Year: 2025
Type: article
Abstract: This paper presents an Artificial Intelligence (AI)-based approach to train localization through the detection of railway infrastructure objects using convolutional neural networks. The proposed system identifies key visual landmarks such as traffic lights, level crossings, tunnels, bridges, and passenger platforms directly from live video streams captured by onboard cameras during train operation. This enables accurate and continuous localization without relying on satellite navigation systems or additional trackside infrastructure. The object detection model is based on the You Only Look Once (YOLOv11) architecture. It is trained using high-performance Graphics Processing Unit (GPU) resources and subsequently converted and optimized for deployment on the energy-efficient RK3588 neural processing unit (NPU). The system achieves a mean average precision of [email protected]:0.95 = 0.52 and operates in real time at approximately 35 frames per second, meeting the practical requirements for onboard applications. Compared to traditional Global Navigation Satellite System (GNSS)-based solutions, the proposed method is inherently resilient to signal jamming and spoofing while significantly reducing infrastructure costs. Its low power consumption and high-speed inference make it especially well-suited for integration into modern railway systems operating at higher automation levels. The results confirm the feasibility of this AI-driven approach as a scalable and robust solution for train localization in diverse operational conditions.
Author: Vladimir A. Fedorov
Institution: Ural Federal University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4410522980