Instance Segmentation for Accurate Lane Detection and Fitting with Hour Glass Network

Abstract

A novel approach is proposed to improve lane marking identification in autonomous driving systems by combining deep learning-based segmentation with traditional lane detection methods. This approach aims to address challenges faced by each technique individually, such as CNNs struggling with precise localization and traditional methods facing scalability issues. By integrating segmentation with handcrafted features and specialized fitting, the proposed method enhances network convergence speed and location accuracy. A unique lane fitting method based on convergent line prediction is introduced, particularly beneficial for challenging highway conditions. Experimental evaluations on four datasets demonstrate the effectiveness of this approach, showcasing notable improvements in robustness and accuracy in lane marking detection.