Automatic registration of unordered point clouds acquired by Kinect sensors using an overlap heuristic

Abstract This paper proposes and evaluates a pipeline to automatically register point clouds captured by depth sensors like the Microsoft Kinect. The method neither makes assumptions about the view order of the sensors, nor uses any kind of other task-dependent prior knowledge. All point clouds within the input set are aligned in a common, global coordinate system by a successive application of pairwise registration steps. The order of the individual transformations is automatically derived from a global point cloud graph, which uses the overlap of two individual point clouds to establish a weighted link between them. The experiments prove the generality of the proposed approach by applying it to data from a single but moving sensor, multiple Kinects that run simultaneously, as well as laser scanning data. The obtained accuracies in terms of the mean nearest point neighbor distance are below 0.01% of the maximum point distance of the reference data in all cases.