End-to-end railway obstacle detection enhanced by point cloud segmentation

Railway transportation is an essential component of the transportation system. The crucial point to keep railway transportation safe is an accurate and efficient railway obstacle detection system that intrudes into the railway tracks. Thanks to the development of image processing, object detection in railways uses color image data to find obstacles. However, it may lead to problems such as low-quality images in low illumination and bad weather. In recent years, light detection and ranging sensors have shown potential in railway obstacle detection. In this paper, we propose a novel railway obstacle detection framework enhanced by a point cloud segmentation model based on deep learning. High-resolution railway point cloud data is collected by light detection and ranging sensors, and each point gets a predicted label by the segmentation model. Based on isolated track points, a track region is generated. The railway background is compared with the real-time point cloud to output the different points, which are detected obstacles. Extensive experiments have demonstrated the advantages of our approach, including improved detection accuracy with a detection rate of 90% for cubic obstacles measuring 15 centimeters per side in a range of 50 meters, significantly reduced sensitivity to rain noise, removing 81.61% of rain-induced noise points, and lower computational cost, achieving segmentation of 1.6 million points in just 0.8 s. Our method also offers a new benchmark for the detection of railway obstacles. Furthermore, our framework has been successfully deployed and validated in real-world railway scenarios, demonstrating practical feasibility and robust operational performance.