Dense Map Construction by Stereo Camera with Removal of Dynamic Points

Dense point cloud maps are the basis for constructing other forms of maps, and the quality of the point cloud is also relevant to the safety of tasks such as navigation and path planning. Dynamic objects in the scene often greatly reduce the spatial representation capability of the map. This paper proposes a complete dense point cloud mapping system, based on a stereo camera, with the capability of cleanup point clouds from dynamic objects. In particular, the framework introduces a learning-based stereo-matching method with noise suppression based on edge detection of the depth map to obtain a cleaner point cloud. Further, a dynamic point cloud removal method of improved Octomap based on ground segmentation and inert update strategy is proposed to eliminate the ghosting, and improve the map update efficiency. The experimental results demonstrate the remarkable effect of the dynamic removal method on the SemanticKITTI dataset. More importantly, the practical effectiveness of the proposed framework to build clean dense point cloud map on stereo sequences is verified in the KITTI Odometry dataset.