Runway Lane Marking Semantic Segmentation Using Neural Networks

Considers the use of a convolutional neural network to solve the problem of semantic segmentation of horizontal runway lane markings based on camera data. This task is of great practical importance for improving the safety and autonomy of aircraft during landing and taxiing. To train and validate the models, a specialized data set was formed including a variety of runway images obtained from open sources and marked up manually. The article compares two architectures of convolutional neural networks – U-Net and YOLO – in terms of segmentation accuracy, inference speed, and resistance to changes in lighting conditions, weather factors, and visual noise. The applicability of each architecture for real-time tasks including on devices with limited computing resources is evaluated.