Constructing models of articulating objects: range data partitioning

Abstract

In this paper we consider one aspect of the problem of automatically building shape models of articulating objects from example range images. Central to the model construction problem is the registration of range data, taken from different vantage points, into a common coordinate frame. This involves determining a transformation for each set of range data which aligns overlapping surface points in the common frame. Current registration algorithms have been developed specifically for rigid objects, but it is not obvious how these can be extended to articulated or more generally deformable objects. Here, we propose that range images of articulated objects are first segmented into their rigid subcomponents. Each subcomponent can then be registered in isolation using the existing algorithms designed specifically for rigid parts and the final model formed by reassembling all of the submodels. This has motivated the development of a rigid part segmentation algorithm which is described and demonstrated here. The algorithm is currently limited to non-umbilic surfaces, but in this more restricted domain is shown to work well.