Reliable Correlation Tracking Via Dual-Memory Selection Model

Guiji Li,Manman Peng,Ke Nai,Zhiyong Li,Keqin Li

DOI: https://doi.org/10.1016/j.ins.2020.01.015

IF: 8.1

2020-01-01

Information Sciences

Abstract:Correlation-filter-based trackers have shown favorable accuracy and efficiency in visual tracking. However, most of these trackers are prone to drift in cases of heavy occlusions and temporal tracking failures because they only maintain the short-term memory of target appearance via a highly adaptive update mode. In this paper, we propose a reliable visual tracking method based on a dual-memory selection (DMS) model to alleviate tracking drift. Considering that long-term memory is robust to heavy occlusions while short-term memory performs well in rapid appearance changes, the proposed DMS model combines these two memory patterns of the target appearance and adaptively selects a reliable memory pattern to handle the current tracking challenges via a memory selector. For each memory pattern, a memory tracker is established based on discriminative correlation filters. The short-term tracker aggressively updates the target model to capture recent appearance changes via a linear interpolation update model, while the long-term tracker conservatively updates the target model to maintain historical appearance characteristics with a memory-improved update model and a dynamic learning rate. Furthermore, a novel memory evaluation criterion (MEC) is developed to evaluate the reliability of each tracker for memory selection. From credibility and discriminability measurements considering the temporal context, the memory tracker with the highest reliability score is selected to determine the target location in each frame. Extensive experiments on public benchmark datasets demonstrate that the proposed tracking method performs favorably compared to multiple recent state-of-the-art methods.