Robust Human Detection in Low-Visibility Environments Using Multimodal Thermal and RGB Fusion with YOLOv8

Abstract

Thermal-based human detection is a fundamental computer vision task for applications such as night-time surveillance, border monitoring, rescue operations, and security systems, where reliable performance under low-visibility conditions is critical. Thermal imaging presents some difficulties, such as limited texture representation, background thermal noise, scale fluctuations, and a higher chance of false and duplicate detections, yet it is intrinsically invariant to changes in illumination, unlike traditional RGB imaging. Based on the YOLOv8m object detection architecture, this study suggests a robust person detection system that is enhanced by multi-scale feature extraction and post-detection validation methods.The suggested method suppresses redundant and false bounding boxes while successfully detecting persons at different distances, orientations, and environmental conditions. While post-processing validation enhances detection confidence and spatial consistency, multi-scale detection increases sensitivity to small and far-off targets. In comparison to baseline models, experimental assessments on a proprietary thermal image dataset show significant gains in detection accuracy and precision while retaining high recall rates. The results confirm the effectiveness and reliability of the proposed framework for real-world thermal surveillance and safety-critical applications.