Transformed polynomials for global registration of point clouds

In this paper, we introduce a novel approach for global registration of partially overlapping point clouds. The approach identifies feature points of matching objects based on surface-approximating polynomials and finds an initial transformation depending on these polynomials. We compute an extended set of rotationally-invariant features for polynomials. In contrast to purely feature-based approaches, we do not only compute transformations based on the invariant properties of polynomials, but actually transform the polynomials into a common coordinate system and compare the transformed coefficients. This results in an improved correspondence analysis of local surfaces. Hence, using transformed polynomials, we gain more discriminating information about different structures. Therefore, the approach can handle partial scans of different objects simultaneously. Each partial scan is assigned to one of the objects and registered accordingly. Moreover, the approach is robust against noise and can process real data.