Exploration of Vision-based Railway Turnout Recognition and Application

Railway turnouts are crucial components of railroad transportation, responsible for changing the direction of trains. Turnouts can aid in train positioning systems based on railway geographical information by accurately recognizing and classifying turnouts and predict the path of the train in advance, thereby enhancing the safety and efficiency of railway transit. However, the present turnout recognition algorithms perform poorly in real-world applications since there are few turnout-related datasets and the features between different classes are highly similar. This work systematically explores the problems that turnout recognition may face and corresponding improvement solutions. We customized a dataset and based on this, we conduct extensive experiments to explore the impact of factors such as data augmentation, image resolution, and training size on turnout recognition performance. The best solution can achieve 89.03% Top-1 accuracy and 93 FPS speed, enabling high accuracy while achieving real-time performance. Our work provides great promise for improving the performance of train environment perception and positioning systems and has the potential to be widely used in real-world rail transit.