Research On Three-dimensional Reconstruction Based On Kinect Depth Map

With the advances in computer vision theory and the development of hardwaretechnology, the three-dimensional reconstruction technology has been appliedincreasingly in people’s lives and production. The three-dimensional reconstructionbased on Kinect, because of its rapid and convenient features, has become a hotspot thatresearchers have concerned about. In this paper, three-dimensional reconstructionresearch of the scene and rotating object based on Kinect depth camera, is studied. Themain contents are as follows:First, the three-dimensional scene reconstruction method is studied in this paper.The registration is the key step of3D reconstruction. In the target error function of theprevious ICP (Iterative Closest Point) registration algorithm, each corresponding pointpair was assigned the same weight. However, the characteristics of the correspondingpoint pairs are different, so the roles of point pairs are different in the registrationprocess. The better corresponding point pair should be preferably given a higher weight.The feature points reflect the depth image features and their reliability is higher than thecommon points, so they should be given a higher weight. The ORB (Oriented FAST andRotated BRIEF) texture method is famous for high efficiency. In this paper, the ORBfeature points are extracted and assigned a higher weight than common points. Based onthe above weighted strategy, an improved registration algorithm is formed to improvethe accuracy of registration process.Secondly, in the three-dimensional reconstruction of the rotating object, thebackground of depth image needs to be removed. After extraction the image of rotatingobject image, the subsequent processing can be carried on. In this paper, the framedifference method is firstly used to find the difference of two adjacent frames topreliminarily extract the information of rotating object. Then according to the statisticsof the distribution of the depth values, the depth threshold is determined for foregroundobjects and background image segmentation, and the extraction of rotating object iscomplete. In the conventional registration algorithm, the same target error function wasused for all corresponding point pairs. In this paper, the points in the point cloud arefirstly classified. After the corresponding point pairs were found, the different targeterror functions are applied depending on the type of the point, to improve theregistration accuracy. The experimental results show that when this method was used for scene androtating objects three-dimensional reconstruction, the registration accuracy hasimproved, and there is no significant decrease in time performance. In addition, thismethod also has good effect for removing background in rotating objectthree-dimensional reconstruction.