STIR-YOLO: Small Target IR-YOLO for Long-range LWIR Object Detection Under Foggy Conditions

Low-visibility conditions such as fog significantly impair vision-based pedestrian detection, posing safety challenges for autonomous driving, and surveillance. While electro-optical (Eo) cameras are highly susceptible to scattering, contrast loss, night-time glare, infrared (IR) imaging remains robust by preserving thermal signatures independent of illumination. This study provides a systematic comparison of EO and IR modalities for pedestrian detection under day and night fog conditions. An evaluation protocol is established to account for low-visibility and domain shifts, with analysis of modality-dpecific degradation effects. For IR, specialized detection strategies are investigated, including high-resolution (stride-4) detection heads, mulit-scale feature fusion, small-object reweighting (via hard-example mining), and an ILNet-style, background-suppression module to enhance thermal contrast. Benchmarking general-purpose and IR-specific detectors under consistent conditions, reveals key failure modes and robustness trends. Results show that IR-based approaches deliver more stable performance in fog, with tailored, detectors further improving sensitivity to pedestrians, particularly in severe night-time conditions. The findings provide practical guidance for sensor selection and detector design in foggy environments.