A Safe Vision Method for Train Positioning

Train positioning is crucial for safe operation of a train. In this paper, we propose a novel safe vision method to achieve train positioning by recognizing wayside kilometer posts. The proposed method contains three parts: kilometer posts detection, kilometer posts semantic recognition, and runtime safe monitoring. A YOLO v5 model is developed to detect kilometer posts from the wayside while a train is running. A Convolutional Recurrent Neural Network (CRNN) model is used to recognize the information on the kilometer posts indicating an absolute position of the railway line. In addition, a safety monitor is proposed to avoid machine learning-based vision methods resulting incorrect recognition results. The monitor verifies whether the recognition results satisfy rule-based properties at runtime. Following the fail-safe principle, the monitor prevents the recognition results from being used for train positioning whenever a violation is detected. We evaluate our method on several benchmarks on a semi-physical simulation platform. Experiment results show that our models efficiently recognize kilometer posts with 88.32% accuracy. With the safety monitor, the safe positioning rate reaches 99.50%, providing a new idea for train visual positioning.