Risk level assessment of railway obstacles by integrating scene parsing and spatial distance

Abstract

Existing vision-based detection techniques place excessive emphasis on object recognition. They largely ignore the assessment of intrusion risks. To tackle this problem, the present paper introduces a risk level evaluation method that incorporates lateral distance. The method first employs YOLORS to perform railway track segmentation and obstacle detection. Its backbone network integrates CSP modules, an SPPF module, and the ResNet50 architecture. Lateral displacements are then classified into three risk levels: high, medium, and low. Experiments conducted on a self-built dataset show that YOLORS achieves 84.9% mAP at a processing speed of 154 frames per second. Notably, its segmentation branch maintains strong edge consistency even in highly complex environments. On this basis, the lateral distance measurements are used to quantify the intrusion risk of foreign objects on the track. This enables a graded early warning system. The approach provides a quantitative decision-making basis for active train protection. As a result, the intelligence level of safety protection is significantly enhanced.