Pedestrian Detection in Urban Rail Transit Based on Deep Learning

To address the issues of reduced accuracy and increased missed detection rates caused by occlusion and crowd density in pedestrian detection for urban rail transit scenarios, this paper proposes a deep learning-based pedestrian detection method tailored to such environments. The proposed approach introduces a small-object detection head into the network architecture to enhance the model's sensitivity to small-scale pedestrians. Additionally, an improved feature fusion structure, AFPN-4H, is designed to optimize inter-layer information interaction and improve the model's adaptability and accuracy across multiple scales. Furthermore, the integration of the GIoU loss function significantly improves the localization accuracy of occluded targets. Through the joint optimization of network structure and loss function, the model demonstrates enhanced robustness under complex rail transit conditions. Experimental results on the WiderPerson dataset show that, compared with the original YOLOv11n model, the proposed method achieves performance improvements of 1.7 % in precision (P), 2.6 % in recall (R), 2.7 % in [email protected], and 2.7 % in [email protected]: 0.95.