DMSC-Net: a multimodal pedestrian detection network based on infrared and visible image fusion

Abstract Pedestrian detection is essential in computer vision for enhancing traffic safety. Traditional pedestrian detection methods often struggle with target distortion and detail blurring in complex environments, leading to missed detections and false positives. We propose the Dual-Stream Convolutional Channel Network, which leverages a CMFENet backbone integrated with a squeeze-and-excitation block to effectively fuse visible and infrared images. In the neck, the cross-modal multiscale fusion module is employed to enhance multiscale feature representation. This design effectively addresses challenges such as occlusion, overlap, and the detection of small edge objects. We implement extensive experiments on the public LLVIP and M3FD datasets, and compared with previous methods, our method achieves better performance. On the LLVIP dataset, the proposed algorithm achieves an mAP@50 of 98.3%, with reductions in parameters and computational costs of 51.5% and 35.6%, respectively, compared to the baseline model YOLOv8n. This algorithm improves pedestrian detection under adverse conditions while its lightweight design supports practical deployment.