Leveraging 2D Semantic Information for Dynamic Object Removal in Static 3D Point Cloud Map Construction

Construction of static 3D point cloud map is essential for the localization and navigation of mobile robots. However, dynamic objects observed during the mapping process leave long traces on the map, impeding the performance of subsequent tasks. To tackle this problem, this paper proposes a framework that leverages 2D semantic images to remove dynamic objects from 3D point cloud map and generate a static map. While implementing the framework, viewpoint differences between the Camera and LiDAR sensors can cause occlusion problem. We address this problem by utilizing an occlusion pixel mask. Through experiments, we validated that our proposed method effectively removes traces caused by dynamic objects and constructs a 3D map that accurately represents the surrounding environment.