Pedestrian Target Detection in Low-Light Conditions Using DSCCA-Enhanced YOLOv10

Pedestrian detection in low-light conditions is critical for applications like nighttime surveillance and disaster rescue. However, image degradation severely compromises feature extraction, leading to significant performance drops in existing detectors. To address this, we propose a Dual-Modal Spatial-Channel Cooperative Attention (DSCCA) module integrated into YOLOv10. Our key innovations include: (1) A lightweight modal-interactive gating mechanism that uses infrared features to preserve visible-light textures while enhancing infrared edges with visible cues; (2) A dual-branch attention structure (channel + spatial) to amplify discriminative features; (3) An illumination-adaptive fusion strategy that dynamically balances branches based on light intensity. Experiments on LLVIP and CVC14 datasets show state-of-the-art results: 69.2% mAP50–95 on LLVIP and 59.4% mAP50–95 on CVC14, outperforming existing fusion methods (CFT, ICAFusion) while maintaining low computational cost (7.2 M parameters).