Normal segmentation of geometric range data has been a common practice integrated in the building blocks of point cloud registration. Most well-known point to plane and plane to plane state-of-the-art registration techniques make use of normal features to ensure a better alignment. However, the latter is influenced by noise, pattern scanning and difference in densities. Consequently, the resulting normals in both a source point cloud and a target point cloud will not be perfectly adapted, thereby influencing the alignment process, due to weak inter surface correspondences. In this paper, a novel approach is introduced, exploiting normals differently, by clustering points of the same surface into one topological pattern and replacing all the points held by this model by one representative point. These particular points are then used for the association step of registration instead of directly injecting all the points with their extracted nor-mals. In our work, normals are only used to distinguish different local surfaces and are ignored for later stages of point cloud alignment. This approach enables us to overcome two major shortcomings; the problem of correspondences in different point cloud densities, noise inherent in sensors leading to noisy normals. In so doing, improvement on the convergence domain between two reference frames tethered to two dissimilar depth sensors is considerably improved leading to robust localization. Moreover, our approach increases the precision as well as the computation time of the alignment since matching is performed on a reduced set of points. Finally, these claims are backed up by experimental proofs on real data to demonstrate the robustness and the efficiency of the proposed approach.