High-speed railway track components inspection framework based on YOLOv8 with high-performance model deployment

Abstract

Railway inspection poses significant challenges due to the extensive use of various components in vast railway networks, especially in the case of high-speed railways. These networks demand high maintenance but offer only limited inspection windows. In response, this study focuses on developing a high-performance rail inspection system tailored for high-speed railways and railroads with constrained inspection timeframes. This system leverages the latest artificial intelligence advancements, incorporating YOLOv8 for detection. Our research introduces an efficient model inference pipeline based on a producer-consumer model, effectively utilizing parallel processing and concurrent computing to enhance performance. The deployment of this pipeline, implemented using C++, TensorRT, float16 quantization, and oneTBB, represents a significant departure from traditional sequential processing methods. The results are remarkable, showcasing a substantial increase in processing speed: from 38.93 Frames Per Second (FPS) to 281.06 FPS on a desktop system equipped with an Nvidia RTX A6000 GPU and from 19.50 FPS to 200.26 FPS on the Nvidia Jetson AGX Orin edge computing platform. This proposed framework has the potential to meet the real-time inspection requirements of high-speed railways.