Track Fastener Detection Based on Improved YOLOv4-Tiny Network

Abstract

Track fasteners on the railway tracks are extremely critical to ensure the safe operation of the railway transportation system. Fast and accurate fastener detection is of great significance for improving the inspection efficiency of railway tracks. However, the existing fastener detection methods have the problem that the detection accuracy and detection speed of the model cannot be well balanced. In this paper, we present a track fastener detection method, which is based on the YOLOv4-Tiny deep convolution neural network. Specifically, data augmentation technology is applied to resolve imbalanced samples, the swish activation function is applied to track fastener detection, and the optimized detection model is deployed in Jetson Xavier NX embedded platform. The experimental results show that the proposed method can effectively improve the accuracy and speed of fastener detection. It paves the way for the real-time track inspection tools to reduce track inspection cost and improve track safety.