A Rail Detection Method Using Bird's Eye View of LiDAR Point Clouds

Rail detection is an important part of intelligent rail transportation, and research on rail recognition is of significant importance for both detecting foreign object intrusion in railways and realizing unmanned driving of trains. This paper proposes a new method for rail detection based on bird's-eye view (BEV) of LiDAR point clouds, which aims to solve the problems of traditional point cloud algorithms, such as weak robustness and low accuracy, as well as the problems of large computation and poor real-time performance in point cloud deep learning networks. The proposed method constructs a BEV by projecting the rail point cloud onto a 2D image and retaining the point cloud features, and then uses an anchor-based rail detection convolutional neural network (CNN) to detect the rail. Finally, the predicted rail are mapped back to the original point cloud to label the rail. The experiments on a self-built dataset show that the proposed method achieves good results in different rail scenarios.