[Japanese | Thesis | Researches in Minoh Lab | Minoh Lab]
We propose a method for reconstructing human body surface from point clouds.
Recently, dence digital surface data of human bodies are required for computer-aided design of clothes, furniture and seats of vehicles and so on. To obtain 3D geometric data, modeling tools such as CAD software are usually used. However, to obtain geometric data for complicated shapes like human bodies, it is more efficient to use data measured by 3D devices such as laser range finders. The data obtained by such devices are, in principle, point clouds. Therefore, there arises the problem how to reconstruct the surfaces of original human bodies from the point clouds.
A typical way of reconstructing surfaces from point clouds is to generate a polygonal meshes by Delaunay triangulation. Although to compute Delaunay triangulation on 2D plane Delaunay is very reliable method, in 3D space it is pretty fragile because the method cannot give accurate surface for concave object.
In our proposal, we take the structure of human bodies into account. Roughly, a human body consists of six parts; a head, a trunk, two arms and two legs. Each part has a cylinder-like shape. Hence, most cloud points of each part of human bodies locate radially around its principal axis.
We assume that the global structure of a human body is given. That is to say, following information is given: the sections that divide the point clouds according to the six parts and the rough directions of parts.
We project the point clouds of each part onto an appropriate cylindrical surface that are obtained by the given information. The coordinate system of the projected points represent the location along the axis and the angle around the axis. On this 2D plane, we compute Delaunay triangulation.
However, two problems force us to revise the strategy above. The first one, even when a human stands upright, some of the principal axes of parts are not always straight because some parts contain joints, for examples, elbows, wrists and so on. The second problem is that human bodies contain portions that are not cylindrical: the ends of the parts (e. g., the tops of heads) and the joints of the parts (e. g., shoulders and crotches).
For the first problem, we divide each part into several sub-parts that are cylindrical. For the second problem, instead of cylindrical surfaces we use hemisphere surfaces for such portions. On each projected plane, we perform Delaunay triangulation and unify the all results. For the unification, we use following method; (i) we locate adjacent cylinders and hemispheres so that they overlap each other; (ii) we compute the Delaunay triangulation for all cylinders and hemispheres; and (iii) we remove surplus polygonal meshes. By using the unified results we can reconstruct the whole surface.
In our experiment, we used point clouds of a human body and reconstructed surfaces of all parts. Our method generated polygonal mesh of human body surface containing no irregular polygons. We also obtained an experimental result that our method gives better results comparing with the method based on cylindrical projection using a single axis for each part.