Lightweight of SiamCAR Network for UAV Single Target Track

UAV single target tracking is one of the hot research directions in UAV field. Since the trackers based on deep learning usually have complex network structure and need a lot of computing and memory resources in the process of running the model, the realtime requirement cannot be guaranteed when it is applied to the low-cost on-board computer. In this paper, a reasonable network lightweight method is proposed based on the simple target tracking framework Siam CAR, and a lightweight full-convolution target tracking algorithm combined with multi-feature fusion is proposed. In this method, the GHI-block (Ghost expansion for Hybrid Inverted residual convolution block) proposed by this paper is introduced into the feature extraction network to achieve lightweight and feature fusion. The experimental results tested on UAV123 benchmark show that compared with the original algorithm, our algorithm reduces parameters by 800%, increases speed by 195% (700% increase in speed on low-power platforms). While having fewer parameters and FLOPs(Floating Point Operations), the improved algorithm achieves competitive tracking accuracy, and can meet the realtime requirements of UAV.