Scan registration can estimate the pose of the vehicle based on information acquired by range sensors. Those techniques could obtain optimal results when applying in indoor environments. Nevertheless, their performance decrease in unstructured environments because of the vast range of operating conditions. This work provides a computational approach to improve the results of the well-know iterative closes point (ICP) approach and its variants in an urban scenario. The proposed method describes a pre-processing approach where the point cloud information was divided into several groups. Then, the rigid matrix associated with vehicle motion was obtained by minimizing the sum squared registration error among the most significant groups. This methodology was validated using the Ford and Kitti datasets. The results showed that the proposal performed better in the long-term for the point-to-point version in comparison with the original implementation. Meanwhile, when applying the proposal with the point-to-plane version, similar results to the original implementation were obtained. Nevertheless, the consistency analysis of the Z-axis showed a better performance for the cluster-based proposal in all the point-to-plane implementations. These outcomes suggests that the proposed approach could improve the performance of localization techniques in urban scenarios based on separable groups of data.