Semantic Line Map: A Novel Visual Map Representation Scheme for Autonomous Driving Systems

This is the authors' abstract. We don't add key points for this paper.

High-precision maps are a crucial require-ment for autonomous driving systems, yet traditional mapping approaches often encounter issues related to storage, efficiency, and positioning speed. To address these issues, a new map representation scheme called “Semantic Line Map (SEML)” has been proposed, which combines deep learning techniques with an efficient storage method based on line segments, using road linear landmarks as carriers to establish a semantic line segment map of the road environment. To achieve visual line segment semantic segmentation mapping, we utilize camera data and employ the COLMAP algorithm for 3D scene reconstruction based on motion inference structure. We also use the PSPNet semantic segmentation network to obtain end-to-end semantic segmentation of 2D images. The resulting segmented images undergo line segment extraction and computation of 3D line segment positions, enabling us to obtain the final 3D semantic line graph. Our experimental results demonstrate that our algorithm is efficient and capable of achieving high-fidelity mapping solutions.