Conditional Adversarial Networks Based Road Marking Identification and Vectorization for High Definition Map

Road markings are the primary road feature of High-Definition (HD) maps in autonomous vehicles and play a critical role in traffic safety. Methods are mostly sensitive to the intensity value of data captured by laser and the type of road scene. To solve this problem, this paper proposes a method of using the image-to-image translation based on conditional adversarial networks to achieve the extraction, recognition, and identification of road marking based on laser data. This method contains three steps: (1) the generation of 3D intensity images based on the extracted ground surface, (2) an automated road marking extraction based on conditional adversarial networks of image-to-image method, (3) the identification and vectorization based on modified Normalized Cross Correlation (NCC) template matching algorithm. Quantitative and qualitative analysis based on experimental data for different road scenarios are used to verify the robustness of the method and the accuracy of the extraction results. The experimental result based on different road scenes is promising and valuable for the update of road feature database. The proposed method makes the extraction, recognition, and identification of road markings more robust and accurate while also delivering a valuable solution for the HD map used by intelligent and connected vehicles.