Research on video-based character recognition method for train cargo cars

Abstract

After the train reaches the destination station, the actual arriving train number needs to be confirmed to ensure that the arriving vehicle is accurate. The traditional method of obtaining car numbers is by manually transcribing car numbers, which has the disadvantages of large workload, low efficiency and error-prone. Therefore, this paper designs a video-based train carriage character recognition system, and proposes a method of locating and recognizing train carriage characters based on YOLOv5s and CRNN. The four target detection models, SSD, Faster -RCNN, YOLOV5m, and YOLOv5s, are trained until the models converge completely using 10,000 train carriage datasets manually labeled with Labelimg software. The experimental results show that YOLOv5s outperforms the other models in terms of recognition accuracy and recognition efficiency, with YOLOv5s achieving a detection accuracy of 0.99 for the carriage character region and an average detection speed of 31 frames. The average accuracy of the algorithm is 0.96. Finally, the fusion of YOLOv5s and CRNN models results in a stable average frame rate of more than 20 frames for the detection of complete train carriages. The method in this paper can automatically identify the carriage numbers, which can reduce the labor of workers, avoid the errors caused by manual transcription of records, increase the real-time recording, and is of great significance to the development of railroad transportation system.