LWIR-HPE: a high-variance keypoint perception network for human pose estimation in weakly textured LWIR scenes

Human pose estimation under adverse imaging conditions remains challenging for visible-light systems, where rain, snow, haze, and low illumination often cause severe contrast degradation and background interference. Although long-wave infrared (LWIR) imaging is less sensitive to illumination changes and better suited to privacy-sensitive scenarios, its weak texture and blurred structural cues still hinder reliable foreground-background separation, particularly for keypoints with large positional variation. To address this problem, we propose LWIR-HPE, a dedicated human pose estimation framework for weak-texture LWIR scenes, and construct FI-POSE, a dedicated LWIR human pose estimation dataset for training and evaluation. LWIR-HPE incorporates two task-specific components: a Dual-Channel Aggregated Convolutional Attention Module (DACAM), which enhances contour-aware feature responses while suppressing background distractions, and an enhanced Universal Inverted Bottleneck module (UIB-E), which improves multi-scale structural representation and contextual aggregation for keypoints with large positional variation. A multi-task prediction strategy is further introduced to improve localization robustness. Experimental results show that LWIR-HPE achieves 90.8% PCKh on the FI-POSE test set, outperforming the baseline by 16.4 percentage points. Additional experiments on an infrared-domain validation set provide complementary evidence of its applicability under another infrared imaging setting.