Self-supervised rigid transformation equivariance for accurate 3D point cloud registration

Aiming at the insufficient accuracy and the low efficiency of 3D point cloud registration, a point cloud registration method based on subgraph matching and reinforcement learning was proposed to achieve the accurate and fast registration of low-quality point cloud. Firstly, the 3D point cloud registration can result from a series of discrete rigid transformation actions, and this work used a reinforcement learning strategy to train an end-to-end model to iteratively predict the rigid transformation actions. Then, for the model architecture, a Siamese backbone was used to extract the local feature information of the source point cloud and the target point cloud, respectively. Similar nodes in the source graph and the target graph were associated through a proposed cross-graph attention module. The aggregation of graph nodes was designed to extract global features of two graphs, by using the weighted sum with gating vectors. Finally, the global features of the source graph and the target graph were fused, and the discrete rigid transformation action was predicted based on the fused feature. The reinforcement learning strategy significantly improves the generalization of point cloud registration. The cross-graph attention module further improves the accuracy and efficiency of point cloud registration. Extensive experiments on both synthetic and real-scanned datasets, ModelNet40 and ScanObjectNN, demonstrate that, compared with the latest point cloud registration method, ReAgent, the proposed method can reduce the mean average error of rotation by at least 0.16 and the isotropic rotation error by at least 0.16, effectively improving the accuracy of registration on low-quality point clouds.

3D point cloud registration in remote sensing field has been greatly advanced by deep learning based methods, where the rigid transformation is either directly regressed from the two point clouds (correspondences-free approaches) or computed from the learned correspondences (correspondences-based approaches). Existing correspondences-free methods generally learn the holistic representation of the entire point cloud, which is fragile for partial and noisy point clouds. In this paper, we propose a correspondences-free unsupervised point cloud registration (UPCR) method from the representation separation perspective. First, we model the input point cloud as a combination of pose-invariant representation and pose-related representation. Second, the pose-related representation is used to learn the relative pose wrt a "latent canonical shape" for the source and target point clouds respectively. Third, the rigid transformation is obtained from the above two learned relative poses. Our method not only filters out the disturbance in pose-invariant representation but also is robust to partial-to-partial point clouds or noise. Experiments on benchmark datasets demonstrate that our unsupervised method achieves comparable if not better performance than state-of-the-art supervised registration methods.

RoCNet++: Triangle-based descriptor for accurate and robust point cloud registration

The Iterative Closest Points (ICP) algorithm is the mainstream algorithm used in the process of accurate registration of 3D point cloud data. The algorithm requires a proper initial value and the approximate registration of two point clouds to prevent the algorithm from falling into local extremes, but in the actual point cloud matching process, it is difficult to ensure compliance with this requirement. In this paper, we proposed the ICP algorithm based on point cloud features (GF-ICP). This method uses the geometrical features of the point cloud to be registered, such as curvature, surface normal and point cloud density, to search for the correspondence relationships between two point clouds and introduces the geometric features into the error function to realize the accurate registration of two point clouds. The experimental results showed that the algorithm can improve the convergence speed and the interval of convergence without setting a proper initial value.

Three-dimensional face analysis is now a hot research topic in computer vision. The acquisition of 3D faces requires acquiring multi-view face and obtaining them by point cloud registration. In this paper, a deep learning method is introduced in the multi-view RGBD depth sensor system to acquire matching point pairs of image data. Thus, 3D spatial point pairs are obtained. Then, the initial poses between the point clouds of multi-view faces are obtained by rigid body transformation estimation, and the coarse registration of the point clouds is completed. On this basis, for fine registration of multi-view 3D face point clouds using the Iterative Closest Point (ICP) algorithm, this paper proposes a method to convert partially overlapping point cloud registration into sub-region point cloud registration through facial region positioning. In this letter, the maximum value of RMSE for the experimental results of point cloud coarse registration is 2.26 mm with the mean and standard deviation of 1.18 ± 0.33 mm; the maximum value of RMSE for the experimental results of point cloud fine registration is 1.47 mm with the mean and standard deviation of 0.91 ± 0.14 mm. It is confirmed that the proposed method is very robust and effective for the registration of multi-view 3D face point clouds.

Point cloud registration from laser scanning data is a technique to establish the mapping relationship between source and target point clouds, which has been widely used in automatic 3D reconstruction, pose estimation, localization, and navigation. While algorithms like Super4PCS and MSSF-4PCS can achieve registration without initial poses, they are relatively slow, less accurate, and require iterations. To address these issues, we propose a 3D point cloud registration algorithm based on interval segmentation and multi-dimensional feature. Firstly, the source and target point clouds are segmented internally and the point cloud curvature is designed to narrow down the search range for the registration between the segmented point clouds. Secondly, the corresponding four-point sets in the segmented areas of the source and target point clouds are determined using affine invariance constraints. Finally, a multi-dimensional feature vector based on curvature features and fast point feature histogram is established to determine the unique corresponding four-point set pairs, and the rigid body transformation matrix is solved accordingly. Our algorithm is tested on publicly available 3D point cloud data models Bunny, Dino, Dragon, and Horse from Stanford University. Results showed that our algorithm improved registration accuracy by 24.39% and registration efficiency by 46.21% compared to the MSSF-4PCS point cloud registration algorithm. Multiple sets of experimental results confirmed the effectiveness of our algorithm. The proposed 3D point cloud registration is proved to be fast with high accuracy, which can be utilized for automatic segmentation, reconstruction, and modelling from Laser Scanning Data.

Point cloud registration aims to find a rigid spatial transformation to align two given point clouds; it is widely deployed in many areas of computer vision, such as target detection, 3D localization, and so on. In order to achieve the desired results, registration error, robustness, and efficiency should be comprehensively considered. We propose a deep learning-based point cloud registration method, called DOPNet. DOPNet extracts global features of point clouds with a dynamic graph convolutional neural network (DGCNN) and cascading offset-attention modules, and the transformation is predicted by a multilayer perceptron (MLP). To enhance the information interaction between the two branches, the feature interaction module is inserted into the feature extraction pipeline to implement early data association. We compared DOPNet with the traditional method of using the iterative closest point (ICP) algorithm along with four learning-based registration methods on the Modelnet40 data set. In the experiments, the source and target point clouds were generated by sampling the original point cloud twice independently; we also conducted additional experiments with asymmetric objects. Further evaluation experiments were conducted with point cloud models from Stanford University. The results demonstrated that our DOPNet method outperforms these comparative methods in general, achieving more accurate and efficient point cloud registration.

This paper proposes a deep learning model for point cloud registrations of different sizes. 3D point clouds play a very important role in various fields. They have been widely studied, and recently deep learning has also started to deal with point clouds. PointNet was the first deep learning model for point cloud classification and semantic segmentation. Since then, methods based on PointNet for tasks like point cloud registration have also been proposed. However, these methods are only suitable for identical or nearly identical point clouds. However, in practice, the sizes of point clouds vary depending on the capture distance, sensor type, the environment, and many other factors. Therefore, it is often the case that point clouds that need to be registered are of very different sizes. For example, point clouds captured in the same environment by an omnidirectional LiDAR and an RGB-D camera will have very different sizes. Conventional methods cannot cope with such situations. In this paper, we propose ‘PointpartNet', a new deep neural network based on partial feature extraction. This network enables feature extraction of partial point clouds by partitioning the point clouds. It uses the features of partial point clouds to search for matching regions between point clouds of different sizes. This makes it capable of registering point clouds of different sizes. In qualitative experiments, we demonstrate its high robustness and accuracy for point cloud registration of different sizes in comparison to previous research.

Terrestrial laser scanning (TLS) technology has the advantages of wide range, high efficiency, and low cost in spatial information collection, so it is widely used in infrastructure monitoring and measurement. During TLS application, the registration and positioning of the point cloud have a direct impact on the quality of the data and the validity of the results. The linear distribution of the tunnel structure and the lack of significant features present challenges in the registration and positioning of 3D point clouds in railway tunnels. The commonly used registration methods are difficult to achieve high registration accuracy and are prone to propagation errors, which reduce the accuracy and effectiveness of results. To achieve accurate registration and positioning of multistation clouds in railway tunnels, we propose a coordinate-based global registration method. To determine the coordinates of scan points in the reference coordinate system and the direction of the reference coordinate system, a few fixed control points are used during the data collection stage. Consequently, each station cloud can be precisely positioned and automatically registered in the reference coordinate system without accumulating or propagating errors. In addition, the coordinate-based registration method eliminates the introduction of errors due to artificial target setting and feature point extraction, as well as the problem of accurately positioning the entire point cloud in the reference coordinate system, thereby enhancing the accuracy, efficiency, and automation levels of cloud registration. The experiment demonstrates that the coordinate-based global registration method is robust and applicable in complex scenes, and it is suitable for the accurate positioning and registration of multistation clouds in linear and curved railway tunnels. The coordinate-based registration method reduces the amount of error in the global registration link by 65% when compared to the point-based registration method, and the point cloud accuracy has reached fine registration, ensuring that fine-grained inverse modeling of the tunnel structure can be performed.

Registration of 3D point clouds is an important issue in 3D scanning domain. In this work, we developed a turntable-base structured light system for the automatically 3ED scanning purpose. To realize the fully automatically 3D point clouds registration, a planar surface is placed on the turntable and scanned with different rotation angles. The turntable axis direction vector is calculated by averaging the intersection lines of the reconstructed planes firstly. And the linear minimization function is constructed for the robust and accurate estimation of the turntable origin. A variety of objects are used in the experiment. And the results show that, point clouds under different scanning angles can be precisely registered. In addition, there are no iterative procedures in the proposed algorithm, and that makes the registration procedure more efficiently.?

Point cloud registration is the basis of real-time environment perception for robots using 3D LiDAR and is also the key to robust simultaneous localization and mapping (SLAM) for robots. Because LiDAR point clouds are characterized by local sparseness and motion distortion, the point cloud features of coal mine roadway environments show a weak texture and degradation. Therefore, for these environments, the traditional point cloud registration method to register directly will lead to problems, such as a decline in registration accuracy, z-axis drift, and map ghosting. To solve the above problems, we propose a point cloud registration method based on IMU preintegration with the sensor characteristics of LiDAR and IMU. The system framework of this method mainly consists of four modules: IMU preintegration, point cloud preprocessing, point cloud frame matching and point cloud registration. First, IMU sensor data are introduced, and IMU linear interpolation is used to correct the motion distortion in LiDAR scanning, and the IMU preintegration error function is constructed. Second, the point cloud segmentation is performed using the ground segmentation method of RANSAC to provide additional ground constraints for the z-axis displacement and to remove the unstable flawed points from the point cloud. On this basis, the LiDAR point cloud registration error function is constructed by extracting the feature corner points and feature plane points. Finally, the Gaussian Newton solution is used to optimize the constraint relationship between the LiDAR odometry frames to minimize the error function, complete the LiDAR point cloud registration and better estimate the position and pose of the mobile robot. The experimental results show that compared with the traditional point cloud registration method, the proposed method has a higher point cloud registration accuracy, success rate and computational efficiency. The LiDAR odometry constructed using this method can better reflect the authenticity of the robot trajectory and has higher trajectory accuracy and smaller absolute position and pose error.

3D Point cloud registration is still a very challenging topic due to the difficulty in finding the rigid transformation between two point clouds with partial correspondences, and it’s even harder in the absence of any initial estimation information. In this paper, we present an end-to-end deep-learning based approach to resolve the point cloud registration problem. Firstly, the revised LPD-Net is introduced to extract features and aggregate them with the graph network. Secondly, the self-attention mechanism is utilized to enhance the structure information in the point cloud and the cross-attention mechanism is designed to enhance the corresponding information between the two input point clouds. Based on which, the virtual corresponding points can be generated by a soft pointer based method, and finally, the point cloud registration problem can be solved by implementing the SVD method. Comparison results in ModelNet40 dataset validate that the proposed approach reaches the state-of-the-art in point cloud registration tasks and experiment resutls in KITTI dataset validate the effectiveness of the proposed approach in real applications.

3D point cloud registration is a critical technology in the fields of visual measurement and robot automation processing. In actual large-scale industrial production, the accuracy of point cloud registration directly affects the quality of automated welding processes. However, most existing methods are confronted with serious challenges such as the failure of partial-to-partial point cloud model registration when facing robot automatic processing guidance and error analysis work. Therefore, this paper proposes a novel two-stage network architecture for point cloud registration, which aims at robot pose adjustment and visual guidance in the field of automated welding by using 3D point cloud data. Specifically, we propose a neighborhood-based multi-head attention module in the coarse registration stage. The neighborhood information of each point can be aggregated through focusing on different weight coefficients of multi-head inputs. Then the spatial structure features which is used to establish the overlapping constraint of point clouds are obtained based on above neighborhood information. In the fine registration stage, we propose the similarity matching removal module based on multiple attention fusion features to explore deeper features from different aspects. By using deep fusion features to guide the similarity calculation, the interference of non-overlapping points is removed to achieve the finer registration. Eventually, we compare and analyze the proposed method with the SOTA ones through several error metrics and overlap estimation experiments based on the ModelNet40 dataset. The test results indicate that our method, relative to other mainstream techniques, achieves lower error rates and the most superior accuracy of 98.61% and recall of 98.37%. To demonstrate the generalization performance of proposed algorithm, extensive experiments on the Stanford 3D Scanning Repository, 7-Scenes and our own scanning dataset using partially overlapping point clouds individually under clean and noisy conditions show the validity and reliability of our proposed registration network.

The present work proposes a new method Reg3DFacePtCd for registration of point clouds. The key contribution of the present method is that an unknown face in 3D point cloud form is given to the system and is registered to the already existing known 3D face point clouds using a fast 3D face registration method. The novelty of the present technique is that at first the alignment and registration parameters are found out by initially registering eight key points of the unknown source model to that of the known model. Next, the rest of the point clouds of the unknown model are registered to that of the known model using the same parameters found as above. The main method used for alignment is iterative closest point (ICP) using point-based technique followed by registration in the least squares sense. Mainly there are two significant contributions. Firstly, we have developed a new mathematical model facial landmark point based model across poses to obtain the target or the known model to which all the unknown models will be registered. Secondly, a novel way to accelerate point cloud matching by reducing the number of points has been proposed. Using a small number of points necessarily would speed up the registration process but may inculcate errors. So, to determine the registration quality of the fundamental eight key points on which the entire registration process is based, a new robust metric namely ICV (ICP certainty vector) consisting of several key components have been used. Finally, we have addressed several important face registration issues like pre-processing, convergence and quality of registration of the entire facial point cloud based on the eight key points. Extensive experimentation on Frav3D, GavabDB, and the Bosphorus databases on a high-performance computing environment show the novelty and robustness of the method.

Three‐dimensional (3D) point cloud registration is a fundamental key issue in 3D reconstruction, 3D object recognition and augmented reality. In this study, the authors propose a novel local feature descriptor called local angle statistics histogram (LASH) for efficient 3D point cloud registration. LASH forms a description of local shape geometries by encoding their properties on angles between the normal vector of the point and the vector formed by the point and other points in its local neighbourhood. In addition, they propose a 3D point cloud registration algorithm based on LASH. The registration algorithm firstly detects triangle matching points with consistent similarity ratios, and then aggregates each pair of triangular matching points into a set of matching points. They can use these matching sets to calculate multiple transformations between two point clouds. Finally, they use the error function to identify the best transformation and to achieve coarse alignment of the two point clouds. Experiments and comparisons with other global algorithms demonstrate that the proposed approach can be applied to register point clouds with considerable or limited overlaps and is robust to noise.