Detecting, tracking and counting objects in videos

1. A method of detecting and tracking objects of interest comprising:

obtaining, by a computing device, multiple frames of images from an image capturing device;

obtaining, using a neural network, detections of objects of interest in each frame, each detection including a bounding box and a score that describes a level of confidence that an object of interest was detected in the bounding box;

accumulating, by the computing device, multiple batches of the detections, where each batch of the detections includes detections of objects of interest in two or more frames of the multiple frames of images;

creating, by the computing device, object tracks for the objects of interest in each batch of detections; and

associating, by the computing device, the object tracks over the multiple batches of detections.

2. The method of claim 1 , further comprising:

comparing, by the computing device, the bounding boxes against avatars; and

filtering out false positive detections based on a result of the comparing.

3. The method of claim 2 , wherein the false positive detections are filtered out based on unlikely locations of the detections.

4. The method of claim 2 , wherein the false positive detections are filtered out based on unlikely orientations of the detections.

5. The method of claim 2 , wherein the false positive detections are filtered out based on unlikely scales of the detections.

6. The method of claim 1 , wherein the detections are obtained using a deformable parts model which has been trained on example instances of the objects of interest.

7. The method of claim 6 , wherein the example instances of the objects of interest include humans observed from a top down view.

8. The method of claim 1 , wherein the detections are obtained using a convolutional neural network that simultaneously regresses to an optimal bounding box location and also classifies a region within the bounding box as one of a plurality of categories of objects of interest.

9. The method of claim 1 , wherein creating, by the computing device, object tracks for the objects of interest in each batch of detections, further comprises:

converting the batches of detections into a weighted graph having vertices, where the vertices of the weighted graph depict an association between a pair of object detections over consecutive or temporally proximal frames.

10. The method of claim 9 , wherein the weighted graph is optimized as a solution to a binary integer linear program that finds a minimum number of paths with least weight from source vertices to sink vertices, such that every vertex is part of the same path and no vertex is included in more than one path.

11. A system for detecting and tracking objects of interest, comprising:

an image capture device configured to obtain multiple frames of images;

an object detection module configured to:

obtain, using a neural network, detections of objects of interest in each frame, each detection including a bounding box and a score that describes a level of confidence that an object of interest was detected in the bounding box; and

accumulate multiple batches of the detections, where each batch of detections includes detections of objects of interest in two or more frames of the multiple frames of images; and

an object tracking module configured to:

create, by the computing device, object tracks for the objects of interest in each batch of detections; and

associate, by the computing device, the object tracks over the multiple batches of detections.

12. The system of claim 11 , wherein the object detection module is further configured to:

compare the bounding boxes against avatars; and

filter out false positive detections based on a result of the comparing.

13. The system of claim 12 , wherein the false positive detections are filtered out based on unlikely locations of the detections.

14. The system of claim 12 , wherein the false positive detections are filtered out based on unlikely orientations of the detections.

15. The system of claim 12 , wherein the false positive detections are filtered out based on unlikely scales of the detections.

16. The system of claim 11 , wherein the detections are obtained using a deformable parts model which has been trained on example instances of the objects of interest.

17. The system of claim 16 , wherein the example instances of the objects of interest include humans observed from a top down view.

18. The system of claim 11 , wherein the detections are obtained using a convolutional neural network that simultaneously regresses to an optimal bounding box location and also classifies a region within the bounding box as one of a plurality of categories of objects of interest.

19. The system of claim 11 , wherein the object tracking module is further configured to:

convert the batches of detections into a weighted graph having vertices, where the vertices of the weighted graph depict an association between a pair of object detections over consecutive or temporally proximal frames.

20. The system of claim 19 , wherein the weighted graph is optimized as a solution to a binary integer linear program that finds a minimum number of paths with least weight from source vertices to sink vertices, such that every vertex is part of the same path and no vertex is included in more than one path.