Improved Algorithm with YOLOv5s for Obstacle Detection of Rail Transit

As an infrastructure for urban development, it is particularly important to ensure the safe operation of urban rail transit. Foreign object intrusion in urban rail transit area is one of the main causes of train accidents. To tackle the obstacle detection challenge in rail transit, this paper introduces the CSYOLO urban rail foreign object intrusion detection model. It utilizes the improved YOLOv5s algorithm, incorporating an enhanced convolutional attention CBAM module to replace the original YOLOv5s algorithm's backbone network C3 module. In addition, the KM-Decoupled Head is proposed to decouple the detection head, and SIoU is applied as the loss function. Tested on the WZ dataset, the average accuracy increased from 0.844 to 0.893 .The research method in this paper provides a reference basis for urban rail transit safety detection, which has certain reference value.