Edge-Computing Oriented Real-Time Missing Track Components Detection

Abstract

Track integrity is critical for railroad safety. Traditional track inspections are either labor-intensive or require centralized data processing, making them susceptible to human error and lapses between data collection and situation awareness. The advent of deep learning and computer vision provides promising potential for automated track inspections. However, few existing systems are edge-computing oriented or provide inspection results in real time. In this study, a novel ultra-portable system for real-time detection of track components, such as spikes, bolts, and clips, is developed by integrating the cutting-edge YOLOv8 object detection model with a tailored template matching algorithm. In this system, YOLOv8 serves to recognize track components, while the template matching algorithm discerns missing components based on predefined patterns. Field blind testing results verified the exceptional performance of the model in detecting track components and a remarkable speed of 98.12 frames per second. Leveraging these detection results, the proposed template matching technique displayed an impressive recall rate of 90% and an accuracy rate of 90.77% in identifying missing components. The proposed system provides an affordable and versatile solution for track inspection, aiming to improve railway safety.