A Graph Neural Network Based Obstacle Detection for Radar Point-Cloud in Rail Transit

Rail Trainsit operates in complex environments, such as darkness, rain and fog, which pose significant risks to operational safety. Therefore, detecting obstacles in front of the train with high accuracy and robustness is crucial for ensuring safe transit operation. Radar sensors possess abilities to work in low illumination, harsh weather conditions, and being low-cost and compact, making them highly preferable for obstacle detection in complex environment. However, point clouds produced by radar are normally sparse, irregular and unstructured in nature, known as non-Euclidean structure data which cannot be adopted by conventional Convolutional Neural Networks (CNNs) directly. To address this challenge, we propose a novel detection network architecture called Rail-RadarGNN, a radar-only object detection method based on Graph Neural Networks (GNNs). Our approach integrates an embedding radar point into high-dimensional feature space, auto-feature registration (autoreg) and graph attention mechanisms to enhance feature extraction and improve detection accuracy. Experiments conducted on the Radarscene dataset demonstrate that our proposed method outperforms existing detection methods on the benchmark dataset, highlighting its superior performance in terms of accuracy and robustness. Furthermore, a self-constructed rail traffic dataset demonstrate that our proposed method effectively extracts features at an early stage, leading to more accurate obstacle detection. This advancement provides a promising solution for real-time obstacle detection in rail transit systems, enhancing the safety and efficiency of train operations under various environmental conditions.