[Japanese | Thesis | Researches in Minoh Lab | Minoh Lab]
We propose a method for generating a complete human shape model from the partial data measured by range sensors.
Recently, it has become easy to measure the shape of human bodies by the range sensors. The human shape models of individuals become necessary in various fields, such as digital analysis of human shape, virtual reality in computer world, and so on.
On the other hand, various methods for generating the geometric shape models of objects by integrating measured data have been proposed. These methods assume that the whole surface data of the object are obtained. In the case of humans, the occlusion makes it impossible to measure some region of human body, for example, chin (the region beneath lower jaw), region under arms, crotch, and so on, and the light reflectivity also makes it difficult to measure hair.
In our proposal, we make use of a standard human model with a priori constraints. A complete shape model of an individual is obtained by deforming the standard model according to the partial measured data. In this report, we deal with the standing up straight human data. The standard model is divided into seven parts (shoulder, chest, two arms, waist, and two legs), and has seven principal axes in each part.
Our method consists of two processes; (i) obtaining the correspondence between the partial measured data and the standard model, and (ii) deforming the standard model according to the measured data. In the process (i), users give the corresponding points on the model and the data, and the location of the data for the model is calculated from these points. In the process (ii), there are two kinds of deformation; similarity-based-deformation and local-copy.
In the similarity-based-deformation, a global proportion (tall or short, fat or thin) of an individual is estimated. For this estimation, similarity ratios of the parts are calculated separately. The standard model is deformed proportionally by these similarity ratios. The similarity ratio is calculated from the correspondence between the standard model and measured data, if the part has corresponding measured data. If it does not, the average of the similarity ratios is used. In this deformation, first, each part is deformed along its axis. Second, the cross section of each part is enlarged (or reduced).
In the local-copy, we deform the closed regions on the standard model so that the regions have the same shape as their corresponding measured data. In order to preserve the connectivity and the smoothness of the standard model, the regions around the closed region is also deformed proportional to the distance from the boundary of the closed region.
Furthermore, we take the advantage of the symmetry of the human. There is the case that a region does not have corresponding measured data, and the symmetric region has the corresponding measured data. In this case, the region is also deformed as well as the symmetric region by the rules above: the similarity-based-deformation and the local-copy.
In our experiment, we used two different human shape models. We extracted some regions from one model (object model) as the partial measured data and used the other model as the standard model. We obtained the shape model of an individual by fitting this standard model onto these measured data.
We evaluated our method by comparing the deformed shape model with the object model. Our experimental results showed the validity of our method. Moreover, if we used more measured data, we obtained better results.