Sistema de inspeção de talas de junção ferroviárias via computação de borda

Abstract

Railway accidents can cause significant financial and environmental damage, in addition to putting people’s lives at risk. A crucial component for railway safety is the rail joint, which are steel pieces adjusted and fixed in pairs at the rail tracks to ensure the continuity of the track’s rolling surface. However, their inspection is currently carried out manually, requiring considerable human and financial resources. This project involves the development of a system for the identification and analysis of rail joints using computer vision and deep learning techniques. The system consists of a camera mounted at the front of the locomotive, a GPS system, and a processing unit, all embedded on the locomotive, where all captured data is processed. Additionally, various optimization techniques for deep learning models were employed to enable the implementation of the system on low-energy-cost devices. Ultimately, the system was validated with real data from a railway network, achieving an accuracy above 90% in the geographical localization of the rail joints and an accuracy above 90% for the analysis of errors in the tracked rail joints.