Research On RGB-D Based Point Clouds Registration In 3D Reconstruction

With the rapid development of computer software and hardware technology and the continuous expansion of the computer vision application fields,the 3D reconstruction which is able to quickly and accurately model the real world has caused serious concern of scholars at home and abroad.It has broad application prospects in the field of pattern recognition,heritage restoration,film special effects and reverse engineering.In recent years,after Microsoft introduced a cheap RGB-D depth sensor Kinect,rapid and accurate 3D reconstruction becomes a hot research point in computer vision.However,the reconstruction process must integrating multi 3D data sets(point clouds)from different viewpoints to fuse a complete surface information of real scene.So 3D point clouds registration is the key technology in 3D reconstruction,which directly affects the accuracy of reconstructed model.In the completion of this paper,I have completed the research of 3D point clouds registration.The main work includes the following aspects:1)Applying Kinect as a starting point,I have studied its internal structure,color and depth image synchronous acquisition principle and the relevant development library such as OpenNI,Kinect for Windows SDK and PCL to achieve the acquisition and preprocessing tasks.2)By the analysis and implements of the most widely used 3D registration algorithm ICPs,I compared their advantages and disadvantages.Then I proposed the improved method in this paper.The experiment results indicate that this paper's registration algorithm can overcome the local minima in ICPs.3)On the basis of the results obtained from different viewpoints,I use a close loop detection framework g2 o to reduce the impact of error accumulation.All these algorithms above are implemented in the VS2010 IDE under Window7 system.I applied the generic point cloud models as experiment data.The validity of this paper's algorithm is verified through our contrast experiments.