Robust Visual Cooperative Tracking Using Constrained Adaptive Sparse Representations and Sparse Classifier Grids

We present a novel computational framework that is capable of dealing with many real-world visual tracking problems. A novel spatio-temporal weighting scheme is introduced to maximize the separation between target and background, improving classification accuracy. These weights are then used to define a norm over which a constrained adaptive sparse representation (CASR) of target and background patches is computed. This representation defines a similarity metric that is used by a novel particle-based NormalHedge (NH) algorithm to identify the target on subsequent frames. If the NH algorithm is successful in cleanly identifying the target, the target and background dictionaries are updated according to an adaptive algorithm, which avoids the addition of aberrant or redundant atoms and deletes atoms that have become uninformative. The spatio-temporal weights are then updated and the weighting-CASR-NH-dictionary selection loop starts over again. If the NH algorithm is unsuccessful in cleanly identifying the target, a computationally efficient sparse classifier grid is used for target retrieval. In this paper, we discuss the details of the techniques proposed and compare the accuracy and computational efficiency of the resulting algorithm with that of several existing algorithms. These comparisons demonstrate the value of the proposed algorithm to the solution of real-world online tracking problems.