An Operational Framework for Reconstructing Lane-Level Road Maps Using Open Access Data

Abstract

Lane-level road maps are crucial for urban traffic management, autonomous driving, and vehicle navigations. Optical remote sensing image suffers from trees and buildings occlusion for lane-level road mapping due to the top-down view. While street view images (SVIs) have been used for road detection, however, most of the previous articles focused on extracting road in image space. The reconstruction of lane-level road maps with measurability in geographic space remains challenging. Hence, this article proposed an operational framework for extracting and reconstructing lane-level road maps from urban open access data. First, a sample strategy was used to collect SVIs based on OpenStreetMap (OSM) road central lines. Then, a deep-learning-based method was adopted to identify lanes accurately, and road width was extracted based on design knowledge and OSM information. Finally, the lane-level road map was reconstructed by integrating the lane and its width information. The proposed framework achieves the transformation from image space to geographic space. The case study shows that 82.43% of the roadway is accurately reconstructed in lane-level. The difference between the reconstructed width of the roadway and the reference true value is within the m-level and the RMSE is 0.32 m. The proposed method is cost-effective and accurate-acceptable for acquiring lane-level road datasets in cities.

Keywords

• Deep learning (DL)
• open access data (OAD)
• street view image (SVI)
• urban road
• width measurement