Enabling Context-aware Collision Avoidance for Autonomous Railway Maintenance Vehicle

Abstract

Compared to the recent advancements in self-driving cars, There is a significant gap for autonomous railway vehicles, which still rely on a basic “start-stop” approach without being aware of their surrounding environment. This research developed a context-aware collision avoidance system for autonomous railway vehicles. It addressed the gap by enabling the vehicle to make appropriate decisions by distinguishing obstacles’ types and locations. A key challenge was obtaining an annotated dataset for railway object detection, alongside balancing detection accuracy with precise distance estimation using cameras and LiDAR. The model utilised Transfer Learning by pre-training models on large datasets like COCO and RailSem19. It then leveraged this knowledge by fine-tuning it with a custom dataset tailored to railway maintenance scenarios. A sensor fusion approach was implemented to combine the strengths of both sensors. This multimodal approach extended 2D object tracking results into 3D by aligning the camera and LiDAR’s coordinate systems. The collision avoidance system was integrated into a robotic framework built on ROS2, which supported real-time processing and communication between each sensor. The validation experiments included both controlled indoor and outdoor scenarios. Indoor validation evaluated the system’s performance under various lighting conditions; outdoor validation was conducted on a test track to ensure a high Technology Readiness Level. The indoor validation identified the optimal operating range for illumination and provided guidelines for handling extreme conditions, such as tunnels or night-time environments. The outdoor validation achieved a high success rate of 80% by making correct actuation decisions based on the detected obstacles’ types and positions.