Enhanced Safety in Multi-Lane Automated Driving Through Semantic Features

Abstract

Accurate lane detection is crucial for the safety and reliability of multi-lane automated driving, where the complexity of traffic scenarios is significantly heightened. Leveraging the semantic segmentation capabilities of deep learning, we develop a modified U-Net architecture tailored for the precise identification of lane lines. Our model is trained and validated on a robust dataset from Kaggle, comprising 2975 annotated training images and 500 test images with masks. Empirical results demonstrate the model's proficiency, achieving a peak accuracy of 95.19% and a Dice score of 0.928, indicating exceptional precision in segmenting lanes. These results represent a notable contribution to the enhancement of safety in automated driving systems.