3D Point Cloud Registration Algorithm Using Deep Learning

3D Point cloud registration technology has always been the research focus of the field of computer vision.It is the basis of 3D reconstruction,and has wide applications in medical imaging,autonomous driving,robotics,etc.Traditional point cloud registration algorithms are usually performed using local or global features,which tend to generate local optima and are time-consuming.In recent years,with the development of deep learning,3D point cloud registration algorithms based on learning have emerged one after another and largely improved the registration performance.In order to further improve the accuracy of registration,this thesis deeply studies the 3D point cloud registration methods using the correspondence estimation network and the attention mechanism.The main research contents are summarized as follows:First,the thesis proposes a network model based on mixed attention mechanism and correlation estimation.Considering that the point clouds have complex internal features and random transformations,a hybrid attention mechanism was proposed to focus on the detail features of the point cloud data,extract the key feature information,and enhance the learning ability of the network to important information.The source point cloud and the target point cloud are input at the same time.Through residual connections,the internal correlation of the point cloud as well as the spatial relation between different point clouds can be obtained,thus the more robust point cloud features can be achieved.Subsequently,using the correspondence estimation network for nonlinear excitation,the point cloud features in various space can be better expressed and maintained,thus a closer correlation between point cloud pairs can be achieved.The experimental results on the artificially synthesized dataset Model Net40 and real dataset ICL-NUIM show a significant improvement in registration accuracy in terms of noisy and unseen point clouds with a large affine transformation.Second,the thesis proposes a local feature extraction network with high correspondences to obtain more abundant local information.The key to 3D point cloud registration is fully mining the internal correlation of complex point cloud data and the spatial connection between different points,while the existing works still have shortages in these areas.To solve this problem,this thesis proposes a subtract attention network,which uses several subtract attention modules to generate point-wise feature representation,and extracts the key points of feature space on this basis.In addition,for the direct input point cloud local coordinates,a position encoding network is proposed to determine the spatial correlation between different points.After combining the spatial features of different dimensions,the reliable connections of key points in the feature space are generated.We use the two networks in parallel and take the neighborhood point sets of the sampling points of both source and the target point clouds as input,thus the local correspondences between different points can be obtained and the registration accuracy can be improved.The experimental results on the four different types of data of the widely used dataset Model Net40 show the superiority of the proposed method.