Robust and accurate partial surface registration based on variational implicit surfaces for automatic 3D model building

Abstract

Three-dimensional models are often assembled from several partial reconstructions from unknown viewpoints. In order to provide a fully automatic, robust and accurate method for aligning and integrating partial reconstructions without any prior knowledge of the relative viewpoints of the sensor or the geometry of the imaging process, we propose a 4-step registration and integration algorithm based on a common Variational Implicit Surface (VIS) representation of the partial surface reconstructions. First, a global crude registration without a priori knowledge is performed followed by a pose refinement of partial reconstruction pairs. Pair-wise registrations are converted into a multi-view registration, before a final integration of the reconstructions into one entity or model occurs. Furthermore, making use of the smoothing properties of the VIS representations, the algorithm proves to be robust against noise in the reconstruction data. Experimental results on real-live, as well as noiseless and noisy simulated data are presented to show the feasibility, the accuracy and robustness of our registration scheme.