Dense Point Cloud Mapping Based on RGB-D Camera in Dynamic Indoor Environment

With the continuous progress of science and technology, the dense point cloud mapping has gradually replaced the traditional laser mapping and sparse point cloud mapping because of the rich environmental information and depth information. To address the ghosting problem of moving objects in dynamic environment, this paper introduce Mask R-CNN to extract moving objects mask in input frame. After the positioning accuracy are met, we proposes a robust mask truncated signed distance function (MTSDF) model to build dense map in dynamic indoor environment, which identifies the dynamic voxels by the truncation distance of the voxel and reset them, then, the dense point mapping is updated. We respectively tested on TUM RGB-D dataset and our lab, the results show that our method not only achieve better localization precision, but remove the ghosting generated by moving object in dense point cloud mapping well.