3D Modeling Of The Rotationally Symmetric Objects Using Kinect

As one of the hot points in computer vision,three-dimensional modeling is widely used in the fields of virtual reality,virtual dressing and unmanned driving.Due to the rapid development of depth sensor technology,the popularization of 3D object modeling based on depth sensors has become possible.The traditional 3D modeling based on depth sensors mainly uses the ICP(Iterative Closest Point)algorithm to register the point cloud.When modeling the complete symmetric objects such as cylinders and cones,it does not work well due to the lack of structural features.Therefore,in this thesis three-dimensional modeling of symmetric objects is studied with the depth sensor.The detailed work is as follows:For fully symmetric objects,the ICP algorithm based on multi-view depth data collection fails due to similarity of data.In this thesis the symmetry of object is transmitted into the same boundary from all views,and then the 3D point cloud is vertical projected to the planar point cloud.Then the boundary of the 2D point cloud is extracted,and the boundary is rotated to solve the problem of the model which can not be closed during the ICP registration procedure.To deal with the discrete points on the model surface,the MLS(Moving Least Square)algorithm is designed to fit the initial boundary,making the reconstructed model surface smoother and more symmetrical.Experimental results show that this method can quickly model the the fully symmetric objects,and the modeling accuracy can reach above 98%.For the fully symmetrical object with a hole on the side,rotating the boundary directly will ignore the gap feature.In this thesis twice boundary fitting is performed,and the size of hole is converted into the relative angle.This angle is used to calculate the rotation matrix,to adjust the initial position of the two boundaries.Then the two boundaries are rotated by a certain angle and superimposed to complete the model reconstruction.When dealing with the problem of data collection from a single view,in this thesis the transformation matrix between the floor normal vector and the camera coordinate axis is computed,and the whole point cloud is rotated to adjust its direction.Experimental results show that this method can reconstruct modeling of symmetrical objects with hole on the side at any shooting view with high accuracy and fast speed.