Redefined 3D point cloud normal estimation method based on sample selection

Abstract

The invention discloses a redefined 3D point cloud normal estimation method based on sample selection, and the method comprises the steps: collecting three-dimensional point cloud data, and carrying out the evaluation and screening of the three-dimensional point cloud data through a confidence estimation strategy; performing random sampling on the screened three-dimensional point cloud data through a k-nearest neighbor algorithm and a probability-based sampling strategy to obtain local point cloud blocks and global point cloud blocks; a PCA algorithm is introduced to carry out alignment processing on the point cloud blocks to obtain a point cloud data set; constructing a multi-constraint joint optimization solving network; inputting a point cloud data set for training, optimizing the network through a loss function to obtain an optimized multi-constraint joint optimization solving network, and outputting a predicted point cloud normal; a neural gradient field module is introduced to carry out directional correction on the predicted point cloud normal to obtain a final point cloud normal; according to the method, a network with higher robustness is constructed through multi-constraint integration, data of various noise levels can be processed, and normal estimation results with consistent directions and higher precision are obtained.

Description

Technical Field

[0001] This invention relates to the field of point cloud normal estimation technology, and in particular to a method for redefining 3D point cloud normal estimation based on sample selection. Background Technology

[0002] Point cloud normal estimation is a core task in computer graphics. Due to the inherent disorder and non-uniform distribution of point clouds, normals, as fundamental features, provide crucial geometric information for downstream tasks. Efficient normal estimation methods have proven to be valuable for various applications, including odometry and mapping, 3D reconstruction, point cloud denoising, and semantic segmentation.

[0003] In recent years, deep learning methods have attracted much attention in normal estimation tasks. Compared with traditional PCA methods and jet fitting, deep learning techniques have shown advantages in handling diverse data and robustness to noise, and do not require complex parameter tuning. Typical methods follow a unified paradigm: sampling local neighborhood points as input, using neural networks to extract features, and finally estimating the normal or fitting a local surface directly through a regressor. Such algorithms trained on the PCPNet dataset have achieved excellent performance, and their generalization ability can be directly transferred to real-world scanning datasets such as Semantic3D, SceneNN, and NYU depth v2.

[0004] Thanks to the feature extraction capabilities of neural networks, learning-based normal estimation algorithms (especially the current state-of-the-art SHS-Net) have made groundbreaking progress. However, existing methods still rely on training on the PCPNet dataset, which contains multi-scale noise. While noise diversity can improve network generalization, as noise intensity increases, the sampled training samples may deviate significantly from the potentially clean surface, causing the patch pattern to deviate from the ideal state of the true normal, introducing ambiguity into the training process. This limitation constrains algorithm performance, causing the trained model to produce biased results on noise-free data, leading to a decrease in the accuracy of the final result. Summary of the Invention

[0005] This invention provides a method for redefining the normal estimation of 3D point clouds based on sample selection, in order to overcome the above-mentioned technical problems.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A method for redefining the normal of a 3D point cloud based on sample selection includes: S1: Collect raw 3D point cloud data, construct a confidence estimation strategy, evaluate and filter the raw 3D point cloud data through the confidence estimation strategy, and obtain the filtered 3D point cloud data; S2: Introduce the k-nearest neighbor algorithm and a probability-based sampling strategy. Randomly sample the filtered 3D point cloud data using the k-nearest neighbor algorithm and the probability-based sampling strategy respectively to obtain local point cloud blocks and global point cloud blocks. S3: Introduce the PCA algorithm to align local point cloud blocks and global point cloud blocks to obtain a point cloud dataset; S4: Introduce QSTN and SHS-Net networks to construct a multi-constraint joint optimization solution network; input the point cloud dataset into the multi-constraint joint optimization solution network for training, construct a loss function with confidence as weight, optimize the parameters of the multi-constraint joint optimization solution network through the loss function, obtain the optimized multi-constraint joint optimization solution network, and output the predicted point cloud normals; S5: Introduce the neural gradient field module to perform orientation correction on the predicted point cloud normals, and obtain the final corrected point cloud normals.

[0007] Furthermore, the multi-constraint joint optimization solution network includes a QSTN network and an SHS-Net network connected in sequence; The QSTN network is used to obtain the rotation matrix of local point cloud blocks in the point cloud dataset. The rotation matrix is ​​used to perform secondary alignment of local point cloud blocks and global point cloud blocks in the point cloud dataset to obtain the local point cloud block training set and the global point cloud block training set. The SHS-Net network includes a local branch neural network, a global branch neural network, and a feature splicing layer connected to the local branch neural network and the global branch neural network; The local branch neural network is used to extract local geometric feature vectors of point cloud blocks in the training set of local point cloud blocks; The global branch neural network is used to capture the shape-level contextual information of point cloud blocks in the global point cloud block training set and output a global feature vector. The feature splicing layer is used to splice local geometric feature vectors and global feature vectors to obtain multi-scale fused features, i.e., predicted point cloud normals.

[0008] Furthermore, the confidence estimation strategy involves evaluating and filtering the original 3D point cloud data using confidence based on a surface inclusion strategy and confidence based on a normal difference strategy, retaining data points whose confidence values ​​are both higher than the preset confidence values, thus obtaining the filtered 3D point cloud data.

[0009] Furthermore, the original 3D point cloud data is evaluated and filtered using confidence scores based on surface inclusion strategies and normal difference strategies, including: Step 1: Denoise the original 3D point cloud to obtain a noise-free point cloud; Step 2: For any data point in the original 3D point cloud data, calculate the distance from that point to the bottom surface of the noise-free point cloud, as shown in formula (1). (1) in, The distance from this point to the bottom surface of the noise-free point cloud. For any data point in the original 3D point cloud data, noise-free point clouds Any point in it; The confidence level based on the surface inclusion strategy is calculated as shown in Equation (2). (2) in, Based on the confidence level contained in the surface, The scale value of the original 3D point cloud data. It is a constant; Calculate data points The approximate surface point normal is shown in formula (3). (3) in, for The approximate surface point normal; Step 3: Calculate the difference between the normal of the data point and the normal of the approximate surface point, as shown in formula (4). (4) in, For the difference value, Represents the dot product of two vectors. The function calculates the arccosine value and returns the angle in radians. For data points The true normal line; The confidence level based on the normal difference strategy is calculated as shown in Equation (5). (5) in, It is a constant; Step 4: Evaluate and filter the original 3D point cloud data, retaining two data points whose confidence values ​​are both higher than the preset confidence value, to obtain the filtered 3D point cloud data.

[0010] Furthermore, a loss function is constructed using confidence scores as weights, and the parameters of the multi-constraint joint optimization network are optimized using the loss function, including: Transformation regularization loss and z-direction transformation loss are introduced to constrain the rotation matrix output by the QSTN network, as shown in equations (6) and (7). (6) (7) in, This represents the regularization loss function. Represents the loss function for the z-direction transformation; It is the identity matrix. express Unit vector along the axis Point cloud No. The rotation matrix output by QSTN operation on each sample. express The true normal line; Minimize the predicted point cloud normal With true normal The sinusoidal loss between the two sides is used to obtain the central loss of the fusion confidence, as shown in Equation (8). (8) in, , where is either the confidence level based on the surface inclusion strategy or the confidence level based on the normal difference strategy. This represents the predicted point cloud normal. Represents the true normal; The weighted loss of the fusion confidence is constructed as shown in formula (9). (9) in, This represents the predicted weight of each point in the sampled point cloud block after the multi-constraint joint optimization network performs downsampling on the local point cloud block. This represents the number of sampling points in the sampling point cloud. The true weight of each point in the sampled point cloud block is shown in formula (10). (10) in, Represents the first sampling point in the cloud block. One point, Point cloud The true normal line, As shown in formula (11), the intermediate variables are... (11) The neighborhood consistency loss function for fusing confidence values ​​is constructed as shown in formula (12). (12) in, For the predicted sampling point normal; For the true sampling point normal; The final loss function is constructed as shown in Equation (13). (13) in, arrive All are weighting factors.

[0011] Furthermore, the k-nearest neighbor algorithm and a probability-based sampling strategy are introduced. The filtered 3D point cloud data are randomly sampled using the k-nearest neighbor algorithm and the probability-based sampling strategy, respectively, to obtain local point cloud blocks and global point cloud blocks, including: A point is randomly selected from the filtered 3D point cloud data. The k-nearest neighbor algorithm is used to select neighboring points centered on this point until a set value is reached, thus forming a local point cloud patch. ; A probability-based sampling strategy is introduced, which samples data points... The density gradient is used as the sampling probability for global point sampling, and the density gradient is shown in Equation (14). (14) in, This indicates a numerical truncation operation. For point clouds A random sampling index set containing One element, For point clouds The first in Data points, It is any point in the point cloud; Furthermore, a neural gradient field module is introduced to perform orientation correction on the predicted point cloud normals, resulting in the final corrected point cloud normals, including: The predicted point cloud normals are oriented and corrected using a neural gradient field module, as shown in Equation (15). (15).

[0012] Beneficial effects: This invention provides a method for redefining the normal of 3D point clouds based on sample selection, which has the following advantages: 1. By introducing a confidence strategy, the collected data is filtered, and the confidence score is used as a loss weight to update the parameters of the constructed network, so as to achieve reasonable sample selection and abnormal sample suppression. This can establish a more robust estimation network than directly using the correction normal. 2. Based on QSTN and SHS-Net networks, a dual-branch multi-constraint joint optimization solution network is constructed. By integrating multiple constraints, a more robust network is built, which can handle data with various noise levels and achieve higher accuracy. 3. By performing orientation correction on the predicted point cloud normals, we can obtain normal estimation results with consistent direction and higher accuracy. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart of a method for redefining 3D point cloud normal estimation based on sample selection provided by the present invention; Figure 2 This is a diagram of the overall process. Figure 3 This is a schematic diagram illustrating the area under the curve (AUC) analysis on PCPNet and the Famous Shape dataset. Figure 4 A visual comparison diagram of PCPNet and the well-known FamousShape dataset; Figure 5 A schematic diagram illustrating the qualitative comparative analysis of the SceneNN dataset; Figure 6 A schematic diagram illustrating the qualitative comparative analysis of the Semantic3D dataset; Figure 7 This is a schematic diagram comparing the reconstruction effects of Poisson surfaces based on different normal estimation methods. Figure 8 A schematic diagram illustrating the qualitative analysis results of point cloud denoising. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] This embodiment provides a method for redefining 3D point cloud normal estimation based on sample selection, such as... Figure 1 As shown, it includes: S1: Collect raw 3D point cloud data, construct a confidence estimation strategy, evaluate and filter the raw 3D point cloud data through the confidence estimation strategy, and obtain the filtered 3D point cloud data; S2: Introduce the k-nearest neighbor algorithm and a probability-based sampling strategy. Randomly sample the filtered 3D point cloud data using the k-nearest neighbor algorithm and the probability-based sampling strategy respectively to obtain local point cloud blocks and global point cloud blocks. S3: Introduce the PCA algorithm to align local point cloud blocks and global point cloud blocks to obtain a point cloud dataset; S4: Introduce QSTN and SHS-Net networks to construct a multi-constraint joint optimization solution network; input the point cloud dataset into the multi-constraint joint optimization solution network for training, construct a loss function with confidence as weight, optimize the parameters of the multi-constraint joint optimization solution network through the loss function, obtain the optimized multi-constraint joint optimization solution network, and output the predicted point cloud normals; S5: Introduce the neural gradient field module to perform orientation correction on the predicted point cloud normals, and obtain the final corrected point cloud normals.

[0017] Specifically, two potential problems in the training phase were first identified: (1) the labeled normals of some training samples are unreliable; (2) there may be noise blocks lacking effective underlying surface structures. Therefore, a confidence estimation strategy was proposed after data collection. The strategy can effectively suppress the negative impact of damaged samples on the robustness of model training. Secondly, the k-nearest neighbor algorithm and a probability-based sampling strategy are introduced. The selected 3D point cloud data are randomly sampled using the k-nearest neighbor algorithm and the probability-based sampling strategy to obtain local point cloud blocks and global point cloud blocks. The PCA algorithm is then introduced to align the local point cloud blocks and global point cloud blocks to obtain a point cloud dataset. Through sampling, the local information of each sample point and the global information related to the local point can be effectively obtained. The input of data at two scales ensures the stability of the algorithm. The alignment operation using PCA can reduce the degree of freedom of the input data pattern. Furthermore, QSTN and SHS-Net networks are introduced to construct a multi-constraint joint optimization solution network. The point cloud dataset is input into the multi-constraint joint optimization solution network for training. A loss function is constructed using confidence as weights. The parameters of the multi-constraint joint optimization solution network are optimized using the loss function to obtain the optimized multi-constraint joint optimization solution network, which outputs the predicted point cloud normals. The multi-branch inputs are aligned by a shared deformable transformable network (QSTN) to eliminate local geometric deformation interference. By introducing confidence weights to construct the loss function, the estimation bias of the network on the task caused by bad samples can be reduced. Finally, a neural gradient field module is introduced to orient the predicted point cloud normals, resulting in the final corrected point cloud normals and obtaining normal estimation results with consistent orientation and higher accuracy.

[0018] In a specific embodiment, the following scheme is used to collect raw 3D point cloud data, construct a confidence estimation strategy, evaluate and filter the raw 3D point cloud data using the confidence estimation strategy, and obtain the filtered 3D point cloud data: A confidence estimation strategy is constructed, which uses confidence based on the surface inclusion strategy and confidence based on the normal difference strategy to evaluate and filter the original 3D point cloud data, and retains data points whose confidence values ​​are both higher than the corresponding preset confidence values, thus obtaining the filtered 3D point cloud data. The raw 3D point cloud data is evaluated and filtered using confidence scores based on surface inclusion and normal difference strategies, including: The surface-inclusion-based strategy primarily achieves this by evaluating the distance from the query point in the sampling block to the underlying surface. Clearly, points far from the underlying surface struggle to effectively capture local surface structure information. Given a query point... The objective is to calculate the distance from this point to the bottom surface. However, models affected by noise Estimating its underlying surface is challenging. Therefore, a corresponding noise-free point cloud is employed. Assuming that it can represent the complete underlying surface, the distance from this point to the underlying surface of the noise-free point cloud is calculated, and then the confidence score based on the surface inclusion strategy is obtained. The specific steps are as follows: Step 1: Denoise the original 3D point cloud to obtain a noise-free point cloud. ; Step 2: For any data point in the original 3D point cloud data, calculate the distance from that point to the bottom surface of the noise-free point cloud, as shown in formula (16). (16) in, The distance from this point to the bottom surface of the noise-free point cloud. For any data point in the original 3D point cloud data, noise-free point clouds Any point in it; The confidence level based on the surface inclusion strategy is calculated as shown in Equation (17). (17) in, Based on the confidence level contained in the surface, The scale value of the original 3D point cloud data. It is a constant; in this scheme, The value is set to 0.05; Noisy points may be far from the potential surface, and their normals may differ significantly from the normals of the nearest point on the potential surface. Training samples from such points have low reliability. Therefore, the normal difference-based strategy mainly estimates the confidence level by comparing the current normal of each point with the normal of the nearest point on the potential surface. The specific steps are as follows: For the labeled normal vector is query point In noise-free clouds Search for its nearest neighbor. This point It can be approximated as a curved surface The projection points on the surface. Therefore, calculate the data points. The approximate surface point normal is shown in formula (18). (18) in, for The approximate surface point normal; The difference between the normal to the data point and the normal to the approximate surface point is calculated as shown in formula (19). (19) in, For the difference value, Represents the dot product of two vectors. The function calculates the arccosine value and returns the angle in radians. For data points The true normal line; The confidence level based on the normal difference strategy is calculated as shown in formula (20). (20) in, It is a constant; in this scheme The value is set to 0.06; Step 4: Evaluate and filter the original 3D point cloud data, retaining two data points whose confidence values ​​are both higher than the preset confidence value, to obtain the filtered 3D point cloud data.

[0019] In this scheme, the confidence scores based on surface inclusion and normal difference are used to quantify the reliability of training samples, screen reliable training samples, effectively suppress the negative impact of damaged samples on the robustness of model training, improve the stability of subsequent training models, and thus improve the accuracy of the model's normal estimation.

[0020] In a specific embodiment, such as Figure 2As shown, the k-nearest neighbor algorithm and a probability-based sampling strategy are introduced. The scheme for randomly sampling the filtered 3D point cloud data using the k-nearest neighbor algorithm and the probability-based sampling strategy respectively to obtain local and global point cloud blocks is as follows: A point is randomly selected from the filtered 3D point cloud data. The k-nearest neighbor algorithm is used to select neighboring points centered on this point until a set value is reached, thus forming a local point cloud patch. ; A probability-based sampling strategy is introduced, which samples data points... The density gradient is used as the sampling probability for global point sampling, and the density gradient is shown in Equation (14). (14) in, This indicates a numerical truncation operation. For point clouds A random sampling index set containing One element, For point clouds The first in Data points, It is any point in the point cloud.

[0021] Specifically, a point-by-point estimation strategy is used to estimate the normal vector of a given point on the model. During the testing phase, each point is estimated, while during the training phase, random sampling is performed, sampling a certain number of points for training. For a given point, the local point cloud patch set at that point and the global sampled point set centered on that point are obtained: (1) For sampling local point cloud blocks, the k-nearest neighbor search technique is used to obtain the k nearest points to form a local point cloud block. Specifically, a predetermined number (preferably 700) of spatial points within its neighborhood are extracted to construct the local point block; (2) For the global sampling point set, the algorithm samples the global point cloud from the point cloud based on the density gradient, ensuring that the sampling density is higher closer to the center of the sampling point and lower further away.

[0022] By collecting local and global point cloud blocks, local information of each sample point and global information related to the local point can be effectively obtained. The input of data at two scales ensures the stability of the algorithm, making the prediction network more stable in estimating non-uniform sampling and noisy data.

[0023] In a specific embodiment, such as Figure 2 As shown, the scheme for aligning local and global point cloud blocks using the PCA algorithm to obtain the point cloud dataset is as follows: For global and local sampling point sets, preprocessing is required first, namely, using PCA to perform rotation operations to reduce the problem of excessive freedom of the input data; This scheme employs a shared PCA operation, which involves performing PCA calculations on local point cloud blocks to obtain the rotation transformation matrix, and then simultaneously applying the rotation transformation matrix to both local and global point cloud blocks, resulting in a point cloud dataset.

[0024] Alignment operations using PCA can reduce the degrees of freedom of the input data pattern, ensuring that the neural network focuses more on the potential surface distribution of the sampled block and improve the estimation accuracy of local normals.

[0025] In a specific embodiment, QSTN and SHS-Net networks are introduced to construct a multi-constraint joint optimization solution network. The point cloud dataset is input into the multi-constraint joint optimization solution network for training. A loss function is constructed using confidence scores as weights. The parameters of the multi-constraint joint optimization solution network are optimized using the loss function to obtain the optimized multi-constraint joint optimization solution network, which then outputs the predicted point cloud normals. like Figure 2 As shown, the multi-constraint joint optimization solution network includes a QSTN network and an SHS-Net network connected in sequence; The QSTN network is used to obtain the rotation matrix of local point cloud blocks in the point cloud dataset. The rotation matrix is ​​used to perform secondary alignment of local point cloud blocks and global point cloud blocks in the point cloud dataset to obtain the local point cloud block training set and the global point cloud block training set. The SHS-Net network includes a local branch neural network, a global branch neural network, and a feature splicing layer connected to the local branch neural network and the global branch neural network; The local branch neural network is used to extract local geometric feature vectors of point cloud blocks in the training set of local point cloud blocks; The global branch neural network is used to capture the shape-level contextual information of point cloud blocks in the global point cloud block training set and output a global feature vector. The feature splicing layer is used to splice local geometric feature vectors and global feature vectors to obtain multi-scale fused features, i.e., predicted point cloud normals; Specifically, the two branches of the SHS-Net network use a shared QSTN (Quaternion Spatial Transformer Network) operation to rotate the input local point cloud block and global point cloud block again to ensure that the data is properly aligned. The two branches extract global and local features respectively, as follows: The aligned local point cloud blocks are input into the local branch neural network, and the aligned global point cloud blocks are input into the global branch neural network, where: Local branch neural networks incorporate multilayer perceptrons (MLPs) and graph attention mechanisms to extract local geometric features; The global branch network adopts a point Transformer architecture to capture shape-level contextual information and output a global feature vector; The two branches achieve multi-scale feature fusion through a feature concatenation layer; Both QSTN and SHS-Net networks are commonly used by technicians in the field of point cloud data processing, so they will not be described in detail. A loss function is constructed using confidence scores as weights. This loss function is then used to optimize the parameters of the multi-constraint joint optimization network, including: Introducing transformation regularization loss and Direction transformation loss for point clouds No. The rotation matrix output by QSTN operation on each sample The goal of this type of regularization loss is to achieve a rigid transformation, making the sampling block follow the shape as closely as possible. The axes are vertically aligned, thus reducing the additional degrees of freedom of the point cloud blocks; Transformation regularization loss and z-direction transformation loss are introduced to constrain the rotation matrix output by the QSTN network, as shown in equations (15) and (16). (15) (16) in, This represents the regularization loss function. Represents the loss function for the z-direction transformation; It is the identity matrix. express The unit vector along the axis in this scheme , Point cloud No. The rotation matrix output by QSTN operation on each sample. express The true normal line; Minimize the predicted point cloud normal With true normal The sinusoidal loss between the two sides is used to obtain the central loss of the fusion confidence, as shown in Equation (17). (17) in, , where is either the confidence level based on the surface inclusion strategy or the confidence level based on the normal difference strategy. This represents the predicted point cloud normal. Represents the true normal; After training and testing, choosing any confidence level to construct the loss function can improve the prediction accuracy of the network. Therefore, any confidence level can be chosen when constructing the loss function. The weighted loss of the fusion confidence is constructed as shown in formula (18). (18) in, This represents the predicted weight of each point in the sampled point cloud block after the multi-constraint joint optimization network performs downsampling on the local point cloud block. This represents the number of sampling points in the sampling point cloud. The true weight of each point in the sampled point cloud block is shown in formula (19). (19) in, Represents the first sampling point in the cloud block. One point, Point cloud The true normal line, As an intermediate variable, as shown in formula (20), (20) The neighborhood consistency loss emphasizes the importance of local points near the query center. A neighborhood consistency loss function that integrates confidence values ​​is constructed, as shown in formula (21). (twenty one) in, For the predicted sampling point normal; For the true sampling point normal; The final loss function is constructed as shown in Equation (22). (twenty two) in, arrive All are weighting factors; in this scheme , , , , ; The network parameters are solved by using back gradient optimization. The optimization objective is to minimize the loss function defined by the weighted average of the five losses. When the loss function is minimized, i.e., when it tends to converge, the final network parameters are obtained.

[0026] In this scheme, the QSTN operation can further reduce the degrees of freedom of the input data, making the subsequent SHS-Net more focused on the potential surface structure. In addition, the introduction of confidence weights can make the network focus on reasonable sample data during training, reducing the estimation bias of the network on the task caused by bad samples.

[0027] In a specific embodiment, a neural gradient field module is introduced to perform orientational correction on the predicted point cloud normals, resulting in the following scheme for obtaining the final corrected point cloud normals: Normal estimation methods based on direct regression using neural networks often struggle to simultaneously balance directional accuracy and geometric precision. To address this issue, this solution decouples the two: like Figure 2 As shown in the estimation stage, the network only focuses on the regression of normal geometry accuracy. However, by introducing NeuralGF and using the neural gradient function to estimate the normal direction of the entire input model, the direction estimation results of each query point can be obtained.

[0028] The predicted point cloud normals are oriented and corrected using the neural gradient field module, as shown in Equation (23). (twenty three).

[0029] By performing orientation correction on the predicted point cloud normals, we can obtain normal estimation results with consistent orientation and higher accuracy.

[0030] Specific implementation examples are as follows; 1. Dataset and Experiment Setup All comparative experimental models in this embodiment were trained on the same training set of the PCPNet dataset, which provides true normals with consistent orientation. Multiple scales of noise and distribution density variations were applied during training.

[0031] To evaluate the generalization ability of the method, both synthetic and real datasets were used: the synthetic datasets included PCPNet and Famous, while the real datasets included Semantic3D and SceneNN. The test dataset configuration was completely consistent with SHS-Net. Similar to SHS-Net, the root mean square error of angle (RMSE) and area under the curve (AUC) were used for quantitative comparison. During training, the batch size was set to 145, the number of training epochs to 800, the Adam optimizer was used, and the base learning rate was set to 0.0009. The network was trained on a single NVIDIA RTX 3090 GPU. In terms of network configuration, the local point block encoding randomly selected the query point and its 700 neighboring points to construct point blocks; the global shape encoding sampled 1200 points from the shape point cloud.

[0032] 2. Normal estimation performance The experimental results on the synthetic dataset are shown in Tables 1 and 2. Table 1 compares the root mean square error of the angle estimation for non-directional normals, evaluating the performance of our method, classical geometric methods, and deep learning methods on the PCPNet and FamousShape datasets. Table 2 compares the root mean square error of the corrected non-directional normal estimation, evaluating the performance of our method, classical geometric methods, and deep learning methods on the improved PCPNet dataset. Table 1

[0033] Table 2

[0034] As shown in Tables 1 and 2, our proposed method, using two confidence levels, is competitive with existing deep learning methods and traditional methods on the synthetic datasets PCPNet and Famous. Here, '1' represents a strategy based on surface inclusion, and '2' represents a strategy based on normal difference. In particular, our method achieves optimal performance on low-noise and variable-density point clouds. In high-noise scenarios, our accuracy is comparable to mainstream methods, indicating that our sample selection strategy enhances the mining of local point patch patterns, thereby reducing interference from noisy samples during training and achieving more accurate local normal estimation.

[0035] Furthermore, the accuracy of state-of-the-art algorithms was compared on the PCPNet dataset for corrected normals, as shown in Table 3. Table 3 compares the root mean square angle error of the oriented normal estimation, evaluating the performance of our method and other methods on PCPNet and the well-known shape dataset (FamousShape). Table 3

[0036] As can be seen from Table 3, our method provides the most accurate normals under almost all noise levels and density variations in both datasets. Our method still maintains the strongest competitiveness on high-noise point clouds, which confirms the robustness of the normal estimation architecture. Figure 3 The AUC results for each method are presented, from... Figure 3 As can be seen from the results, this method exhibits the best performance and maintains significant stability at different angle thresholds.

[0037] Figure 4 The qualitative comparison results of the angular errors of various points in the point cloud are presented through heatmaps. Figure 4 It can be seen that the method has the smallest error in areas of density variation, complex geometry, and local details.

[0038] Experimental results show that by combining the regression network with the global directional normal estimation method NeuralGF, the accuracy of normal estimation is significantly improved, which is in stark contrast to SHSNet's use of a single network for normal and orientation estimation.

[0039] To evaluate the generalization ability of the proposed method, quantitative and qualitative experiments were conducted on the indoor SceneNN and outdoor Semantic3D real datasets. The SceneNN dataset was acquired using a depth camera, and its ground truth normals were calculated from the reconstructed mesh. Table 1 shows the RMSE scores of each method under the same test model as SHS-Net. The proposed method shows a significant improvement over SHS-Net and remains competitive with state-of-the-art methods such as CMG-Net and MESCNet, verifying the enhancement effect of the confidence-based selection strategy proposed in this invention on the generalization ability of the prediction network. Figure 5 The visualization results of the angle error are shown, demonstrating that this method achieves more accurate estimates at local details. On the Semantic3D laser scan point cloud dataset without true normals, Figure 6 By using RGB normal mapping to present visual contrast, this method achieves sharper estimation results in detailed areas.

[0040] 3. Ablation Research Parameter settings: Given that the confidence values ​​(including confidence values ​​based on surface inclusion and normal differences) in this invention are used to suppress unreasonable samples during training, their selection directly affects model accuracy. The selection of confidence values ​​is influenced by the sigma value, which will be discussed in detail in this section. The value setting strategy differs for the PCPNet dataset. Table 4 shows the experimental analysis of non-directional normal ablation under different σ values ​​on the PCPNet dataset. Table 4

[0041] Table 4 shows that the optimal method for calculating confidence levels based on surface inclusion is... =0.05, while the confidence level based on the difference in normals is optimal. The value is 0.06. Too high. Values ​​that are too low can lead to convergence in sample confidence levels, making it difficult to distinguish between superior and inferior samples; A value that makes it difficult for noisy samples to participate in training, therefore a moderate value is needed. Setting the value is crucial.

[0042] To verify the effectiveness of the multi-constraint joint optimization solution network of this invention, ablation experiments were conducted on the PCPNet dataset. The experimental results are shown in Table 5, which presents the ablation experiment results under different configurations of the PCPNet dataset. Table 5

[0043] (A) is a single-branch architecture based on local point blocks (without directional normals), and (B) is a dual-branch architecture that merges global and local point blocks. Tables 5(A)-(B) compare the baseline models of single-branch (local point block input) and dual-branch (global + local input). The results show that without the introduction of additional loss and the QSTN module, the addition of global information may impair model accuracy (average RMSE increases from 9.89 to 9.97). Table 5 shows that regardless of the single / dual-branch framework, the introduction of the QSTN module, z-direction transformation loss, and neighborhood consistency loss significantly improves accuracy (by 0.61% and 2.51% respectively compared to the baseline). With enhanced representational capabilities and the introduction of effective constraints, the global branch can provide useful information for normal estimation. The dual-branch framework was ultimately chosen as the final model. The last three columns of Table 5 evaluate the reweighted sample selection strategy. The results show that directly using corrected normals for training (such as CMG-NET) provides limited improvement, while the loss function proposed in this invention, based on surface inclusion and normal difference confidence suppression, brings significant improvement (more significant under the multi-branch architecture), indicating that the reweighted strategy can provide more flexible constraints.

[0044] 4. Application of the method Poisson Reconstruction: To explore the potential of this method in other tasks, we investigate the application of normals in point cloud surface reconstruction. The classic Poisson reconstruction method is employed. Figure 7 This image shows the surface reconstruction results using different methods. It can be seen from the image that the surface reconstructed using this method is more accurate, with sharper local details and boundaries.

[0045] Point cloud denoising: The accuracy of normal estimation is verified by applying a normal-based denoising method. Figure 8 The results of denoised point cloud, reconstructed surface, and feature detection are presented. Visualization demonstrates that the normal estimation method effectively improves denoising quality, generating smoother surfaces in flat areas while preserving sharp features at edges.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for redefining 3D point cloud normal estimation based on sample selection, characterized in that, The method comprises the following steps: S1: collecting original three-dimensional point cloud data, constructing a confidence estimation strategy, evaluating and screening the original three-dimensional point cloud data through the confidence estimation strategy, and obtaining screened three-dimensional point cloud data; S2: introducing a k-nearest neighbor algorithm and a probability-based sampling strategy, respectively sampling the screened three-dimensional point cloud data through the k-nearest neighbor algorithm and the probability-based sampling strategy, and obtaining local point cloud blocks and global point cloud blocks; S3: introducing a PCA algorithm to align the local point cloud blocks and the global point cloud blocks, and obtaining a point cloud dataset; S4: introducing a QSTN network and a SHS-Net network to construct a multi-constraint joint optimization solving network; inputting the point cloud dataset into the multi-constraint joint optimization solving network for training, constructing a loss function with the confidence as the weight, optimizing the parameters of the multi-constraint joint optimization solving network through the loss function, obtaining an optimized multi-constraint joint optimization solving network, and outputting a predicted point cloud normal; S5: introducing a neural gradient field module to correct the predicted point cloud normal, and obtaining a final corrected point cloud normal.

2. The method of claim 1, wherein, The multi-constraint joint optimization solving network comprises a QSTN network and a SHS-Net network connected in sequence; The QSTN network is used to obtain a rotation matrix of the local point cloud blocks in the point cloud dataset, and the rotation matrix is used for secondary alignment of the local point cloud blocks and the global point cloud blocks in the point cloud dataset, so as to obtain a local point cloud block training set and a global point cloud block training set; The SHS-Net network comprises a local branch neural network, a global branch neural network, and a feature splicing layer connected with the local branch neural network and the global branch neural network; The local branch neural network is used to extract a local geometric feature vector of the point cloud blocks in the local point cloud block training set; The global branch neural network is used to capture shape-level context information of the point cloud blocks in the global point cloud block training set, and output a global feature vector; The feature splicing layer is used to splice the local geometric feature vector and the global feature vector, and obtain a multi-scale fusion feature, i.e., a predicted point cloud normal.

3. The method of claim 1, wherein, The confidence estimation strategy is to evaluate and screen the original three-dimensional point cloud data by using a confidence based on a surface inclusion strategy and a confidence based on a normal difference strategy, and to retain data points with both confidence values higher than a preset confidence value, so as to obtain screened three-dimensional point cloud data.

4. The method of claim 3, wherein, Evaluating and screening the original three-dimensional point cloud data by using a confidence based on a surface inclusion strategy and a confidence based on a normal difference strategy comprises the following steps: Step 1: denoising the original three-dimensional point cloud to obtain a noise-free point cloud; Step 2: for any data point in the original three-dimensional point cloud data, calculating the distance of the point to the underlying surface of the noise-free point cloud, as shown in formula (1), （1） wherein, is the distance of the point to the underlying surface of the noise-free point cloud, is any one data point in the original three-dimensional point cloud data, is the noise-free point cloud any one point in calculating the confidence based on the surface inclusion strategy, as shown in formula (2), （2） wherein, is a confidence based on the surface containing, is a scale value of the original three-dimensional point cloud data, is a constant; Computing the approximate surface point normal of the data point as shown in equation (3), （3） wherein is approximated surface point normal; Step 3: calculating the difference between the normal of the data point and the normal of the approximate surface point, as shown in formula (4), （4） wherein, is a difference value, denotes the inner product of two vectors, denotes the inverse cosine function that computes the arc sine value and returns the angle in radians, is the true normal of the data point . calculating the confidence based on the normal difference strategy, as shown in formula (5), （5） wherein is a constant; Step four: evaluating and screening the original three-dimensional point cloud data, retaining the data points with both confidence values higher than the preset confidence value, and obtaining the screened three-dimensional point cloud data.

5. The method of claim 4, wherein, A loss function is constructed with the confidence as the weight, and the network for solving the multi-constraint joint optimization is optimized through the loss function, including: The transformation regularization loss and the z-direction transformation loss are introduced to constrain the rotation matrix output by the QSTN network, as shown in formulas (6) and (7), （6） （7） wherein, denotes a regularization loss function, denotes a z-direction transformation loss function; is an identity matrix, denotes an axis direction unit vector, denotes a point cloud The rotation matrix output by the QSTN operation of the first denotes the true normal of Minimizing predicted point cloud normals with real normals sinusoidal loss between the predicted point cloud normals and the ground truth normals to get a center loss of the fusion confidence as shown in equation (8), （8） wherein, is any one of a confidence based on a surface inclusion strategy and a confidence based on a normal difference strategy, denotes a predicted point cloud normal, denotes a true normal; A weight constraint loss function is constructed by fusing the confidence, as shown in formula (9), （9） wherein, represents the prediction weight of each point in the sampled point cloud block after the network solves the multi-constraint joint optimization, and represents the number of sampled points in the sampled point cloud block; represents the real weight of each point in the sampled point cloud block, as shown in formula (10), （10） wherein denotes the i-th point in the block of sampled point cloud, denotes the true normal of the point cloud, is an intermediate variable as shown in equation (11),​​ （11） A neighborhood consistency loss function is constructed by fusing the confidence value, as shown in formula (12), （12） wherein, is the predicted sample point normal; is the real sample point normal; The final loss function is constructed, as shown in formula (13), （13） wherein to are weight factors.

6. The method of claim 1, wherein, The k-nearest neighbor algorithm and the probability-based sampling strategy are introduced, and the screened three-dimensional point cloud data is randomly sampled through the k-nearest neighbor algorithm and the probability-based sampling strategy to obtain local point cloud blocks and global point cloud blocks, including: A point in the screened three-dimensional point cloud data is randomly selected, and the k-nearest neighbor algorithm is used to select the neighboring points with the point as the center until the set value is reached to stop, forming a local point cloud block; A probability-based sampling strategy is introduced that globally samples data points with a density gradient as the sampling probability, the density gradient being shown in equation (14), （14） wherein, denotes a numerical truncation operation, is a set of random sampling indices of the point cloud containing elements, is the th data point in the point cloud , and is an arbitrary point in the point cloud.

7. The method of claim 5, wherein the redefined 3D point cloud normal estimation based on sample selection is characterized by, A neural gradient field module is introduced to directionally correct the predicted point cloud normal to obtain the final corrected point cloud normal, including: The neural gradient field module is used to directionally correct the predicted point cloud normal, as shown in formula (15), （15）。

Images