3D Reconstruction Of Indoor Scenes Based On Kinect

This article is aimed at3D reconstruction of indoor scenarios based on the Microsoftmotion-sensing camera Kinect with the access to depth and color informationsimultaneously and a continuous output at a rate of30frames per second. These featuresmake3D reconstruction technology more practical applicability.Firstly, RGB-D camera is driven to get the3D scenarios data flow with OpenNI andcalibrated with the checkerboard calibration algorithm based on PCL (Point Cloud Library)Also OpenCV, a computer vision Library, is included during this process. Lastly theresulting indoor scene of3D Point Cloud structure is obtained.Because of the high noise from original depth map, the depth data of point cloudcontains a large number of invalid values, which results in a lot of "holes" in the gray imageand are marked with "nan". In order to improve the quality of3D reconstruction, the depthmap needs to be repaired. While the grey value of3D indoor scene is continuous, thechanges are not so obvious and the edge noise is high, the traditional image filteringmethods are not available for the repairing work. As a result, an modified depth map repairmethod is introduced in this work, which includes the following steps: edge cutting,inter-frame background filtering and dynamic scene repair based on background, small arearepair based on RGB image using Gaussian kernel weighted operator. Lastly,3D spacepoints in the form of k-d tree are searched, filtered by down-sampling and removed outliers.The next procedure is about point clouds registration from different perspectives.Firstly, the transformations of different point cloud coordinate systems are analyzed. Thenthe comparison for traditional point cloud registration algorithms: ICP and SAC-IA is givenand the improved flow is proposed, combining with these two methods and completing theregistration of many frames.The results illustrate that the improved filter and registration algorithms can effectivelyimprove the quality of depth image, the matching accuracy and speed, so they can beeffectively applied to3D indoor scene reconstruction. At the same time, the low price ofKinect makes3D indoor scene reconstruction more widely used in our daily life, which isof great significance to the popularization and application of computer vision.