Structural spatio-temporal transform for robust visual tracking

This paper presents a tracking method which decouples tracking process into translation and scale estimation steps. A coarse estimation step in translation is implemented with particle filter first. Then, a two layers of correlation filters is proposed to robustly estimate object variation in translation and scale accurately. The appearance of object is divided into several local blocks. Each block is a basic unit for data updating and it is capable of accurately locating the sub-context of target based on the trained block filters. A local weight vector is developed to structurally and flexibly formulate spatial-temporal transform feature map with online learning framework. The block-updated filters are assembled to a final tracker for the accurate translation estimation. To handle the adaptive scale variation, a sample pyramid based tracker is built to estimate the scale accurately. Experiments on the public benchmark demonstrate the advantage of proposed algorithm over the state-of-the-art approaches.