Transformer-Based Memory Guided Thermal Infrared Target Tracking Framework for Traffic Assistance

Due to the nature of infrared imaging, accurately tracking targets in traffic environments crowded with vehicles and pedestrians poses considerable challenges for TIR trackers. In scenarios with similar interfering objects or target occlusions, trackers tend to drift toward distractors, leading to tracking failures. To address this challenge, this study proposes MGTrack, a Memory Guided network for TIR Target Tracking. MGTrack efficiently leverages temporal context to enhance the discriminative capacity of TIR target features. In this study, an Adaptive Recurrent Memory Network (ARMN) is proposed to implicitly encode spatiotemporal information implicitly during tracking, thereby maintaining memory updates and leveraging memory to effectively enhancing spatial features. Furthermore, a Memory Filter module is introduced to eliminate redundant information during the memory encoding process, which further enhances tracking robustness. Comparative experimental results on two widely used TIR target-tracking benchmarks demonstrate that the MGTrack is both efficient and effective, outperforming existing methods and achieving superior performance across various traffic scenarios.