inventors: BARANOWSKI, Mateusz Jerzy, GUPTA, Shalini, ALBUZ, Elif
• Granted Patent: US 9,971,959 B2
• Time in Patent Office: Days
• US Class Current: 382/226
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/24323 Tree-organised classifiers
• CPC Class Codes: G06F 18/245 relating to the decision surface
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G06V 10/774 Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 10/955 using specific electronic processors

Abstract

In one embodiment of the present invention, a graphics processing unit (GPU) is configured to detect an object in an image using a random forest classifier that includes multiple, identically structured decision trees. Notably, the application of each of the decision trees is independent of the application of the other decision trees. In operation, the GPU partitions the image into subsets of pixels, and associates an execution thread with each of the pixels in the subset of pixels. The GPU then causes each of the execution threads to apply the random forest classifier to the associated pixel, thereby determining a likelihood that the pixel corresponds to the object. Advantageously, such a distributed approach to object detection more fully leverages the parallel architecture of the PPU than conventional approaches. In particular, the PPU performs object detection more efficiently using the random forest classifier than using a cascaded classifier.

Claims

1. A computer-implemented method for identifying an object in one or more images, the method comprising:

selecting a first subset of pixels included in a first image;

associating a first execution thread with a first pixel included in the first subset of pixels; and

causing the first execution thread to apply a first decision tree included in a plurality of decision trees to the first pixel to determine a first likelihood that the first pixel is associated with a first object and thereby indicating a first probability that the first object is included in the first image, wherein the plurality of decision trees comprises a random forest classifier.

2. The method of claim 1, further comprising:

associating a second execution thread with a second pixel included in the first subset of pixels; and

causing the second execution thread to apply the first decision tree to the second pixel to determine a second likelihood that the second pixel is associated with the first object and thereby indicating a second probability that the first object is included in the first image.

3. The method of claim 2, wherein the first execution thread applies the first decision tree to the first pixel substantially in parallel with the second execution thread applying the first decision tree to the second pixel.

4. The method of claim 1, further comprising:

causing the first execution thread to apply a second decision tree included in the plurality of decision trees to the first pixel to determine a second likelihood that the first pixel is associated with the first object and thereby indicating a second probability that the first object is included in the first image; and

computing an aggregated likelihood that the first pixel is associated with the first object based on the first likelihood and the second likelihood.

5. The method of claim 1, further comprising:

selecting a second subset of pixels included in a second image;

associating a second execution thread with a second pixel included in the second subset of pixels; and

causing the second execution thread to apply the first decision tree to the second pixel to determine a second likelihood that the second pixel is associated with the first object and thereby indicating a second probability that the first object is included in the second image.

6. The method of claim 5 wherein the second image is a scaled version of the first image.

7. The method of claim 6, further comprising:

determining that the first likelihood and the second likelihood represent spatially overlapping detections included in a plurality of overlapping detections; and

computing an aggregated likelihood that the first object is included in the first image based on the weighted sum of all likelihoods that are associated with the plurality of overlapping detections and the number of overlapping detections included in the plurality of overlapping detections.

8. The method of claim 1, wherein the random forest classifier includes multi-block local binary pattern (MB-LBP) features.

9. The method of claim 8, wherein applying the first decision tree to the first pixel comprises:

computing a first set of MB-LBP features that are associated with a first pixel window that specifies a predetermined number of pixels that are proximally located to the first pixel;

selecting a first path in the first decision tree based on the first set of MB-LBP features; and

identifying the first likelihood based on the endpoint of the first path.

10. A computer-readable storage medium including instructions that, when executed by a processing unit, cause the processing unit to identify an object in one or more images by performing the steps of:

selecting a first subset of pixels included in a first image;

associating a first execution thread with a first pixel included in the first subset of pixels; and

causing the first execution thread to apply a first decision tree included in a plurality of decision trees to the first pixel to determine a first likelihood that the first pixel is associated with a first object and thereby indicating a first probability that the first object is included in the first image, wherein the plurality of decision trees comprises a random forest classifier.

11. The computer-readable storage medium of claim 10, further comprising:

associating a second execution thread with a second pixel included in the first subset of pixels; and

causing the second execution thread to apply the first decision tree to the second pixel to determine a second likelihood that the second pixel is associated with the first object and thereby indicating a second probability that the first object is included in the first image.

12. The computer-readable storage medium of claim 11, wherein the first execution thread applies the first decision tree to the first pixel substantially in parallel with the second execution thread applying the first decision tree to the second pixel.

13. The computer-readable storage medium of claim 10, further comprising:

causing the first execution thread to apply a second decision tree included in the plurality of decision trees to the first pixel to determine a second likelihood that the first pixel is associated with the first object and thereby indicating a second probability that the first object is included in the first image; and

computing an aggregated likelihood that the first pixel is associated with the first object based on the first likelihood and the second likelihood.

14. The computer-readable storage medium of claim 10, further comprising:

selecting a second subset of pixels included in a second image;

associating a second execution thread with a second pixel included in the second subset of pixels; and

causing the second execution thread to apply the first decision tree to the second pixel to determine a second likelihood that the second pixel is associated with the first object and thereby indicating a second probability that the first object is included in the second image.

15. The computer-readable storage medium of claim 14 wherein the second image is a scaled version of the first image.

16. The computer-readable storage medium of claim 15, further comprising:

determining that the first likelihood and the second likelihood represent spatially overlapping detections included in a plurality of overlapping detections; and

computing an aggregated likelihood that the first object is included in the first image based on the weighted sum of all likelihoods that are associated with the plurality of overlapping detections and the number of overlapping detections included in the plurality of overlapping detections.

17. The computer-readable storage medium of claim 10, wherein the random forest classifier includes multi-block local binary pattern (MB-LBP) features.

18. The computer-readable storage medium of claim 17, wherein applying the first decision tree to the first pixel comprises:

computing a first set of MB-LBP features that are associated with a first window of pixels that includes a fixed-sized number of pixels that are proximally located to the first pixel;

selecting a first path in the first decision tree based on the first set of MB-LBP features; and

identifying the first likelihood based on the endpoint of the first path.

19. A computing device configured to train random forest classifiers for object detection, the system comprising:

a first memory that includes a plurality of decision trees, wherein the plurality of decision trees comprises a random forest classifier; and

a processing unit coupled to the memory and configured to;

select a first subset of pixels included in a first image;

associate a first execution thread with a first pixel included in the first subset of pixels; and

cause the first execution thread to apply a first decision tree included in the plurality of decision trees to the first pixel determine a first likelihood that the first pixel is associated with a first object and thereby indicating a first probability that the first object is included in the first image.

20. The computing device of claim 19, wherein causing the first execution thread to apply the first decision tree comprises:

identifying a plurality of windows of pixels that is associated with the first subset of pixels, wherein each window of pixels specifies a predetermined number of proximally located pixels;

copying a memory patch from a second memory to the first memory, wherein the memory patch includes image data associated with the plurality of windows of pixels;

causing the first execution thread to identify a first window of pixels included in the plurality of window of pixels based on the first pixel;

causing the first execution thread to compute a first set of features based on the image data included in the memory patch and associated with the first window of pixels; and

causing the first execution thread to process the first decision tree based on the first set of features.