Traveling by high-speed rail and railway transportation have become an important part of people's life and social production. Track is the basic equipment of railway transportation, and its performance directly affects the service lifetime of railway lines and vehicles. The anomaly detection of rail fasteners is in a priority, while the traditional manual method is extremely inefficient and dangerous to workers. Therefore, this paper introduces efficient computer vision into the railway detection system not only to locate the normal fasteners, but also to recognize the fasteners states. To be more specific, this paper mainly studies the rail fastener detection based on improved You can Only Look Once version 5 (YOLOv5) network, and completes the real-time classification of fastener states. The improved YOLOv5 network proposed contains five sections, which are Input, Backbone, Neck, Head Detector and a read-only Few-shot Example Learning module. The main purpose of this project is to improve the detection precision and shorten the detection time. Ultimately, the rail fastener detection system proposed in this paper is confirmed to be superior to other advanced algorithms. This model achieves on-line fastener detection by completing the "sampling-detection-recognition-warning" cycle of a single sample before the next image is sampled. Specifically, the mean average precision of model reaches 94.6%. And the model proposed reaches the speed of 12 ms per image in the deployment environment of NVIDIA GTX1080Ti GPU.

No abstract