关键词: Kinect, 三维点云, 点云特征, 点云配准

Abstract: Aiming at the problem of point cloud registration in 3D reconstruction, this paper presents an automatic registration method based on the feature of point cloud. Firstly, it utilizes Microsoft Kinect sensor to capture depth images in several different views and the interest regions are extracted and converted to 3D point cloud. Secondly, point clouds are filtered and the fast point feature histograms are estimated, then the bidirectional fast approximate nearest neighbor algorithm and random sample consensus are employed to search the final corresponding points. Finally, after computing the initial transformation matric applying singular value decomposition, the iterative closest point algorithm is used to get refined result on the base of initial registration. Experiments show that this registration method can not only ensure the quality of point cloud registration, but reduce the computation complexity, and achieve higher maneuverability and better robustness.

Key words: Kinect, 3D point cloud, feature of point cloud, point cloud registration