Convolutional Neural Network Based 3D Point Cloud Registration With Structured Light Projection

As an important research direction of computer vision,three-dimensional model reconstruction is widely used in automatic robots,cultural relics protection,medical treatment,augmented reality,space navigation and other fields.With the continuous development of 3D point cloud acquisition technology in recent years,the research on 3D point cloud processing has become a new hotspot in the field of computer vision.The model reconstruction based on threedimensional point clouds is especially important and challenging.This paper focuses on the threedimensional point cloud acquired by structured light,and studies the three-dimensional model reconstruction in the aspects of point cloud local feature descriptor,point cloud registration and multi-view point cloud registration.The main contents and achievements of this paper are as follows:(1)In the aspect of local feature descriptors of point clouds,the feature extraction algorithm framework of convolution neural network is adopted.Through self-supervised learning of local feature descriptors of point clouds by convolution neural network.The descriptors have good discriminability and robustness.(2)In point cloud registration,based on the local feature descriptors of point clouds are extracted by convolutional neural network,feature matching is carried out to obtain homonymous point pairs of point clouds to achieve rough registration of point clouds.ICP algorithm is improved as a precise registration algorithm to further optimize the registration parameters.The experimental results show that the proposed algorithm can achieve pairwise point cloud registration with high accuracy.(3)In multi-view point cloud registration,in order to reduce the cumulative errors in multi view cloud registration,a LRS multi-view point cloud registration method proposed,which optimizes the results of multi-view pointcloud registration globally.The experimental results show that this method can achieve multi-view point cloud registrationsuccessfully.A three-dimensional model reconstruction platform based on structured light is built.Thereafter,the three-dimensional reconstruction of the target model is realized using the proposed methods.