A Real-time Train Perception Method for Obstacle Intrusion Based on Front View Projection

Abstract

In order to improve the obstacle perception ability of automatic train operation in rail transit, it is necessary to increase the train's ability to perceive obstacle intrusion in the operational scenarios. Aiming at addressing the limitations in the commonly used multi-sensor fusion algorithm, which include inadequate clearance analysis, poor real-time performance and high computing power demands, this paper proposes a method to perceive encroachment obstacles in real time with cameras and LiDAR as sensors. The proposed method, which applies the information fusion approach based on the two-dimensional projection plane of the front view (FV), involves the establishment of a projection matrix through offline joint calibration, projection of the LiDAR point cloud onto the FV plane, and extraction of track images from cameras, to enable clearance calculation. By incorporating correction of sensor synchronization errors based on point cloud prediction, a judgment regarding track intrusion can be made, dependent on the projected obstacle point cloud and calculated clearance. Data collection and experimental verification were carried out using a train equipped with sensors and a perception system. The experimental results show the practicality of the proposed method on low-power embedded devices, achieving an average algorithm time of 16.2 ms. Moreover, the method is proven effective in real-time detection of track intrusion by obstacles ahead of trains, while simultaneously acquiring their locations and sizes.

Keywords

• autonomous driving
• obstacle perception
• multi-sensor fusion
• camera
• lidar
• rail transit