A voxelize structured refinement method for registration of point clouds from Kinect sensors

Abstract

3D scanning of objects has been widely used for many years in computer graphics and computer vision. There are a variety of solutions in this area, such as the motion or multiple sensors for scanning. In this study, we propose an approach that generates a scan with a natural motion of the user, through a fixed Kinect sensor whose usage is more practical and cost-effective than conventional 3D scanners. Local voxelized structure based on (LVS) is proposed for efficient 3D point cloud, captured by Kinect as low-quality. The approach allows the generation of full point cloud data in a wide range of indoor and short-range 3D objects. The developed system for object scanning is easy to set up, generating simple and impressive results. The 3D object standing on the turntable facing a single fixed Kinect sensor is rotated at specific angles (e.g. 90°) to obtain multiple point cloud scan data. Afterward, the center of gravity of each scanned point cloud data is shifted into (0,0,0) origin position for merging and aligning operations. So subsequent scans are obtained. The point cloud data obtained from the second and subsequent scans are transformed in the y-axis direction with respect to the center point (0,0,0), respectively. In some case, the axis-x and axis-z can be used for rotating too. The transformed point cloud data obtained from the different angles are aligned with respect to each other, shifted according to the determined merging key points. An algorithm that runs on the sections of point cloud for refinement operation is performed on a complete 3D point cloud data. Thus, the resulting scan has a 3D, clean and orderly structure free from the data crowd. Our approach has verified over a large number of users and different 3D objects and compared with a reference scan according to metric specifications.