Real-time railroad track components inspection based on the improved YOLOv4 framework

Abstract

According to the Federal Railroad Administration (FRA) database, track component failure is one of the major factors causing train accidents. To improve railroad safety and reduce accident occurrence, tracks need to be regularly inspected. Many computer-aided track inspection methods have been introduced over the past decades, however, inspecting missing or broken track components still heavily relies on manual inspections. To address those issues, this study proposes a real-time and cost-effective computer vision-based framework to inspect track components quickly and efficiently. The cutting-edge convolutional neural network, YOLOv4 is improved trained, and evaluated based on the images in a public track components image database. Compared with other one-stage object detection models, the customized YOLOv4-hybrid model can achieve 94.4 mean average precision (mAP) and 78.7 frames per second (FPS), which outperforms other models in terms of both accuracy and processing speed. It paves the way for developing portable and high-speed track inspection tools to reduce track inspection cost and improve track safety.