Unsupervised Deep Learning-Based Point Cloud Detection for Railway Foreign Object Intrusion

Railway foreign object intrusion detection (FOID) is a crucial sensing task for avoiding collisions between trains and obstacles ahead, ensuring operational safety. However, the complexity of the railway operating environment makes existing camera-based FOID susceptible to weather and lighting conditions, thereby reducing reliability. LiDAR has become one of the important sensing sources onboard because it is less affected by the environment. Therefore, this paper proposes a deep learning approach for railway FOID based on point cloud. This approach adopts an unsupervised learning framework that includes masked point prediction and anomaly detection, overcoming the difficulties of point cloud annotation under fully supervised learning. Additionally, considering the absence of anomalies in the collected data, a pseudo-anomaly sample generation algorithm is designed for generating test data. Experiments are conducted on the point cloud collected from the onboard view of the Beijing Yanfang Line. The results show that the proposed method can achieve a detection rate of 92.2% without labeled data.