3D Reconstruction Of Large-Scale Scene Based On Kinect

3D reconstruction is an important branch of Computer Graphics and Computer Vision.It not only plays an indispensable role in scientific researches,it also has a great influence on our daily life.This thesis is going to present how to reconstruct a 3D model of a large-scale scene.All of the research is based on Microsoft Kinect,and in order to achieve the goal,we also need the help of some 3rd party libraries like PCL and OpenCV.As we all know,the quality of the point cloud is closely related to our depth images.Because of the blank hole in our original depth images,we can’t generate good point cloud.So we need to take some pre-processing steps before generating the point cloud.This thesis proposes a method of depth images filtering in time domain.Compared with traditional median filtering and mean filtering,the new method can effectively fill up the blank hole.To generate the point cloud with depth images,we need the camera intrinsic parameters to compute the coordinate of the point cloud.So this thesis also introduces the knowledge of camera calibration and some frequently-used coordinate systems.One single point cloud in the thesis has more than 100,000 points,using traditional ICP algorithm to registrate two point clouds is very inefficient.In order to improve ICP algorithm’s performance,this thesis proposes an improved ICP algorithm called ICP algorithm Based on Keypoint.Firstly,we will introduce the knowledge of SIFT and RANSAC.With the help of RANSAC,this thesis will show us how to get some high-quality matched points.With these points,we use ICP algorithm to compute the transforming matrix of two point clouds.The new improved ICP algorithm will let us reduce our data number from 100,000 to 1000,which help us improve greatly the efficiency of our project.At the same time,this thesis shows how to down sampling a point cloud.It also describes how to remove points which are out of the range.Lastly,it shows how to generate mesh for a point cloud.This thesis shows that consumers can use low-cost device like Microsoft Kinect,to reconstruct some large-scale scenes.Though there are still some details need to be improved,because of the color difference between images,but the result can basically satisfy people’s demands.