A point cloud registration method based on a fast point pair feature attention module

Abstract

The application discloses a point cloud registration method based on a fast point pair feature attention module. The method first collects point cloud data, and then matches the point cloud by using a point cloud registration model. The point cloud registration model is based on an encoder-decoder architecture, and fuses a fast point pair feature attention module FPFA, an Inception deep convolution module IDCM, a lightweight cross-frame attention aggregation module LCAAM and a two-stage-based matching matrix calculation module TSMM to obtain a final point cloud registration result. Through comparative experiments and ablation experiments on public data sets and self-built workpiece data sets, it is verified that the method has excellent accuracy, speed and robustness, especially in low overlap rate scenarios and industrial workpiece registration, and provides an efficient solution for factory pipeline automation and precise mechanical arm grabbing, and has important application value for industrial detection scenarios with limited computing resources.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of point cloud processing, and particularly relates to a point cloud registration method based on a fast point pair feature attention module, which is suitable for scenes such as industrial manufacturing, robot vision, three-dimensional reconstruction, etc. that need to process point cloud data. BACKGROUND

[0002] In modern industrial production, scientific research exploration and many frontier technical fields, point cloud data processing has become a key link to realize accurate operation and in-depth analysis. As one of the core technologies of point cloud processing, point cloud registration aims to accurately align point cloud data obtained from different perspectives, different times or different acquisition devices, and construct a complete and accurate three-dimensional model, thereby providing a reliable data basis for subsequent operation and analysis. In the industrial manufacturing field, especially in the factory assembly line production scene, the importance of point cloud registration is particularly prominent. With the acceleration of manufacturing towards intelligence and automation, robotic arms are taking on more and more fine operation tasks on the production line, such as workpiece grabbing, part assembly, etc. However, many complex factors in the actual production environment bring great challenges to point cloud data acquisition and registration. In order to overcome these problems, in recent years, automated technologies based on deep learning have been gradually introduced to realize more intelligent and efficient leather production.

[0003] From the perspective of lighting conditions, the lighting in the factory is often uneven, with insufficient lighting in some areas, resulting in missing information in the collected point cloud data, making it difficult to accurately reflect the true geometric shape of the workpiece; and on the surface of some metal workpieces, due to their reflective properties, the collected point cloud will have noise and deviation, further interfering with the accuracy of the point cloud data. At the same time, the placement of workpieces on the assembly line is usually chaotic and there is occlusion between them, which makes the geometric structure of the original point cloud collected seriously incomplete. In the field of robot navigation and positioning, point cloud registration is also a key technology to realize autonomous navigation and accurate operation of robots. When mobile robots work in complex indoor or outdoor environments, they need to collect point cloud data of the surrounding environment through devices such as laser radars, and register point clouds collected at different times to construct an accurate map and determine their own position and attitude. However, factors such as dynamic object interference in the actual environment, scene similarity and sensor error make point cloud registration face many difficulties. In the field of three-dimensional reconstruction, point cloud registration is the basis for constructing high-quality three-dimensional models. However, in actual operation, due to the precision limitations of acquisition devices, the complexity of the acquisition environment and the similarity of object surface textures, the accuracy and efficiency of point cloud registration often cannot meet the needs. SUMMARY

[0004] The application provides a point cloud registration method based on a fast point pair feature attention module, which improves the registration speed while ensuring the registration accuracy, so as to balance the accuracy and efficiency.

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

[0006] Collecting point cloud data, preprocessing the point cloud data, obtaining source point cloud P and target point cloud Q;

[0007] Building a point cloud registration model FPFAM-Reg, registering 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-based 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 structure 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 different frame point clouds extracted by the encoder, and outputs enhanced features

[0011] The decoder gradually recovers the enhanced features and the high-level structure features to the target space, and inputs the recovered point cloud features and

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

[0013] Further, the encoder includes three identical encoder blocks, each of which down-samples the input point cloud data, then extracts multi-scale features through an Inception deep convolution module IDCM, then performs feature interaction enhancement through two serial fast point pair feature attention modules FPFA, finally normalizes the features through a normalization layer, uses residual connection to retain the original feature information, and then introduces nonlinearity through a ReLU activation function to obtain the output of each encoder block.

[0014] Further, the decoder comprises three identical decoder blocks, each of which comprises, in sequence, an up-sampling 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 a skip connection.

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

[0016]

[0017] Ls is a loss function of coarse matching, used to solve the global matching problem of point clouds at a coarse granularity, and the formula is as follows:

[0018]

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

[0020]

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

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

[0023] A data acquisition module is configured to acquire point cloud data and perform preprocessing;

[0024] A point cloud registration module is configured to perform registration on the preprocessed point cloud data by using a point cloud registration model FPFAM-Reg.

[0025] In a third aspect, the present application 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 application provides a computer readable storage medium having a computer program stored thereon, wherein the program implements the steps of the above method when executed by a processor.

[0027] The present application has the following advantages:

[0028] 1. The model provided by the application constructs a fast point-to-feature attention module (FPFA), which fuses pose-independent feature representation and adaptive attention mechanism, and realizes point-to-geometry relationship coding and feature interaction enhancement through a double-path architecture. Compared with traditional modules, FPFA significantly improves the geometric perception ability while reducing the calculation, successfully balances the calculation efficiency and feature discrimination, greatly speeds up the point cloud registration speed, improves the registration accuracy, and solves the problem of low registration efficiency in industrial production.

[0029] 2. The application introduces an Inception deep convolution module, which combines multi-scale feature fusion and deep convolution optimization. By dividing the feature map in the channel dimension and performing different convolution operations, the long-distance dependence of the point cloud is effectively captured. While reducing the calculation cost, the spatial perception ability of the model is maintained, making up for the shortcomings of traditional methods in multi-scale feature extraction.

[0030] 3. For the problem of occlusion and noise in point cloud registration. This module focuses on key geometric features through cross-frame attention mechanism, and the application constructs a lightweight cross-frame attention aggregation module (LCAMM), which combines geometric-guided attention mechanism and parameter sharing strategy to simplify the calculation process and enhance the robustness of the model in complex scenes, improving the matching accuracy.

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

[0032] Figure 1 The FPFA-Reg network architecture diagram of the application;

[0033] Figure 2 The encoder structure diagram of the application;

[0034] Figure 3 The decoder structure diagram of the application;

[0035] Figure 4 The fast point-to-feature attention module (FPFA) structure diagram of the application;

[0036] Figure 5 The Inception deep convolution module (IDCM) structure diagram of the application;

[0037] Figure 6 The lightweight cross-frame attention aggregation module (LCAMM) structure diagram of the application;

[0038] Figure 7 Fig. 1 is a structural diagram of a two-stage matching matrix calculation module (TSMM) according to the present application. DETAILED DESCRIPTION

[0039] The present application will be further described in conjunction with the embodiments and the accompanying drawings.

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

[0041] Step 1, data acquisition and preprocessing: using two three-dimensional cameras to collect workpiece point cloud data under the same scene from different angles. Since the actually collected point cloud data may contain noise, outliers and scale differences, etc., it needs to be preprocessed to obtain the source point cloud P and the target point cloud Q.

[0042] The preprocessing is specifically: first, a statistical filtering method is used to remove noise points, by calculating the statistical information of each point and its neighborhood points in the point cloud, the points deviating from the normal range are regarded as noise points and are removed. Then, normalization operation is performed to unify the coordinate range of the point cloud data to the [0, 1] interval, so as to eliminate the differences in scale of different point cloud data and ensure the consistency of subsequent processing.

[0043] Step 2, constructing a point cloud registration model FPFAM-Reg and registering the source point cloud P and the target point cloud Q. As shown in Figure 1 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 includes three encoder blocks with the same structure, as shown in Figure 2 Taking one of the encoder blocks as an example, the encoder block includes a down-sampling layer, an Inception deep convolution module IDCM, a first fast point pair feature attention module FPFA, a second fast point pair feature attention module FPFA, a normalization layer, an addition and activation function layer, which are sequentially cascaded.

[0045] In this encoder block, first, the input point cloud is down-sampled by the farthest point sampling (FPS) method. Assuming that the original input point cloud contains tens of thousands of points, it can be down-sampled to thousands of points by the FPS method, ensuring that the sampling points are uniformly distributed in the entire point cloud, making the sampled point cloud have stronger robustness to noise and density changes.

[0046] The down-sampled point cloud enters the Inception deep convolution module (IDCM), as shown in Figure 5 The IDCM divides the input point cloud data into four groups in the channel dimension through the Split layer, i.e.:

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

[0048] Where X is the input point cloud data, X hw ,X w ,X h Used for different convolution branches, X id The original features are directly preserved to reduce computational redundancy.

[0049] After being divided into four groups, each group enters a separable convolutional layer of different depths 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] DWConv, or Depthwise Separable Convolution, decomposes standard convolution into two independent operations: depthwise convolution and pointwise convolution. The former learns features in the spatial dimension using a single-channel convolution kernel, while the latter uses a 1×1 convolution to achieve cross-channel feature fusion. Then, Y... hw Y w Y h Y id The data is concatenated, normalized, and then input into a multilayer perceptron (MLP) for processing. This process is then fused with the initial input point cloud data to obtain multi-scale point cloud features. These multi-scale point cloud features contain multiple anchor points. That is, the center point of the feature that needs to be updated; anchor point. The corresponding normal vector is The anchor point Having a support point set {P S}

[0055] After processing by the IDCM module, the multi-scale point cloud features enter two serial fast point pair feature attention modules FPFA, as shown in Figure 4 The first FPFA accepts output from the IDCM and downsampling. In the FPFA module, point pair relationships are constructed within a local spherical neighborhood, fine-grained geometric correlations are captured using a relative coordinate encoding strategy, inter-point structural information is efficiently extracted through a separable spatial attention layer, and robust, rotation-invariant geometric descriptors are generated in combination with a feature reuse mechanism.

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

[0057] The Pose-Agnostic layer provides rotation-invariant PPF (Point Pair Feature) coordinates through local geometric calculation. For an anchor point and its support points , the PPF coordinates are defined as:

[0058]

[0059] where ||d||2 is the Euclidean distance between the neighborhood point and the anchor point , is the normal vector of the anchor point, is the normal vector of the neighborhood point , and ∠(·) is a vector angle function. In point pair feature-based geometric coding, in addition to the distance between points, the angle between a point and a normal vector is also used to ensure stable local geometric description under rotation or other attitude changes. Since the four components of the PPF coordinates are completely based on local relative geometric relationships and are independent of the global coordinate system, the PPF coordinates naturally have rotation invariance.

[0060] In addition, feature extraction is performed on the support points to obtain their feature vectors , and the neighborhood feature matrix and

[0061] The output of the Pose-Agnostic layer is input to a linear layer (Linear) to obtain E S , i.e., the PPF coordinate matrix of all support points; and and are input to another linear layer to obtain x S and X S ; subsequently, ES , x S , X S in the input 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, and the coordinate embedding and projection decompose the PPF into:

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

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

[0065]

[0066] where, is the query vector of the anchor feature projection; K = X S W k is the key vector of the support feature projection; is a learnable parameter.

[0067] After obtaining the message m through the attention layer, message passing and feature updating are performed, and the message m is weightedly fused by the geometric cues and the context value V = X S W V After layer normalization and linear projection, the anchor feature is updated.

[0068] The second FPFA further enhances the features.

[0069] After that, the features are standardized through the normalization layer (Norm), and the batch normalization (Batch Normalization) technology is used to stabilize the training process and accelerate the convergence. Then, the output of the first FPFA is added to the output after the Norm processing through the residual connection (Add), the relatively original feature information is retained, the gradient vanishing problem is avoided, and finally the ReLU activation function is introduced to introduce nonlinearity and enhance the expression ability of the model. In the same way, after processing through three encoder blocks, the high-level structure features P' and Q' of the point cloud data are obtained, and and Input two lightweight cross-frame attention aggregation modules (LCAMM) respectively, and simultaneously... and Input decoder.

[0070] LCAMM structure 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 This reduces the number of parameters and improves computational consistency. A geometric descriptor E based on point-pair features is introduced to inject rotation invariance into the attention computation. The attention weight a′ is calculated through cross-frame cross-attention computation:

[0071]

[0072] Input For example, for Perform local feature extraction to obtain the source frame P' S and target frame P' T After the weight matrix is ​​normalized by Softmax, it is compared with the value vector V of the target frame. T Weighted summation generates the source frame enhancement feature X. Then, through residual connections and an MLP layer, the final source frame enhancement feature is obtained.

[0073] Obtain by following the same steps Features extracted during the encoder stage and Its output super point features simultaneously integrate local details and global semantics, forming 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. The decoder block has a similar structure to the encoder block, but it replaces downsampling with upsampling, using the nearest neighbor interpolation method. The decoder receives features from different levels of the encoder through skip connections, gradually reconstructing the abstract features output by the encoder into the target space. After processing by three decoder blocks, the output is... and

[0075] Finally, and Input is based on the two-stage matching matrix calculation module (TSMM), such as Figure 7The hierarchical abstraction is performed in the coarse matching stage to generate super-points covering the key regions. By farthest point sampling (FPS), tens of thousands of original point clouds are down-sampled to hundreds of super-points. Specifically, for the super-point set of two frames of point clouds and The matrix S ∈ R M×N is calculated, where each element S i,j represents the feature similarity of the super-points:

[0076]

[0077] where, is the feature matrix of the super-point , is the feature matrix of the super-point .

[0078] To alleviate the mismatch caused by scene symmetry or repeated structure, double Softmax normalization is used for independent normalization of rows and columns:

[0079]

[0080] where Softmax j (S i,: ) is the normalization of i rows, and Softmax i (S :,j ) is the normalization of j columns.

[0081] Then, the matching pair screening and geometric pre-verification are performed, and high-confidence matching pairs are screened from the normalized matrix , usually using a dynamic Top-K strategy to adaptively select the number of matches according to the point cloud overlap rate.

[0082]

[0083] where, is the sparse matching pair output by coarse matching. Through super-point abstraction, rotation-invariant feature encoding, double normalization and geometric verification, coarse matching achieves the goal of quickly extracting robust corresponding relationships from massive point clouds.

[0084] In the fine matching stage, the rotation-invariant features output by the decoder in FPFAM-Reg and the sparse matching pairs generated by coarse matching are accepted. For each coarse matching pair (s p , s q ), its neighborhood point set is extracted in the original point cloud. The points within a radius r centered on s p are represented asq For each point in the center, within the radius r, is represented as For each point in the local region and The feature similarity is calculated for each point pair (p k , q l ):

[0085]

[0086] To enhance the robustness of matching, the points in the row and column are respectively normalized by Softmax:

[0087]

[0088] The Sinkhorn algorithm is used for matching optimization, and the similarity matrix C' is converted into a doubly stochastic matrix by iteratively normalizing the row and column (usually 3 to 5 times), which satisfies:

[0089]

[0090] Select the points as candidate matches, and generate a preliminary matching set where τ is the inlier judgment threshold, usually set to 0.1. Based on the purified matching pairs, a refined rigid transformation is solved using a weighted least squares optimization framework, a covariance matrix is constructed, and the optimal rotation matrix R and translation vector t are solved by singular value decomposition (SVD). For each matching pair , the decentralized coordinates are calculated and the covariance matrix is constructed:

[0091]

[0092] where and are the centroids of the source point set and the target point set, respectively. After SVD decomposition of H, R = VU T is obtained, and the translation is calculated by , and finally the refined transformation matrix is synthesized. Using this transformation on the source point cloud , the high-precision alignment result with the target point cloud

[0093] To further verify the effectiveness of the point cloud registration model FPFAM-Reg of the application, comparative experiments are carried out 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 data sets used include the 3DMatch data set, the 3DLoMatch data set and the self-built workpiece data set. The 3DMatch data set integrates multiple indoor and outdoor scene RGB-D data, contains more than 1500 scene segments, and the overlap rate between point clouds is usually high (more than 30%). The 3DLoMatch data set is an extended version of 3DMatch, which is specifically designed for low overlap point cloud registration problems. The overlap rate of the test set is between 10% and 30%, and complex scenes are introduced to increase the difficulty. The self-built workpiece data set is collected from the third phase workshop of a certain elevator company in Hangzhou, and there are 3800 3D models belonging to 5 categories (gaskets, bent parts, U-shaped cards, stainless steel angle codes and stainless steel plates), of which 150 are used for verification, 200 are used for testing, and the rest are used for training.

[0094] In the experiment, three evaluation indexes of feature matching recall rate (FMR), inlier ratio (IR) and registration recall rate (RR) are used to measure the performance of the model. FMR is used to measure the description ability of features. The higher the FMR, the more distinctive the feature descriptors extracted by the algorithm, and more true corresponding point pairs can be effectively matched. IR is used to measure the correctness of the corresponding relationship. The higher the IR, the higher the reliability of the matching pair, and it is more likely to obtain accurate results by solving the pose through RANSAC or SVD. RR is used to measure the final registration performance, that is, the proportion of point pairs that can be correctly matched by the registration algorithm. High RR indicates that the algorithm has strong robustness in complex scenes (such as low overlap, noise interference, occlusion, etc.).

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

[0096]

[0097] As shown in Table 1, FPFAM-Reg model shows significant advantages on the 3DMatch dataset. From the scene fine-grained comparison, in the Home2 scene, the RR of FPFAM-Reg reaches 77.7%, significantly surpassing SpinNet and Predator, highlighting its adaptability to complex spatial structures. In the Kitchen scene, FPFAM-Reg has an RR of 97.8%, only 0.3% less than CoFiNet, showing comparable competitiveness with the leading methods. In terms of cross-scene stability, FPFAM-Reg has an RR of 98.3% and 95.8% in the Home1 and Hotel1 scenes, respectively, both ranking in the top three. In the Lab scene, FPFAM-Reg has an RR of 82.2%, significantly better than D3Feat's 70.5% and CorsNet's 71.6%, indicating its robustness advantage in low-texture and weak geometric feature environments. In terms of overall performance, FPFAM-Reg has an average RR of 90.2% in all scenes, only slightly lower than CoFiNet's 90.8%.

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

[0099]

[0100]

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

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

[0103]

[0104] As shown in Table 3, FPFAM-Reg also performs outstandingly on the 3DLoMatch dataset in multiple key indicators. In the FMR indicator, FPFAM-Reg reaches 80.1% on the 3DLoMatch dataset, which is lower than CoFiNet's 83.7%, but better than CorsNet's 70.7% and Predator's 78.3%, indicating that it has strong feature expression ability in the small overlap area matching task. In the IR indicator, FPFAM-Reg reaches 38.9% on the 3DLoMatch dataset, significantly surpassing all comparison methods, with an increase of 1.6 percentage points over the second-place Predator's 37.3%, which verifies its innovation in noise suppression and geometric consistency, providing more reliable geometric constraints for complex three-dimensional registration tasks.

[0105] In terms of RR, FPFAM-Reg is 61.1% on the 3DLoMatch dataset, which is lower than CoFiNet's 67.9%, but in combination with the IR indicator for comprehensive analysis, FPFAM-Reg sacrifices a small amount of registration success rate to obtain higher inlier quality, and this trade-off strategy is more valuable in practical applications that focus on registration accuracy. In the sub-scene, FPFAM-Reg achieves 68.9% and 66.1% in RR in the Hotel2 and Hotel3 scenes, respectively, significantly better than SpinNet and CoFiNet, reflecting its strong representation ability for local geometric structures in complex indoor environments. In the Lab scene, FPFAM-Reg's RR reaches 51.9%, an increase of more than 10 percentage points over SpinNet's 41.4% and Predator's 39.7%, fully verifying the model's robustness in low overlap rate scenarios.

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

[0107]

[0108] As shown in Table 4, FPFAM-Reg shows a significant competitive advantage in comprehensive performance on the self-built workpiece dataset. In the FMR index, FPFAM-Reg ranks first with an excellent performance of 98.9%, which is 0.4 percentage points higher than 98.5% of the optimal model CoFiNet in the comparative method, fully verifying its robustness in complex workpiece feature extraction and matching tasks. In the inner point ratio index dimension, FPFAM-Reg ranks second with a score of 63.5%, which is lower than 66.2% of Predator, but still has a significant advantage compared to other models. In the registration recall rate index, FPFAM-Reg shows good scene adaptability with a result of 88.3%. Although CoFiNet's 89.3% and Predator's 88.9% perform better in this indicator, the comprehensive analysis of FMR and IR shows that FPFAM-Reg balances the performance between key indicators while maintaining the stability of registration recall rate.

[0109] Table 5 shows the performance of different models on the RR index on the self-built workpiece dataset

[0110]

[0111] As shown in Table 5, from the analysis of specific workpiece categories, FPFAM-Reg reaches the optimal level of all models with an RR value of 98.3% in the steel plate registration task, which is 0.8 percentage points higher than the second place Predator's 97.5%, indicating that it has a significant advantage in geometric feature extraction capability for large-scale plane workpieces. In the U-shaped card category, FPFAM-Reg leads other methods with an RR value of 93.9%, which is 2.6% and 1.8% higher than Predator's 91.3% and CoFiNet's 92.1%, respectively, highlighting the model's good robustness to local features of curved structures. In the gasket detection task, FPFAM-Reg ranks third with an RR value of 95.6%, which is only 0.9 percentage points lower than the optimal method RoReg's 96.5%, still maintaining high competitiveness. In the two types of irregular workpieces, bending pieces and angle codes, FPFAM-Reg does not reach the optimal value, but through the introduction of a lightweight cross-frame attention module, it realizes efficient feature aggregation, and its performance is still better than SpinNet and D3Feat in the comparative experiment.

[0112] The above are only preferred embodiments of the present application, and it should be pointed out that the above preferred embodiments should not be regarded as limiting the present application, and the protection scope of the present application should be limited by the scope defined by the claims. For ordinary skilled persons in the art, several improvements and refinements can be made without departing from the spirit and scope of the present application, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. A point cloud registration method based on a fast point-to-feature attention module, characterized in that, The method comprises the following steps: Collecting point cloud data, preprocessing the point cloud data, obtaining source point cloud P and target point cloud Q; Constructing a 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 comprises an encoder, a decoder, a lightweight cross-frame attention aggregation module LCAMM and a two-stage-based matching matrix calculation module TSMM. The encoder extracts features of the source point cloud P and the target point cloud Q, and converts them into high-level structure features respectively and Then, the encoder is input into the lightweight cross-frame attention aggregation module LCAMM and the decoder respectively; the encoder includes three identical encoder blocks; each encoder block performs down-sampling on the input point cloud data, then extracts multi-scale features through an Inception deep convolution module IDCM, then performs feature interaction enhancement through two serial fast point pair feature attention modules FPFA, finally performs standardization processing on the features through a normalization layer, uses a residual connection to retain the original feature information, then introduces nonlinearity through a ReLU activation function, and obtains the output of each encoder block. The lightweight cross-frame attention aggregation module LCAMM integrates high-level feature information of different frame point clouds extracted by the encoder and outputs enhanced features and ; The decoder directly transmits the features of different levels of the encoder to corresponding layers of the decoder through skip connections in the decoding process, and restores high-level structure features and to the target space step by step, and inputs the restored point cloud features and ; the decoder comprises three identical decoder blocks, and each of the decoder blocks comprises, in sequence, an upsampling layer, an Inception depth convolution module (IDCM), two fast point pair feature attention modules (FPFA) connected in series, a normalization layer, a residual connection and an activation function layer. The two-stage based matching matrix calculation module TSMM receives the enhanced features and and point cloud features and , the point cloud registration is optimized through coarse matching and fine matching, and a final point cloud registration result is output.

2. The point cloud registration method of claim 1, wherein, The lightweight cross-frame attention aggregation module LCAMM performs linear projection on input features to generate query, key and value vectors, adopts a shared weight matrix, introduces a point pair feature-based geometric descriptor E to inject rotation invariance into attention calculation, and calculates attention weights by cross-frame cross-attention calculation, in which the query vector of a source frame is multiplied by the key vector of a target frame and a bias term of the geometric descriptor is added ; After the weight matrix is normalized by Softmax, it is weighted and summed with the value vector of the target frame, and then passes through a residual connection and an MLP layer to obtain the final source frame enhancement feature; wherein the calculation formula of the attention weight is as follows: 。 3. The point cloud registration method of claim 1, wherein, In the two-stage-based matching matrix calculation module TSMM, the coarse matching process comprises the following steps: The original point cloud is hierarchically abstracted to generate super points covering key areas and down-sampled; similarities between super point features are calculated, and double Softmax normalization is used to alleviate false matches; and a dynamic Top-K strategy is used to screen matching pairs to obtain sparse matching pairs as the output of coarse matching .

4. The point cloud registration method of claim 3, wherein, In the two-stage-based matching matrix calculation module TSMM, the fine matching process comprises the following steps: Limiting fine matching to the local area corresponding to coarse matching to reduce the amount of calculation; calculating the feature similarity of the point pairs in the local area, performing Softmax normalization, using the Sinkhorn algorithm for matching optimization, and screening out candidate matching pairs; based on the purified matching pairs, using a weighted least squares optimization framework, solving the optimal rotation matrix R and translation vector t through singular value decomposition (SVD), and synthesizing a refined transformation matrix Ĉ.

5. The point cloud registration method of claim 1, wherein, The following loss function is used in the point cloud registration model training process: Wherein Ls is the loss function of coarse matching, and the formula is as follows: wherein is a sparse match pair generated for coarse matching, is a superpoint pair is a normalized similarity score, is a number of superpoints; Lp is the loss function formula of fine matching as follows: wherein is the set of preliminary matches, is the set of preliminary superpoint pairs, is the set of correct point pairs, is the set of points coordinates after applying the predicted rigid transformation, is the inlier threshold, λ is a hyperparameter used to control the balance between loss values; α is the weight balancing the geometric error and the inlier ratio.

6. A point cloud registration system implementing the method of any one of claims 1-5, characterized by It comprises: A data acquisition and preprocessing module responsible for collecting point cloud data and preprocessing; A registration processing module for registering the preprocessed point cloud data using the point cloud registration model FPFAM-Reg.

7. A computer-readable storage medium, characterized in that, The computer program is stored on the computer and causes the computer to execute the method of any one of claims 1-5 when executed in the computer.

8. A computing device comprising a memory and a processor, wherein, The memory stores executable code, and the processor executes the executable code to implement the method of any one of claims 1-5.

Images