Dual-stream collaborative tracking algorithm combined with reliable memory based update

Abstract

Most visual tracking algorithms based on Siamese network adopt a fixed template in the whole tracking process, which sacrifices the available target appearance information. To address this issue, we propose a dual-stream collaborative tracking algorithm combined with reliable memory based update in this paper. Specifically, we first conduct a multi-task convolutional network consisting of convolutional kernel fusion and semantic integration, hence enhancing the discriminability of the convolutional model. Secondly, a reliable template storage policy is developed to establish a dynamic memory queue that carries available appearance information. In addition, to improve the template adaptability against appearance variation, we train a dynamic update network by serving the memory queue as samples, thereby realizes the adaptive update of the object template. Finally, a dual-stream collaborative tracking framework is proposed to integrate the matching results of the initial branch and the updated branch in the response layer, which alleviates the limited representation problems with the fixed template. We perform a thorough performance evaluation on the OTB2013, OTB2015, DTB70, Temple-Color-128, VOT2016, VOT2018, UAV20L and LaSOT benchmarks. Extensive experimental results show that our proposed method outperforms several state-of-the-art algorithms and achieves real-time operations.