The Research Of 4PCS Point Cloud Registration Parallel Algorithm Based On GPU

With the rapid development of the information age,3D technology is becoming mature,and as an important research topic in computer graphics,it is used widely and involved in every aspect of the life.Because of partial occlusion and other issues,it needs multiple scans to splice and registration to obtain a complete point cloud data.As a key technology of 3D technology,3D point cloud registration has received a lot of attention and research.In recent years,computer is rapidly becoming an era of parallel computing,and more and more computers apply multi-core processors,and the improvement of the requirement of performance results in higher demand of the software speed.Based on previous researches,the thesis uses GPU parallel technology to accelerate the 4PCS to improve the efficiency.The main work of the thesis is as follows:Aiming at the problem of inefficient implementation of the approximate congruent four points searching algorithm,the thesis transplants it to GPU through CUDA platform and distributes the large-scale matrix operation to appropriate threads for parallel.Finally,the algorithm is proved by several experiments,which improves the searching efficiency and the efficiency of the 4PCS.Because of the complexity of tree building and searching,the kd-tree algorithm in the approximate congruent four points searching algorithm and LCP influences the parallel algorithm seriously.Aiming at the problem,the thesis proposes to use distance comparison to replace tree building and searching by tree and transplant it to GPU,and it is proved to increase the number of groups of approximate congruent four points and improve the accuracy of the 4PCS.For the problem of high time complexity of the LCP,the thesis transplants the cyclic operation part to GPU by CUDA,and the multi-thread parallel execution is used instead of the cyclic calculation of registration accuracy.Its feasibility and effectiveness are proved.The final comparison analysis shows that the parallel registration algorithm has better time-to-speed ratio than before.