In the complex working environment of urban rail transit, unsafe behavior of workers is the main factor causing safety accidents. In view of the problem that operators do not wear helmets and fluorescent clothing as required in the construction of urban rail transit, an improved algorithm based on YOLOv5 is proposed to detect whether operators wear helmets and fluorescent clothing. First, YOLOv5 is combined with multi-headed self-attention to enhance the extraction of small target features from deep network and improve the expression ability of network features; Secondly, the convolutions in YOLOV5 are replaced by depthwise separable convolutions, which can further reduce the network parameters; Finally, context information extraction module is fused for improvement to fully capture the relationship between pixels. The experimental results show that the improved algorithm mAP reaches 88.5, 3.5% higher than the original YOLOv5 algorithm, and can basically meet the needs of daily safety supervision.