Abstract: In view of problems of unsatisfactory detection accuracy and less information reported existed in pedestrian crossing detection methods, an underground pedestrian crossing detection system based on video was designed. The system is based on Gaussian mixture background modeling (GMM), and uses pedestrian crossing detection algorithm to recognize moving pedestrian objects in video streaming and mark them. Further, the state buffer will calculate the crossing trend and direction on the basis of the recognition results. The actual application result shows that the system can realize active safety detection and overcome interference such as flashing lights, and quickly identify the pedestrian location and its crossing direction, which has good stability and accuracy. The system also can achieve on-line and real-time video processing in high definition monitoring with D1 resolution at 25 fps.