System and method for visually tracking with occlusions
Described herein are tracking algorithm modifications to handle occlusions when processing a video stream including multiple image frames. Specifically, system and methods for handling both partial and full occlusions while tracking moving and non-moving targets are described. The occlusion handling embodiments described herein may be appropriate for a visual tracking system with supplementary range information.
1. A method of visually tracking targets in which full occlusions of the targets are encountered, comprising:
identifying target regions in an image;
assigning target regions with old tracks and assigning target regions with new tracks;
detecting and identifying fully occluded tracks;
modifying track discard rules for tracks identified as fully occluded that are unassigned to target regions in order to maintain the occluded tracks for a greater duration than tracks not identified as fully occluded that are unassigned to target regions;
adjusting a motion model used to predict tracks identified as fully occluded to maintain the occluded tracks along a last observed path of the occluded tracks and gradually increase a prediction uncertainty; and
deciding which tracks to maintain and which tracks to discard.
2. The method of claim 1 , wherein the assigning step includes predicting a location in a current video frame that includes the image.
3. The method of claim 2 , wherein predicting step predicts the location by approximating a location with an observation of a target's current position and applying a Kalman filter.
4. The method of claim 1 , wherein the assigning step includes using a cost minimization algorithm.
5. The method of claim 1 , further comprising processing all tracks to determine if the tracks are assigned target regions.
6. The method of claim 1 , further comprising creating new tracks from unassigned target regions.
7. The method of claim 1 , wherein the deciding step includes not assigning target regions to a target for N consecutive images before discarding the target if the target meets a minimum number of target regions.
8. The method of claim 7 , wherein the deciding step includes immediately discarding a target upon not being assigned a target region if the target has less than the minimum number of target regions.
9. The method of claim 1 , further comprising modifying algorithms used in the adjusting and deciding steps to increase the likelihood that a single track is maintained when a real target becomes temporarily occluded when supplemental range information is available.
10. The method of claim 9 , wherein the range information is provided by computational stereo algorithms applied to stereo sensors.
11. A method of visually tracking targets in which occlusions of the targets are encountered, comprising:
identifying target regions in an image;
assigning target regions with old tracks and assigning target regions with new tracks;
determining if a track is assigned a target region;
if the track is assigned a target region, determining if the target region is partially occluded, and if yes, adjusting a centroid estimation; and
if the track is not assigned a target region, identifying the track as fully occluded, determining if a predicted location is occluded, and modifying track discard rules for the track identified as fully occluded.
12. A system for visually tracking targets in which full occlusions of the targets are encountered, comprising:
an input device for receiving an image;
a memory comprising instructions for:
identifying target regions in the image;
assigning target regions with old tracks and assigning target regions with new tracks;
detecting and identifying fully occluded tracks;
modifying track discard rules for tracks identified as fully occluded that are unassigned to target regions in order to maintain the occluded tracks for a greater duration than tracks not identified as fully occluded that are of unassigned to target regions;
adjusting a motion model used to predict tracks identified as fully occluded to maintain the occluded tracks along a last observed path of the occluded tracks and gradually increase a prediction uncertainty; and
deciding which tracks to maintain and which tracks to discard; and
a processor for executing the instructions.
13. A system for visually tracking targets in which occlusions of the targets are encountered, comprising:
an input device for receiving an image;
a memory comprising instructions for:
identifying target regions in an image;
assigning target regions with old tracks and assigning target regions with new tracks;
determining if a track is assigned a target region;
if the track is assigned a target region, determining if the target region is partially occluded, and if yes, adjusting a centroid estimation; and
if the track is not assigned a target region, identifying the track as fully occluded, determining if a predicted location is occluded, and modifying track discard rules for the track identified as fully occluded; and
a processor for executing the instructions.
14. A non-transitory computer readable medium for visually tracking targets in which occlusions of the targets are encountered, comprising instructions for performing the method of claim 1 .
15. A non-transitory computer readable medium for visually tracking targets in which occlusions of the targets are encountered, comprising instructions for performing the method of claim 11 .