Visual Tracking via Spatial-Temporal Regularized Correlation Filters with Advanced State Estimation

Discriminative correlation filter (CF) based visual trackers achieves outstanding performance with the handcrafted feature in visual tracking. In this work, based on the discriminative correlation filter, we propose a new Spatial-Temporal regularized correlation filters with advanced state estimation (CFASE) to achieve more significant tracking performance. First, we propose a new method to estimate correlation filters more precisely using prediction from the previous two filters, considering the drift during the tracking process. Second, we train two correlation filters models to obtain scale estimation and object location, respectively. The separated two correlation filter models help to reduce the adverse effects of scale changes on object location. Third, our tracker introduces average peak-to-correlation energy (APCE) to evaluate the accuracy of scale estimation and object location. Experimentally, the proposed tracker (CFASE) achieves outstanding and real-time performance for the challenging benchmark sequence (OTB2013, OTB2015, and TC128).