The former rail transit plays an important role in the urban public transportation system, and with the rapid development of urban rail transit, the subway is a proprietary, high-density, high-capacity urban rail transit system that covers all kinds of underground and aboveground rights-of-way in urban areas. China has opened subway mileage of 6,000 km, as an important means of transportation for urban population travel, it greatly relieves the pressure of urban ground transportation. Therefore, ensuring the safety of subway trains not only helps maintain the normal operation of the subway, but also plays a vital role in safeguarding the lives and properties of passengers. The obstacles that may exist in the subway track environment are one of the important factors affecting the safety of subway trains, and the driving environment of the subway may make it difficult for the train driver to react effectively in time, which undoubtedly poses a threat to the safe driving of subway trains. Therefore, the realization of real-time detection of subway track obstacles can, to a certain extent, guarantee the safety of subway train driving.
In this paper, based on the current research of obstacle detection technology based on machine vision, an obstacle detection system for subway environment with real-time processing is researched according to the actual characteristics of the subway driving environment and the actual needs of obstacle detection. The system adopts a vehicle-based monocular machine vision method. The system uses monocular machine vision to analyze the acquired video frames, identify the tracks present in the area, and detect the targets in that target area. Information about the obstacle is derived by fusing and analyzing the algorithm results. The system applies deep learning algorithms to the system, which reduces the effect of light on the vision sensor and improves the stability and accuracy of the system. And the video decoding optimization and algorithm optimization are carried out on the embedded device, so that the frame rate of the algorithm processing video is comparable to the frame rate of the camera, which reduces the target misses due to the fast train speed and further improves the accuracy of obstacle detection.