Research On Point Cloud Registration Method Based On Improved Convolutional Neural Network

With the rapid development of modern information technology and graphics application technology,3D reconstruction technology has received more and more attention,especially in the medical and industrial field.The development of this technology has played an important role in the realization of the industry digitalization.The 3D reconstruction technology can virtualize objects into digital models processed by computers in the real world.How to extract key information from these objects to achieve 3D reconstruction quickly has become a new research hotspot and difficulty.The three-dimensional reconstruction technology mainly includes texture reconstruction and geometric processing.The geometric processing needs accurate registration of objects,and the accuracy of registration affects the effect of the threedimensional reconstruction directly.The traditional registration method has higher requirements on the initial point set selection generally,and the time complexity of search registration is higher,especially for the point cloud matching with rotation translation or misalignment,the matching error rate will increase significantly which reduces the accuracy of registration.In order to solve this problem,many scholar have proposed new improvements,but there is still room to improve the accuracy of registration.This paper takes data collection and data classification as the starting point,filtering and matching data sets,and then proposes some improvement methods by combining convolutional neural network technology.The main tasks are:(1)Under the background of the practical significance of ICP point cloud registration algorithm and its research,this paper elaborated the research status of point cloud registration all over the world,and proposed a new idea of point cloud registration technology.This paper introduces the convolutional neural network,and introduces its basic principle and practical application briefly,and then proposes the theoretical basis for the following improvement methods.(2)Point cloud data collection is a basic step before point cloud registration.Traditional data acquisition methods require a good initial posture of the point cloud and slow selection of corresponding points.To order to solve these problems,three orthogonal view CNN structures distinguished the target from other areas in the 3D point cloud candidate is proposed which can provide more accurate point cloud data.This article uses the plastic bottle as the research object,making the detailed assessment and the comparison by simulation experiment.(3)For the large amount of computation and time-consuming problem of registration,an improved depth auto-encoder compression coverage set(LORAX)algorithm to perform point cloud registration is proposed.This method can stratify the collected point cloud which obtain a more results compared with traditional methods.Finally,the improved algorithm reduces the computational complexity greatly,and the registration error and the registration timeconsuming decrease.