Research Of 3D Reconstruction Of Indoor Scene Based On Kinect

With the development of computer vision technology, 3D reconstruction technology has become one of the hotspots. Traditionally, 3D reconstruction technique is achieved by laser scanning device or a plurality of color camera, but its disadvantage are expensive, complex operation, the amount of data and the poor of real-time, so the development of 3D reconstruction has been restricted and a new device needs to be urgent appear, which can solve some traditional problems of 3D reconstruction technology simply and easily. In order to solve the above problem, this paper collects the original data of indoor scene based on Kinect and presents a 3D reconstruction of indoor scene based on relative coordinate ICP(Iterative Closest Point). Our works are as follows*:1) In order to obtain the raw data of the scene, the paper sets up a platform of data collection system based on Kinect sensor in the software of VS2010 + PCL(Point Cloud Library) + OpenNI. The platform uses a construction method of PCL source code and CMake to compile the PCL source code and configure, produce libraries under the PC and Kinect sensor hardware. Then it uses OpenNI interface to read the original depth images and color images. Finally, in order to process the data for the late comfortably. We fuse the two information to the 3D point cloud data for registering.2) For the problem that the center coordinate of point cloud data which is under different angle of view for multi-frames is not uniform., this paper presents an registration algorithm of point cloud data based on relative coordinate ICP. The algorithm calculates transformation matrix of adjacent frames combine SIFT(Scale Invariant Feature Transform) and RANSAC as the initial value of ICP iteration; then it uses ICP algorithm to establish the center of initial point cloud data based on relative coordinates relations for matching between frames; finally registration of point cloud data based on top and down layers and obtain visual 3D reconstruction model of indoor scene. Experimental results show that the proposed algorithm has better accuracy and faster speed for reconstruction of indoor scene with good results. In consideration of the large amount of per frame data, and give a method of interlaced decimation. Experimental results show that the proposed method can reduce the number of iterations and the registration time in the entire process of registration.