HiPATrack: Hierarchical Dependency and Position-Aware for TIR Object Tracking

Abstract

Most current thermal infrared (TIR) object tracking methods are adapted from the RGB domain, and their feature fusion strategies typically emphasize high-level semantic representations. As a result, the already sparse low- and mid-level discriminative cues (e.g., texture and boundaries) in TIR images are further attenuated, leading to degraded performance under challenging conditions such as similar-object interference and occlusion. To address this issue, we propose HiPATrack, a TIR tracking framework that combines hierarchical dependency modeling with position-consistency constraints. Specifically, the Hierarchical Dependency-aware Fusion Module (HD-FM) enlarges the receptive field via cross-layer dilated convolutions and leverages self-attention to explicitly model global dependencies across feature hierarchies, thereby reinforcing weak low- and mid-level discriminative responses. In addition, the Position-Aware Gating Module (PAGM) introduces absolute position encoding and depthwise separable convolutions to impose geometric constraints and improve spatial alignment, mitigating positional drift caused by insufficient semantic cues in TIR imagery. Extensive experiments on five mainstream TIR tracking benchmarks demonstrate that HiPATrack achieves an average success rate (Suc) gain of 4.8% over the state-of-the-art Transformer-based tracker OSTrack. Note to Practitioners—Thermal Infrared (TIR) object tracking is increasingly critical in various intelligent automation systems, including robotic navigation and surveillance systems, particularly under challenging conditions such as low visibility or night-time operations. In the domain of robotic navigation, TIR object tracking enables systems to reliably detect and monitor targets in low-light environments, thereby enhancing autonomous capabilities and the decision-making process. Within surveillance systems, TIR tracking is essential for detecting intruders or objects based solely on thermal signatures, providing an additional security layer. This paper introduces HiPATrack, an innovative framework designed to improve tracking accuracy and speed in TIR environments. The HiPATrack framework incorporates hierarchical dependency feature and position-aware feature alignment techniques, specifically optimized to address the unique challenges of TIR object tracking, such as similar target interference and partial occlusion. Extensive evaluations of HiPATrack in multiple TIR benchmark datasets demonstrate its leading advancements in accuracy and robustness.