Locality constrained low-rank sparse learning for object tracking

摘要

In this paper, we present a locality constrained low rank sparse learning algorithm for object tracking under the particle filter framework. Locality should be as important as the sparsity. It can further exploit spatial relationship among particles and increase the consistency of low rank coding. Locality information among the training data and dictionary is mined. This can be achieved by using the local constraints as the regularization term. Combined the low rank and sparse criteria, the total objective function is constructed for locality constrained low rank sparse learning. It can be solved by a sequence of closed form update operations. The best target candidate is chosen by jointly evaluating the reconstructive error and classification error. Extensive experimental results on challenging video sequences demonstrate that the proposed tracking method achieves state-of-the-art performance in term of accuracy and robustness.