Research Of High-precision Point Cloud Registration Algorithm Based On Deep Learning

Point cloud registration is essential in 3D computer vision and graphics and has important applications in 3D reconstruction,3D data fusion,simultaneous localization and mapping(SLAM),and many other research fields.With the development of autonomous driving,artificial intelligence,virtual reality technology etc.,deep learning in point cloud registration has also achieved many inspiring research findings.However,point cloud data collected by external sensors often have redundancy and quality defects.On the one hand,the registration process takes up a lot of memory resources and processing time.On the other hand,noise interference will also lead to abnormal correspondences generated by feature matching,thereby influencing the accuracy of point cloud registration.In the field of deep learning,point cloud registration methods represented by PCRNet only focus on the global information of point cloud without considering local features,which makes it difficult to characterize the complex space and restricts the ability of scene understanding,thereby reducing the accuracy of point cloud registration.At the same time,PCRNet only connects the features of two groups of point cloud data in dimensions based on concatenation,which cannot fully use the complementarity of different levels and cannot well fuse the local features with global information,thereby destroying the accuracy of point cloud registration.Aiming at the aforementioned problems,this thesis mainly focuses on studying and improving the following three issues:(1)Aiming at the redundancy and quality defects in point cloud data collected by external sensors,this thesis proposes a point cloud data preprocessing scheme based on geometric features,mainly containing point cloud filtering and point cloud segmentation.The point cloud filtering part utilizes the random sampling method to complete a specified number of collections for the original point cloud data and then utilizes a statistical filter to remove outliers from the downsampled point cloud.The point cloud segmentation part utilizes the euclidean clustering algorithm to separate the target object better,which can lay the foundation for the next point cloud registration using effective point cloud information.(2)Aiming at the problem that the deep-learning-based methods for point cloud registration represented by PCRNet only focus on the global information of point cloud without considering local features,this thesis improves the feature extraction module of the primary PCRNet,and uses position adaptive convolution to construct our network PACNet.Compared with the primary PCRNet,the experimental results demonstrate that PACNet has brought 25.6%,13.7%,21.6%,39.4%,22.1%,and19.1% improvements in the six criteria: mean squared error(MSE),root mean square error(RMSE)and mean absolute error(MAE)of the rotation matrix and translation vector,respectively.Compared with other state-of-the-art algorithms,the method proposed in the thesis also has higher accuracy in point cloud registration.(3)Aiming at the problem that PCRNet only connects the features of two groups of point cloud data in dimension based on concatenation,which cannot well fuse the features of different levels,this thesis improves the feature extraction module of the primary PCRNet,and utilizes the dual attention mechanism to form our point cloud registration network PACNet-att.Compared with the primary PCRNet,the experimental results show that PACNet-Att brings 30.3%,16.5%,23.4%,40.7%,22.9%,and 24% improvement in the six criteria mentioned in issue(2).Compared with other state-of-the-art algorithms,the method proposed in the thesis has higher accuracy,better generalization performance,and stronger noise robustness in point cloud registration tasks.