Automatic Obstacle Detection Method for the Train Based on Deep Learning

Automatic obstacle detection is of great significance for improving the safety train operation. However, existing autonomous operation trains mainly depends on signaling control system and lacks extra equipment to perceive environment. To further enhance efficiency widely deployed fully automatic (FAO) systems train, this study proposes an intelligent based deep learning. It collects perceptual information from industrial cameras light ranging (LiDAR), implements functionality including rail region detection, visual–LiDAR fusion. Specifically, first two parts adopt convolutional neural network (CNN) algorithms semantic segmentation object pixel-wisely identify track area ahead detect potential obstacles track, respectively. The fusion part integrates visual data with LiDAR achieve environmental perception all weather conditions. can also determine geometric relationship between decide whether trigger a warning alarm. Experimental results show that proposed in has strong performance robustness. rate (precision) 99.994% recall reaches 100%. system, applied metro Hong Kong Tsuen Wan line, effectively improves urban