SSiamTrack: Infrared Target Tracking via Spatiotemporal Siamese Network

Thermal infrared (TIR) target tracking is a critical task in computer vision with substantial practical value. Due to the unique imaging principles employed in TIR, these images can effectively mitigate the adverse effects of illumination changes in complex environments. Consequently, TIR has been widely applied in all-weather environment perception for unmanned aerial vehicles (UAVs), especially in disaster rescue, battlefield reconnaissance, and security monitoring. However, the inherent lack of sufficient texture information in TIR images can limit the accuracy of UAV target tracking. To address this challenge, we propose a Spatiotemporal Siamese Network Tracker based on a Convolutional Attention Mechanism (SSiamTrack). Firstly, our method introduces CIoU Loss, which simultaneously accounts for both the center-point distance and the diagonal distance of bounding boxes. Secondly, a lightweight detection head is constructed using Depthwise Separable Convolutions to reduce the parameter. Finally, we design an efficient plug-and-play multi-scale feature fusion module, that utilizes an attention mechanism to effectively integrate shallow spatial features with deep semantic features. Experimental results demonstrate that, SSiamTrack reduces parameters by 9.5% while achieving a 9.0% and 6.5% increase in success rate and precision on the LSOTB-TIR, respectively, and a 9.7% and 9.0% improvement on the PTB-TIR.