Abstract: Aiming at the problem of crowded and serious mutual occlusion among passengers in the process of passenger flow density detection of urban rail transit train carriage, this paper proposed a passenger flow density detection method of train carriage based on improved YOLOv5s model, designed a detection model of passenger flow density in the train compartment based on CCTV (Closed Circuit Television) system monitoring for real-time target detection. In order to solve the problem of crowd density and occlusion, the paper optimized YOLOv5s, used BiFPN (Bi directional Feature Pyramid Network) structure to strengthen network feature fusion, designed a loss function calculation method, and improved NMS (Non Maximum Suppression) method to avoid the false deletion of candidate boxes. The paper conducted experiments on the standard pedestrian detection dataset and the self-made subway carriage passenger data set. The results show that the detection accuracy of the improved model is improved compared with the original model on the two types of datasets.