Object tracking based on locality-constrained linear coding joint sparse representation appearance model

摘要

Focusing on the robust object tracking, we propose an algorithm combining discriminative model with generative model. In the discriminative model, we exploit the prior visual information to learn an over-complete dictionary, and use the locality constrained linear (LLC) coding to represent the object. Then use the linear SVM classifier to separate the foreground from the background to implement object tracking. In the generative model, we propose a sparse generative model to partition the object into patches and take the occlusion factor into account to construct object templates. Then use the particle filter to evaluate the target position. Finally joint the two models to acquire final tracking result. In addition, in order to handle the object appearance variation caused by occlusion, fast motion, illumination change and background clutter, we make a simple yet effective update scheme. Both qualitative and quantitative evaluations on challenging image sequences demonstrate that the proposed algorithm performs favorably against several state-of-the-art methods.