A New Obstacle Detection Approach for Railway Transit Using Cooperative Deep Learning Models

Abstract

In railway lines, the safety of the line and the line surroundings is of great importance for the operation of the train. An obstacle around the line causes accidents in railway transportation and endangers line safety. Ensuring a safe operating condition for the train and warning the driver early is an important need in this respect. Although deep learning-based object detection methods are recommended for detecting obstacles in the rail and surrounding objects, distance detection is not performed in these methods. In distance detection studies, additional cameras and sensors are needed. In this study, it is aimed to detect the rail and surrounding objects with a two-stage network. The first network determines the rail and its surroundings by performing semantic segmentation in real time. Then, if there is an obstacle in the segmented rail area, the distance of the obstacle will be estimated with a deep convolutional regression network. The modified Unet model will be used for segmentation. RailSem19 dataset is used to segment the railway and its surroundings. The object around railway is detected by using YoloV8 object detection model. The ZoeDepth model is activated only when there is an obstacle and the distance of the obstacle is measured. For this purpose, the network is trained on the KITI dataset and the distances of the images detected during the testing phase are determined.