Fast compressive tracking based on multi-feature weighted appearance model

Visual tracking is a hot but challenging research direction in computer vision field. The appearance of target changes continuously over space and time, which makes it difficult to develop a robust and efficient tracking algorithm. Most existing tracking algorithms describe the target with single feature which may make the tracker easily influenced by various intrinsic and extrinsic changes and even lead to trace failure. In this paper, in order to describe the complex changes of the target and background in real scene more accurately and to improve the flexibility of statistical model, an effective and efficient tracking algorithm is proposed based on an appearance model of weighted multiple features extracted from blocked images in the compressed domain. Two complementary matrices are generated to measure multi-dimension intensity and texture features. Mahalanobis distance is employed as the weights of the various features. Then a naive Bayesian classifier is utilized to make a binary classification for the candidates. A coarse-to-fine search strategy is adopted to reduce the computational complexity. Experimental results demonstrated that the robustness and accuracy of our algorithm are greatly improved compared with several state-of-theart tracking algorithms.