A new registration method to robustly align a series of sparse 3D data

Optical measurement tasks often require an acquisition of several partial 3D views to collect complete 3D information of an object surface. We consider a sensor that acquires the surface information by taking a series of sparse partial 3D views while being freely moved around the object. In order to obtain a dense 3D model of the surface an alignment of all partial views is required. Existing methods usually detect common surface features and map them onto each other. However, in case of sparse data these methods fail, because insufficient or no neighborhood surface information of points is available to find common features. We propose a method that is specifically tailored for the robust registration of sparse 3D data. In a first step, consecutive images are aligned successively. The approach focuses on detecting corresponding points instead of surface features. The successive registration leads to an accumulation of the error along the series of views. Therefore, in a second step, the global registration error is minimized. The combination of these two steps provides a robust registration method. In this paper we will present the new registration method and give some results.