None Realtime 3D Reconstruction Based On Point Cloud Generated By Kinect

The developing of visualization, visual reality, and multi-media, and the rising of 3D printing are widely used in traditional manufacture, medical industry and digital entertainment that requires generation of 3D models rapidly and efficiently. In these situation, the traditional method that is low efficient, although it has the feature of less polygon, optional vertex and clear structure can not satisfy the growing requirement. The existed 3D reconstruction methods those are researched sufficiently including reconstruction based on signal image and based on multi-images are aimed at calculate the depth through the image itself and generated scene and graphic. They do not provide 3D ploygon grid that widely used. After that, the point cloud technology’s rising gives us a new efficient path to generation 3D models. Until Microsoft pushing out the Kinect, a consumer grade somesthetic equipment that be used in point cloud generation, we had to use the expensive laser scanner to get the point cloud data.Our article is aimed at probing into a series of methods those are precise for application, cheap for affording and simple for operation.Through experiment, we propose the main steps for 3D model reconstruction.1. Generate point cloud data. Use convert algorithm to find the corresponding relationship between depth image and color image those are gotten via Kinect and depth image and space point. Finally, we get cloud for usage.2. Segment object from original point cloud. Segment object point cloud from whole scene point cloud via RANSAC (random sample consensus).3. Rectify object point cloud. In experiment, we find a method to transform point cloud from camera coordinate system that is used in Kinect to world coordinate and align local coordinate to world coordinate.4. Compensate the deficiency face. Use the rotation angle getting by measure the rotation angle of cloud deck to rotate the original object point for the purpose that compensate the deficiency.5. Smooth the point cloud. Smooth the object point cloud via MLS (Method of Least Square) and estimate the normal.6. Down sampling. Spare the object point cloud through voxel grid.8. Reconstruction the 3D ploygon grid. In the experiment, three kind of reconstruction algorithm are compared and defined there scope of application:1)Fast greedy tranglize fits into high density point cloud and object holding many planes and without less various changs in edge.2)B-Spline method fits into unregulated and continuous normal and curvature in point cloud.3)Poisson method have better performance in average condition both in plane and curve, but must control the scale of point cloud efficiently.