This paper presents a system for automated 3D reconstruction of unorganized RGB-D images. Our algorithm is based on image feature matching and graph theory. We use a multiple-view registration scheme based on graph connectivity in order to reduce propagation errors found in motion. We estimate motion between two-views with 3D points back-projected from visual features. Furthermore, we apply a noise model in the pose estimation routine to improve the results. The noise model for a three-dimensional measured point is calculated with variance forward propagation. Using simulated data we prove that the pose estimation in the registration of two point clouds is the Maximum Likelihood Estimator. Equally, we select the minimum number of points for the matching strategy from statistical tests with simulated and real data. Finally, we validate the entire system with real data.