Edge aware detail enhanced pedestrian detection network

Achieving real-time pedestrian detection in complex scenarios is crucial for autonomous driving and video surveillance. Repeated downsampling causes pedestrian edge information to be lost, and channel-wise weighting is insufficient to capture spatial relationships among pedestrians. Adding feature extraction modules increases computational overhead and complicates edge deployment. To reduce detector complexity, lightweight detection heads are adopted, but may weaken discriminative features in crowded and occluded scenes. To address these issues, this paper proposes the Edge Aware Detail Enhanced Pedestrian Detection Network (YOLO-EADENet). To preserve pedestrian edge information, we introduce a Global Edge Information Transfer (GEIT) module that extracts multi-scale edge features from shallow representations and fuses them with backbone features. Coordinate Attention (CA) is incorporated to complement the spatial information missing in channel-wise weighting by leveraging directional cues. We also employ a Lightweight Shared Detail Enhanced Convolutional Detection (LSDECD) head to reduce computational overhead while improving detection accuracy. Experiments on CityPersons and CrowdHuman demonstrate that YOLO-EADENet outperforms mainstream detectors and is suitable for edge deployment. Specifically, our medium-scale model improves AP50 by 3.0 percentage points over YOLOv12-m while using 3.6M fewer parameters. Hardware deployment and evaluations on VisDrone, RTTS, and KITTI further demonstrate its applicability and deployment feasibility.