Enhancing Infrared Tracking for Visual Semantic Navigation via Parameter-Efficient Adaptation of AI Foundation Models

Navigation systems increasingly rely on semantic perception, where object-level segmentation and tracking provide structured information for mapping, localization, and decision-making. Recent foundation models such as Segment Anything Model 2 (SAM2), building on the original Segment Anything framework (SAM), deliver strong segmentation and tracking performance on RGB imagery, but their robustness deteriorates in low-light or visually degraded environments. Thermal infrared (IR) sensing provides complementary, illumination-invariant measurements in such conditions, yet IR-focused foundation models remain scarce because large, densely annotated IR datasets are rarely available. To address this gap, we adapt SAM2 to the IR domain using a parameter-efficient Low-Rank Adaptation (LoRA) strategy that injects trainable low-rank modules into early encoder layers while freezing all original backbone parameters. In addition, we construct pixel-wise supervision from only axis-aligned bounding boxes by iteratively refining pseudo masks in a BoxSup-style pipeline and using them to update only the LoRA parameters. Evaluations on the LSOTB-TIR benchmark show that, whereas baseline SAM2 often loses the target or produces noisy masks under challenging attributes such as distractors and occlusions, the proposed LoRA-adapted model maintains more stable target trajectories and sharper segmentation boundaries. These results demonstrate that lightweight domain adaptation guided by pseudo masks can effectively extend RGB-trained foundation models to IR perception, providing stronger and more reliable semantic cues for navigation applications including semantic SLAM and cooperative localization.