Correlation filter for object tracking with temporal-spatial constraint

Abstract

In this work, inspired by the Passive-Aggressive learning (PA), we proposed a Temporal-Spatial Constraint Correlation Filter (TSCF) model to simultaneously constrain the spatial mask and the update direction of the filter. Firstly, the spatial regular term ensures that the background redundancy information does not interfere with the filter update during the tracking process. Secondly, the temporal regular term ensures that the spatial mask and the filter do not change dramatically. Thirdly, our proposed TSCF model can be effectively solved based on the alternate direction method of multiplier (ADMM), where each sub-problem has a closed solution. Finally, our experiments on the OTB100 benchmark shows that our tracker has efficient performance compare with many advanced algorithms, which get an AUC score of 0.599 and an accuracy of 0.794.