Real-time object detection technology in railway operations

Abstract

Development of new technology to enhance train operability, in particular during manual driving by real-time object detection on track, is one of the rising trends in the railway industry. The function of object detection can provide train operators with reminder alerts whenever there is an object detected close to a train, e.g. a defined distance from the train. In this paper, a two-stage vision-based method is proposed to achieve this goal. At first, the Targets Generation Stage focuses on extracting all potential targets by identifying the centre points of targets. Meanwhile, the Targets Reconfirmation Stage is further adopted to re-analyse the potential targets from the previous stage to filter out incorrect potential targets in the output. The experiment and evaluation result shows that the proposed method achieved an Average Precision (AP) of 0.876 and 0.526 respectively under typical scenario sub-groups and extreme scenario sub-groups of the data set collected from a real railway environment at the methodological level. Furthermore, at the application level, high performance with the False Alarm Rate (FAR) of 0.01% and Missed Detection Rate (MDR) of 0.94%, which is capable of practical application, was achieved during the operation in the Tsuen Wan Line (TWL) in Hong Kong.