Point cloud registration method based on fast point pair feature attention module

Abstract

The invention discloses a point cloud registration method based on a fast point pair feature attention module, and the method comprises the steps: firstly collecting point cloud data, and then carrying out the matching of point clouds through a point cloud registration model. The point cloud registration model is based on an encoder-decoder architecture, a fast point pair feature attention module FPFA, an Inception deep convolution module IDCM, a lightweight cross-frame attention aggregation module LCAAM and a matching matrix calculation module TSMM based on two stages are fused, and a final point cloud registration result is obtained. According to verification of a contrast experiment and an ablation experiment on a public data set and a self-built workpiece data set, the method is excellent in accuracy, speed and robustness, and especially has remarkable advantages in low-overlapping-rate scenes and industrial workpiece registration; an efficient solution is provided for factory assembly line automation and mechanical arm accurate grabbing, and the method has important application value for industrial detection scenes with limited computing resources.

Description

Technical Field

[0001] The present invention belongs to the field of point cloud processing technology, and specifically relates to a point cloud registration method based on a fast point-to-feature attention module, which is suitable for scenarios such as industrial manufacturing, robotic vision, and three-dimensional reconstruction that require processing point cloud data. Background Art

[0002] In modern industrial production, scientific research and exploration, and many cutting-edge technology fields, point cloud data processing has become a key link in achieving precise operations and in-depth analysis. Point cloud registration, as one of the core technologies of point cloud processing, aims to accurately align point cloud data acquired from different perspectives, at different times, or with different acquisition devices to construct a complete and accurate three-dimensional model, thereby providing a reliable data foundation for subsequent operations and analysis. In the field of industrial manufacturing, especially in factory assembly line production scenarios, the importance of point cloud registration is particularly prominent. As the manufacturing industry accelerates towards intelligence and automation, robotic arms are taking on more and more refined operation tasks on the production line, such as workpiece grasping and parts assembly. However, many complex factors in the actual production environment have brought huge challenges to point cloud data collection and registration. In order to overcome these problems, deep learning-based automation technology has gradually been introduced in recent years to achieve more intelligent and efficient leather production.

[0003] Regarding lighting conditions, factories often experience uneven illumination, with some areas experiencing insufficient illumination. This results in missing information in the collected point cloud data, making it difficult to accurately reflect the workpiece's true geometry. Furthermore, the reflective properties of some metal workpiece surfaces can introduce noise and deviation into the collected point cloud, further compromising its accuracy. Furthermore, workpieces on the assembly line are often arranged in a chaotic manner and often obstruct one another, resulting in significant geometric incompleteness in the original point cloud. In the field of robotic navigation and positioning, point cloud registration is also a key technology for enabling autonomous navigation and precise operation. When operating in complex indoor or outdoor environments, mobile robots need to collect point cloud data of their surroundings using devices such as LiDAR. These point clouds, collected at different times, must be registered to construct an accurate map and determine their own position and posture. However, factors such as interference from dynamic objects, scene similarity, and sensor errors in real-world environments pose numerous challenges to point cloud registration. In the field of 3D reconstruction, point cloud registration is fundamental to building high-quality 3D models. However, in actual operation, due to the accuracy limitations of the acquisition equipment, the complexity of the acquisition environment, and the similarity of the surface texture of the object, the accuracy and efficiency of point cloud registration are often difficult to meet the requirements. Summary of the Invention

[0004] The present invention proposes a point cloud registration method based on a fast point-to-feature attention module, which improves the registration speed while ensuring the accuracy of point cloud registration, so as to achieve a balance between accuracy and efficiency.

[0005] In a first aspect, the present invention provides a point cloud registration method based on a fast point-to-feature attention module, comprising the following steps:

[0006] Collect point cloud data, pre-process the point cloud data, and obtain source point cloud P and target point cloud Q;

[0007] Construct the point cloud registration model FPFAM-Reg to register the source point cloud P and the target point cloud Q;

[0008] The point cloud registration model FPFAM-Reg includes an encoder, a decoder, a lightweight cross-frame attention aggregation module LCAMM, and a two-stage matching matrix calculation module TSMM.

[0009] The encoder extracts features from the source point cloud P and the target point cloud Q and converts them into high-level structural features. and Input into the lightweight cross-frame attention aggregation module LCAMM, input into the decoder;

[0010] The lightweight cross-frame attention aggregation module LCAMM integrates the high-level feature information of point clouds from different frames extracted by the encoder and outputs enhanced features

[0011] The decoder will enhance the features and high-level structural features Gradually restore to the target space and input the restored point cloud features and

[0012] The two-stage matching matrix calculation module TSMM receives point cloud features and The point cloud registration is optimized through coarse matching and fine matching, and the final point cloud registration result is output.

[0013] Furthermore, the encoder includes three identical encoder blocks, each of which downsamples the input point cloud data, then extracts multi-scale features through the Inception deep convolution module IDCM, and then enhances feature interaction through two serially connected fast point pair feature attention modules FPFA. Finally, the features are standardized through a normalization layer, and the original feature information is retained using a residual connection. Nonlinearity is then introduced through a ReLU activation function to obtain the output of each encoder block.

[0014] Furthermore, the decoder includes three identical decoder blocks, each of which includes an upsampling layer, an Inception deep convolution module IDCM, two fast point-to-feature attention modules FPFA connected in series, a normalization layer, a residual connection and an activation function layer; the decoder receives features of different levels output by the encoder through jump connections.

[0015] Furthermore, the following loss function is used in the point cloud registration model training process:

[0016]

[0017] Among them, Ls is the loss function of coarse matching, which is used to solve the problem of global matching of point clouds at coarse granularity. The formula is as follows:

[0018]

[0019] Lp is the loss function of fine matching, which is used to optimize the local geometric consistency at the point level under the guidance of coarse matching. The formula is as follows:

[0020]

[0021] λ is a hyperparameter used to control the balance between loss values ​​and is usually set to 1; α is a weight for balancing geometric error and inlier ratio and is usually set to 0.5.

[0022] In a second aspect, the present invention provides a point cloud registration system for implementing the above method, comprising:

[0023] Data acquisition module, used to collect point cloud data and perform preprocessing;

[0024] The point cloud registration module is used to register the preprocessed point cloud data through the point cloud registration model FPFAM-Reg.

[0025] In a third aspect, the present invention provides an electronic device comprising a memory, a processor and a computer program stored in the memory, wherein the processor implements the steps of the above method when executing the program.

[0026] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.

[0027] Beneficial effects of the present invention:

[0028] 1. The model proposed in this paper constructs a Fast Point-to-Feature Attention (FPFA) module, which integrates pose-independent feature representation and an adaptive attention mechanism. Leveraging a dual-path architecture, FPFA achieves both point-to-point geometric relationship encoding and feature interaction enhancement. Compared to traditional modules, FPFA significantly improves geometric perception while reducing computational overhead. It successfully balances computational efficiency with feature discrimination, significantly accelerating point cloud registration and improving registration accuracy, alleviating the problem of low registration efficiency in industrial production.

[0029] 2. This paper introduces the Inception deep convolution module, integrating multi-scale feature fusion with deep convolution optimization. By partitioning the feature map along the channel dimension and performing different convolution operations, it effectively captures long-range dependencies in the point cloud. While reducing computational costs, it maintains the model's spatial perception capabilities and addresses the shortcomings of traditional methods in multi-scale feature extraction.

[0030] 3. To address occlusion and noise issues in point cloud registration, this module focuses on key geometric features through a cross-frame attention mechanism. The present invention constructs a lightweight cross-frame attention aggregation module (LCAMM), which combines the geometry-guided attention mechanism with a parameter sharing strategy to simplify the computational process, enhance the model's robustness in complex scenarios, and improve matching accuracy.

[0031] In summary, the point cloud registration method based on the fast point-to-feature attention module proposed in the present invention effectively solves the problems of accuracy and efficiency faced by point cloud registration in complex industrial scenes. The present invention meets the requirements of industrial production and other fields for high-precision and real-time point cloud registration, and has important practical value and promotion significance for industrial scenes with limited computing resources and complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is the FPFAM-Reg network architecture diagram of the present invention;

[0033] Figure 2 This is a structural diagram of the encoder of the present invention;

[0034] Figure 3 This is a structural diagram of the decoder of the present invention;

[0035] Figure 4 This is the structure diagram of the fast point-pair feature attention module (FPFA) of the present invention;

[0036] Figure 5 This is the structural diagram of the Inception deep convolution module (IDCM) of the present invention;

[0037] Figure 6 This is the structure diagram of the lightweight cross-frame attention aggregation module (LCAMM) of the present invention;

[0038] Figure 7 This is a structural diagram of the dual-stage matching matrix calculation module (TSMM) of the present invention. DETAILED DESCRIPTION

[0039] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0040] The present invention provides a point cloud registration method based on a fast point-to-feature attention module, comprising the following steps:

[0041] Step 1: Data Collection and Preprocessing: Two 3D cameras are used to collect point cloud data of the workpiece in the same scene from different angles. Because the actual collected point cloud data may contain noise, outliers, and scale differences, it needs to be preprocessed to obtain the source point cloud P and target point cloud Q.

[0042] The preprocessing process involves first removing noise points using statistical filtering. By calculating statistical information for each point in the point cloud and its surroundings, points that deviate from the normal range are considered noise points and removed. Next, normalization is performed to unify the coordinate range of the point cloud data to the [0, 1] interval to eliminate scale differences between different point cloud data and ensure consistency in subsequent processing.

[0043] Step 2: Construct the point cloud registration model FPFAM-Reg and register the source point cloud P and the target point cloud Q. Figure 1 As shown in Figure 3, the model includes an encoder, a decoder, a lightweight cross-frame attention aggregation module (LCAMM), and a two-stage matching matrix calculation module (TSMM).

[0044] The encoder contains three encoder blocks with the same structure, such as Figure 2 Taking one of the encoder blocks as an example, the encoder block includes a downsampling layer, an Inception deep convolution module IDCM, a first fast point-to-feature attention module FPFA, a second fast point-to-feature attention module FPFA, a normalization layer, an addition layer, and an activation function layer, which are cascaded in sequence.

[0045] In this encoder block, the input point cloud is first downsampled using the Farthest Point Sampling (FPS) method. Assuming the original input point cloud contains tens of thousands of points, the FPS method can downsample it to thousands of points, ensuring that the sampled points are evenly distributed throughout the point cloud, making the sampled point cloud more robust to noise and density variations.

[0046] The downsampled point cloud enters the Inception deep convolution module (IDCM), such as Figure 5 As shown in Figure 2, IDCM divides the input point cloud data into four groups in the channel dimension through the Split layer, namely:

[0047] X hw ,X w ,X h ,X id =Split(X) #Formula (4)

[0048] Among them, X is the input point cloud data, X hw ,X w ,X h Used for different convolution branches, X id Directly retain the original features to reduce computational redundancy.

[0049] After being divided into four groups, they enter different depth-separable convolutional layers for feature extraction. The multi-branch feature extraction process is as follows:

[0050] Y hw =DWConv 3×3 (X hw ) # Formula (5)

[0051] Y w =DWConv 1×7 (X w ) #Formula (6)

[0052] Y h =DWConv 7×1 (X h ) #Formula (7)

[0053] Y id =X id #Formula (8)

[0054] Among them, DWConv is a depth-wise separable convolution. Its core idea is to decompose the standard convolution into two independent operations: depth-wise convolution and point-wise convolution. The former uses a single-channel convolution kernel to learn features in the spatial dimension, and the latter uses 1×1 convolution to achieve cross-channel feature fusion. hw 、Y w 、Y h 、Y id Concatenate (Concat), then normalize (Norm) and input into the multi-layer perceptron (MLP) for processing, and fuse with the initial input point cloud data to obtain multi-scale point cloud features; the multi-scale point cloud features include multiple anchor points That is, the center point of the feature needs to be updated; the anchor point The corresponding normal vector is The anchor point With support point set {P S}.

[0055] After being processed by the IDCM module, the multi-scale point cloud features enter two serial fast point pair feature attention modules FPFA, such as Figure 4 As shown in Figure 2, the first FPFA accepts the outputs from the IDCM and downsampling. In the FPFA module, point-pair relationships are constructed within a local spherical neighborhood. A relative coordinate encoding strategy is used to capture fine-grained geometric associations. A separable spatial attention layer efficiently extracts inter-point structural information. A feature reuse mechanism is then used to generate a robust, rotation-invariant geometric descriptor.

[0056] Specifically, the FPFA module includes a Pose-Agnostic layer, a first linear layer, an attention layer, a second linear layer, an Add&Norm layer, and a third linear layer connected in series;

[0057] The Pose-Agnostic layer provides rotation-invariant PPF (Point Pair Feature) coordinates through local geometric calculations. and its supporting points Defined as:

[0058]

[0059] Among them, ||d||2 is the neighborhood point and anchor points The Euclidean distance between two points, is the normal vector of the anchor point, Neighborhood points The normal vector of the point is ∠(·), and ∠(·) is the vector angle function. In the geometric encoding based on point pair features, in addition to the distance between points, the angle information between the point and the normal vector is also used to ensure that a stable local geometric description can be obtained under rotation or other posture changes. Since the four components of the PPF coordinates are completely based on the local relative geometric relationship and have nothing to do with the global coordinate system, they are naturally rotationally invariant.

[0060] In addition, the support points Perform feature extraction and obtain its feature vector The neighborhood feature matrix and

[0061] The output of Pose-Agnostic Input to a linear layer (Linear), get E S , that is, the PPF coordinate matrix of all support points; and Input to another linear layer to get x S and X S ; E is then output by the linear layerS 、x S 、X S Input into the attention layer (Attention).

[0062] In the attention layer, the attention mechanism uses five learnable matrices W G 、W E 、W Q 、W K and W V To project the input, coordinate embedding and projection decompose the PPF into:

[0063] G=E S W G , E=E S W E #Formula (10)

[0064] in, is the learnable projection matrix, and c is the feature dimension. Then the pose-independent attention weight calculation is performed, and the attention weight is generated by fusing geometric clues and context features:

[0065]

[0066] in, is the query vector projected by the anchor feature; K = X S W k is the key vector of the support point feature projection; are learnable parameters.

[0067] After obtaining the message m through the attention layer, message transmission and feature update are performed. The message m is composed of geometric clues and context value V = X S W V Weighted fusion, after layer normalization and linear projection, updates the anchor features.

[0068] The second FPFA further enhances the features.

[0069] After that, the features are normalized through the normalization layer (Norm), and the batch normalization (BatchNormalization) technology is used to stabilize the training process and accelerate convergence. The output of the first FPFA is then added to the output after the Norm process through the residual connection (Add), retaining the relatively original feature information and avoiding the gradient vanishing problem. Finally, the ReLU activation function is used to introduce nonlinearity to enhance the expressive power of the model. In the same way, after processing through the three encoder blocks, the high-level structural features P' and Q' of the point cloud data are obtained. and Input two lightweight cross-frame attention aggregation modules (LCAMM) respectively, and at the same time, and Enter the decoder.

[0070] The LCAMM structure is as follows Figure 6 As shown. LCAMM performs linear projection on the input features to generate query, key, and value vectors, using a shared weight matrix W Q 、W K and W V , reducing the number of parameters and improving computational consistency. A geometric descriptor E based on point pair features is introduced to inject rotation invariance into attention calculation. Attention weight a′ is calculated by cross-frame cross-attention calculation:

[0071]

[0072] To input For example, Perform local feature extraction to obtain the source frame P' S and target frame P' T , the weight matrix is ​​normalized by Softmax and the value vector V of the target frame T The weighted sum is performed to generate the source frame enhanced feature X. Then, through the residual connection and MLP layer, the final source frame enhanced feature is obtained.

[0073] Follow the same steps to obtain Features extracted in the encoder stage and The super-point features it outputs simultaneously fuse local details and global semantics to form a feature representation that is robust to rotation, translation and partial occlusion.

[0074] The decoder consists of three identical decoder blocks, such as Figure 3 As shown in Figure 2. The decoder block has the same structure as the encoder block, but replaces downsampling with upsampling, and upsampling uses the nearest neighbor interpolation method. The decoder receives features from different levels of the encoder through skip connections and gradually restores the abstract features output by the encoder to the target space. After processing by three decoder blocks, the output and

[0075] Finally, and The input is based on a two-stage matching matrix calculation module (TSMM), such as Figure 7As shown in the figure. In the coarse matching stage, hierarchical abstraction is performed to generate super points covering key areas. Through the farthest point sampling (FPS), tens of thousands of original point clouds are downsampled to hundreds of super points. Specifically, for the super point set of two frames of point clouds and Calculate the matrix S∈R M×N , where each element S i,j Represents the feature similarity of super points:

[0076]

[0077] in, For super point The characteristic matrix of For super point The feature matrix of .

[0078] To alleviate mismatching caused by scene symmetry or repeated structures, double Softmax normalization is used to normalize rows and columns independently:

[0079]

[0080] Among them, Softmax j (S i,: ) is the normalization of row i, Softmax i (S :,j ) is the normalization of column j;

[0081] Then, matching pair screening and geometric pre-verification are performed, and the normalized matrix To screen high-confidence matching pairs, a dynamic Top-K strategy is usually adopted to adaptively select the number of matches based on the point cloud overlap rate.

[0082]

[0083] in, The sparse matching pairs are output by coarse matching. Coarse matching achieves the goal of quickly extracting robust correspondences from massive point clouds through super-point abstraction, rotation-invariant feature encoding, dual normalization, and geometric verification.

[0084] In the fine matching stage, the rotation invariant features output by the decoder in FPFAM-Reg are accepted and sparse matching pairs generated by coarse matching For each rough matching pair (s p ,s q ), extract its neighborhood point set in the original point cloud. China-Israel p The point within the radius r is represented by China-Israelq The point within the radius r is represented by For local areas and Each point pair (p k ,q l ) Calculate feature similarity:

[0085]

[0086] To enhance the matching robustness, Softmax normalization is performed on the points in the rows and columns respectively:

[0087]

[0088] Use the Sinkhorn algorithm to optimize the matching and transform the similarity matrix C′ into a double random matrix by iterative row and column normalization (usually 3 to 5 times) satisfy:

[0089]

[0090] choose The points are taken as candidate matches to generate a preliminary matching set Where τ is the threshold for internal point determination, which is usually set to 0.1. Based on the purified matching pairs, the refined rigid transformation is solved, and the weighted least squares optimization framework is adopted to construct the covariance matrix and solve the optimal rotation matrix R and translation vector t through singular value decomposition (SVD). For each matching pair Compute the decentralized coordinates and construct the covariance matrix:

[0091]

[0092] in, and are the centroids of the source and target points respectively. After SVD decomposition of H, we get R = VU T , and then through Calculate the translation amount and finally synthesize the refined transformation matrix Apply this transformation to the source point cloud You can get the target point cloud High-precision alignment results

[0093] In order to further verify the effectiveness of the point cloud registration model FPFAM-Reg of the present invention, a comparative experiment was conducted to compare the model FPFAM-Reg with the mainstream methods in the field of point cloud registration in recent years. The comparison methods include SpinNet, RoReg, D3Feat, Predator, CoFiNet and CorsNet. The datasets used include the 3DMatch dataset, the 3DLoMatch dataset and the self-built workpiece dataset. The 3DMatch dataset integrates RGB-D data of multiple indoor and outdoor scenes, containing more than 1,500 scene fragments, and the overlap rate between point clouds is generally high (more than 30%). The 3DLoMatch dataset is an extended version of 3DMatch, specifically for the problem of low overlap point cloud registration. The overlap rate of its test set is between 10% and 30%, and complex scenes are introduced to increase the difficulty. The self-built workpiece dataset was collected from the Phase III workshop of a certain elevator company in Hangzhou. It contains 3,800 3D models belonging to five categories (gaskets, bent parts, U-shaped cards, stainless steel angle codes, and stainless steel plates). Among them, 150 are used for verification, 200 for testing, and the rest for training.

[0094] In the experiment, three evaluation metrics were used to measure model performance: feature match recall (FMR), inlier ratio (IR), and registration recall (RR). FMR measures the descriptive power of features. A higher FMR indicates that the feature descriptors extracted by the algorithm are more discriminative and can effectively match more true corresponding point pairs. IR measures the accuracy of correspondences. A higher IR indicates a more reliable matched pair, and subsequent pose calculations using RANSAC or SVD are more likely to yield accurate results. RR measures the final registration performance, i.e., the percentage of point pairs that the registration algorithm can correctly match. A high RR indicates that the algorithm is robust in complex scenarios (such as low overlap, noise interference, and occlusion).

[0095] Table 1 Performance of different models in registration recall (RR) on the 3DMatch dataset

[0096]

[0097] As shown in Table 1, the FPFAM-Reg model demonstrates significant advantages on the 3DMatch dataset. In a fine-grained comparison of scenes, in the Home2 scene, FPFAM-Reg achieves a 77.7% RR, significantly outperforming SpinNet and Predator, highlighting its adaptability to complex spatial structures. In the Kitchen scene, FPFAM-Reg achieves a 97.8% RR, only 0.3% lower than CoFiNet, demonstrating competitiveness with state-of-the-art methods. In terms of cross-scene stability, FPFAM-Reg achieves RRs of 98.3% and 95.8% in the Home1 and Hotel1 scenes, respectively, ranking in the top three. In the Lab scene, FPFAM-Reg achieves a RR of 82.2%, significantly outperforming D3Feat's 70.5% and CorsNet's 71.6%, demonstrating its robustness in environments with low texture and weak geometric features. Overall, FPFAM-Reg achieves an average RR of 90.2% across all scenes, second only to CoFiNet's 90.8%.

[0098] Table 2 Comparison of registration recall (RR) of different algorithms on 3DMatch dataset at different corresponding sampling points

[0099]

[0100]

[0101] As shown in Table 2, FPFAM-Reg also performs exceptionally well on the 3DMatch dataset at different sampling points. At 2500 sampling points, FPFAM-Reg achieves a RR of 93.3%, significantly outperforming RoReg's 92.1%, SpinNet's 91.1%, and CoFiNet's 87.8%, demonstrating its excellent balance between feature matching efficiency and accuracy. At a low sampling point of 500, FPFAM-Reg maintains a RR of 80.8%, significantly higher than SpinNet's 72.7% and Predator's 62.3%, demonstrating its robustness under limited information conditions. Although FPFAM-Reg achieves a RR of 70.6% at the extremely low sampling point of 250, lower than CoFiNet's 75.8%, the decrease is significantly smaller than that of D3Feat and SpinNet, further demonstrating the high stability of the algorithm design.

[0102] Table 3 Comparison of FMR, IR, and RR of different models on 3DMatch and 3DLoMatch datasets

[0103]

[0104] As shown in Table 3, FPFAM-Reg also performed outstandingly across multiple key metrics on the 3DLoMatch dataset. In terms of the FMR metric, FPFAM-Reg achieved 80.1% on the 3DLoMatch dataset. While lower than CoFiNet's 83.7%, it outperformed CorsNet's 70.7% and Predator's 78.3%, demonstrating its strong feature representation capabilities in small overlapping region matching tasks. In terms of the IR metric, FPFAM-Reg achieved 38.9% on the 3DLoMatch dataset, significantly surpassing all compared methods and increasing by 1.6 percentage points from the second-place Predator's 37.3%. This demonstrates its innovative approach to noise suppression and geometric consistency, providing more reliable geometric constraints for complex 3D registration tasks.

[0105] In terms of RR, FPFAM-Reg achieved 61.1% on the 3DLoMatch dataset, lower than CoFiNet's 67.9%. However, a comprehensive analysis of the IR metric shows that FPFAM-Reg trades a small amount of registration success rate for higher inlier quality. This trade-off strategy is more valuable in practical applications where registration accuracy is crucial. In segmented scenarios, FPFAM-Reg achieved RR of 68.9% and 66.1% in the Hotel2 and Hotel3 scenarios, respectively, significantly outperforming SpinNet and CoFiNet, demonstrating its powerful ability to represent local geometric structures in complex indoor environments. In the Lab scenario, FPFAM-Reg achieved RR of 51.9%, an improvement of over 10 percentage points over SpinNet's 41.4% and Predator's 39.7%, fully demonstrating the model's robustness in low-overlap scenarios.

[0106] Table 4 Comparison of FMR, IR, and RR of different models on the self-built artifact dataset

[0107]

[0108] As shown in Table 4, FPFAM-Reg demonstrates significant competitive advantages in overall performance on a self-built artifact dataset. In terms of the FMR metric, FPFAM-Reg takes the lead with an outstanding 98.9%, surpassing the 98.5% achieved by CoFiNet, the best competing method, by 0.4 percentage points, demonstrating its robustness in complex artifact feature extraction and matching tasks. In terms of the inlier ratio metric, FPFAM-Reg ranks second with a score of 63.5%. While lower than Predator's 66.2%, it still holds a significant advantage over other models. In terms of registration recall, FPFAM-Reg demonstrates good scene adaptability with a score of 88.3%. Although CoFiNet's 89.3% and Predator's 88.9% perform better in this metric, a comprehensive analysis of both FMR and IR demonstrates that FPFAM-Reg achieves a balanced performance across key metrics while maintaining stable registration recall.

[0109] Table 5 Performance of different models on the self-built artifact dataset in terms of RR index

[0110]

[0111] As shown in Table 5, FPFAM-Reg achieves the best performance among all models in the steel plate registration task, achieving a 98.3% RR (Reference Rate) score, a 0.8 percentage point improvement over the second-place Predator (97.5%), demonstrating its significant advantage in extracting geometric features from planar, large-scale workpieces. In the U-shaped card category, FPFAM-Reg leads all other methods with a 93.9% RR (Reference Rate), a 2.6% improvement over Predator's 91.3% and a 1.8% improvement over CoFiNet's 92.1%, respectively. This demonstrates the model's robustness to local features of curved surfaces. In the gasket inspection task, FPFAM-Reg ranks third with a 95.6% RR (Reference Rate), only 0.9 percentage points behind the best-performing method, RoReg (96.5%), maintaining its competitive edge. Although FPFAM-Reg did not reach the optimal value on two types of irregular workpieces, namely bent parts and corner codes, it achieved efficient feature aggregation by introducing a lightweight cross-frame attention module, and its performance was still better than SpinNet and D3Feat in the comparative experiments.

[0112] The above are merely preferred embodiments of the present invention. It should be noted that the above preferred embodiments should not be construed as limiting the present invention, and the scope of protection of the present invention should be determined by the scope defined in the claims. Persons skilled in the art will appreciate that improvements and modifications may be made without departing from the spirit and scope of the present invention, and such improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A point cloud registration method based on a fast point-to-feature attention module, characterized in that: The following steps are involved: Collect point cloud data, pre-process the point cloud data, and obtain source point cloud P and target point cloud Q; Construct the point cloud registration model FPFAM-Reg to register the source point cloud P and the target point cloud Q; The point cloud registration model FPFAM-Reg includes an encoder, a decoder, a lightweight cross-frame attention aggregation module LCAMM, and a two-stage matching matrix calculation module TSMM. The encoder extracts features from the source point cloud P and the target point cloud Q and converts them into high-level structural features. and Then input them into the lightweight cross-frame attention aggregation module LCAMM and the decoder respectively; The lightweight cross-frame attention aggregation module LCAMM integrates the high-level feature information of point clouds from different frames extracted by the encoder and outputs enhanced features and During the decoding process, the decoder directly transfers the features of different levels of the encoder to the corresponding layers of the decoder through skip connections, and Gradually restore to the target space and input the restored point cloud features and The dual-stage matching matrix calculation module TSMM receives enhanced features and And point cloud features and The point cloud registration is optimized through coarse matching and fine matching, and the final point cloud registration result is output.

2. The point cloud registration method according to claim 1, characterized in that: The encoder includes three identical encoder blocks, each of which downsamples the input point cloud data, then extracts multi-scale features through the Inception deep convolution module IDCM, and then enhances feature interaction through two serially connected fast point feature attention modules FPFA. Finally, the features are standardized through a normalization layer, and the original feature information is retained using a residual connection. The nonlinearity is then introduced through the ReLU activation function to obtain the output of each encoder block.

3. The point cloud registration method according to claim 2, characterized in that: The decoder includes three identical decoder blocks, each of which includes an upsampling layer, an Inception deep convolution module IDCM, two fast point-to-feature attention modules FPFA connected in series, a normalization layer, a residual connection and an activation function layer.

4. The point cloud registration method according to claim 1, wherein: The lightweight cross-frame attention aggregation module LCAMM performs linear projection on the input features to generate query, key, and value vectors, using a shared weight matrix. It introduces a geometric descriptor E based on point pair features to inject rotation invariance into the attention calculation. Through cross-frame cross-attention calculation, the query vector of the source frame is multiplied by the key vector of the target frame, and the bias term of the geometric descriptor is added to calculate the attention weight a'. After the weight matrix is ​​normalized by Softmax, it is weighted and summed with the value vector of the target frame, and then passed through the residual connection and MLP layer to obtain the final source frame enhanced features. The calculation formula of the attention weight is as follows:

5. The point cloud registration method according to claim 1, characterized in that: In the two-stage matching matrix calculation module TSMM, the coarse matching process includes the following steps: Perform hierarchical abstraction on the original point cloud to generate super points covering key areas and perform downsampling. Then calculate the similarity between super point features and use double Softmax normalization to alleviate mismatching. Use dynamic Top-K strategy to filter matching pairs and obtain sparse matching pairs of coarse matching output.

6. The point cloud registration method according to claim 5, characterized in that: In the two-stage matching matrix calculation module TSMM, the fine matching process includes the following steps: The fine matching is limited to the local area corresponding to the coarse matching to reduce the amount of calculation; the feature similarity is calculated for the point pairs in the local area, and Softmax normalization is performed. The Sinkhorn algorithm is used for matching optimization to screen out candidate matching pairs; based on the purified matching pairs, a weighted least squares optimization framework is used to solve the optimal rotation matrix R and translation vector t through singular value decomposition (SVD) to synthesize the refined transformation matrix 7. The point cloud registration method according to claim 1, characterized in that: The following loss function is used in the training process of the point cloud registration model: Among them, Ls is the loss function of coarse matching, and the formula is as follows: in is the sparse matching pair generated by coarse matching, S(i,j) is the normalized similarity score of the superpoint pair (i,j), and N is the number of superpoints; Lp is the loss function formula for fine matching as follows: in is the preliminary matching set, (x, y) is the preliminary matching superpoint pair, To correctly match point pairs, T(x) is the coordinate of point x after applying the predicted rigid body transformation, τ is the inlier threshold, λ is a hyperparameter used to control the balance between loss values; α is the weight for balancing geometric error and inlier ratio.

8. A point cloud registration system for implementing the method according to any one of claims 1 to 7, characterized in that: include: Data acquisition and preprocessing module, responsible for collecting point cloud data and preprocessing; The registration processing module is used to register the pre-processed point cloud data using the point cloud registration model FPFAM-Reg.

9. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 7.

10. A computing device comprising a memory and a processor, characterized in that: The memory stores executable code, and when the processor executes the executable code, the method according to any one of claims 1 to 7 is implemented.

Images