Efficient Thermal Infrared Tracking via Multi-Enhancement and Spatial Shift Pyramid

Inherent issues with thermal infrared (TIR) images include blurred object boundaries and insufficient texture features, leading to subpar performance by existing TIR trackers in distinguishing foreground from background and capturing fine-grained details. To overcome these limitations, we introduce an efficient TIR tracking model, called ETIT, based on channel domain multi-group enhancement (CDME) module and spatial shift pyramid (SSP) module. First, the CDME module bridges the gap between high and low layers in the search region, enhancing object features of the search region while suppressing background information, ensuring our tracker can accurately capture detailed features from the search region early on. Second, given the insufficient multiscale characteristics of the original response map, the SSP module introduces a spatial shift module and a multiscale pyramid (MP) structure. This approach leverages the complementarity of semantic and pixel-level information to address the inherent lack of color and texture information in TIR images, significantly boosting the response map’s sensitivity to the object. To validate the superiority of our ETIT model, we conduct extensive comparative and ablation experiments on five challenging benchmarks. Experimental results demonstrate that our model outperforms current tracking models in most cases, achieving a good balance between performance and speed.