In this paper, we address the problem of pair-wise registration of large unstructured 3D/color datasets. Our purpose is to improve the classical ICP (Iterative Closest Point) algorithm by using color information, in order to deal with large datasets and with objects for which the geometric information is not significant enough. After a brief presentation of classical ICP (Iterative Closest Point) algorithm and of the research works developed to improve its performance, we propose a new strategy to improve the selection of points. Color information is used to reduce the search space during the matching step. Experimental results obtained with real range images show that the algorithm provides an accurate estimation of the rigid transformation.