RPD-SLAM: Real-time Semantic SLAM for Partially Dense Mapping with Road-Sign Understanding in Dynamic Outdoor Environments

SLAM provides autonomous vehicles with precise localization, environmental understanding, and mapping capabilities. However, traditional SLAM methods often struggle to handle dynamic elements and irrelevant geometric features, leading to a decrease in localization and mapping accuracy. This paper introduces a real-time dynamic semantic SLAM system-RPD-SLAM, which can remove dynamic elements in real time to improve localization accuracy. Additionally, we project dense road-sign point clouds onto the sparse point cloud map in real time, enriching the map with road information while maintaining lightweight storage, thereby providing critical navigation and localization support for autonomous vehicles. Evaluations on the KITTI dataset show that RPD-SLAM outperforms ORB-SLAM2 in localization accuracy while maintaining real-time performance. Tests conducted in a university campus environment further validate the system's adaptability and efficiency.