Visual Tracking via Robust and Efficient Temporal Regularized Correlation Filters

Abstract

We present a robust and efficient temporal regularized correlation filters (RETCF) model to achieve outstanding performance with hand-crafted features in visual tracking. The model of online learning is overly dependent on the fixed learning rate; When the condition of the object makes a difference, the model cannot learn useful information adaptively. Therefore, we take advantage of temporal regularization to obtain a robust discriminative correlation filters model instead of updating the model of the object. Besides, this correlation filter model can be efficiently optimal via the alternating direction method of multipliers (ADMM). Comparing with other top-ranked of CF trackers, the proposed tracker (RETCF) achieves real-time and outperformance for the challenging benchmark sequences (OTB2013, OTB2015, and TC128).