3D Point Cloud Data Registration And Target Detection Based On Voxel

Registration technology is an important field of computer vision,and it can basically be divided into two types: image-based registration and based on 3D point cloud registration.Registration technology has been widely used in military,industrial,medical,archeology and other fields.3D reconstruction technique is also an important subject of research in computer vision.3D reconstruction of a scene has broad application in unmanned reconnaissance aircraft,intelligent robot detection and other fields.Meanwhile,in computer vision field,the goal target detection especially the plane detection has been widely studied and applied in target perception,robotics navigation,3D mapping and other fields.A classification method based on pyramid voxel to address the high computation cost problem in processing 3D point cloud data has been presented in the paper.Currently 3D reconstruction of the point cloud main use ICP algorithm,and the algorithm needs to have a good initial pose estimation,or easy to fall into local optimum iteration.Extended Gaussian Images are proposed to registration for this problem.First,points cloud are divided into voxels,and then calculate the normal vector of each voxel.Second,map the normals to the Gaussian sphere and calulate the number of each normal to form extended Gaussian Images.Last,use extended Gaussian Images to get the rotation matrix,and see the matrix as the initial rotation matrix of ICP algorithm.For the 3D reconstruction of the scene,the Kinect device is used to capture RGB maps and depth maps.First,align RGB maps and depth maps.Then,RGB maps are matched together and feature points are extracted.Last,correspond these feature points to 3D point clouds and use these 3D point clouds computing rotation matrixes.Collect continuous frames data to get a larger scene.For target detection of 3D point cloud,the plane detection is mainly introduced.The 3D point clouds are first partitioned into pyramid voxels.For each voxel,apply the Eigen value decomposition method to analyze 3D points inside and propose an index for plane detection.The voxel detection results are analyzed and combined for planar object detection based on the geometric relationship between voxels and the planar parameters from voxels.The voxel-wise processing instead of point-wise is used in the proposed method so that it can greatly enhance the computation efficiency with good detection results.