Author(s)

Kai Sun, Shaobin Li, Yahan Yang, Xiaobin Di, Yu Song and Ke Chen

Corresponding Author

Kai Sun

ABSTRACT

In order to solve the current high cost of personnel and equipment, cumbersome process and difficult maintenance of traditional train detection methods at railway level crossings, this paper proposes the YOLOv3DN model, which adds the DenseNet network to the real-time target detection network yolov3 based on the Darknet architecture. It detects trains approaching and leaving the crossing in real time and provides early warning to ensure the safe passage of the crossing.The video samples of trains in various weather scenarios were collected at multiple sites near Beijing South Railway Station and Beijing Railway Station.And the network was trained and verified, which proved the algorithm's good real-time performance, accuracy and robustness. The accuracy reached 98.7%, the recall rate reached 96.86%, and the fps> 40, which can meet the needs of practical applications, and has great significance and broad prospects.

KEYWORDS

Train detection; railway level crossings; YOLOv3DN; DenseNet; Darknet; real-time detection; video