ASMTrack: Thermal Infrared Target Tracking Network Based on Atkinson-Shiffrin Memory Model

Thermal Infrared (TIR) target tracking is a critical technology in intelligent transportation systems (ITSs) due to its resilience to weather and lighting variations. However, existing approaches typically rely on sparse template matching with the current search region, which is insufficient for extracting strongly discriminative features from TIR images that lack texture, color, and other fine details. To address these limitations, this paper presents ASMTrack, a novel TIR tracking framework inspired by the Atkinson-Shiffrin Memory Model. The framework employs a unified memory mechanism that integrates real-time sensory memory, fine-grained short-term memory, and compact long-term memory, enabling exploitation of temporal contexts across hundreds of frames. A Memory Extraction Network captures rich historical information by conducting dense associations between sensory memory and the memory pool. Furthermore, a memory aggregation strategy transforms sensory memory into high-resolution short-term memory, ensuring efficient preservation of visual features from dozens of recent frames, while memory compression periodically distills frequently accessed short-term information into condensed long-term memory. Experiments on two widely used benchmarks demonstrate that ASMTrack achieves state-of-the-art performance, and ablation studies further validate the effectiveness of each component.