ROAD LANE LINE DETECTION USING U-NET ARCHITECTURE FOR SELF DRIVING CARS

Abstract

This project introduces a deep learning-based road lane line detection system using the U-Net architecture, aimed at enhancing autonomous driving and advanced driver assistance systems (ADAS). The model is trained on the CULane dataset, leveraging semantic segmentation to accurately detect lane markings in diverse road conditions. Preprocessing techniques such as resizing, normalization, and data augmentation are applied to improve model robustness. The system employs binary cross-entropy loss and is optimized using the Adam optimizer for efficient learning. A custom data pipeline ensures smooth training and evaluation, with real-time predictions visualized for performance assessment. The trained model effectively identifies lane boundaries, even under challenging lighting and occlusion scenarios. Extensive experiments validate the system’s accuracy and reliability in real-world conditions. The model achieves an accuracy of 96.15%, outperforming many existing lane detections models. The proposed approach contributes to safer and more intelligent vehicle navigation by reducing lane departure risks. Future enhancements include integrating temporal consistency for improved detection stability.