An End-to-End Door Breaching Prediction Framework Based on Improved RAFT Optical Flow Estimation

Metro systems play a vital role in modern urban transportation, but passenger door-breaching behavior during the final seconds of boarding poses severe safety and operational risks. To address this issue, this paper proposes an end-to-end threat prediction framework based on an improved lightweight optical flow estimation network, Tiny-RAFT. The method first employs Tiny-RAFT to extract pixel-level motion vectors between consecutive video frames, obtaining real-time speed and directional information of passengers approaching the platform doors. These optical flow features are then fused with temporal information representing the remaining door-closing time to generate a dynamic threat value through a fully connected prediction layer. When the predicted value exceeds a set threshold, the intelligent door control system activates visual and auditory warnings and coordinates with train and platform doors to prevent accidents. Experimental results on the Human Flow dataset and a self-constructed metro door-breaching dataset demonstrate that the proposed method achieves higher accuracy and faster inference speed compared to existing approaches, ensuring both real-time performance and reliability for metro safety management.