| Three-dimensional reconstruction mainly discusses how to obtain three-dimensionalinformation of objects in space with some given scene images. It helps the real-worldscenes represent in computer and owns broad application prospects in many fields. Thispaper uses color information lists and correlative depth information lists capturing fromKinect to build three-dimensional model of the scene. We will discuss and solve severalproblems, such as scene images acquisition, depth map repairmen, camera calibration,matching algorithms among multi-scenes, et al, so that can obtain3D point cloud model ofthe scene efficiently and fast. This paper is mainly divided into the following parts.Stage of image capturing and depth map filtering: At the beginning, the paper gives abrief description about the physical structure of Kinect and the principle of obtainingdepth information. In consideration of the missing depth information in original depthmaps, several filtering methods are used to handle the problem and analyzing results aregiven afterwards, and finally the paper chooses JBU as the depth map filtering algorithm.End of the paper, we gives the analysis of the filtering methods performance, and proposesan improved JBU method.Stage of camera calibration: We introduce the principles of camera calibration and acalibration process of a depth and color camera pair. Then we use the Kinect CalibrationToolbox to perform Kinect calibration experiment and obtain cameras parameters finally.Stage of3D reconstruction: In the stage, the SURF method is mainly described atfirst. We confirm that SURF method is quicker and more robust in image matchingprocess compared with SIFT method through experiments. Then a point cloud registrationcalled ICP is introduced briefly. Finally the paper succeeds in forming3D reconstructionof multi-scenes by ICP with SURF method. |