Dual-backbone dynamic fusion of visible and thermal features for precise low-light pedestrian detection
Guanglei Zhu, Md Abdus Samad Kamal, Kou Yamada · IET conference proceedings. · 2026
Low-light conditions remain a significant challenge for pedestrian detection due to the inherent limitations of both visible-light and thermal images, which can lead to reduced detection accuracy. Existing studies primarily focus on directly fusing the two modalities, often overlooking the potential interference between them. To address this issue, we propose a dual-stream YOLOv8 architecture that simultaneously processes visible and thermal images while enabling dynamic fusion to mitigate inter-modality interference. To avoid excessive model complexity, we introduce a lightweight Modulation Fusion Module (MFM) for adaptive feature integration. Furthermore, we redesign the original C2f module in YOLOv8 by incorporating the Inverted Residual Mobile Block (iRMB), resulting in a novel and efficient C2f_iRMB module. The proposed model is trained on the LLVIP dataset, and its effectiveness is evaluated through ablation studies. Comparative experiments with existing state-of-the-art methods demonstrate that our approach achieves superior performance in low-light pedestrian detection tasks.