Abstract: Aiming at the problem of anti-collision warning during underground electric locomotive running, a deep learning-based obstacle location method for underground track is proposed. Combining Yolov5 target detection and UFLD track detection algorithms, the monocular distance measurement is carried out by using the property of constant actual track width. First, the y coordinate at the bottom of the Yolov5 detection frame is used to locate the target obstacle and the track on the same horizontal line. Secondly, the track width in the pixel coordinate system of the image window is calculated. Finally, the distance of the target obstacle is measured through the principle of small-hole imaging and the conversion of multiple coordinate systems. The experimental results show that the error rate of the system is less than 5% in the visual range, and the average running time of the video frame is 35 ms, which meets the real-time requirement of the intelligent control of underground track electric locomotive.