Fusing binocular vision and deep learning to detect dynamic wheel-rail displacement of high-speed trains

Abstract

Large displacements between the wheel and rail lead to track damage or even train derailment. Detecting dynamic wheel-rail displacement (DWRD) can provide a technical means for discriminating the operational health status of rail vehicles and formulating active control strategies. In this paper, a new method, which fuses binocular vision (BV) and deep learning (DL), is proposed to detect the DWRD online, the method is abbreviated as BV-DL. First, a binocular camera is used to capture the dynamic wheel-rail contact video online, while two lasers are equipped to focus on the wheel-rail contact region to locate some keypoints in the wheel-rail contact images. Then, an ROI (region of interest) detection network is proposed to recognize the wheel-rail contact region automatically, and a keypoint detection network is proposed to detect the wheel-rail pixel displacement (WRPD). Finally, a 3D coordinate transformation (3DCT) model is built to calculate the wheel-rail world displacement in real-time, i.e., the DWRD between the keypoints. To test the performance of the method, a laser displacement sensor-based method is used to verify the reliability of the 3DCT model. Meanwhile, a traditional image processing method, a MobileNet V3-based method, and a YOLOX-s-based method are introduced to predict the WRPD. Experimental results show that compared with the four methods, the proposed BV-DL method can achieve optimal results in pixel coordinate calculation and 3D coordinate computation.