Method for tracking objects in videos using forward and backward tracking
View Patent ↗A method tracks an object in a sequence of frames of a video. The method is provided with a set of tracking modules. Frames of a video are buffered in a memory buffer. First, an object is tracked in the buffered frames forward in time using a selected one of the plurality of tracking module. Second, the object is tracked in the buffered frames backward in time using the selected tracking module. Then, a tracking error is determined from the first tracking and the second tracking. If the tracking error is less than a predetermined threshold, then additional frames are buffered in the memory buffer and the first tracking, the second tracking and the determining steps are repeated. Otherwise, if the error is greater than the predetermined threshold, then a different tracking module is selected and the first tracking, the second tracking and the determining steps are repeated.
1. A computer implemented method for tracking an object in a sequence of frames of a video, comprising the steps of:
providing a set of tracking modules;
buffering frames of a video in a memory buffer;
first tracking an object in the buffered frames forward in time using a selected one of the plurality of tracking module
second tracking the object in the buffered frames backward in time using the selected tracking module; and
determining a tracking error from the first tracking and the second tracking, and if the tracking error is less than a predetermined threshold, then buffering additional frames in the memory buffer and repeating the first tracking, the second tracking and the determining steps, and otherwise if the error is greater than the predetermined threshold, then selecting a different tracking module and repeating the first tracking, the second tracking and the determining steps,
wherein each tracking module has an associated complexity, and wherein the set of tracking modules is selected from a group comprising a mean-shift tracker, a particle filter tracker, a covariance tracker, a Kalman filter tracker, an ensemble tracker, an exhaustive search tracker, and an online classifier tracker.
2. The method of claim 1 , further comprising:
decreasing a size of the buffer if the frame difference error is greater than a predetermined threshold, and otherwise increasing the size of the buffer if the frame difference error is less than a predetermined threshold.
3. The method of claim 2 , in which the buffer size is a non linear function.
4. The method of claim 1 , in which the frame difference error is a frame difference based global motion between consecutive frames in the buffer.
5. The method of claim 1 , in which each tracking module has an associated complexity, and further comprising:
arranging the plurality of tracking modules in a high to low complexity order.
6. The method of claim 1 , in which the different tracking module is selected based on the tracking error.
7. The method of claim 1 , in which the tracking error is based on a location of the object before tracking, and after the first forward tracking followed by the second backward tracking.
8. The method of claim 1 , in which the tracking error is based on an object model determined at a location of the object before tracking, after the first tracking, and after the second tracking.
9. The method of claim 1 , in which the complexity is a computational load of the tracking module.
10. The method of claim 1 , in which the complexity for the particle filter tracker is a number of particles.
11. The method of claim 1 , in which the complexity for the mean-shift filter tracker is a size of the kernel.
12. The method of claim 1 , in which file complexity for the Kalman filler tracker and the exhaustive search tracker is a size of the search region.
13. The method of claim 1 , in which the complexity for the covariance tracker is a number of features.
14. The method of claim 1 , in which the complexity for the online classifier tracker and the ensemble tracker is a number of classifiers.
15. The method of claim 1 , further comprising:
changing the complexity of the selected tracker module based on the tracking error.