Improved UTOPIC method for partially overlapped point cloud registration

Abstract

The invention relates to an improved UTOPIC method for partially overlapped point cloud registration, and belongs to the field of computer vision and point cloud registration. According to the method, the UTOPIC with uncertainty perception is taken as a basic framework, and a double-branch feature extractor is introduced in a feature extraction part, so that richer and deeper local feature information of the point cloud is acquired, and the local feature information of the point cloud is acquired; in the uncertainty perception feature enhancement part, a double-constraint matching module is introduced, and a Euclidean distance matching matrix and an affinity matching matrix are utilized to realize double constraints on the point cloud, so that a point corresponding relation matrix is obtained, and the number of wrong matching points is reduced; in the overlapping guide registration part, an improved overlapping point prediction module is introduced to accurately identify points with relatively high overlapping probability in a source point cloud and a target point cloud, and an overlapping region of the two point clouds is found to weaken the influence of a non-overlapping region on registration, so that the point cloud registration precision is effectively ensured.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and point cloud registration, specifically involving an improved UTOPIC method for registration of partially overlapping point clouds. Background Technology

[0002] Point clouds are datasets composed of massive discrete 3D points. Rigid point cloud registration, a crucial step in point cloud data processing, aims to find the transformation matrix between point clouds from two different viewpoints, thereby aligning them to the same coordinate system. Currently, point cloud registration technology is developing rapidly and has begun to emerge in many practical engineering applications, including but not limited to workpiece surface defect detection, pose estimation, autonomous driving, and 3D reconstruction. However, due to external environmental factors and the accuracy of visual measurement equipment, point clouds inevitably contain noise and outliers during acquisition. Furthermore, two point clouds obtained from different viewpoints of the same object are likely to have non-overlapping regions and significant initial pose deviations, all of which pose challenges to the registration operation between point cloud data. Therefore, how to further improve the registration accuracy and robustness of future point cloud registration algorithms under complex environmental conditions has become a hot topic of widespread research interest.

[0003] Currently, point cloud registration methods can be broadly categorized into two main types: traditional registration algorithms and deep learning-based registration algorithms. Among traditional registration algorithms, the Iterative Closest Point (ICP) algorithm is the most representative. The core idea of ​​ICP is to find the closest corresponding point in the target point cloud for each point in the source point cloud. Then, based on these correspondences, using Euclidean distance as a metric, iteratively estimates and updates the rigid transformation relationship (rotation and translation) between the two point clouds to minimize the spatial pose deviation between the source and target point clouds. However, ICP and its various improved versions generally suffer from a significant drawback—a strong dependence on the initial pose of the point clouds. If the source and target point clouds differ greatly in their initial positions or poses, such algorithms are highly likely to get trapped in local optima, thus failing to obtain accurate global point cloud registration results.

[0004] In recent years, with the successful application of deep learning technology in the field of image vision, researchers have begun to introduce various deep neural network models into 3D point cloud registration tasks. For some existing deep neural network models used for point cloud registration (such as Iterative Distance-Aware Convolutional and Mutually Supervised Point Removal Network (IDAM) and Deep Nearest Point Network (DCP), the core steps are to first construct descriptors that can characterize local or global features, then match source and target points based on these descriptors, and finally estimate the rigid transformation parameters between the two sets of point clouds through regression algorithms or singular value decomposition (SVD). Compared with traditional ICP, although such point cloud registration methods have improved in accuracy and speed, they still have many shortcomings in the feature extraction stage. Moreover, when there is only partial overlap between point clouds, a large number of irrelevant or noisy points can easily lead to incorrect matching results, making it difficult for them to cope well with various complex and changing working scenarios in the future.

[0005] Therefore, how to improve the point cloud feature extraction capability, the ability to perceive overlapping regions, and the registration robustness of future deep neural network models in the case of partial point cloud overlap has become a key technical problem that urgently needs to be solved in the field of point cloud registration. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the prior art by providing an improved UTOPIC method for registration of partially overlapping point clouds. This method uses an Uncertainty-aware Overlap Prediction Network (UTOPIC) as its basic framework. In the feature extraction part, it introduces a dual-branch feature extractor to obtain richer and deeper local feature information from the point cloud. In the uncertainty-aware feature enhancement part, a dual-constraint matching module is introduced, utilizing Euclidean distance matching matrix and affinity matching matrix to implement dual constraints on the point cloud, thereby obtaining a point correspondence matrix and reducing the number of incorrectly matched points. In the overlap-guided registration part, an improved overlap point prediction module is introduced to accurately identify points with a high probability of overlap in the source and target point clouds, finding the overlapping region between the two point clouds to reduce the influence of non-overlapping regions on registration, thus effectively ensuring the accuracy of point cloud registration.

[0007] To achieve the above objectives, the technical solution of this invention is: an improved UTOPIC method for registration of partially overlapping point clouds, using the uncertainty-aware overlap prediction network UTOPIC as its basic framework; a dual-branch feature extractor is introduced in the feature extraction part to capture local feature information of the point cloud; in the uncertainty-aware feature enhancement part, a dual-constraint matching module is introduced to obtain the point correspondence matrix; in the overlap-guided registration part, an improved overlapping point prediction module is introduced to distinguish the overlapping regions of the point clouds to be registered, ultimately achieving point cloud registration. The method includes the following steps:

[0008] Step S1: Read point cloud data from the obtained point cloud dataset to obtain the three-dimensional coordinates of each point cloud sample. After downsampling, use random rotation and random translation operations to generate two point clouds with different initial poses and the true value of rigid transformation. Then, perform a cropping operation on the two point clouds to generate a partially overlapping source point cloud S and a target point cloud Q.

[0009] Step S2: Construct an improved UTOPIC network model for point cloud registration, including a feature extraction module, an uncertainty-aware feature enhancement module, and an overlap-guided registration module. The feature extraction module captures local feature information of the point cloud, the uncertainty-aware feature enhancement module obtains the point correspondence matrix, and the overlap-guided registration module identifies the overlapping areas of the point cloud to be registered.

[0010] Step S3: Train the point cloud registration model. The loss function used for training is consistent with the improved UTOPIC network model. Use the trained point cloud registration model for registration of partially overlapping point clouds.

[0011] Compared with the prior art, the present invention has the following beneficial effects:

[0012] In situations involving noise and partial overlap of point clouds, most traditional point cloud registration algorithms and some existing learning-based algorithms often fail to achieve the desired registration results. To address this issue, this invention improves the traditional UTOPIC model: First, a dual-branch structure is introduced in the feature extraction section to extract point cloud features in a multi-level manner, capturing point cloud geometric information at different scales, thereby significantly enhancing the perception of point cloud shape features. Second, a dual-constraint matching module is introduced in the uncertainty-aware feature enhancement section to evaluate the point cloud correspondence matrix from two dimensions, effectively suppressing erroneous matching caused by outliers. Finally, in the overlap-guided registration section, an improved overlap point prediction module is introduced, which can more accurately locate points in the overlapping region, reducing interference from non-overlapping points, noise, and outliers. These improvements effectively enhance the model's registration accuracy and robustness in situations with partial point cloud overlap. Attached Figure Description

[0013] Figure 1This is the improved UTOPIC point cloud registration network model proposed in this invention;

[0014] Figure 2 This is a structural diagram of the dual-branch feature extractor of the present invention;

[0015] Figure 3 This is a structural diagram of the efficient local attention mechanism of the present invention;

[0016] Figure 4 This is a structural diagram of the geometric edge convolution module of the present invention;

[0017] Figure 5 This is a structural diagram of the geometric transformer module of the present invention;

[0018] Figure 6 This is a structural diagram of the improved overlap point prediction module proposed in this invention;

[0019] Figure 7 This is a visualization of the point cloud registration effect of the present invention, where blue represents the source point cloud and yellow represents the target point cloud. Detailed Implementation

[0020] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] This invention provides an improved UTOPIC method for registration of partially overlapping point clouds, using the uncertainty-aware overlap prediction network UTOPIC as its basic framework. In the feature extraction part, a dual-branch feature extractor is introduced to capture local feature information of the point cloud. In the uncertainty-aware feature enhancement part, a dual-constraint matching module is introduced to obtain the point correspondence matrix. In the overlap-guided registration part, an improved overlap point prediction module is introduced to identify the overlapping regions of the point clouds to be registered, ultimately achieving point cloud registration. The method includes the following steps:

[0022] Step S1: Read point cloud data from the obtained point cloud dataset to obtain the three-dimensional coordinates of each point cloud sample. After downsampling, use random rotation and random translation operations to generate two point clouds with different initial poses and the true value of rigid transformation. Then, perform a cropping operation on the two point clouds to generate a partially overlapping source point cloud S and a target point cloud Q.

[0023] Step S2: Construct an improved UTOPIC network model for point cloud registration, including a feature extraction module, an uncertainty-aware feature enhancement module, and an overlap-guided registration module. The feature extraction module captures local feature information of the point cloud, the uncertainty-aware feature enhancement module obtains the point correspondence matrix, and the overlap-guided registration module identifies the overlapping areas of the point cloud to be registered.

[0024] Step S3: Train the point cloud registration model. The loss function used for training is consistent with the improved UTOPIC network model. Use the trained point cloud registration model for registration of partially overlapping point clouds.

[0025] The following is a detailed implementation process of the present invention.

[0026] This invention proposes an improved UTOPIC method for registration of partially overlapping point clouds, such as... Figure 1 As shown, it includes the following steps:

[0027] Step S1: Read point cloud data from the obtained point cloud dataset to obtain the three-dimensional coordinates of each point cloud sample. After downsampling, generate two point clouds with different initial poses and the true value of rigid transformation using random rotation and random translation operations. Then, perform a cropping operation on the two point clouds to generate a partially overlapping source point cloud S and a target point cloud Q.

[0028] Step S2: Construct an improved UTOPIC network model for point cloud registration, such as... Figure 1 As shown, it mainly consists of three parts: a feature extraction module, an uncertainty-aware feature enhancement module, and an overlap-guided registration module.

[0029] The feature extraction module mainly utilizes a two-branch structure composed of "adaptive graph convolution + attention mechanism" and "geometric edge convolution" (e.g., Figure 2 As shown), to extract features F from the source point cloud S. S Features F of the target point cloud Q Q The following section will focus on extracting the features F of the source point cloud S. S Taking this as an example, the dual-branch structure will be described in detail. Figure 2 In the adaptive graph convolution branch, four AdaptConv layers of sizes 64, 64, 128, and 256 are sequentially connected in series. The features output from each layer are concatenated to obtain a 512-dimensional output feature. For each AdaptConv layer, let its input source point cloud be... The corresponding features are Where N is the number of points contained in the source point cloud, and D is the feature dimension. A directed graph is constructed using the k nearest neighbors (kNN) of each point. in It is a set of points. Let x represent the set of edges. i The center point of the graph convolution is... It is the set of index points in its neighborhood. For the m-th (m = 1, 2, ..., M) channel of the M-dimensional features output by each layer of adaptive graph convolution, the point feature (f) is used. i ,f j Dynamically generate an adaptive kernel:

[0030]

[0031] Where, Δf ij =cat[f i ,f j -f i ] represents the input features of the adaptive kernel, where cat[·,·] is the concatenation operation; g m (·) represents the feature mapping function, which is a multilayer perceptron (MLP) used here. Each adaptive kernel corresponds to a filter defined in AdaptConv.

[0032] Using an adaptive kernel With Δx ij Perform convolution to obtain

[0033]

[0034] Where, point x i With point x j Spatial coordinate relationship between <·,·> represents the inner product of two vectors, and σ is a non-linear activation function.

[0035] As shown in equation (2), the m-th adaptive kernel With Δx ij In combination, this means that the kernel size should satisfy the dot product dimension requirement, that is, the above feature mapping function should be... In this way, the spatial location of the input point cloud can be effectively incorporated into each layer, combined with the dynamically extracted feature correspondences. The h values ​​of each channel are stacked. ijm Forming connection point (x) i ,x j Edge features between ) Thus, we obtain point x. i Local features within the neighborhood are Extending to the entire point cloud means

[0036] Subsequently, an efficient local attention mechanism (ELA) is used to enhance the expressive power of the extracted local features, the structure of which is as follows: Figure 3 As shown. Specifically, given input features (C, H, and W represent the feature channel dimension, height, and width, respectively), and one-dimensional average pooling is performed on each channel in both the horizontal and vertical spatial directions to obtain the corresponding output features z_h and z_w. The output value of feature z_h at the c-th channel at the corresponding height h is... It can be represented as:

[0037]

[0038] Where, x c (h,i w ) represents the feature value of the input feature X in the c-th channel at the corresponding position.

[0039] Similarly, the output of the c-th channel of feature z_w at the corresponding width w It can be represented as:

[0040]

[0041] Where, x c (j h ,w) represents the feature value of the input feature X in the c-th channel at the corresponding position.

[0042] The two transformations described above aggregate features along two spatial directions, generating a pair of bidirectional perceptual complementary feature maps. These maps allow attention blocks to capture long-distance dependencies along one spatial direction while preserving precise positional relationships along the other, enabling a larger receptive field and encoding precise positional information. To effectively utilize features z_h and z_w, one-dimensional convolution is applied to enhance positional information in the horizontal and vertical directions. This enhanced positional information is then processed by group normalization (denoted as G_n) to obtain positional attention representations in the horizontal and vertical directions, as shown in Equation (5).

[0043]

[0044] Where σ is a non-linear activation function, and F_h and F_w are one-dimensional convolutions.

[0045] We obtain two attention representations y h and y w Then, Equation (6) is applied to weight the input features to emphasize the feature representations of interest.

[0046] Y = x c ×y h ×y w (6)

[0047] Finally, by using point x i The center point x is obtained by applying the max pooling function to the edge features of the neighborhood. i The output features (i.e., the output features of this layer obtained by max pooling in the channel dimension) are extracted layer by layer using 4 layers of AdaptConv. After extracting features layer by layer, the output features of each layer are concatenated in the channel dimension to obtain the features extracted by the adaptive graph convolution branch.

[0048] The geometric edge convolution branch consists of four GEConv layers of sizes 64, 64, 128, and 256. The outputs of each layer after Maxpooling are concatenated to obtain a 512-dimensional output feature. Figure 4 The main implementation process of GEConv is demonstrated. It still uses kNN for neighborhood search to determine the local neighborhood around each center point. For center point u... i ( Figure 4 Each point u in the local neighborhood of the red dot j (j = 1, 2, ..., k), calculate the geometric features, which are obtained by calculating the point u. j With center point u i and the average value m of all points in the neighborhood i ( Figure 4 The distances between the points (in the middle yellow dot) are obtained using the Euclidean and cosine distances. j geometric features It consists of five parts, including two Euclidean distances. and three angles The cosine values ​​constitute the geometric features of the extracted local neighborhood, defined as follows:

[0049] Subsequently, using the geometric feature set X g To construct the neighborhood geometric edge set By connecting the geometric features of each point in the neighborhood Geometric characteristic mean of all points in the neighborhood ( Figure 4 The geometric edges are constructed using the light pink dots. The formula for calculating the geometric edges is as follows:

[0050]

[0051] Meanwhile, in order to create deeper models through the layered stacking of GEConv layers, it is necessary to propagate the features of points from the previous layer to the current layer. Therefore, feature edges within the neighborhood are generated using equation (8), and the set of feature edges is represented as follows: Since the first layer does not exist, therefore It is an empty set

[0052]

[0053] The final edge information is obtained by concatenating the geometric edges with the feature edges. Right now

[0054] The features of each edge are extracted using equation (9), and feature upscaling is performed using a shared MLP to raise the features of each edge to a higher-dimensional feature space (this process is denoted as γ). Then, nonlinear activation is performed using σ, i.e.

[0055]

[0056] The output features of the final center point are obtained by aggregating using the Maxpool maximum pooling function.

[0057] In summary, the output of GEConv in the first layer can be described by equation (10), since there is no previous layer, and therefore no features from the previous layer. For subsequent layers, the features from the previous layer are used to calculate the features of the current layer, as shown in equation (11). Finally, the features extracted layer by layer are concatenated along the channel dimension to obtain the features extracted by the geometric edge convolution branch.

[0058]

[0059] Features extracted by convolutional branches of splicing adaptive graphs Features extracted by geometric edge convolution branches The source point cloud features extracted from the two branches can then be obtained. For target point cloud Feature extraction (where J is the number of points contained in the target point cloud) follows the same steps to obtain the target point cloud features.

[0060] Furthermore, if the complete shape of two partially overlapping point clouds can be recovered, registering them becomes much easier. Therefore, in the feature extraction module, after extracting the point cloud features, an additional point cloud completion decoder from the PCN (Point Cloud Completion Network) is added. The completion loss function of this part is also included in the overall network loss function. This ensures that the features extracted by the feature extractor not only register the two point clouds but also contain information for restoring "local fragments" into "complete objects." This completion module incorporates rich geometric knowledge into the feature extractor, enhancing feature representation. It's important to note that this completion module is only used during the training phase and will be removed during the testing phase.

[0061] The structure of the uncertainty perception feature enhancement module is as follows: Figure 1 The middle section, as shown, mainly consists of a geometric transformer module, an uncertainty quantification module, a dual-constraint matching module, and an uncertainty-aware weighting module. The geometric transformer module is as follows: Figure 5As shown, it includes three parts: geometric relation embedding, geometric self-attention, and cross-attention. The geometric relation embedding part encodes the distance and angle between points. This information varies at different locations in the point cloud. Specifically, given two points p... i and p j Their geometric relationship can be described as follows:

[0062] (1) Point spacing: The point spacing is the distance between points p in Euclidean space. i and point p j The distance ρ between i,j =||p i -p j ||2.

[0063] (2) Point angle: The search point p in this invention i two nearest neighbors and Then calculate the vector. and The sum (counted as) If the angle α between the points is α, then... i,j Through calculation get.

[0064] (3) Difference in the perimeter of the triangle: Search point p i two nearest neighbors and The three points together form a triangle Search point p j two nearest neighbors and The three points together form a triangle The difference in perimeter between two triangles is defined as and

[0065] The final geometric relationship is obtained by combining the point spacing, point angle, and triangle perimeter difference, and embedding them into g. i,j As shown in equation (12):

[0066] g i,j =cat[ρ i,j ,α i,j ,η i,j (12)

[0067] The geometric self-attention module accurately captures the inherent structure of a single point cloud (source or target point cloud) by leveraging embedded geometric relationships, and extracts contextual information within the single point cloud using the internal spatial structure and characteristics of the data. The cross-attention module promotes feature interaction and fusion between the source and target point clouds, which helps enhance the global consistency of features. These two modules iterate alternately R times, ultimately generating attention features that include global context.

[0068] For two point clouds that partially overlap, a one-to-one correspondence cannot be found between each point. Most existing registration methods, when determining the overlapping region of two point clouds, ignore the inherent uncertainty between overlapping and non-overlapping regions, and are easily influenced by other points with similar features but located in non-overlapping regions. Based on these considerations, the uncertainty quantization module of this invention utilizes a probability-based overlap uncertainty quantization scheme to generate more discriminative features. Specifically, for the source point cloud S, this invention assigns a Gaussian distribution to each point, with a mean μ. i and variance σ i Parameterization, these parameters are... The predictions are derived from different MLP layers. Then, each point s in the source point cloud can be sampled from the learned distribution. i The overlap fraction, i.e., o i ~N(μ) i ,σ i The sampling operation consists of two steps: first, a random sample, ε, is obtained from the standard Gaussian distribution N(0,I). i ~N(0,I), and through μ i +ε i σ i Obtain the desired sample; then, to measure the overlap uncertainty at each point, resample the overlap score for each point K times from the learned distribution, denoted as O. S ={o (1) ,...,o (K)}. O S Treat it as an empirical sample from an approximate predicted distribution, and measure the uncertainty by calculating the variance:

[0069] U S =norm(var(O S ))(13)

[0070] in The expression represents overlap uncertainty, norm(·) is the minimum-maximum normalization, and var(·) is used to calculate the variance.

[0071] Similarly, the overlap uncertainty U of the target point cloud Q QIt was determined in the same way.

[0072] The dual-constraint matching module generates a soft correspondence matrix based on the features of the input source and target point clouds. To obtain the soft correspondence matrix This invention considers two aspects: Firstly, the feature descriptors between corresponding points should be similar, and the Euclidean distance between their features should be relatively small. Therefore, this invention calculates the Euclidean distance matrix C1 between the source point cloud features and the target point cloud features, that is, the element in the i-th row and j-th column of the C1 matrix is... On the other hand, the more similar the features of a pair of points are, the larger their values ​​in the affinity matrix. Therefore, the affinity matrix C2 between features is calculated, that is, the element at the i-th row and j-th column of the C2 matrix has a value of...

[0073] W is a learnable parameter matrix; after normalizing the instances C1 and C2, the two matrices are combined by summing and averaging to obtain the soft correspondence matrix.

[0074]

[0075] IN stands for instance normalization.

[0076] To handle outliers, a slack variable is added to both the rows and columns of the soft correspondence matrix. The Sinkhorn algorithm is then used to alternately normalize the rows and columns of the soft correspondence matrix, making the sum of the row directions and the sum of the column directions approach 1, until the Sinkhorn algorithm converges, yielding the final soft correspondence matrix.

[0077] By employing this dual-constraint approach, a more balanced and robust matching score can be obtained, especially when the matching matrices of the two calculation methods differ significantly. Measuring the correspondence from two complementary directions—feature affinity and point-to-feature difference—helps to minimize misjudgments of outliers and improve overall matching accuracy.

[0078] Once the overlap uncertainty and soft correspondence matrix are obtained, the features learned by the network for each point are reweighted to highlight regions with low uncertainty. and correspondence matrix and overlapping uncertainty U S U Q As input, uncertainty-perceived features and The calculation formula is as follows:

[0079]

[0080] Where φ(·) is the multilayer perceptron. ⊙ represents matrix multiplication, while ⊙ represents element-wise multiplication.

[0081] To reduce the importance of features in uncertain regions, a feature random masking strategy based on uncertainty is adopted. This strategy directly incorporates uncertainty information into the training process, effectively guiding the network to focus on learning the features of uncertain points. It is worth noting that this random masking strategy will be disabled during the testing phase to ensure that the network makes inferences based on complete features in real-world applications.

[0082] The overlap-guided registration module is mainly used to predict the transformation parameters between the source point cloud and the target point cloud. Its structure is as follows: Figure 1 As shown on the right, this module uses features weighted by overlapping uncertainties. and As input, it undergoes a second feature interaction via a geometric transformer to obtain... and Subsequently, the path proceeds through two branches. One branch passes through the improved overlap point prediction module to predict the overlap score O. S With O Q (The overlap score represents the probability that the point lies in the overlapping region of two point clouds); the other branch predicts the hard correspondence matrix of the final point pair. It is a binary matrix, where an element with a value of 1 represents two corresponding points, and an element with a value of 0 represents two non-corresponding points.

[0083] In partially overlapping registration tasks, the presence of non-overlapping points can affect the estimation of point correspondences. Therefore, this invention needs to utilize the local similarity between two point clouds to determine the overlapping region, thereby improving the registration accuracy of subsequent point clouds. Specifically, this invention proposes an improved overlapping point prediction module (e.g., Figure 6 (As shown) to evaluate the overlap score of two point clouds. First, the input features are calculated. and The cosine similarity matrix between them is obtained by using the Proj module through steps such as one-dimensional convolution Conv1d, one-dimensional batch normalization BatchNorm1d, and ReLU activation. and Then, using the similarity matrix similarity, respectively with... and Interactive features are obtained by transmitting information across point clouds. and Original features Projection features and interaction features Concatenate along the channel dimension to extract the original features. Projection features and interaction features The points are concatenated along the channel dimension; then, the concatenated features are non-linearly transformed and fused using the first Conv module. The fused features are then passed through a second Conv module and activated with a sigmoid function to obtain the final prediction score O for the overlapping region of the source and target point clouds. S With O Q .

[0084] Hard correspondence matrix From input features and After obtaining the soft correspondence matrix through the second double-constraint matching module, it is calculated using the Linear Assignment Problem Solver (LAP).

[0085] When a hard correspondence is obtained and overlap score O S With O Q Then, the weighted singular value decomposition (SVD) algorithm can be used to solve for the rigid body transformation parameters between the source point cloud and the target point cloud, and the rotation matrix R and translation vector t between the two point clouds can be predicted.

[0086] Step S3: Train the point cloud registration model. The loss function used for training is consistent with that of the traditional UTOPIC model. Use the trained point cloud registration model for registration of partially overlapping point clouds.

[0087] Figure 7 This is a visualization of the point cloud registration effect of the present invention, where blue represents the source point cloud and yellow represents the target point. Specific implementation examples:

[0089] To verify the effectiveness of the method of this invention, it was compared with common point cloud registration methods (ICP, FGR, IDAM, MFGNet, ROPNet, RPMNet) on the ModelNet40 dataset (which contains 40 categories of point cloud data, including airplanes, cars, chairs, and people). Training was performed on the first 20 categories, and testing was conducted on the last 20 categories. This registration test on unseen data categories better reflects the model's generalization ability. The rotation matrix is ​​randomly sampled using three Euler angles within the range of [0°, 45°], and the translation vector is randomly sampled within the range of [-0.5, 0.5] on each axis. The target point cloud Q is obtained by applying the above-described random rigid transformation to the source point cloud S. To simulate partial overlap, a half-space with a random direction was generated, and two point clouds were cropped along this direction according to the cropping ratio, thus constructing point clouds to be registered with different overlapping regions. The cropping ratio was set to 70% in the experiment, and random noise sampled from the normal distribution N(0,0.01) with a size controlled within [-0.05,0.05] was added to the point clouds.

[0090] The proposed point cloud registration model was trained 200 times with a batch size of 8. The SGD optimizer was used with an initial learning rate of 0.001. The number of overlapping samples K was set to 50, the number of repetitions R in the geometric transformer was set to 3, and the number of nearest neighbors k was set to 20.

[0091] The network evaluation metrics used are: the anisotropic errors of the three-dimensional rotation matrix R and the three-dimensional translation vector t between two point clouds, including the root mean square error RMSE, the mean absolute error MAE, and the isotropic error Error.

[0092] As shown in Table 1, the method of this invention has the lowest values ​​for all evaluation indicators in the rotation matrix R and translation vector t, and the obtained transformation parameters are close to the actual transformation parameters. Furthermore, its registration effect on partially overlapping point clouds with abnormal noise is significantly better than other comparative methods. To further support the above conclusions, Figure 5 The point cloud registration visualization results of a portion of the objects in the dataset obtained using the method of this invention are presented.

[0093] Table 1. Registration results using different methods

[0094]

[0095] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. An improved UTOPIC method for registration of partially overlapping point clouds, characterized in that, The UTOPIC network with uncertainty awareness is used as the basic framework. A dual-branch feature extractor is introduced in the feature extraction part to capture the local feature information of the point cloud. In the uncertainty awareness feature enhancement part, a dual-constraint matching module is introduced to obtain the point correspondence matrix. In the overlap-guided registration part, an improved overlap point prediction module is introduced to identify the overlapping area of ​​the point cloud to be registered, and finally the point cloud registration is achieved.

2. The improved UTOPIC method for registration of partially overlapping point clouds according to claim 1, characterized in that, Includes the following steps: Step S1: Read point cloud data from the obtained point cloud dataset to obtain the three-dimensional coordinates of each point cloud sample. After downsampling, use random rotation and random translation operations to generate two point clouds with different initial poses and the true value of rigid transformation. Then, perform a cropping operation on the two point clouds to generate a partially overlapping source point cloud S and a target point cloud Q. Step S2: Construct an improved UTOPIC network model for point cloud registration, including a feature extraction module, an uncertainty-aware feature enhancement module, and an overlap-guided registration module. The feature extraction module captures local feature information of the point cloud, the uncertainty-aware feature enhancement module obtains the point correspondence matrix, and the overlap-guided registration module identifies the overlapping areas of the point cloud to be registered. Step S3: Train the point cloud registration model. The loss function used for training is consistent with the improved UTOPIC network model. Use the trained point cloud registration model for registration of partially overlapping point clouds.

3. The improved UTOPIC method for registration of partially overlapping point clouds according to claim 2, characterized in that, The feature extraction module utilizes a dual-branch structure consisting of adaptive graph convolution + attention mechanism and geometric edge convolution to extract features F of the source point cloud S. S Features F of the target point cloud Q Q .

4. The improved UTOPIC method for registration of partially overlapping point clouds according to claim 3, characterized in that, For extracting features F from the source point cloud S S The specific implementation method is as follows: The adaptive graph convolutional branch consists of four AdaptConv layers of sizes 64, 64, 128, and 256, strung together. The features output from each layer are concatenated to obtain a 512-dimensional output feature. For each AdaptConv layer, let the input source point cloud be... The corresponding features are Where N is the number of points contained in the source point cloud, and D is the feature dimension; a directed graph is constructed using kNN of each point's k nearest neighbors. in It is a set of points. Let x represent the set of edges; i The center point of the graph convolution is... It is the set of index points in its neighborhood; for the m-th channel of the M-dimensional features output by each layer of adaptive graph convolution, m = 1, 2, ..., M, the point feature (f) is used. i ,f j Dynamically generate an adaptive kernel: Where, Δf ij =cat[f i ,f j -f i ] represents the input features of the adaptive kernel, where cat[·,·] is the concatenation operation; g m (·) represents the feature mapping function; each adaptive kernel corresponds to a filter defined in AdaptConv; Using an adaptive kernel With Δx ij Perform convolution to obtain Where, point x i With point x j Spatial coordinate relationship between <·,·> represents the inner product of two vectors, and σ is a non-linear activation function; Stacking h for each channel ijm Forming connection point (x) i ,x j Edge features between ) Get point x i Local features within the neighborhood are Extending to the entire point cloud means Subsequently, an efficient local attention mechanism is used to enhance the expressive power of the extracted local features. Specifically, given input features... C, H, and W represent the feature channel dimension, height, and width, respectively. One-dimensional average pooling is performed on each channel in both the horizontal and vertical spatial directions to obtain the corresponding output features z_h and z_w. The output value of feature z_h at the c-th channel at the corresponding height h is... Represented as: Where, x c (h,i w ) represents the input feature X at the corresponding position (h, i) w The eigenvalue of the c-th channel; The output of the c-th channel of feature z_w at the corresponding width w express: Where, x c (j h (j, w) represents the input feature X at the corresponding position (j) h The eigenvalues ​​of the c-th channel (w) are: One-dimensional convolution is applied to enhance the horizontal and vertical positional information, followed by group normalization operation G_n to process the enhanced positional information, resulting in positional attention representations in the horizontal and vertical directions: Where F_h and F_w are one-dimensional convolutions; We obtain two attention representations y h and y w Then, the input features are weighted using the following formula: Y=x c ×y h ×y w Where, x c The input feature X is the feature value in the c-th channel; Finally, by using point x i The center point x is obtained by applying the max pooling function to the edge features of the neighborhood. i After extracting features layer by layer using 4 layers of AdaptConv, the output features of each layer are concatenated along the channel dimension to obtain the features extracted by the adaptive graph convolution branch. The geometric edge convolution branch consists of four GEConv layers with sizes of 64, 64, 128, and 256. The outputs of each layer after Maxpooling are concatenated to obtain a 512-dimensional output feature. kNN is used for neighborhood search to determine the local neighborhood around each center point; for center point u i Each point u in the local neighborhood j For j = 1, 2, ..., k, calculate the geometric features, which are obtained by calculating the geometric features of point u. j With center point u i and the average value m of all points in the neighborhood i The distances between them are obtained using the Euclidean and cosine distances; u j geometric features It consists of five parts, including two Euclidean distances. and three angles The cosine values ​​constitute the geometric features of the extracted local neighborhood, defined as follows: Subsequently, using the geometric feature set X g To construct the neighborhood geometric edge set By connecting the geometric features of each point in the neighborhood Geometric characteristic mean of all points in the neighborhood The geometric edges are constructed using the following formula: The feature edges within the neighborhood are generated using the following formula, and the set of feature edges is represented as follows: Since the first layer does not exist, therefore It is an empty set The final edge information is obtained by concatenating the geometric edges with the feature edges. Right now The features of each edge are extracted using the above formula, and then the features are upscaled using a shared multilayer perceptron (MLP) to raise the features of each edge to a higher-dimensional feature space. This process is denoted as γ. Then, nonlinear activation is performed using σ, i.e. The output features of the final center point are obtained by aggregating using the Maxpool maximum pooling function. The output of GEConv in the first layer is described by the following formula, since there is no previous layer, there are no features of the previous layer; For subsequent layers, the features from the previous layer are used to calculate the features of the current layer, as shown in the following formula: Finally, the features extracted layer by layer are concatenated along the channel dimension to obtain the features extracted by the geometric edge convolution branch. Features extracted by convolutional branches of splicing adaptive graphs Features extracted by geometric edge convolution branches Source point cloud features extracted from two branches are obtained. For target point cloud Feature extraction yields target point cloud features. J represents the number of points contained in the target point cloud.

5. An improved UTOPIC method for registration of partially overlapping point clouds according to claim 4, characterized in that, In the feature extraction module, after the point cloud features are extracted, an additional point cloud completion decoder from the point cloud completion network PCN is added, and the corresponding completion loss function is also added to the overall network loss function. The point cloud completion decoder is only used during the training phase and removed during the testing phase.

6. An improved UTOPIC method for registration of partially overlapping point clouds according to claim 2, characterized in that, The uncertainty-aware feature enhancement module includes a geometric transformer module, an uncertainty quantization module, a dual-constraint matching module, and an uncertainty-aware weighting module; wherein... The geometry transformer module consists of three parts: geometric relation embedding, geometric self-attention, and cross-attention. The geometric relationship embedding part encodes the distance and angle between points, given two points p. i and p j Their geometric relationship can be described as follows: (1) Point spacing: The point spacing is the distance between points p in Euclidean space. i and point p j The distance ρ between i,j =||p i -p j ||2; (2) Angle between points: Search point p i two nearest neighbors and Then calculate the vector. and The sum of, is counted as Then the included angle α i,j Through calculation get; (3) Difference in the perimeter of the triangle: Search point p i two nearest neighbors and The three points together form a triangle Search point p j two nearest neighbors and The three points together form a triangle The difference in perimeter between two triangles is defined as and The final geometric relationship is obtained by combining the point spacing, point angle, and triangle perimeter difference, and embedding them into g. i,j As shown in the following formula: g i,j =cat[ρ i,j ,a i,j ,or i,j ] The geometric self-attention component precisely captures the inherent structure of a single point cloud by leveraging embedded geometric relationships, and extracts contextual information within the single point cloud using the internal spatial structure and characteristics of the data. The cross-attention component facilitates feature interaction and fusion between the source and target point clouds. The geometric self-attention component and the cross-attention component alternately iterate R times, ultimately generating attention features that contain global context. The uncertainty quantization module utilizes a probability-based overlapping uncertainty quantization scheme to generate more discriminative features; specifically, for the source point cloud S, a Gaussian distribution is assigned to each point, with a mean μ. i and variance σ i Parameterization, these parameters are... The data is derived from predictions made by different Multilayer Perceptron (MLP) layers, and then each point s in the source point cloud is sampled from the learned distribution. i The overlap fraction, i.e., o i ~N(μ) i ,σ i The sampling operation consists of two steps: first, a random sample, ε, is obtained from the standard Gaussian distribution N(0,I). i ~N(0,I), and through μ i +ε i σ i Obtain the desired sample; then, resample the overlap score for each point K times from the learned distribution, denoted as O. S ={o (1) ,...,o (K) }; to O S Treat it as an empirical sample from an approximate predicted distribution, and measure the uncertainty by calculating the variance: U S =norm(var(O S )) in It represents the overlap uncertainty, norm(·) is the minimum-maximum normalization, and var(·) is used to calculate the variance; Simultaneously, determine the overlap uncertainty U of the target point cloud Q. Q ; The dual-constraint matching module generates a soft correspondence matrix based on the features of the input source and target point clouds. Calculate the Euclidean distance matrix C1 between the source point cloud features and the target point cloud features. The element at the i-th row and j-th column of matrix C1 is... They are respectively and The i-th and j-th feature points are identified; the affinity matrix C2 between the features is calculated, where the element at the i-th row and j-th column of matrix C2 has a value of [value missing]. W is a learnable parameter matrix; after normalizing the instances C1 and C2, the two matrices are combined by summing and averaging to obtain the soft correspondence matrix. Where IN is instance normalization; The uncertainty-aware weighting module, after obtaining the overlapping uncertainty and soft correspondence matrix, reweights the features learned by the network for each point to highlight regions of low uncertainty; using features... Soft correspondence matrix and overlapping uncertainty U S U Q As input, uncertainty-perceived features and The calculation formula is as follows: Where φ(·) is the multilayer perceptron. ⊙ represents matrix multiplication, while ⊙ represents element-wise multiplication.

7. An improved UTOPIC method for registration of partially overlapping point clouds according to claim 6, characterized in that, To handle outliers, a slack variable is added to both the rows and columns of the soft correspondence matrix. The Sinkhorn algorithm is then used to alternately normalize the rows and columns of the soft correspondence matrix, making the sum of the rows and the sum of the columns approach 1, until the Sinkhorn algorithm converges, yielding the final soft correspondence matrix.

8. An improved UTOPIC method for registration of partially overlapping point clouds according to claim 6, characterized in that, To reduce the importance of features in uncertain regions, a feature random masking strategy based on uncertainty is adopted, which directly incorporates uncertainty information into the training process. During the testing phase, the random masking strategy is disabled.

9. An improved UTOPIC method for registration of partially overlapping point clouds according to claim 2, characterized in that, The overlapping guided registration module uses the overlapping uncertainty-weighted features obtained after processing by the uncertainty-aware feature enhancement module. and As input, it undergoes a second feature interaction via a geometric transformer to obtain... and Subsequently, the path proceeds through two branches. One branch passes through the improved overlap point prediction module to predict the overlap score O. S With O Q The overlap score represents the probability that the corresponding point lies in the overlapping region of two point clouds; the other branch predicts the hard correspondence matrix of the final point pair. It is a binary matrix, where an element with a value of 1 represents two corresponding points, and an element with a value of 0 represents two non-corresponding points.

10. An improved UTOPIC method for registration of partially overlapping point clouds according to claim 9, characterized in that, The improved overlap point prediction module first calculates the input features. and The cosine similarity matrix between them is obtained using the Proj module through steps including one-dimensional convolution Conv1d, one-dimensional batch normalization BatchNorm1d, and ReLU activation. and Then, using the similarity matrix similarity, respectively with... and Interactive features are obtained by transmitting information across point clouds. and Will and splicing along the channel dimension, and The points are concatenated along the channel dimension; then, the concatenated features are non-linearly transformed and fused using the first Conv module. The fused features are then passed through a second Conv module and activated with a sigmoid function to obtain the final prediction score O for the overlapping region of the source and target point clouds. S With O Q ; Hard correspondence matrix From input features and After obtaining the soft correspondence matrix through the double-constraint matching module, it is calculated using the linear assignment problem solver LAP. When a hard correspondence is obtained and overlap score O S With O Q Then, the weighted singular value decomposition (SVD) algorithm is used to solve for the rigid body transformation parameters between the source point cloud and the target point cloud, and the rotation matrix R and translation vector t between the two point clouds are predicted.

Images