Abstract: Objective: Accurately monitoring the status of large passenger flow in metro system is the basis for recognizing and managing passenger flow risks. Metro stations have limited spatial capacity and are prone to risky incidents such as passenger flow congestion and conflicts, making it necessary to develop a method for real-time recognition of PMT (passenger flow movement trajectories) in metro stations. Method: A method is proposed for recognizing metro station PMT based on SORT (simple on-line and real-time tracking) algorithm and optical flow model, which serves as an auxiliary means for metro safety protection and control. Firstly, the SORT algorithm is adopted to achieve real-time passenger flow recognition in high-density scenarios; secondly, the PWC-Net (pyramidal, warping and cost volume optical convolutional neural network) optical flow model is employed for real-time passenger flow locating and PMT recognition; finally, the proposed method is validated using real station video data of Shanghai Metro. Result & Conclusion: The above-mentioned recognition method effectively associates targets, recognizes overlapping passenger flows, and improves the real-time nature of recognition. By employing the PWC-Net optical flow model for recognizing PMT, the model size and training time are 94.12% and 50.00% of the FlowNet2 (network based on dataset training sequence) model respectively. Therefore, this method can meet the scenario requirements of real-time recognition of high-density PMT in metro stations.