Research On Point Cloud Big Data Registration Algorithm Based On Parallel Processing

With the rise of 3D laser point cloud technology,more and more fields need to use point cloud processing technology,such as: unmanned driving,robot recognition,face recognition and other artificial intelligence fields,as well as scanning and repairing precision mechanical parts,heritage building protection The field of 3D reconstruction is inseparable from the participation of 3D point cloud processing technology.As one of the 3D point cloud processing technologies,point cloud registration plays a crucial role in 3D reconstruction.In the collection of point cloud data,due to the size and surface of the object,it is impossible to obtain all the point cloud data of the measured object at one time.It is necessary to collect data from multiple stations,and then splicing the clouds of each station.The point cloud data in each coordinate system is integrated into a unified coordinate system by means of rotation,translation,scaling,etc.This process is called point cloud registration.Among them,the iterative closest point(ICP)algorithm proposed by BESL is still the mainstream,but because the algorithm does not have a good initial position,it is easy to fall into the local optimal solution.Therefore,most of the current algorithms first obtain a good initial position through coarse registration,and then use the ICP algorithm for fine registration.In this paper,the coarse registration algorithm and the fine registration algorithm are respectively studied.Aiming at the shortcomings of the current traditional algorithms,some optimizations have been made in the registration accuracy and efficiency.The main work of this paper:Firstly,the research significance of point cloud registration technology in 3D point cloud reconstruction technology is expounded,the research status of rough registration algorithm and fine registration algorithm at home and abroad is introduced,and the shortcomings of the algorithm are summarized.Then,the fast point feature histogram(FPFH)is discussed in depth,and the extraction and matching of feature points are added on the basis of FPFH,which further improves the registration accuracy.In this method,the feature points of the point cloud data are first extracted by the scale-invariant feature transform(SIFT)algorithm,and then the corresponding point pairs are found by calculating the FPFH,and the random sample consensus algorithm(RANSAC)is used.The correct point pairs are retained,and the transformation matrix is calculated by the singular value decomposition(SVD)algorithm to complete the rough registration.The Stanford dataset is used in the experiment,and the final results show that the extraction and matching of feature points can greatly improve the accuracy of the initial matrix.Secondly,in view of the low efficiency of point cloud big data processing by traditional ICP,an ICP algorithm based on parallel processing,referred to as CUDA-ICP,is proposed.The parallel algorithm is mainly implemented based on the CUDA framework developed by NVIDIA.The GPU is used to accelerate the operation of the ICP algorithm.Due to its own parallelism,the GPU can greatly accelerate the iterative process of the ICP,thereby achieving the purpose of accelerating the registration efficiency.In finding matching point pairs,KD-tree is introduced for fast point-to-point search,which further improves the efficiency of the ICP algorithm.The Stanford dataset and vehicle-mounted point cloud data are used in the experiment to demonstrate the advantages of the CUDA-ICP algorithm in big data point cloud registration processing.Finally,summarize the work carried out in this paper,consider the shortcomings of this paper,and look forward to the future research direction.