YOLO-B2DV: An Integrated Real-Time Framework for Detection, Tracking, and Recognition of Catenary Pole Number Plates in Railway Video Streams

In order to realize the automatic recognition and tracking of the catenary-pillar number plate in the intelligent railway system, this paper proposes an integrated real-time video stream analysis framework YOLO-B2DV. This study is based on a self-built and publicly available complex scene dataset, which provides a reliable benchmark for algorithm performance evaluation. Aiming at the challenges of low detection accuracy and easy target loss faced by traditional methods in railway video streaming scenes, this framework innovatively integrates three core modules: (1) The improved YOLO-B2 detection model significantly improves the ability to locate number plates under complex backgrounds by introducing Bidirectional Feature Pyramid Network (BiFPN) and P2 small target detection layer. (2) The task-optimized DeepSORT tracker effectively overcomes the problem of cross-frame target loss caused by train vibration and occlusion; (3) The end-to-end RBC character recognition framework, based on the ResNet-BiLSTM-CTC cascade structure, achieves high-precision recognition without explicit character segmentation. Experimental verification shows that YOLO-B2DV achieves an average recognition accuracy of 89.12% on the self-built data set, which is about 7% higher than the mainstream benchmark methods. It shows significant advantages in detection accuracy, tracking stability and system real-time performance, which provides an effective technical way for the intelligent operation and maintenance of railway infrastructure.