RailVID: A Dataset for Rail Environment Semantic

At present, rail transit is becoming the main means of urban and intercity fast passenger and freight transportation. Its safe operation is of great significance to protect the life of people, security of property and maintain social stability. Though the degree of intelligent traffic has been improved , there are still many safety risks in the current system by using manual hazard monitoring in the railway. Limited by the particularity and complexity of railway scenes, there are few studies on the perception and understanding of rail transit environment. In this paper, we propose a new rail transit dataset – RailVID. We use the AT615X Infrared thermography from InfiRay to collect data and record different railway scenarios, including carport, depot, and straight. We then propose an improved BiSeNet realtime semantic segmentation network for evaluation. Based on this dataset, we carry out environment perception, environment understanding, and safety decisions on the track area in front of the train, and we propose a solution for fully automatic train operation of rail transit.The dataset we provide compensates for the infrared data that is not in the existing dataset, and our data covers special weather and various conditions. Experiments show that our method achieves a higher Mean Pixel Accuracy in the collected dataset, and the processing speed also meets the real-time requirement.