Vision and Inertial Sensors Fusion for Train Positioning in GNSS-Denied Environments

Abstract

The accurate train positioning is essential for ensuring safety and operational efficiency in modern rail systems. Traditional methods based on trackside infrastructure or satellite signals often suffer from limited precision or high cost, especially in Global Navigation Satellite Systems (GNSS)-denied environments. To address these challenges, this article proposes a hybrid vision–inertial train positioning method that combines the visual absolute positioning with inertial measurement unit (IMU)-based relative positioning. An enhanced you only look once (YOLO)-based object detection algorithm and an end-to-end text recognition network are employed to identify and interpret railway landmarks. The absolute position of the train is then retrieved by matching recognized text with a preconstructed database. To achieve continuous and robust localization, a differential evolution Kalman filter (DE-KF) is introduced to adaptively fuse IMU data with the vision-derived observations, dynamically tuning the process noise covariance in response to environmental variation. The proposed method was validated at Beijing National Railway Experimental Center. Experimental results demonstrate that the system maintains positioning errors within 3.5 m and achieves high recognition performance, with an mAP50 of 98.0%. These findings confirm the effectiveness of the proposed fusion framework for real-time, accurate, and resource-efficient train localization.