Method for segmenting steel rail lines in video frames using the Rail-UNet model

As the importance of train operation safety becomes increasingly prominent, rail detection technology plays a crucial role in the field of signal processing. This paper is dedicated to accurately identifying railway areas in video frames, in order to output segmentation maps that mark out these areas. The main challenges faced in rail detection include variations in lighting conditions, the inherent diversity of the rails themselves, and the demand for high-speed detection. We have designed a new Rail-UNet model based on an improved U-Net model, with the goal of achieving optimal recognition of rail track lines in captured images even under changing lighting conditions. For the real-time semantic segmentation of complex rail lines such as compound turnouts and cross-over tracks, the model aims to maintain good image processing results. Under high-speed and dynamic image acquisition conditions, the model seeks to enhance the positive segmentation effects of rail lines in video frames in three aspects: real-time collection of rail track lines, improved handling of complex rail structures, and enhanced performance under dynamic conditions. We anticipate that our method will achieve excellent performance in both speed and accuracy, with the lightweight version targeting an accuracy rate of 90% and high-precision rail line segmentation processing at a rate of over 300 frames per second.