Point cloud registration optimization method based on spatial feature coding and geometric perception attention calculation

Abstract

The invention relates to the technical field of point cloud data registration, in particular to a point cloud registration optimization method based on spatial feature coding and geometric perception attention calculation, and effectively overcomes the problems of insufficient matching precision and low efficiency under low overlapping and complex spatial relationships. The method comprises the following steps: firstly, performing efficient feature extraction and encoding on point cloud data by using a spatial feature encoder to obtain feature representation containing rich spatial information; thirdly, designing a self-attention mechanism based on geometric perception, and weighting a key region in the point cloud through a self-adaptive attention weight, so that the model can focus on the most significant and highly recognizable features in the point cloud data; and then, through multi-stage feature fusion and local geometric modeling, the precision and robustness of point cloud registration are further improved. Experimental results show that in various complex environments, especially high noise, rotation, deformation and the like, the method shows registration performance superior to that of a traditional method.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud data registration, and particularly to a point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation. Background Art

[0002] Point cloud registration aims to accurately align point cloud data from different perspectives, different times, or different sensors, and is a key foundation for 3D computer vision tasks such as autonomous driving, robot navigation, augmented reality (AR), and medical image analysis. However, due to the disordered and unstructured nature of point cloud data itself, combined with external factors such as sensor noise, perspective changes, and partial occlusion, achieving efficient and accurate 3D point cloud registration is still a highly challenging task.

[0003] Traditional point cloud registration methods are mainly divided into two categories: correspondence-based methods and deep learning-based methods. Correspondence-based methods, such as the Iterative Closest Point (ICP) and its derivative algorithms, although widely used, they rely on an iterative optimization process, are easily affected by the problems of initial alignment deviation and local optimal solutions, and have a high computational complexity. In contrast, deep learning-based methods, such as RoITr and GeoTransformer, directly extract point cloud features through end-to-end learning, greatly improving the robustness and efficiency of matching. However, existing deep learning models still face many challenges in processing large-scale point cloud data, extracting multi-scale features, and modeling complex spatial relationships. Especially in scenarios dealing with rotational changes, low overlap regions, and high-noise point clouds, it is difficult to ensure stable and accurate configuration performance. Summary of the Invention

[0004] The technical problem to be solved by the present invention is: In order to overcome the above technical problems, the present invention provides a point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation.

[0005] The technical solution adopted by the present invention to solve its technical problems is: A point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation. First, a spatial feature encoder is used to extract and encode point cloud data to obtain a feature representation containing spatial information. Then, key regions in the point cloud are weighted by adaptive attention weights, and the self-attention mechanism is based on a geometric perception spatial attention module, enabling the model to focus on the most significant and discriminative features in the point cloud data. The geometric perception spatial attention module combines the geometric information and spatial relationships of the point cloud. Finally, multi-stage feature fusion and local geometric modeling are performed.

[0006] A point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation of the present invention specifically includes the following steps:

[0007] Step 1: Extract features from the input point cloud data, and process the noise and rotational transformation in the data through downsampling and data augmentation. On this basis, use K-nearest neighbor search to determine the local neighborhood of each point, estimate the normal vector of each point based on the local geometric structure within the neighborhood, and adjust it to the unified viewing direction, providing accurate input for subsequent local and global attention analysis and ensuring high-precision point cloud registration and matching.

[0008] Step 2: Based on the point cloud input preprocessed in Step 1, extract multi-scale features through a spatial feature encoder, and use sine position encoding to enhance the spatial perception ability of the point cloud, providing rich feature information for the registration process.

[0009] Step 3: Use a geometric perception spatial attention module to adaptively assign different weights to the point cloud data in the local and global ranges, enabling the model to focus more on the feature matching in the key areas and improving the pertinence of registration.

[0010] Step 4: Combine the global Transformer structure to achieve the fusion of cross-frame context information, further enhancing the global consistency and matching accuracy of the point cloud data.

[0011] Step 5: Adopt a coarse-to-fine matching strategy, first perform a preliminary matching, and then combine the geometric constraints in the loss function to optimize the point cloud alignment result with high precision, thereby improving the accuracy and stability of point cloud registration.

[0012] The input point cloud data is three-dimensional point cloud data. The spatial feature encoder is used to convert the input three-dimensional point cloud data into a high-dimensional feature representation and capture the geometric information and semantic information of the point cloud through a multi-stage feature extraction process.

[0013] The spatial feature encoder mainly consists of two parts: an original point embedding module and a multi-stage feature extraction module. The original point embedding module is used to convert the input three-dimensional point cloud coordinates into high-dimensional features to enhance the model's understanding of spatial geometric information, and the multi-stage feature extraction module is responsible for extracting and fusing the geometric features of the point cloud from multiple scales.

[0014] To obtain a high-dimensional representation that can be used for subsequent geometric feature extraction and attention mechanism, the original point embedding module converts the input three-dimensional point cloud coordinates (B, N, 3) into high-dimensional embedding features through position encoding.

[0015] After obtaining this input tensor, first calculate the embedding dimension feat of each point dim , ensuring that the output dimension outdim is consistent with the input dimension; secondly, create a feature range feat range, whose value is a sequence of integers not exceeding the embedding dimension and is used for subsequent embedding calculations; then, the feature range is calculated to obtain the embedding coefficient dim for each dimension embed , and the specific embedding coefficient formula is as follows:

[0016]

[0017] In Equation (1), α is a hyperparameter. To better capture the spatial structure information, based on the above calculations, the feature tensor is further normalized, and the sine and cosine values of the normalized values are calculated. This process enhances the model's understanding of spatial geometric information. The basic idea of position encoding is to use sine and cosine functions to map the three-dimensional coordinates of each point into a high-dimensional space to generate periodic position embeddings. The formula is as follows:

[0018] z i = [sin(x i ·W sin ), cos(x i ·W cos )] (2)

[0019] In Equation (2), x i ∈R 3 represents the three-dimensional coordinates of the i-th point in the point cloud data, and W sin and W cos represent the periodic transformation matrices corresponding to the sine and cosine components respectively. They independently control the frequency distribution of position encoding in the sine and cosine directions, and then generate the high-dimensional embedding feature z i ∈R d .

[0020] Finally, the sine and cosine values are stacked and flattened to obtain the final position embedding feature. In this way, the spatial information of the original point cloud is converted into high-dimensional features, which will help capture and process the geometric information of the point cloud in the subsequent steps.

[0021] After passing through the original point embedding module, a multi-stage feature extraction module is used to capture the geometric structure of the point cloud through layer-by-layer feature aggregation:

[0022] First, representative points are selected through the farthest point sampling (FPS) algorithm, then the k-nearest neighbor (kNN) algorithm is used to aggregate their neighborhood features, and then local geometric aggregation (LGA) is used to further process the local geometric information. The specific calculation process is as follows:

[0023] The point farthest from the current point is selected as the new representative point through the farthest point sampling algorithm to maximize the coverage of the point cloud. Given the input point cloud P = {p1, p2,..., p n}, where n is the number of points, the output of FPS is a subset P FPS :

[0024]

[0025] Among them, \(p\in P\) represents the currently selected representative point, and \(p'\in P\) represents other points in the candidate set. After selecting representative points through the farthest point sampling algorithm, the k-nearest neighbor algorithm is used to aggregate the features of points in the neighborhood. By calculating the Euclidean distance between points, the k nearest points are selected and feature aggregation is performed. For any two points \(p\) i and \(p\) j in the point cloud, the Euclidean distance between them is:

[0026]

[0027] where \(l\) represents the dimension index in three-dimensional space, and \(l\) takes natural numbers from 1 to 3. \(p\) i,l and \(p\) j,l respectively represent the coordinate values of \(p\) i and \(p\) j on the \(l\)-th dimension. After extracting the features and coordinates of adjacent points through the k-nearest neighbor algorithm, they are passed into the local geometric aggregation module, and the geometric information of the local point cloud is encoded to capture the local geometric structure: First, the mean and standard deviation of each neighborhood feature are calculated to normalize the local point coordinates; then, the local features and the features of neighborhood points are concatenated, that is, the features corresponding to the local features are extended to all neighborhood points; then, the input neighborhood coordinates and features generate position encodings through the \(\sin\) and \(\cos\) functions, and their concatenation and flattening are used to adjust the dimensions. Finally, the position encoding is used;

[0028] Geometric processing of the point cloud in the neighborhood is performed through position embedding, and these local features are fused into the global representation, and weighted aggregation is performed through the following formula:

[0029]

[0030] In formula (5), \(k\) represents the number of points included in the neighborhood of the current point, that is, the neighborhood size set in the k-nearest neighbor algorithm; \(w\) i is the weight of the \(i\)-th neighborhood point, indicating the contribution of this point to the aggregation result, and is obtained by calculating the rotation of the embedding space; \(t\) i is the feature vector of this point; \(X\) agg represents the aggregated local feature vector, which is the result of the weighted average of the features of points in the neighborhood and is used for subsequent point cloud representation.

[0031] The geometric perception spatial attention module is based on the geometric information G, position encoding E, and spatial feature S of the spatial features. By calculating the similarity between the source point cloud Q and the target point clouds O and V, and combining additional constraints of the geometric structure, it improves the accuracy of point cloud registration. This module can effectively model the geometric relationships between points in the point cloud. Especially when dealing with complex scenes with different pose changes, the model can maintain high robustness. Among them, G and E are generated through geometric structure embedding, and the geometric structure embedding generation process is as follows:

[0032] First, calculate the geometric distance d between each pair of points in the point cloud indices and the geometric angle a between point triples indices , and the calculation steps include distance embedding and angle embedding;

[0033] After further processing the distance embedding and angle embedding through the position encoding, the final geometric information G and position encoding E are generated. Among them, the geometric information G represents the geometric structure information constructed based on the Euclidean distance between points; the position encoding E represents the relative spatial position information constructed based on the included angle of normal vectors; both the geometric information G and the position encoding E are embedded and transformed through their respective independent linear projection layers, and the projection matrices W G and W E are used to map them into representations of a unified dimension; the spatial feature S is linearly transformed through W S and is used to participate in the attention calculation together with the above embeddings.

[0034] Then, calculate the attention scores through the dot product, and then perform weighted summation on the attention scores and use normalization to control the attention scores within a reasonable range. The specific calculation method is as follows:

[0035]

[0036] In Equation (6), qE T represents the attention score between the source point and the geometric feature, qS T represents the attention score between the source point and the spatial feature, qO T represents the attention score of the source point itself; c0 represents the normalization factor, which is used to scale the result after weighted summation. When calculating the similarity of multiple points, as the dimension increases, the value will also increase; to prevent the numerical value from being too large and causing gradient explosion during softmax normalization, normalization makes the result of the dot product have an appropriate scale, which helps the stable propagation of the gradient.

[0037] By calculating the dot product between different features, the model can introduce geometric information and spatial features into the matching process. These attention scores help the model weight the point cloud features according to the spatial structure and relative position of the point cloud, ensuring the influence of geometric relationships on point cloud matching.

[0038] The distance embedding is specifically as follows: The goal of distance embedding is to calculate the geometric distance between each pair of points in the point cloud; for the given point cloud data X ∈ R B×N×3 , first calculate the Euclidean distance between each pair of points in the point cloud to obtain a distance matrix with the shape of (B, N, N). The formula is as follows:

[0039]

[0040] where dist[b, i, j] represents the Euclidean distance between the i-th point and the j-th point in the b-th batch, and σ d is a temperature parameter used to scale the distance and control the sensitivity of the distance; by calculating the Euclidean distance between each pair of points in the point cloud and then using σ d for scaling, the distances of different point pairs are standardized to obtain the distance embedding features of each pair of points. This feature will be used as input in subsequent calculations to help the model understand the geometric relationship between points.

[0041] The angle embedding is specifically as follows: Angle embedding is used to calculate the geometric relationship between point triples in the point cloud. By calculating the included angle information between adjacent vectors, the local structure of the point cloud is captured, and this angle information helps to describe the local geometric structure in the point cloud. For each reference point, a point triple is formed by the reference point, its adjacent point, and the adjacent point of its adjacent point. The included angle tensor anhles[b, i, j, m] formed by this triple is then multiplied by the scaling factor factor d to generate the angle embedding vector a indices . The formula is as follows:

[0042] a indices = angles[b, i, j, m] × factor d (8)

[0043] where angles[b, i, j, m] represents the set of geometric included angles between the point triples formed by the reference point i, its adjacent point j, and the nearest neighbor point m of j in the b-th batch. This angle is calculated through the included angle between the vector formed by point i and point j and the vector formed by point j and point m; factor d is a scaling factor used to adjust the range of angle embedding.

[0044] The beneficial effects of the present invention are as follows. The point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation of the present invention effectively overcomes the problems of insufficient matching accuracy and low efficiency under low overlap and complex spatial relationships. First, a spatial feature encoder is used to efficiently extract and encode point cloud data to obtain a feature representation containing rich spatial information. Then, a geometric perception-based self-attention mechanism is designed to weight key regions in the point cloud through adaptive attention weights, enabling the model to focus on the most significant and distinguishable features in the point cloud data. Finally, through multi-stage feature fusion and local geometric modeling, the accuracy and robustness of point cloud registration are further improved. Experimental results show that in various complex environments, especially in high-noise, rotation, and deformation scenarios, the method of the present invention exhibits better registration performance than traditional methods. Description of the Drawings

[0045] The present invention will be further described below with reference to the drawings and embodiments.

[0046] Figure 1 is the overall flowchart of the point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation of the present invention.

[0047] Figure 2 is the point cloud registration model diagram based on spatial feature encoding and geometric perception attention calculation of the present invention.

[0048] Figure 3 is the module diagram of the spatial feature encoder of the present invention.

[0049] Figure 4 is the standard attention calculation diagram and the spatial geometric perception attention calculation diagram.

[0050] Figure 5 is the comparison diagram of the point cloud registration results of the present invention and GeoTrans under different complex scenarios selected from the 3DMatch dataset; Figure 5 In (a) is the input point cloud data, (b) is the true transformation of the input point cloud data, (c) is the point cloud registration result of GeoTrans, and (d) is the point cloud registration result of the present invention.

[0051] Figure 6 is the comparison diagram of the point cloud registration results of the present invention and GeoTrans under different complex scenarios selected from the 3DLoMatch dataset; Figure 6 In (a) is the input point cloud data, (b) is the true transformation of the input point cloud data, (c) is the point cloud registration result of GeoTrans, and (d) is the point cloud registration result of the present invention. Detailed Embodiments

[0052] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, and therefore only showing the components related to the present invention.

[0053] As Figure 1-2 shown, an optimized point cloud registration method based on spatial feature encoding and geometric perception attention calculation of the present invention first uses a spatial feature encoder to extract and encode point cloud data to obtain a feature representation containing spatial information; then weights key regions in the point cloud through adaptive attention weights, and the self-attention mechanism is based on a geometric perception spatial attention module, enabling the model to focus on the most significant and discriminative features in the point cloud data, and the geometric perception spatial attention module combines the geometric information and spatial relationship of the point cloud; finally, multi-stage feature fusion and local geometric modeling are performed.

[0054] An optimized point cloud registration method based on spatial feature encoding and geometric perception attention calculation of the present invention specifically includes the following steps:

[0055] Step 1: Extract features from the input point cloud data, and process the noise and rotation transformation in the data through downsampling and data augmentation; on this basis, use K-nearest neighbor search to determine the local neighborhood of each point, estimate the normal vector of each point based on the local geometric structure within the neighborhood, and adjust it to a unified viewing direction; providing accurate input for subsequent local and global attention analysis, ensuring high-precision point cloud registration and matching.

[0056] Step 2: Based on the point cloud input preprocessed in Step 1, extract multi-scale features through a spatial feature encoder, and use sine position encoding to enhance the spatial perception ability of the point cloud; providing rich feature information for the registration process.

[0057] Step 3: Use the geometric perception spatial attention module to adaptively assign different weights to the point cloud data in the local and global ranges; enabling the model to focus more on the feature matching of key regions and improving the pertinence of registration.

[0058] Step 4: Combine the global Transformer structure to achieve the fusion of cross-frame context information; further improving the global consistency and matching accuracy of the point cloud data.

[0059] Step 5: Adopt a coarse-to-fine matching strategy, first perform a preliminary matching, and then combine the geometric constraints in the loss function to optimize the point cloud alignment result with high precision, thereby improving the accuracy and stability of point cloud registration.

[0060] The input point cloud data is three-dimensional point cloud data. The spatial feature encoder is used to convert the input three-dimensional point cloud data into a high-dimensional feature representation, and capture the geometric information and semantic information of the point cloud through a multi-stage feature extraction process.

[0061] As Figure 3 shown, the spatial feature encoder mainly consists of two parts: an original point embedding module and a multi-stage feature extraction module. The original point embedding module is used to convert the input three-dimensional point cloud coordinates into high-dimensional features, enhancing the model's understanding of spatial geometric information. The multi-stage feature extraction module is responsible for extracting and fusing the geometric features of the point cloud from multiple scales.

[0062] To obtain a high-dimensional representation that can be used for subsequent geometric feature extraction and attention mechanism, the original point embedding module converts the input three-dimensional point cloud coordinates (B, N, 3) into high-dimensional embedding features through position encoding;

[0063] After obtaining this input tensor, first calculate the embedding dimension feat of each point dim , ensuring that the output dimension outdim is consistent with the input dimension; secondly, create a feature range feat range , whose value is an integer sequence not exceeding the embedding dimension, and is used for subsequent embedding calculations; then, calculate the feature range to obtain the embedding coefficient dim of each dimension embed , and the specific embedding coefficient formula is as follows:

[0064]

[0065] In formula (1), α is a hyperparameter. To better capture the spatial structure information, based on the above calculations, further normalize the feature tensor, and calculate the sine and cosine values of the normalized values. This process enhances the model's understanding of spatial geometric information. The basic idea of position encoding is to map the three-dimensional coordinates of each point to a high-dimensional space using sine and cosine functions to generate periodic position embeddings. The formula is as follows:

[0066] z i =[sin(x i ·W sin ),cos(x i ·W cos )] (2)

[0067] In formula (2), x i ∈R 3 represents the three-dimensional coordinates of the i-th point in the point cloud data, W sin and W cosrespectively represent the periodic transformation matrices corresponding to the sine component and the cosine component, which independently control the frequency distribution of the position encoding in the sine and cosine directions, and then generate the high-dimensional embedding feature z of each point i ∈R d ;

[0068] Finally, the sine and cosine values are stacked and flattened to obtain the final position embedding feature. In this way, the spatial information of the original point cloud is converted into high-dimensional features, which will help capture and process the geometric information of the subsequent point cloud.

[0069] After passing through the original point embedding module, a multi-stage feature extraction module is used to capture the geometric structure of the point cloud through layer-by-layer feature aggregation:

[0070] First, representative points are selected by the farthest point sampling (FPS) algorithm, then the k-nearest neighbor (kNN) algorithm is used to aggregate the neighborhood features, and then local geometric aggregation is used to further process the local geometric information. The specific calculation process is as follows:

[0071] The point farthest from the current point is selected as the new representative point by the farthest point sampling algorithm to maximize the coverage of the point cloud. Given the input point cloud P = {p1, p2, …, p n}, where n is the number of points, the output of FPS is a subset P FPS :

[0072]

[0073] where p ∈ P represents the currently selected representative point, and p′ ∈ P represents other points in the candidate set; after selecting the representative point by the farthest point sampling algorithm, the k-nearest neighbor algorithm is used to aggregate the features of the points in the neighborhood. By calculating the Euclidean distance between points, the k closest points are selected and feature aggregation is performed. The Euclidean distance between any two points p i and p j is:

[0074]

[0075] where l represents the dimension index in three-dimensional space, l takes natural numbers from 1 to 3, p i,l and p j,l respectively represent p i and p jThe coordinate value in the l-th dimension; after extracting the features and coordinates of adjacent points through the K-nearest neighbor algorithm, it is passed into the local geometric aggregation module, and the geometric information of the local point cloud is encoded to capture the local geometric structure: First, calculate the mean and standard deviation of each neighborhood feature to normalize the local point coordinates; then, concatenate the local features and the features of the neighborhood points, that is, expand the features corresponding to the local features to all neighborhood points; then, generate position encodings for the input neighborhood coordinates and features through the sin and cos functions, and splice and flatten them to adjust the dimensions. Finally, use the positive and cosine positions;

[0076] Perform geometric processing on the point cloud within the neighborhood through position embedding, and fuse these local features into the global representation, and perform weighted aggregation through the following formula:

[0077]

[0078] In Equation (5), k represents the number of points included in the neighborhood of the current point, that is, the neighborhood size set in the K-nearest neighbor algorithm; w i is the weight of the i-th neighborhood point, indicating the contribution of this point to the aggregation result, and is obtained by calculating the rotation of the embedding space; t i is the feature vector of this point; X agg represents the aggregated local feature vector, which is the result of the weighted average of the features of the points within the neighborhood and is used for subsequent point cloud representation.

[0079] In summary, the spatial feature encoder effectively captures the local and global spatial features of the point cloud through multi-stage feature extraction, position embedding, farthest point sampling, kNN, and local geometric aggregation and other technologies, providing rich and discriminative geometric information for subsequent point cloud registration tasks. Through this modular design, the model can handle complex geometric structures and enhance the accuracy of point cloud matching.

[0080] After effectively capturing the local and global spatial geometric features of the point cloud through the spatial feature encoder, then adopt the geometric-aware spatial attention module to better understand the spatial relationship between point clouds, especially in the case of large pose changes. The specific calculation method of spatial perception attention is as Figure 4 shown, and the working process of spatial geometric perception attention will be introduced in detail below.

[0081] The geometric-aware spatial attention module of the present invention combines the geometric information and spatial relationship of the point cloud to enhance the registration accuracy. Traditional attention mechanisms usually only focus on the context information in the point cloud and ignore the impact of the spatial geometric structure on matching. To solve this problem, our module enhances the attention mechanism by introducing geometric features and spatial features generated by the spatial feature encoder, enabling point cloud registration to better understand the spatial relationship between point clouds.

[0082] The core idea of the geometric perception spatial attention module of the present invention is based on the geometric information G, position encoding E, and spatial feature S of the spatial features. By calculating the similarity between the source point cloud Q and the target point clouds O and V, and combining with the additional constraints of the geometric structure, the accuracy of point cloud registration is improved. This module can effectively model the geometric relationships between points in the point cloud. Especially when dealing with complex scenes with different pose changes, the model can maintain high robustness. Among them, G and E are generated through geometric structure embedding, and the geometric structure embedding generation process is as follows:

[0083] First, calculate the geometric distance d between each pair of points in the point cloud indices and the geometric angle a between point triples indices , and the calculation steps include distance embedding and angle embedding;

[0084] The distance embedding and angle embedding are further processed through the position encoding to generate the final geometric information G and position encoding E. Among them, the geometric information G represents the geometric structure information constructed based on the Euclidean distance between points; the position encoding E represents the relative spatial position information constructed based on the included angle of normal vectors; both the geometric information G and the position encoding E are embedded and transformed through their respective independent linear projection layers, and the projection matrices W G and W E are used to map them into representations of a unified dimension; the spatial feature S is linearly transformed through W S and is used to participate in the attention calculation together with the above embeddings. The specific calculation method is as Figure 4 shown.

[0085] Then, calculate the attention scores through the dot product, and then perform weighted summation on the attention scores and use normalization to control the attention scores within a reasonable range. The specific calculation method is as follows:

[0086]

[0087] In Equation (6), qE T represents the attention score between the source point and the geometric feature, qS T represents the attention score between the source point and the spatial feature, qO T represents the attention score of the source point itself; c0 represents the normalization factor, which is used to scale the result after weighted summation. When calculating the similarity of multiple points, as the dimension increases, the value will also increase; in order to prevent the numerical value from being too large during softmax normalization and causing gradient explosion, normalization makes the result of the dot product have an appropriate scale, which helps the stable propagation of the gradient.

[0088] By calculating the dot product between different features, the model can introduce geometric information and spatial features into the matching process. These attention scores help the model weight the point cloud features according to the spatial structure and relative positions of the point cloud, ensuring the influence of geometric relationships on point cloud matching.

[0089] Overall, the core role of the spatial geometric perception attention module is to further improve the accuracy of point cloud registration by introducing geometric information and spatial features in the point cloud, especially in complex scenes. This module optimizes the point cloud matching process by calculating the similarity between the source point cloud and the target point cloud, combining spatial features and geometric information.

[0090] The distance embedding is specifically as follows: The goal of distance embedding is to calculate the geometric distance between each pair of points in the point cloud; for a given point cloud data X ∈ R B×N×3 , first calculate the Euclidean distance between each pair of points in the point cloud to obtain a distance matrix with a shape of (B, N, N), and the formula is as follows:

[0091]

[0092] where dist[b, i, j] represents the Euclidean distance between the i-th point and the j-th point in the b-th batch, and σ d is a temperature parameter used to scale the distance and control the sensitivity of the distance; by calculating the Euclidean distance between each pair of points in the point cloud and then using σ d for scaling, so as to standardize the distances of different point pairs, thereby obtaining the distance embedding features of each pair of points. This feature will be used as input in subsequent calculations to help the model understand the geometric relationships between points.

[0093] The angle embedding is specifically as follows: Angle embedding is used to calculate the geometric relationships between point triples in the point cloud. By calculating the included angle information between adjacent vectors, it captures the local structure of the point cloud, and this angle information helps to describe the local geometric structure in the point cloud. For each reference point, it forms a point triple with its adjacent points and the adjacent points of its adjacent points. The included angle tensor angles[b, i, j, m] formed by calculating through this triple is then multiplied by the scaling factor factor d corresponding to different dimensions to generate the angle embedding vector a indices , and the formula is as follows:

[0094] a indices = angles[b, i, j, m] × factor d (8)

[0095] Among them, angles[b,i,j,m] represents the set of geometric angles between point triples formed by the reference point i, its adjacent point j, and the nearest neighbor m of j in the b-th batch. This angle is calculated through the angle between the vector formed by point i and point j and the vector formed by point j and point m; factor d is a scaling factor used to adjust the range of angle embedding.

[0096] To comprehensively evaluate the performance of the proposed method in different complex application scenarios, two indoor real-scene datasets, 3DMatch and 3DLoMatch, were used to evaluate the algorithm and compared with different baseline algorithms. The 3DMatch dataset covers 62 different indoor scenes, among which 46 scenes are assigned for model training, 8 scenes for model validation, and the remaining 8 scenes for test evaluation. The feature point pairs in this dataset have an overlap of approximately 30%. The 3DLoMatch dataset also consists of 62 scenes, but the overlap of its feature point pairs is lower, between 10% and 30%. To ensure the consistency of experimental conditions, standard preprocessing was performed on the two datasets.

[0097] In the experiment, the following three metrics were used to evaluate the performance of the proposed method: (1) Feature Matching Recall (FMR): This metric evaluates the upper limit of the algorithm's ability to correctly match under different point cloud matching conditions by calculating the proportion of point cloud pairs with an inlier ratio greater than a specific threshold in a given point cloud pair. (2) Inlier Ratio (IR): This metric evaluates the proportion of correct point pairs that the point cloud matching algorithm can find under the true transformation by calculating whether the deviation distance of the corresponding point pairs under the action of the true transformation matrix is less than a specific threshold. (3) Registration Recall (RR): This metric measures the success rate of point cloud registration by calculating whether the error between the transformation matrix and the true transformation matrix is less than a specific threshold.

[0098] The comparison results of the feature matching recall, inlier ratio, and registration recall of different methods on the 3DMatch and 3DLoMatch datasets are shown in Tables 1 - 3.

[0099] As can be seen from Table 1, the proposed method shows the highest feature matching recall under different numbers of samples, especially on the 3DLoMatch dataset, demonstrating its advantage in the case of low overlap.

[0100] As can be seen from Table 2, the proposed method shows stable performance in terms of the inlier ratio on the 3DMatch dataset, and can still maintain a high inlier ratio even when the number of samples is small (such as 250 and 500 samples), demonstrating its superiority in the case of sparse data.

[0101] As can be seen from Table 3, the proposed method shows extremely high registration recall rates on both the 3DMatch and 3DLoMatch datasets. Especially on the 3DLoMatch dataset (under low overlap conditions), it far exceeds all baseline methods, further proving its powerful capabilities in complex scenarios.

[0102] Based on the above analysis results, it can be seen that compared with other methods, this method can not only achieve accurate point cloud registration under large rotations and low overlaps, but also maintain efficient and stable performance when dealing with situations containing noise and irregular point cloud distributions. This advantage makes this method suitable for practical applications such as autonomous driving and 3D reconstruction that require high-precision point cloud registration.

[0103] Table 1 Comparison of feature matching recall rates of different methods on 3DMatch and 3DLoMatch datasets

[0104]

[0105] Table 2 Comparison of inlier ratios of different methods on 3DMatch and 3DLoMatch datasets

[0106] <S

[0107] Table 3 Comparison of registration recall rates of different methods on 3DMatch and 3DLoMatch datasets

[0108]

[0109] Figure 5 and Figure 6 shows the point cloud registration results of the present invention and the currently best-performing baseline method GeoTrans under different complex scenarios selected from the 3DMatch dataset and the 3DLoMatch dataset. As can be seen from the figure, in Figure 5 (3DMatch) and Figure 6 (3DLoMatch), the method of the present invention and GeoTrans show similar overall registration effects, but on the more challenging 3DLoMatch dataset, it can still be observed that the method of the present invention is more consistent and stable in terms of detail alignment, boundary fusion, and smooth transition. Especially in the low overlap area, the method of the present invention can better restore the structural continuity and reduce the phenomena of dislocation and cracks.

[0110] Furthermore, to evaluate the contribution of each component in the present invention to the overall performance, we conducted ablation experiments to systematically analyze the roles of modules such as local geometric attention, global geometric attention, and spatial feature encoder. The experiments mainly compared the performance under different configurations, especially the impact of local and global attention mechanisms on the point cloud matching accuracy. The specific quantitative results of the ablation experiments are shown in Table 4, further verifying the effectiveness of the spatial feature encoder and spatial geometric attention mechanism in the registration task. After removing the spatial feature encoder, the performance of the model on both 3DMatch and 3DLoMatch decreased significantly, indicating that this module plays a key role in capturing local point cloud structure features and providing geometric context support. After introducing spatial geometric attention, both local and global modules brought stable performance improvements on 3DMatch and 3DLoMatch, especially in terms of registration accuracy (RR). Among them, the addition of global spatial geometric attention enables the model to further improve the matching consistency between features while maintaining high robustness, verifying its effectiveness and necessity in point cloud registration.

[0111] Table 4 Ablation Experiments Based on 3Dmatch and 3DLomatch Datasets

[0112]

[0113] Taking the ideal embodiments of the present invention described above as an inspiration, through the above description, relevant staff can completely make various changes and modifications without departing from the technical idea of the present invention. The technical scope of the present invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. An optimized method for point cloud registration based on spatial feature encoding and geometric perception attention calculation, characterized in that First, use a spatial feature encoder to extract and encode the point cloud data to obtain a feature representation containing spatial information. Then, weight the key regions in the point cloud through adaptive attention weights. The self-attention mechanism is based on a geometric-aware spatial attention module, which enables the model to focus on the most significant and discriminative features in the point cloud data. The geometric-aware spatial attention module combines the geometric information and spatial relationships of the point cloud. Finally, perform multi-stage feature fusion and local geometric modeling.

2. The point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation according to claim 1, characterized in that Specifically, it includes the following steps: Step 1: Extract features from the input point cloud data, and process the noise and rotational transformations in the data through downsampling and data augmentation. On this basis, use K-nearest neighbor search to determine the local neighborhood of each point, estimate the normal vector of each point based on the local geometric structure within the neighborhood, and adjust it to a unified viewing direction. Step 2: Based on the point cloud input preprocessed in Step 1, extract multi-scale features through a spatial feature encoder, and use sine position encoding to enhance the spatial perception ability of the point cloud. Step 3: Use the geometric-aware spatial attention module to adaptively assign different weights to the point cloud data within the local and global ranges. Step 4: Combine the global Transformer structure to achieve the fusion of cross-frame context information. Step 5: Adopt a coarse-to-fine matching strategy, first perform a preliminary match, and then combine the geometric constraints in the loss function to optimize the point cloud alignment result with high precision.

3. The point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation according to claim 1, wherein The input point cloud data is three-dimensional point cloud data. The spatial feature encoder is used to convert the input three-dimensional point cloud data into a high-dimensional feature representation, and capture the geometric information and semantic information of the point cloud through a multi-stage feature extraction process.

4. The point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation according to claim 3, wherein, The spatial feature encoder mainly consists of two parts: an original point embedding module and a multi-stage feature extraction module. The original point embedding module is used to convert the input three-dimensional point cloud coordinates into high-dimensional features to enhance the model's understanding of spatial geometric information. The multi-stage feature extraction module is responsible for extracting and fusing the geometric features of the point cloud from multiple scales.

5. The point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation according to claim 4, characterized in that The original point embedding module converts the input three-dimensional point cloud coordinates (B, N, 3) into high-dimensional embedding features through position encoding. After obtaining the input tensor, first calculate the embedding dimension feat for each point dim , ensuring that the output dimension outdim is consistent with the input dimension; secondly, create a feature range feat range , whose values are an integer sequence not exceeding the embedding dimension and are used for subsequent embedding calculations; then, calculate the feature range to obtain the embedding coefficient dim for each dimension emded , and the specific embedding coefficient formula is as follows: In Equation (1), α is a hyperparameter. Based on the above calculations, further normalize the feature tensor, and calculate the sine and cosine values of the normalized values. The basic idea of position encoding is to map the three-dimensional coordinates of each point into a high-dimensional space using sine and cosine functions to generate periodic position embeddings. The formula is as follows: z i = [sin(x i ·W sin ), cos(x i ·W cos )] (2) In Equation (2), x i ∈R 3 represents the three-dimensional coordinates of the i-th point in the point cloud data, W sin and W cos respectively represent the periodic transformation matrices corresponding to the sine component and the cosine component, which independently control the frequency distribution of the position encoding in the sine and cosine directions, and then generate the high-dimensional embedding feature z i ∈R d ; Finally, stack and flatten the sine and cosine values to obtain the final position embedding feature. In this way, the spatial information of the original point cloud is converted into high-dimensional features.

6. The point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation according to claim 5, wherein, After passing through the original point embedding module, use the multi-stage feature extraction module to capture the geometric structure of the point cloud through layer-by-layer feature aggregation: First, select representative points through the farthest point sampling algorithm, then use the K-nearest neighbor algorithm to aggregate their neighborhood features, and then use local geometric aggregation to further process the local geometric information. The specific calculation process is as follows: Select the point farthest from the current point as the new representative point through the farthest point sampling algorithm to maximize the coverage of the point cloud. Given the input point cloud P = {p1, p2, …, p n}, where n is the number of points, the output of FPS is a subset P FPS : Among them, p ∈ P represents the currently selected representative point, and p ′ ∈ P represents other points in the candidate set; after selecting representative points through the farthest point sampling algorithm, the k-nearest neighbor algorithm is used to aggregate the features of points in the neighborhood. By calculating the Euclidean distance between points, the k nearest points are selected and feature aggregation is performed. For any two points p i and p j in the point cloud, the Euclidean distance between them is: where l represents the dimension index in three-dimensional space, l takes natural numbers from 1 to 3, p i,l and p j,l respectively represent the coordinate values of p i and p j on the l-th dimension; after extracting the features and coordinates of adjacent points through the K-nearest neighbor algorithm, they are passed into the local geometric aggregation module, and the local geometric structure is captured by encoding the geometric information of the local point cloud: First, calculate the mean and standard deviation of each neighborhood feature to normalize the local point coordinates; then, concatenate the local features and the features of the neighborhood points, that is, extend the features corresponding to the local features to all neighborhood points; next, generate position encodings for the input neighborhood coordinates and features through the sin and cos functions, and concatenate and flatten them to adjust the dimensions. Finally, use the positive and cosine positions; Geometric processing of the point cloud in the neighborhood through positional embedding and fusing these local features into the global representation, with weighted aggregation performed through the following formula: In formula (5), k represents the number of points included in the neighborhood of the current point, that is, the neighborhood size set in the K-nearest neighbor algorithm; w i is the weight of the i-th neighborhood point, indicating the contribution of this point to the aggregation result, which is obtained by calculating the rotation of the embedding space; t i is the feature vector of this point; X agg represents the aggregated local feature vector, which is the result of the weighted average of the features of the points in the neighborhood and is used for subsequent point cloud representation.

7. The point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation according to claim 1, characterized in that The geometric-aware spatial attention module is based on the geometric information G of the spatial features, the positional encoding E, and the spatial features S. By calculating the similarity between the source point cloud Q and the target point clouds O, V, and combining additional constraints of the geometric structure, it improves the accuracy of point cloud registration. Among them, G and E are generated through geometric structure embedding, and the geometric structure embedding generation process is as follows: First, calculate the geometric distance d between each pair of points in the point cloud indices and the geometric angle a between point triples indices , and the calculation steps include distance embedding and angle embedding; After further processing the distance embedding and the angle embedding through the positional encoding, the final geometric information G and the positional encoding E are generated, where the geometric information G represents the geometric structure information constructed based on the Euclidean distance between points; The position encoding E represents relative spatial position information constructed based on the angle between normal vectors; both the geometric information G and the position encoding E are embedded and transformed through their respective independent linear projection layers, using projection matrices W G and W E to map them into representations of a unified dimension; Spatial Feature S is linearly transformed by W S and is used to jointly participate in the attention calculation with the above embedding; Next, the attention scores are calculated through the dot product, and then the attention scores are weighted and summed, and then normalized to control the attention scores within a reasonable range. The specific calculation method is as follows: In formula (6), qE T represents the attention score between the source point and the geometric feature, qS T represents the attention score between the source point and the spatial feature, qO T represents the attention score of the source point itself; c0 represents the normalization factor, which is used to scale the result after weighted summation. When calculating the similarity of multiple points, as the dimension increases, the value will also increase.

8. The point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation according to claim 7, characterized in that The distance embedding is specifically as follows: The goal of distance embedding is to calculate the geometric distance between each pair of points in the point cloud; for a given point cloud data X ∈ R B×N×3 , first calculate the Euclidean distance between each pair of points in the point cloud to obtain a distance matrix with a shape of (B, N, N), and the formula is as follows: Among them, dist[b,i,j] represents the Euclidean distance between the i-th point and the j-th point in the b-th batch, and σ d is a temperature parameter used to scale the distance and control the sensitivity of the distance; by calculating the Euclidean distance between each pair of points in the point cloud, and then using σ d to scale it, so as to standardize the distances of different point pairs, thereby obtaining the distance embedding features of each pair of points.

9. The point cloud registration optimization method based on spatial feature encoding and geometric perception attention calculation according to claim 7, wherein Specifically, the angle embedding is as follows: The angle embedding is used to calculate the geometric relationship between point triples in the point cloud. By calculating the angle information between adjacent vectors, the local structure of the point cloud is captured. For each reference point, it forms a point triple with its adjacent points and the adjacent points of its adjacent points. The angle tensor angles[b, i, j, m] formed by calculating through this triple is then multiplied by the scaling factor factor corresponding to different dimensions d to generate the angle embedding vector a indices The formula is as follows: a indices = angles[b, i, j, m] × factor d (8) Among them, angles[b,i,j,m] represents the set of geometric angles between the point triples formed by the reference point i, its adjacent point j, and the nearest neighbor m of j in the b-th batch. This angle is calculated through the angle between the vector formed by point i and point j and the vector formed by point j and point m; factor d is a scaling factor used to adjust the range of angle embedding.

Images