CNN and Binocular Vision-Based Target Detection and Ranging Framework of Intelligent Railway System

In this work, we propose an efficient convolutional neural network (CNN) and a binocular vision-based target detection and ranging framework for intelligent railway systems to localize and detect the trackside equipment, where the location, classification, and boundary violation of the target are jointly addressed. First, we develop an effective enhanced YOLOX-tiny model with the aid of coordinate attention (CA), CIoU, and focal-loss to improve the detection performance, by ensuring the detection speed. Our model improves mAP by 3.25% compared with the original YOLOX-tiny model. Then, we develop a CNN and binocular vision-based target detection and ranging system by integrating the YOLOX target detection algorithm, binocular ranging algorithm, and OCR character recognition algorithm, and proposing the RGB-d-OCR Matching (RDOM) algorithm to match the results of the three algorithms. Moreover, we deployed the proposed algorithm on edge terminals and built a track inspection vehicle to carry the system. As the vehicle travels on the track, our algorithm can automatically position the trackside equipment on both sides and measure the distance to determine whether the equipment has a boundary violation. The experimental results show that the proposed system can achieve an equipment recognition rate of 91.8%, an ID recognition rate of 88%, and a ranging error within 2%. In particular, RDOM successfully matched the results of the three algorithms, ensuring that the detected trackside equipment corresponds to the correct distance and ID. Our work is a revolutionary method that integrates target recognition, OCR recognition, and binocular ranging while considering detection accuracy and efficiency.