A method of 3D model generation of indoor environment with Manhattan world assumption using 3D camera

In this paper, we propose a face set model generation method from 3D point clouds obtained from 3D camera for high-speed and light-weight storing and showing environment information in tele-operation task for robots. In the proposed method, following procedures run in parallel; 3 dominant orthogonal axis estimation and point cloud grouping by normal vectors based on the Manhattan - world assumption, fast registration using dominant axis grouped point cloud, plane position estimation for each dominant axis group, and face set generation by shape estimation for each plane. Experimental results shows that accuracy of plane position estimation is equivalent to measurement accuracy, registration takes about 0.1[s] for each frame, and storage size is reduced to about 10 - 20% from original 3D point cloud size. We also show some generated environment models as experimental results.