| With the in-depth investigation and development of three-dimensional(3D)scanning technology,the 3D reconstruction technique has received wide attention.However,traditional 3D modeling device is expensive and complicate,so,in this thesis,a low-cost and easy-operated 3D reconstruction method with high precision is proposed based on the Kinect sensor.The main works conducted during this research is as follows.1.The introduction and calibration on Kinect sensor.The principle of the Kinect sensor is mainly introduced,together with its related tools.Then,based on the analysis on the imaging model of camera,the relationship of different coordinate systems during the procedure of 3D reconstruction is depicted.Finally,the internal and external parameters of the camera are obtained by calibrating the Kinect sensor,and the relationship between the deep camera and color camera is established which can provide the related parameters for the following acquisition of the Kinect point cloud.2.Preprocessing on the Kinect data.First,the depth image and color image captured by Kinect sensor are aligned using the calibrated parameters in order to relate the two images.Then,the depth image is filtered using the joint bilateral filtering algorithm by experimental comparison.Finally,the depth image data is combined with color image to generate color point cloud data,the point cloud data is further filtered to obtain high quality point cloud data.3.Registration on Kinect point cloud.The procedure of registration on point cloud is divided into three steps to be investigated.The first step is the rough registration in which the SIFT algorithm is used to roughly register the Kinect point cloud.The second step is the precise registration in which the ICP algorithm is used to achieve precise registration.In the meanwhile,to overcome the shortcoming of ICP algorithm,the down sampling on point cloud data is used,and,the distance between point and plane is utilized to replace the Euclidean distance to act as the searching strategy of nearest point.The registration process is also speed up by the K-D tree to speed up the search of nearest point.The third step is the global registration process on point cloud.In order to resolve the problem of error accumulation in the global registration on point cloud,the global registration on point cloud is optimized based on the closed loop constrain to correct the accumulated error.4.The 3D reconstruction on artificial 3D scene.The experiments of 3D modeling on both the single object and indoor scene using Kinect are conducted.The 3D model with good visualization can be obtained.Besides,when the large scaled reconstruction on the indoor scene is performed,a point cloud registration method based on the upper and lower layers is proposed and the experimental results indicate that the proposal can reconstruct the indoor 3D scene effectively. |
