A Multi-Sensor Fusion System for Rail-Side Detection Using LiDAR, Inertial Measurement Unit, and GPS Data Based on Two-Stage Machine Learning Framework

Rail-side detection is crucial for ensuring the safety and precision of Automatic Train Operation (ATO) systems. Accurate identification of whether a train aligns with the left or right rail enables reliable localization, trajectory estimation, speed regulation, and precise station stopping. This paper proposes a novel Automatic Rail Position Detection (ARPD) system that fuses Light Detection and Ranging (LiDAR) point clouds with Inertial Measurement Unit (IMU) and Global Positioning System (GPS) data for robust rail-side classification. The key contribution is a two-stage Artificial Intelligence framework that improves both accuracy and computational efficiency. The first XGBoost model separates rail points from non-rail objects, while the second classifies them as left or right rail. This staged design reduces classification ambiguity and outperforms conventional single-model approaches. Trained on 61,876,680 annotated instances, the proposed system achieves near-perfect performance, with precision, recall, F1 score, and accuracy close to 100% and an ROC-AUC of 0.99. In addition, the ARPD system achieves extremely low per-instance inference latency on the evaluated workstation platform, demonstrating suitability for real-time application to onboard ATO platforms. The experimental results demonstrate that multi-sensor fusion combined with staged AI processing provides a highly accurate and efficient solution for next-generation autonomous railway systems.