Aiming at the problem that some artificial intelligence-based subway vehicle train inspection products in the current rail transit market have low detection accuracy when the vehicle is shifting, a computer vision-based subway speed measurement method is proposed. First, the corrected subway image is obtained through image perspective transformation, and on the basis of the corrected image, the deep learning target detection method is used to locate the subway area. Second, feature point detection is performed on the subway area located in the adjacent two frames of images, and the matching point pair is obtained by using the strongest feature point matching method in the region. Then, the average pixel distance of the subway movement is calculated according to the matching point pair, and the actual moving distance is converted by combining the pixel size. Then, the current real-time speed of the subway can be obtained through the known time difference between two adjacent frames of images. Finally, according to the speed information detected in real time, the line frequency of the line scan camera is adjusted in real time, so as to reduce the image distortion caused by the speed change or parking of the subway vehicle as much as possible. Further improve the detection accuracy of subway train inspection products when the vehicle is shifting. The actual test results at Zhujiajiao Station of Shanghai Metro Line 17 show that the line scan images of subway vehicles obtained by this technology will not be significantly distorted due to changes in vehicle speed. Moreover, the speed measurement frequency is high, and the test error is within 1.2 km/h.