Robust object tracking via adaptive sparse representation

In this paper, we present a robust object tracking system capable of handling pose and scale variations. The system is based on adaptive sparse representation and dictionary learning. We focus on the problem of automatic tracking with no prior knowledge other than the location of the region to be tracked in the first frame, which could be located by a detector. The detected region, i.e., a bounding box, and some samples near the bounding box are extracted as positive samples. In addition, we select regions outside the bounding box as negative samples. Both, positive and negative samples are used to build the dictionary and we use K-SVD method for dictionary learning in order to decrease the number of atoms and improve the processing speed. One of the main drawbacks in tracking systems is false tracking when the object is not in the frame any more. We overcome this problem by comparing the newly tracked region with previously tracked regions to find out if the object is still in the frame or not. If the object is not in the frame, the algorithm stops tracking and starts searching for the object using the sparse detector in the following frames. Experiments on video sequences demonstrate the effectiveness and robustness of the proposed system for tracking.