Object Tracking Method Combined with Lightweight Hybrid Attention Siamese Network

Abstract

Aiming at the problem that the target tracking method of deep learning has a large number of model parameters and insufficient real-time performance, it is difficult to apply to mobile terminals or embedded devices with insufficient computing power. A lightweight hybrid attention-based twin network tracking algorithm is proposed. Firstly, based on MobileNetv3-Large network, group convolution and channel rearrangement are performed; then, in view of the problem that traditional attention mechanism only considers a single scope, this paper proposes a lightweight group-gated mixed attention (Group-gated mixed attention, GG); finally, GG is embedded in the Siamese network structure of this paper and the hierarchical feature fusion strategy is used to improve the tracking accuracy. Experiments show that the parameters of the proposed GG decrease by 26.2% compared with CBAM, decrease by 6.50% compared with SE, and increase Top-1 by 2.59% and 2.68% respectively; the experiments on the OTB100 and VOT2018 datasets demonstrate that the proposed algorithm is comparable to traditional tracking Compared with the algorithm, the accuracy and real-time performance have great advantages.