Real-Time Positioning of Autonomous High-Speed Trains: Combining AI-based Perception and Multi-Source Sensor Data

Due to the complex environment and crowded traffic density, the real-time positioning of trains in high-speed railways is critical to the safety and efficiency of the whole system. Most existing studies in this field focuses on data fusion techniques with collected data from GPS, Balise, and track circuits, while these data from different sources are usually inaccurate and “discrete”. In our study, we proposed a new approach that combines the multi-source data fusion with AI-based perception from in-cabin videos. Specifically, we developed an attention-based (Convolutional Neural Networks) CNN model. Our model has three modules: 1) CNN achieves visual perception recognition from line-side fixed objects (e.g., positioning poles of contact network); 2) A fuzzy matching module corrects the perceptual data; and 3) an attention module integrates the perceptual data with positioning data from multi-source sensor. We collected the real-world data from EOAS system in Beijing-Shanghai high-speed railway. We trained and test our deep learning model with several benchmarks, which verify the effectiveness of our approach regarding the positioning accuracy of high-speed trains.