Railroad semantic segmentation on high-resolution images

Recent advances in machine learning research could significantly alter the railroad industry by deploying fully autonomous trains. To achieve effective interaction between self-driving trains and the environment, an accurate long-range railway detection should be provided. In this paper, we propose a framework for the rail tracks segmentation on high-resolution images ($2168\times 4096$). The announced approach accelerates inference speed 6 times, by using two neural networks. The proposed architecture and its training approach provide a long-range railway segmentation within 150 meters, achieving 20 fps. Also, we propose an auxiliary algorithm detecting possible paths among all the found ones. To determine which data labeling approach has a higher impact, additional experiments were performed. The proposed framework provides a balanced tradeoff between computing efficiency and performance in the railroad segmentation problem.