U-Net-Based Detection of Road and Lane Markings from High-Resolution Images

Abstract

With technological developments in the field of hardware, many autonomous systems are used in daily life. Autonomous vehicles designed for safe travel in the transportation sector perform dynamic environmental control with the help of sensors and cameras. These vehicles need to process the image data they receive from their cameras and transform them into meaningful information. Artificial intelligence-based approaches are very effective in transforming data into meaningful information. In this study, a U-Net-based system is proposed that can automatically detect and classify areas of road and lane markings from high-resolution images. A publicly available dataset was customized for the model's training, validation, and testing phases. The pre-processing phase designed to include high-resolution images in the training of the U-Net model is explained. Dataset samples are split into 70% training, 20% validation, and 10% testing. The training phase performed using the early stopping function is defined for a maximum of 100 epochs. The numerical data of the training and validation phases, which were carried out in accordance with the multi-class semantic segmentation method, were shared. As a result of the test phase of the proposed model, the lowest 37.14%, the highest 93.65%, and an average of 79.48% Intersection over Union (IoU) have been achieved. With this model, the classification and detection of road and lane markings areas can help the dynamic environment control of autonomous vehicles.