Lane Detection Using U-Net With Spatiotemporal Feature Extraction for Improved Road Segmentation

This paper aims to develop and evaluate a robust lane detection system for Advanced Driver Assistance Systems (ADAS) using deep learning techniques. While advanced deep learning models offer excellent detection accuracy, their overwhelming requests on computational resources, energy, and memory often ruin their deployment on embedded platforms. In our work, a unique approach was utilized to generate a comprehensive dataset that not only captures ideal conditions but also incorporates challenging scenarios with wrong detections. The network was trained on this diverse dataset, which improved its robustness and resulted in high accuracy for lane marking detection. The outcomes of this study offer a promising path toward more dependable and effective lane detection systems that can be integrated into real-time automotive and innovative infrastructure applications, ultimately contributing to improved road safety in autonomous driving technologies.