Natural Interaction 3d Point Cloud Registration For Human Head

3D technology keeps growing up and matures in recent years, such as 3D movie, 3D TV, 3D virtual fitting, virtual web 3D shopping, 3D virtual meeting, natural interaction games. They are accepted by more and more consumers now. The 3D registration technology behind them is one of key components of those applications. Laser or infrared ray scanner scans the object from different view and gains 3D data on different sides, then several 3D point datasets (point cloud) are created. Since every point cloud is in different coordinate systems, how to join these point clouds together in different coordinate systems automatically (registration) and become a whole 3D object model, it is a key and difficult problem in computer vision. Some products for it are always expensive and complex. They are only suited for enterprises and not suited for ordinary consumers.In order to resolve this problem I consider three aspects in this paper, including simpler operation, better registration effect and lower cost. I focus on 3D point cloud registration for human head based on natural interaction device. My main contribution in this paper is in below:1) I analyze and compare most arithmetic in point cloud registration and choose the most appropriate one. those includes Iterative Closest Point (ICP)[1] arithmetic, Fast Point Feature Histograms (FPFH)[2] arithmetic, RANdom SAmple Consensus(RANSAC)[3] arithmetic and Singular value decomposition(SVD)[4][5] arithmetic.2) Because FPFH arithmetic's logic is complex and its feature variables are not intuitive, I propose Triangle Normal Feature Histograms (TNFH) arithmetic. It focuses on the triangle plane's normal variation feature around one point. Its feature only has one variable and its logic is simpler. Test shows TNFH arithmetic's registration success rate approximate with FPFH arithmetic.3) Multi-layer system architecture design. Microsoft Kinect controller [6] is chosen for 3D data collection in the low level. The driver is open source OpenNI framework [7]. I propose a process flow including collection, feature estimation, feature filter, transform estimation and align point cloud. I balance the system performance and registration result and implement a 3D point cloud system prototype base on natural interaction device.4) In system view I implement the 3D point cloud system prototype base natural interaction device based on chosen arithmetic. The key and difficult parts are described and test results are provided in paper.This system acts as a basic module for 3D natural interaction applications, it provides the whole 3D body model information of ordinary user, and so more business applications can provide more valuable experience for their consumers.