YOLOv11-MSDAF: A Multi-Scale Dual-Attention Fusion Network for Pedestrian Detection in Crowded Scenes

Abstract

edestrian detection is a fundamental task in computer vision with broad applicability across intelligent public surveillance, autonomous driving collision avoidance, and human-computer interaction systems. However, it faces two principal challenges: frequent occlusions in dense crowds that obscure discriminative visual features, and severe multi-scale variations arising from diverse camera viewpoints and distances, which undermine the model’s capacity to capture fine-grained target information. To tackle these problems, we propose a novel pedestrian detection algorithm named YOLOv11-MSDAF. Specifically, Multi-Scale Convolution Module is proposed that integrates multi-scale convolution, channel shuffling, and an SE attention mechanism to enhance the fusion and representation of fine-grained features for small and occluded pedestrians in the C3k2 blocks of the backbone and neck networks. In addition, an improved detection head DETECT-ASFF-SGA is introduced to embed learnable scale-gate coefficients and the parameter-free SimAM attention mechanism to enable adaptive multi-scale feature aggregation and refined spatial selectivity in the detection process. Extensive experiments are conducted on the CrowdHuman and WiderPerson datasets to demonstrate the competitive performance of the proposed YOLOv11-MSDAF in the pedestrian detection task.