Recognizing Railway Infrastructure Using CNN and Stereoscopic Vision

In the era of Industry 4.0, achieving Grade of Automation 4 (GoA 4) within the railway sphere is pivotal for the development of autonomous networks. This necessitates intelligent systems capable of real-time perception, precise object identification, and accurate determination of distances. This paper investigates railway object detection using Convolutional Neural Network (CNN), specifically the YOLOv8 architecture, extending prior research by integrating CNN with stereoscopic vision for distance determination to the detected objects. Implementing GoA 4 enables autonomous systems to interpret and react within railway environments. Our evaluation of YOLOv8 demonstrates consistent and robust object detection, achieving an mAP of 0.8 at IoU 0.5 and maintaining 0.54 across IoU 0.5-0.95. Utilizing stereoscopic vision, the determination of distances to detected objects within a 250-meter range exhibits an error margin of less than 10%, ensuring high precision in distance estimation. The integration of YOLOv8 with stereoscopic vision for accurate distance estimation, executed within 80 milliseconds, demonstrates remarkable computational efficiency, which is crucial for real-time applications. This efficient process, facilitated by the Nvidia RTX A5000 graphics accelerator, ensures both precision and high speed in dynamic railway settings. Our assessment emphasizes the adaptability and reliability of YOLOv8 in detecting railway infrastructure objects across diverse conditions, signifying its potential for real-world deployment.