Research On 2D/3D Image Registration Algorithm Based On Spatial 4-points Base

Now,the 3D reconstruction has been widely used in various fields of our daily life.And the registration issues has always been the difficult and hot spots in this technical.There are three combinations registration issues of two datasets in 3D reconstruction: registration between images,point cloud registration,image and point cloud registration.The research contents of this thesis involve the registration between pairwise point cloud and the registration of image and point cloud.In the aspect of 3D alignment,the global registration algorithm 4PCS class algorithm with better performance is analyzed.And we present an improved 4PCS point cloud registration algorithm consisting of two stages based on dynamic and asymmetric base.In the rough registration stage,the DA-4PCS algorithm reduces the number of candidate 4-points sets in each test by selecting asymmetric 4-points base and dynamically increasing the length of base segment.And in the fine registration phase,the RC-ICP algorithm segment the two scans into four corresponding regions according to the previously aligned 4-points base,which speed up the search time of nearest point and cut down the data volume involved in the calculation.The result show that,while ensuring the registration accuracy,the algorithm can effectively speed up the original algorithm in both stages of registration.As for point cloud and image registration,a 2D-3D image registration algorithm based on coplanar 4-points sets on the contour is proposed,abbreviated as C4 PSC.The algorithm uses coplanar 4-points sets as registration primitive.Generating plenty of images by perspective projecting point cloud along multiple perspectives.Then,the 2D-3D registration problem is transformed into finding corresponding relationship between images.Learned from the method of finding corresponding 4-points sets in the 4PCS algorithm,we perform the same idea effectively to find the corresponding coplanar 4-points sets between contours of images.After that,we use Homography to calculate the transformation matrix of image and point cloud,thus completing the initial 2D-3D registration.Experiments show that the proposed 2D-3D registration framework can achieve a certain registration effect with a high degree of automation.