Hydrogen refueling stations are facilities that provide eco-friendly hydrogen fuel. As per the government's hydrogen economy policy, more stations are being constructed each year. However, the number of hydrogen stations is still scarce compared to the number of hydrogen-powered vehicles being sold. Additionally, unlike LPG stations, hydrogen refueling stations have long charging times. It can be depending on the vehicle type, but usual SUVs takes around 10 minutes, and a bus takes at least 20 minutes. As a result, hydrogen vehicle users often experience significant waiting times when visiting refueling stations. To resolve this problem, previous research has been conducted on a system that identifies how many vehicles are waiting at a certain point in time. However, the previous research is not supporting a real-time system, so it is difficult to calculate operational history statistics accurately. In this study, we propose a method to generate more diverse hydrogen refueling station operation information than previous studies by utilizing a deep learning-based multi-vehicle tracking technique. Firstly, we use Deep-SORT, one of the Multi-Object Tracking (MOT) algorithms, to assign unique IDs to waiting vehicles at hydrogen stations and track them in real-time, and use the trained YOLOv4 model to detect and classify objects in the images. With using the collected data, the valid information, such as when each vehicle enters the station, how long it waits, and the cumulative number of vehicles visited, can be determined. The validation dataset categorizes environmental conditions by time of day (night/day) and weather (sunny/rainy) and uses video data with more than 10 waiting vehicles. Accuracy is calculated based on the Mean Absolute Percentage Error (MAPE). The results of the study show that the accuracy of when a vehicle enters a charging station, how long it waits, and the cumulative number of vehicles visiting a charging station are all above 85% on average. Error is mainly caused by two similar cars being recognized as one vehicle when they are close together, or by increased ID switching due to occlusion. The solution proposed in this study is expected to be useful in areas related to hydrogen refueling station infrastructure, such as providing a notification service for refueling station managers to recognize vehicle entry, utilizing attribute data to develop a model to observe future waiting times, and providing historical information on usage statistics for each refueling station.
Ms. Eunjung Roh, Senior Engineer, HL Mando