A point cloud registration method for low overlap rate

Abstract

The application provides a point cloud registration method for a low overlap rate, relates to the technical field of three-dimensional point cloud registration, and comprises the following steps: collecting source point cloud and target point cloud; performing feature extraction and downsampling on the source point cloud and the target point cloud; performing position coding on the downsampled point cloud in a point cloud neighborhood range; adding the position coding result to the features; combining an attention module and optimal transport theory to improve the probability that homonymic point pairs of the point cloud are located in an overlapping part of the point cloud, and to improve point cloud registration precision; based on the optimal transport of a similarity matrix, missing corresponding points can be screened out from key points, and a conversion matrix can be directly solved through a credible point pair set constructed by a score mechanism and a spatial compatibility principle, so that the time spent on point cloud registration is shortened.

Description

Technical Field

[0001] This invention relates to the technical field of three-dimensional point cloud registration, and more specifically, to a point cloud registration method for low overlap rates. Background Technology

[0002] A 3D point cloud is a collection of spatial points in a three-dimensional coordinate system, reflecting the spatial distribution and surface features of a 3D object. Depending on the acquisition device, point clouds may also contain other information besides location information, such as RGB and laser reflection intensity information.

[0003] Point cloud registration refers to the process of estimating a transformation matrix, given a source point cloud P and a target point cloud Q, so that P can be transformed to the coordinate system of Q. Traditional point cloud registration methods include ICP and NDT. The ICP method iteratively calculates a transformation matrix such that each point p in point cloud P is transformed to the coordinate system of Q. i After transformation, it is compared with the point cloud q. i The average distance is minimized. While this method offers high accuracy, it requires a good initial coarse registration value and is prone to getting trapped in local optima, making it suitable only for point cloud registration with high overlap. The NDT method uses normal distribution matching for point cloud registration, offering the advantage of pre-calculating the distribution of each point cloud, thus saving registration time. However, it is sensitive to the density of the point cloud itself and cannot be used for point cloud registration with low overlap.

[0004] In recent years, with the development of deep learning methods, many scholars have used deep neural networks to solve the point cloud registration problem. Currently, common deep learning-based point cloud registration methods include direct methods and feature matching methods. The direct method takes the point cloud pair to be registered as input and directly predicts the transformation matrix as output through a deep neural network. The problem with this method is its poor generalization ability and weak interpretability. The feature matching method, on the other hand, registers point clouds based on the following process: 1. Find key points in two point clouds through network learning. 2. Extract descriptors for the key points. 3. Calculate the corresponding point pairs based on the differences between the descriptors. 4. Calculate the transformation matrix using estimation methods such as RANSAC. For example, a low-overlap point cloud registration method is disclosed in the prior art. Addressing the difficulty of searching for identical point pairs in low-overlap scenarios, it uses a self-attention mechanism to achieve overall perception of the point cloud by aggregating point pairs. Simultaneously, it uses a cross-attention mechanism to explicitly mine overlapping region information, predicting the confidence of all points in the overlapping region. Probabilistic selection is used to sample point pairs in the overlapping region during the matching stage, improving the registration recall rate. Simultaneously, dynamically limiting the receptive field of the convolutional kernel to the overlapping region avoids the extraction of invalid geometric neighborhood information, improving the accuracy and precision of point-by-point features. In this method, outlier points are filtered using the RANSAC method, but this requires multiple iterations, resulting in high time costs and failing to accurately find corresponding keypoint pairs. Under low overlap conditions, most keypoints do not have corresponding points in another point cloud. Furthermore, many methods' proposed features lack good spatial rotation and translation invariance properties; when the spatial position of the input point cloud changes, they cannot provide correct matching point pairs, leading to increased errors and low point cloud registration accuracy. Summary of the Invention

[0005] To address the challenges of registering point clouds with low overlap rates, as well as the issues of long registration times and low accuracy, this invention proposes a point cloud registration method oriented towards low overlap rates. Before point cloud registration, the method leverages the strong expressive power of point cloud feature extraction. During the registration process, noise and outlier interference are minimized. Finally, reliable points are selected for direct registration, avoiding the drawbacks of long registration times and improving the accuracy of point cloud registration.

[0006] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows:

[0007] A point cloud registration method for low overlap rates, the method comprising the following steps:

[0008] S1. Acquire source point cloud data and target point cloud data;

[0009] S2. After downsampling the source point cloud and the target point cloud to the same density, normalize the point cloud, and then perform spatial convolution on the point cloud to extract features, thereby obtaining the downsampled point cloud and the features corresponding to each point cloud.

[0010] S3. Perform position encoding on the downsampled point cloud within the neighborhood of the point cloud, add the position encoding result to the feature, input it into the attention module, and output the features corresponding to the source point cloud and the target point cloud;

[0011] S4. Take the inner product of the features obtained in S3 to obtain the similarity matrix. The elements of the similarity matrix are the inner product of the corresponding point features of the source point cloud and the target point cloud. After expanding the similarity matrix, solve for the optimal transmission of the similarity matrix to obtain the fraction matrix.

[0012] S5. Determine the set of corresponding point pairs between the source point cloud and the target point cloud from the fraction matrix, and further filter out multiple credible sets of corresponding point pairs from the set of corresponding point pairs through spatial compatibility.

[0013] S6. For each set of trusted point pairs, determine the points with scores greater than the score threshold in the set of trusted point pairs by using the score threshold. Then, based on such points, obtain the transformation matrix from the source point cloud to the target point cloud to complete the point cloud registration.

[0014] Preferably, in step S2, after downsampling the source point cloud and the target point cloud to the same density and normalizing the point cloud, spatial convolution is performed on the point cloud to extract features, and the process of obtaining the downsampled point cloud and the features corresponding to each point cloud is as follows:

[0015] S21. Merge the source point cloud and the target point cloud into a single point cloud W, and record whether each point on point cloud W belongs to the source point cloud or the target point cloud;

[0016] S22. Downsample the source point cloud and the target point cloud to the same density, then normalize the point cloud coordinates, introduce a spatial convolutional network, perform feature extraction, and output the downsampled point cloud and the features corresponding to each point cloud.

[0017] S23. Based on whether each point recorded in S21 belongs to the source point cloud or the target point cloud, separate the point cloud output in S22 into downsampled source point clouds. and their corresponding features Target point cloud after downsampling and their corresponding features ,in, x This represents the source point cloud before downsampling. y Let the target point cloud before downsampling be denoted as the source point cloud. x If N points remain after downsampling, then It is an N*3 matrix structure, where "3" represents three-dimensional coordinates. It is an N*d data structure, where d represents the length of the feature vector, which varies depending on the spatial convolutional network.

[0018] Here, feature extraction and downsampling are performed on the input point cloud. First, it reduces the number of points in the point cloud, thereby alleviating the computational pressure in the subsequent point cloud registration process. Second, it adjusts the point cloud density of the source point cloud and the target point cloud to the same level through downsampling, which helps to find corresponding point pairs in the subsequent process.

[0019] Preferably, the spatial convolutional network is a publicly available point cloud feature extraction network.

[0020] Preferably, the method for performing position encoding on the downsampled point cloud within its neighborhood in step S3 is as follows:

[0021]

[0022]

[0023] in, Indicates the midpoint of the downsampled point cloud. The set of points within a neighborhood. Indicates the length of each point feature; Point All point-to-point connections within the neighborhood The distance L2 norm; the positional encoding is added to the feature to obtain .

[0024] Here, when this scheme adopts position encoding, it does not encode the position of the point in the entire point cloud, but encodes it within the neighborhood of the point cloud. Since the point cloud is unordered, directly encoding the position of the point in the point cloud cannot provide effective information for point cloud registration. However, encoding the information within the neighborhood of the point cloud can better obtain the relative position information of the point.

[0025] Preferably, the attention module includes a self-attention mechanism network and a cross-attention mechanism network connected in sequence. The self-attention mechanism network enables each point to interact with all points, and the cross-attention mechanism network searches for structural similarities between the source point cloud and the target point cloud, making the features of the overlapping region of the two point clouds more apparent. The self-attention mechanism process is as follows:

[0026]

[0027]

[0028]

[0029] in, This represents the parameters to be learned in the network. The concatenation operation, for the source point cloud, is implemented in a self-attention mechanism network that satisfies... The implementation process of the cross-attention mechanism network satisfies For the target point cloud, the implementation process in the self-attention mechanism network satisfies... The implementation process of the cross-attention mechanism network satisfies After passing through multiple self-attention and cross-attention network mechanisms, features are obtained. .

[0030] Here, compared to spatial convolution, which can only focus on information around neighboring points, the self-attention mechanism network allows each point to interact with all other points, greatly expanding the receptive field of feature extraction. The cross-attention mechanism network, on the other hand, is used to find structurally similar areas within the source and target point clouds, making the features of the overlapping regions of the two point clouds more prominent.

[0031] Preferably, the features obtained in S3 After performing the inner product, we obtain the similarity matrix. ;feature The inner product of two points can represent the similarity of their features, which can be considered as a rough score of the corresponding points. Therefore, the similarity matrix... elements Source point cloud Point cloud of points and target The scores of each pair of points with the same name are used to obtain the similarity matrix. Next, the similarity matrix is ​​expanded row by row and column by column; then, the Sinkhorn algorithm is used to solve for the similarity matrix. The optimal transmission is obtained to obtain the fractional matrix.

[0032] Preferably, in step S5, the process of determining the set of corresponding point pairs between the source point cloud and the target point cloud from the fraction matrix is ​​as follows:

[0033] From the fractional matrix, select the element whose row and column are both maximum values. Use the row and column numbers corresponding to this element as the indices of the corresponding point pairs in the source and target point clouds. Denote the set of corresponding point pairs as... The value of this element is used as the score for pairs of points with the same name. Let be the score of the pair of points with the same name. ,in, Representing pairs of points with the same name Given the coordinates of the source and target point clouds, a set of reliable point pairs is selected from the set of identical point pairs using the principle of spatial compatibility. If... For true pairs of identical points, since rigid body transformations do not change the shape of the object itself, the distance between two identical points within the point cloud remains unchanged before and after the transformation. The threshold value must be 0. Considering the noise interference in the point cloud, a threshold value is set. ,when hour, and It is spatially compatible. Here, It is a very small parameter that is set according to the point cloud density and noise level.

[0034] Preferably, when selecting a set of reliable point pairs from the set of pairs of identical points using the principle of spatial compatibility, the selection process is enhanced using second-order spatial compatibility, wherein the first-order spatial compatibility matrix... The construction is as follows:

[0035]

[0036] Second-order spatial compatibility matrix The construction is as follows:

[0037] in, Represents the matrix dot product operation. Represents the matrix cross product operation; Representation of point pairs and point pairs The number of other point pairs that simultaneously satisfy spatial compatibility; after obtaining the second-order spatial compatibility matrix, each time select a point pair with the largest score, and then find the top pair in the second-order compatibility matrix. Each pair of points is considered as a set of trustworthy pairs, thus obtaining the set of trustworthy pairs. .

[0038] Here, we consider that even incorrect corresponding points may exist. In smaller cases, when there are too many outlier pairs, this method has a low probability of filtering out incorrect outlier pairs. The introduction of second-order spatial compatibility enhances the filtering process and increases the probability of filtering out incorrect outlier pairs.

[0039] Preferably, when selecting a set of trustworthy point pairs from the set of identical point pairs, a specified threshold or a deep learning-based method is used for filtering.

[0040] Preferably, the transformation matrix from source point cloud to target point cloud The expression is:

[0041]

[0042] in, This refers to the final pair of identically named nodes that meet the criteria. For the points corresponding to the scores, R This represents a 3x3 rotation matrix in three-dimensional space. t Represents a three-dimensional translation vector, with a size of 1*3, obtained through... R , t The ability to represent any translation or rotation in three-dimensional space is crucial; the ultimate goal of point cloud registration is to find a set of... This minimizes the average distance between the new point cloud obtained by acting on the source point cloud and the target point cloud. Then, the transformation matrix is ​​solved using the weighted singular value decomposition algorithm. The expression is used to obtain a transformation matrix for each set of reliable point pairs. These transformation matrices are traversed, and the transformation matrix with the best effect is selected as the registration result from the source point cloud to the target point cloud.

[0043] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0044] This invention proposes a point cloud registration method for low overlap rates. First, source and target point clouds are acquired. Then, features are extracted and downsampled from both source and target point clouds. The downsampled point clouds are then encoded within their neighborhoods to enhance feature representation. The position encoding results are added to the features. Next, an attention module and optimal transfer are combined to ensure that corresponding point pairs are located in the overlapping parts of the point clouds, thereby improving the point cloud registration accuracy. Based on the optimal transfer of the similarity matrix, not only can matching point pairs be found, but key points with missing corresponding points can also be filtered out. Through a scoring mechanism, sufficiently reliable point pairs are selected for direct registration, reducing the time spent on point cloud registration. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the point cloud registration method for low overlap proposed in Embodiment 1 of the present invention.

[0046] Figure 2 This diagram illustrates the network structure for implementing the point cloud registration method proposed in Embodiment 2 of the present invention.

[0047] Figure 3 A schematic diagram illustrating the spatial compatibility principle proposed in Embodiment 2 of the present invention;

[0048] Figure 4 This diagram illustrates the first-order and second-order spatial compatibility matrices proposed in Embodiment 2 of the present invention. Detailed Implementation

[0049] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0050] To better illustrate this embodiment, some parts of the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions;

[0051] It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings.

[0052] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0053] The positional relationships depicted in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0054] Example 1

[0055] This embodiment addresses the following problems existing in current 3D point cloud registration methods and proposes a point cloud registration method for low overlap rates.

[0056] (1) Poor performance of point cloud registration with low overlap rate. Many learning-based 3D point cloud registration methods select corresponding point pairs based on the similarity between features after feature extraction, and then use the RANSAC method to filter out suitable point pairs for estimation. This method cannot guarantee that the selected corresponding point pairs are in the overlapping area, and the error rate is high, resulting in poor performance.

[0057] (2) Most current point cloud feature descriptors are constructed using neighborhood information of points. Learning-based methods generally use 3D convolution to extract features, while traditional point cloud features, such as FPFH feature descriptors, only consider neighborhood information of the point cloud. These methods do not make full use of the overall information of the point cloud and the information between two point clouds, which limits the registration effect.

[0058] (3) Most point cloud registration methods require the use of the RANSAC method to find the optimal transformation matrix through random sampling, which takes a long time and is only suitable for cases where the accuracy of identical point pairs is high.

[0059] See Figure 1 The method includes the following steps:

[0060] S1. Acquire source point cloud data and target point cloud data; In this embodiment, a dedicated point cloud collection device is used to collect source point cloud data and target point cloud data for specific application scenarios, and preprocessing is required when necessary.

[0061] S2. After downsampling the source and target point clouds to the same density, normalize the point clouds, and then perform spatial convolution on the point clouds to extract features, obtaining the downsampled point clouds and the features corresponding to each point cloud; the process corresponding to this step is as follows:

[0062] S21. Merge the source point cloud and the target point cloud into a single point cloud W, and record whether each point on point cloud W belongs to the source point cloud or the target point cloud;

[0063] S22. Downsample the source point cloud and the target point cloud to the same density, then normalize the point cloud coordinates, introduce a spatial convolutional network, perform feature extraction, and output the downsampled point cloud and the features corresponding to each point cloud; in this embodiment, the spatial convolutional network is a publicly available point cloud feature extraction network, such as the backbone network of KPConv, PointNet++, etc.

[0064] S23. Based on whether each point recorded in S21 belongs to the source point cloud or the target point cloud, separate the point cloud output in S22 into downsampled source point clouds. and their corresponding features Target point cloud after downsampling and their corresponding features ,in, x This represents the source point cloud before downsampling. y Let the target point cloud before downsampling be denoted as the source point cloud. x If N points remain after downsampling, then It is an N*3 matrix structure, where "3" represents three-dimensional coordinates. It is an N*d data structure, where d represents the length of the feature vector, which varies depending on the spatial convolutional network.

[0065] S3. Perform position encoding on the downsampled point cloud within the neighborhood of the point cloud, add the position encoding result to the feature, input it into the attention module, and output the features corresponding to the source point cloud and the target point cloud;

[0066] In this embodiment, the attention module includes a self-attention mechanism network and a cross-attention mechanism network connected in sequence. In specific implementation, multiple self-attention mechanism and cross-attention mechanism modules are used in an overlapping manner, which makes full use of the information of the point cloud itself, as well as the information of the source point cloud and the target point cloud. The extracted features are more effective than other feature sub-description methods.

[0067] S4. Perform an inner product of the features obtained in S3 to obtain a similarity matrix. The elements of the similarity matrix are the scores of the same point pairs between the source point cloud and the target point cloud. After expanding the similarity matrix, solve for the optimal transmission of the similarity matrix to obtain the score matrix.

[0068] S5. Determine the set of corresponding point pairs between the source point cloud and the target point cloud from the fraction matrix, and further filter out multiple credible sets of corresponding point pairs from the set of corresponding point pairs through spatial compatibility.

[0069] S6. For each set of trusted point pairs, determine the points with scores greater than the score threshold in the set of trusted point pairs by using the score threshold. Then, based on such points, obtain the transformation matrix from the source point cloud to the target point cloud to complete the point cloud registration.

[0070] Through processes S4-S6, after performing inner product of the features, the optimal transfer problem is solved to make the source and target point clouds more consistent in probability distribution, thus finding corresponding point pairs more accurately. Furthermore, since corresponding point pairs have high accuracy, some outlier point pairs can be filtered out using the spatial compatibility principle. Therefore, the point pairs obtained by this method are more accurate than those obtained by other methods, with minimal discrepancies. After finding the corresponding point pairs, because the point pair relationships are relatively accurate, it is not necessary to use the RANSAC method to select interior point estimation transformation matrices. Instead, multiple point pair sets are selected based on spatial compatibility properties, and then the optimal transformation matrix is ​​directly calculated using the singular value decomposition method's minimum sphere constraint problem, significantly reducing the time spent on point cloud registration.

[0071] Example 2

[0072] The network structure diagram corresponding to the point cloud registration method proposed in Example 1 can be found in [link to example]. Figure 2 ,pass Figure 2 It can be seen that these inventions are mainly divided into three parts. The first part is to extract features and downsample the source point cloud and the target point cloud through the same network. The second part is to perform self-attention and cross-attention on the features to obtain the corresponding point pair relationship. The third part is to solve the transformation matrix through the corresponding point pair.

[0073] In practice, the first part involves downsampling and feature extraction of the source and target point clouds. This is done for two reasons: first, to reduce the number of points in the point cloud, thereby reducing the computational burden in the subsequent point cloud registration process; and second, to adjust the point cloud density of the source and target point clouds to the same level through downsampling, which helps to find corresponding point pairs in the subsequent process.

[0074] In the second part of the implementation, the first step is to perform positional encoding on the features. This part does not encode the point's position within the entire point cloud, but rather within its neighborhood. Because point clouds are unordered, directly encoding the point's position within the cloud does not provide effective information for point cloud registration. Encoding the information within the point cloud's neighborhood, however, allows for better acquisition of the point's relative position information. The method for performing positional encoding on the downsampled point cloud within its neighborhood is as follows:

[0075]

[0076]

[0077] in, Indicates the midpoint of the downsampled point cloud. The set of points within a neighborhood. Indicates the length of each point feature; Point All point-to-point connections within the neighborhood The distance L2 norm; the positional encoding is added to the feature to obtain .

[0078] like Figure 2 As shown, the proposed attention module includes a self-attention mechanism network and a cross-attention mechanism network connected sequentially. The self-attention mechanism network enables each point to interact with all points, while the cross-attention mechanism network searches for structural similarities between the source and target point clouds, making the features of the overlapping regions of the two point clouds more prominent. The process of the self-attention mechanism is as follows:

[0079]

[0080]

[0081]

[0082] in, This represents the parameters to be learned in the network. The concatenation operation, for the source point cloud, is implemented in a self-attention mechanism network that satisfies... The implementation process of the cross-attention mechanism network satisfies For the target point cloud, the implementation process in the self-attention mechanism network satisfies... The implementation process of the cross-attention mechanism network satisfies After passing through multiple self-attention and cross-attention network mechanisms, features are obtained. .

[0083] The features obtained from S3 After performing the inner product, we obtain the similarity matrix. ;feature The inner product of two points can represent the similarity of their features, which can be considered as a rough score of the corresponding points. Therefore, the similarity matrix... elements Source point cloud Point cloud of points and target The scores of each pair of points with the same name are used to obtain the similarity matrix. Next, the similarity matrix is ​​expanded row by row and column by column; then, the Sinkhorn algorithm is used to solve for the similarity matrix. The optimal transmission is obtained to obtain the fraction matrix;

[0084] In practical implementation, let the size of the matrix be... ,in ， These represent the point clouds after downsampling. The number of points, for Expand the matrix by 1 row and 1 column , to make it into, size The extra row and column are used to store points for which no matching point was found; then, the Sinkhorn algorithm is used to solve for the similarity matrix. The optimal transmission yields the fraction matrix. , fraction matrix Delete the (m+1)th row and (n+1)th column to obtain the fraction matrix. .

[0085] In step S5, the process of determining the set of corresponding point pairs between the source point cloud and the target point cloud from the fraction matrix is ​​as follows:

[0086] From the fractional matrix, select the element whose row and column are both maximum values. Use the row and column numbers corresponding to this element as the indices of the corresponding point pairs in the source and target point clouds. Denote the set of corresponding point pairs as... The value of this element is used as the score for pairs of points with the same name. Let be the score of the pair of points with the same name. ,in, Representing pairs of points with the same name Given the coordinates of the source and target point clouds, a set of reliable point pairs is selected from the set of identical point pairs using the principle of spatial compatibility. If... For true pairs of identical points, since rigid body transformations do not change the shape of the object itself, the distance between two identical points within the point cloud remains unchanged before and after the transformation. The threshold value must be 0. Considering the noise interference in the point cloud, a threshold value is set. ,when hour, and It is spatially compatible. Here, It is a very small parameter that is set according to the point cloud density and noise level.

[0087] For a spatial compatibility diagram, please refer to Figure 3 ,exist Figure 3 In the diagram, c1, c2, and c3 are correctly matched point pairs, while c4 is an abnormally matched point pair. and As the distance between two normal point pairs, it can be known from the properties of rigid body transformation. And the distance between abnormal point pairs and normal point pairs, and They are not necessarily equal. This diagram only illustrates that the distance remains constant. In rigid body transformations, apart from distance, information such as angles and normal vectors also remain unchanged. Therefore, this information can also be used for filtering. When the proportion of correct point pairs extracted by our network model structure is high, we can directly filter out incorrect point pairs using the principle of spatial compatibility.

[0088] Depend on Figure 3 It can be seen that even incorrect corresponding points may exist. In smaller cases, when there are too many outlier pairs, this method has a low success rate in filtering out erroneous outliers. To address this issue, such as... Figure 4 As shown, the screening process can be enhanced using second-order spatial compatibility. When using the principle of spatial compatibility to select a set of reliable point pairs from a set of pairs of identical points, the screening process is enhanced using second-order spatial compatibility. The first-order spatial compatibility matrix... The construction is as follows:

[0089]

[0090] Second-order spatial compatibility matrix The construction is as follows:

[0091] in, Represents the matrix dot product operation. Represents the matrix cross product operation; Representation of point pairs and point pairs The number of other point pairs that simultaneously satisfy spatial compatibility; after obtaining the second-order spatial compatibility matrix, each time select a point pair with the largest score, and then find the top pair in the second-order compatibility matrix. Each pair of points is considered as a set of trustworthy pairs, thus obtaining the set of trustworthy pairs. .

[0092] Finally, based on these point pairs and a score threshold, the point pairs with scores greater than the threshold are selected, and the transformation matrix from the source point cloud to the target point cloud is calculated. The expression is:

[0093]

[0094] in, This refers to the final pair of identically named nodes that meet the criteria. For the points corresponding to the scores, R This represents a 3x3 rotation matrix in three-dimensional space. t Represents a three-dimensional translation vector, with a size of 1*3, obtained through... R , t The ability to represent any translation or rotation in three-dimensional space is crucial; the ultimate goal of point cloud registration is to find a set of... This minimizes the average distance between the new point cloud obtained by acting on the source point cloud and the target point cloud. Then, the transformation matrix is ​​solved using the weighted singular value decomposition algorithm. The expression is used to obtain a transformation matrix for each set of reliable point pairs. These transformation matrices are traversed, and the transformation matrix with the best effect is selected as the registration result from the source point cloud to the target point cloud.

[0095] Example 3

[0096] Consistent with the basic implementation in Example 2, when filtering out a set of trustworthy point pairs from the set of identical point pairs, a specified threshold or a deep learning-based filtering method can also be used. Deep learning-based filtering methods include models such as PointDSC.

[0097] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A point cloud registration method for low overlap rates, characterized in that, The method includes the following steps: S1. Acquire source point cloud data and target point cloud data; S2. After downsampling the source point cloud and the target point cloud to the same density, normalize the point cloud, and then perform spatial convolution on the point cloud to extract features, thereby obtaining the downsampled point cloud and the features corresponding to each point cloud. S3. Perform position encoding on the downsampled point cloud within the neighborhood of the point cloud, add the position encoding result to the feature, input it into the attention module, and output the features corresponding to the source point cloud and the target point cloud; The method for encoding the location of the downsampled point cloud within its neighborhood as described in step S3 is as follows: in, Indicates the midpoint of the downsampled point cloud. The set of points within a neighborhood. Indicates the length of each point feature; Point All point-to-point connections within the neighborhood The distance L2 norm; the positional encoding is added to the feature to obtain ; S4. Take the inner product of the features obtained in S3 to obtain the similarity matrix. The elements of the similarity matrix are the inner product of the corresponding point features of the source point cloud and the target point cloud. After expanding the similarity matrix, solve for the optimal transmission of the similarity matrix to obtain the fraction matrix. The features obtained from S3 After performing the inner product, we obtain the similarity matrix. ;feature The inner product of two points can represent the similarity of their features, which can be considered as a rough score of the corresponding points. Therefore, the similarity matrix... elements Source point cloud Point cloud of points and target The scores of each pair of points with the same name are used to obtain the similarity matrix. Next, the similarity matrix is ​​expanded row by row and column by column; then, the Sinkhorn algorithm is used to solve for the similarity matrix. The optimal transmission is obtained to obtain the fraction matrix; S5. Determine the set of corresponding point pairs between the source point cloud and the target point cloud from the fraction matrix, and further filter out multiple credible sets of corresponding point pairs from the set of corresponding point pairs through spatial compatibility. In step S5, the process of determining the set of corresponding point pairs between the source point cloud and the target point cloud from the fraction matrix is ​​as follows: From the fractional matrix, select the element whose row and column are both maximum values. Use the row and column numbers corresponding to this element as the indices of the corresponding point pairs in the source and target point clouds. Denote the set of corresponding point pairs as... The element value serves as the score for pairs of points with the same name; denoted as... ,in, Representing pairs of points with the same name Given the coordinates of the source and target point clouds, a set of reliable point pairs is selected from the set of identical point pairs using the principle of spatial compatibility; if For true pairs of identical points, since rigid body transformations do not change the shape of the object itself, the distance between two identical points within the point cloud remains unchanged before and after the transformation. The threshold value must be 0. Considering the noise interference in the point cloud, a threshold value is set. ,when hour, and It is spatially compatible; here, It is a very small parameter that is set based on the point cloud density and noise levels; S6. For each set of trusted point pairs, determine the points with scores greater than the score threshold in the set of trusted point pairs by using the score threshold. Then, based on such points, obtain the transformation matrix from the source point cloud to the target point cloud to complete the point cloud registration.

2. The point cloud registration method for low overlap rates according to claim 1, characterized in that, In step S2, the source and target point clouds are downsampled to the same density and then normalized. Then, spatial convolution is performed on the point clouds to extract features, resulting in the downsampled point clouds and the features corresponding to each point cloud. S21. Combine the source point cloud and the target point cloud into a single point cloud W, and record whether each point on point cloud W belongs to the source point cloud or the target point cloud; S22. Downsample the source point cloud and the target point cloud to the same density, then normalize the point cloud coordinates, introduce a spatial convolutional network, perform feature extraction, and output the downsampled point cloud and the features corresponding to each point cloud. S23. Based on whether each point recorded in S21 belongs to the source point cloud or the target point cloud, separate the point cloud output in S22 into downsampled source point clouds. and their corresponding features Target point cloud after downsampling and their corresponding features ,in, x This represents the source point cloud before downsampling. y Let the target point cloud before downsampling be denoted as the source point cloud. x If N points remain after downsampling, then It is an N*3 matrix structure, where "3" represents three-dimensional coordinates. It is an N*d data structure, which varies depending on the spatial convolutional network.

3. The point cloud registration method for low overlap rates according to claim 1, characterized in that, The attention module includes a self-attention mechanism network and a cross-attention mechanism network connected in sequence. The self-attention mechanism network enables each point to interact with all points within its own point cloud. The cross-attention mechanism network searches for structural similarities between the source and target point clouds, making the features of the overlapping areas of the two point clouds more prominent. The process of the self-attention mechanism is as follows: in, This represents the parameters to be learned in the network. The concatenation operation, for the source point cloud, is implemented in a self-attention mechanism network that satisfies... The implementation process of the cross-attention mechanism network satisfies For the target point cloud, the implementation process in the self-attention mechanism network satisfies... The implementation process of the cross-attention mechanism network satisfies After passing through multiple self-attention and cross-attention network mechanisms, features are obtained. .

4. The point cloud registration method for low overlap rates according to claim 1, characterized in that, Utilizing the principle of spatial compatibility from When selecting a set of trustworthy point pairs, the selection process is enhanced using second-order spatial compatibility, where the first-order spatial compatibility matrix... The construction is as follows: Second-order spatial compatibility matrix The construction is as follows: ; in, Represents the matrix dot product operation. Represents the matrix cross product operation; Representation of point pairs and point pairs The number of other point pairs that simultaneously satisfy spatial compatibility; after obtaining the second-order spatial compatibility matrix, each time from Select the pair of points with the largest score that has not yet been selected, and then find the top pair in the second-order compatibility matrix. Each pair of points is considered as a set of trustworthy pairs, thus obtaining the set of trustworthy pairs. .

5. The point cloud registration method for low overlap rates according to claim 1, characterized in that, When filtering out a set of trustworthy point pairs from a set of identical point pairs, use a specified threshold for filtering or a deep learning-based filtering method.

6. The point cloud registration method for low overlap rates according to claim 4, characterized in that, Transformation matrix from source point cloud to target point cloud The expression is: in, This refers to the final pair of points with the same name that meet the criteria. For the points corresponding to the scores, R This represents a 3x3 rotation matrix in three-dimensional space. t Represents a three-dimensional translation vector, with a size of 1*3, obtained through... R , t The ability to represent any translation or rotation in three-dimensional space is crucial; the ultimate goal of point cloud registration is to find a set of... This minimizes the average distance between the new point cloud obtained by acting on the source point cloud and the target point cloud; in obtaining Then, the transformation matrix is ​​solved using the weighted singular value decomposition algorithm. The expression is used to obtain a transformation matrix for each set of reliable point pairs. These transformation matrices are traversed, and the transformation matrix with the best effect is selected as the registration result from the source point cloud to the target point cloud.

Images