Memristor-Based Attention Network for Online Real-Time Object Tracking

Abstract

Most existing visual object tracking approaches are implemented based on von Neumann computation systems, which inevitably have the problems of high latency. Additionally, remote server processing of video resources requires a large amount of data transmission over the Internet, which limits real-time tracking performance. The integration of visual object tracking technology into electronic devices has become a new trend. However, current visual object tracking approaches have high algorithm complexity, making it difficult to design circuits to implement the corresponding functions. In this paper, a memristor-based attention network and its corresponding algorithm are proposed to achieve online real-time tracking under parallel computing. Memristors are used to construct attention encoding circuits to record changes of the target in historical frames, and adjust attention signals to the target online and in real-time during the tracking process, avoiding the latency problem of the von Neumann architecture. Inspired by the working process of γ-GABAergic interneuron and tripartite synapse, we propose an attention allocation module to selectively allocate attention values. Combining the Winner-Take-All principle, we design a target localization circuit and an optimal attention zone selection circuit for parallel computation to track the location of the target. Finally, experiments and analyses on OTB-100, NFS, and VOT-RTb2022 benchmark datasets verify that the proposed memristor-based attention network has promising tracking performance and achieves a tracking speed of 1000 FPS, demonstrating superior real-time performance.