Vision Based Stationary Railway Track Monitoring System

This paper presents a system for monitoring rail tracks in mountainous areas by using a stationary camera system to detect obstacles on the railbed. On-board obstacle detection systems on trains are already used for the purpose of obstacle detection, but due to the nature of mountainous areas these aren't suitable for our problem case. Our approach combines a Machine Learning approach using Yolov5 with a traditional Computer Vision approach to be able to detect obstacles that Yolo was not trained for such as avalanches, mudslides, etc. The system was tested on a single example setup which was set up with the help of the Austrian railway track company. Our experimental test runs already showed a promising low number of incorrect reports. The system, however, yet does not cope with specific conditions, such as extreme lighting or bad weather.