A point cloud registration method and system based on statistical local feature description and matching

Abstract

This invention discloses a point cloud registration method and system based on statistical local feature description and matching, relating to the field of point cloud registration. The method includes: determining the local feature descriptor of each query point and the local feature descriptors of its corresponding neighboring points in 3D point cloud data; statistically weighting the local feature descriptor of each query point and the local feature descriptors of its corresponding neighboring points to obtain the statistical local feature descriptor of the corresponding query point; acquiring the statistical local feature descriptors of the source point cloud and the target point cloud; and improving the ICP algorithm based on the feature differences between point pairs in the source point cloud and the target point cloud and the average matching distance to perform point cloud registration. This invention can improve the accuracy and robustness of point cloud registration.

Description

Technical Field

[0001] This invention relates to the field of point cloud registration, and in particular to a point cloud registration method and system based on statistical local feature description and matching. Background Technology

[0002] With the rapid development of high-precision 3D scanning equipment such as LiDAR and Kinect, acquiring point cloud data from these devices has become a crucial representation of 3D models. However, limited by the scanner's own performance and the complex conditions of the scanning scene, the scanned point cloud data suffers from problems such as partial point overlap, missing points, and noise interference, significantly impacting its subsequent applications. Therefore, point cloud preprocessing techniques for improving the quality of scanned point clouds are essential. Point cloud registration, as an important point cloud preprocessing technique, plays a key role in subsequent point cloud processing. The purpose of point cloud registration is to find the transformation matrix within a given metric space to establish the correspondence between one point cloud and another. Point cloud registration techniques have been widely applied in fields such as 3D reconstruction, cultural relic restoration, map creation, and autonomous driving.

[0003] Existing point cloud registration algorithms are divided into Iterative Closest Point (ICP) and its variants, and registration algorithms based on traditional geometric features. The most widely used ICP algorithm in point cloud registration was proposed by Besl PJ et al., which establishes an optimization model based on the nearest point distance between point clouds to establish registration relationships. The original ICP algorithm can obtain accurate registration results when the initial poses correspond well; otherwise, it is prone to getting trapped in local optima. In addition, this algorithm has drawbacks such as poor robustness and slow iteration speed. Wang Bin et al. proposed an improved ICP algorithm based on bidirectional distance ratios in the accurate registration stage, which improved registration accuracy. Zhang J et al. proposed a Fast and Robust Iterative Closest Point (Fast ICP) algorithm that minimizes the ICP using a majorization-minimization (MM) algorithm, improving both robustness and registration speed. ICP and its variants are based on establishing registration relationships based on point distance. Therefore, these methods are prone to getting trapped in local optima when the initial pose is poor. They also have significant defects in registering point clouds with missing data. The robustness and global optimization capabilities of the algorithms are still limited.

[0004] Traditional geometric feature-based registration algorithms establish correspondences between point pairs based on shared geometric features. The Signature of Histogram of Orientation (SHOT) descriptor proposed by Salti S et al. establishes a local coordinate system at the query point, combining spatial location information and geometric feature statistics of neighboring points for feature description. The Local Feature Statistics Histogram (LFSH) proposed by Yang J et al. is a recently proposed geometric feature-based method that statistically weights local depth, point density, and normal angle features to provide a more comprehensive description of local shape geometry. While traditional geometric feature-based registration algorithms are more robust, they are significantly affected by the stability of geometric features. The extraction and correspondence calculation of point cloud features result in relatively poor registration speeds for these methods. The point cloud registration algorithm proposed by Li Xinchun et al. based on neighborhood feature point extraction and matching, the improved ICP point cloud registration algorithm based on fast point feature histogram proposed by Liu Yuzhen et al., and the point cloud registration algorithm based on improved 3D shape context proposed by Li Yuxiang et al. all perform registration by extracting feature points. Feature point extraction can improve the registration speed, but it has a significant impact on the registration accuracy.

[0005] Therefore, in order to address the problems of poor robustness and low registration accuracy of the ICP algorithm under conditions of initial pose difference, partial data loss and noise interference, there is an urgent need to provide a new point cloud registration method or system. Summary of the Invention

[0006] The purpose of this invention is to provide a point cloud registration method and system based on statistical local feature description and matching, which can improve the accuracy and robustness of point cloud registration.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A point cloud registration method based on statistical local feature description and matching includes:

[0009] The local feature descriptors of each query point and its corresponding neighboring points in the 3D point cloud data are determined respectively; the local feature descriptors include: point cloud local density, point cloud fitting plane distance variance, Gaussian curvature, and mean curvature; the 3D point cloud data is the source point cloud;

[0010] The local feature descriptor of each query point and the local feature descriptor of its corresponding neighboring points are statistically weighted to obtain the statistical local feature descriptor of the corresponding query point.

[0011] Obtain statistical local feature descriptors for the source and target point clouds;

[0012] The ICP algorithm is improved based on the feature differences and average matching distance between point pairs in the source and target point clouds to perform point cloud registration.

[0013] Optionally, the step of determining the local feature descriptor of each query point and the local feature descriptor of its corresponding neighboring points in the 3D point cloud data further includes:

[0014] Three-dimensional point cloud data is acquired using a 3D scanner; the 3D point cloud data serves as the source point cloud.

[0015] The KD-tree method is used to determine the neighboring points within the neighborhood range corresponding to each query point in the three-dimensional point cloud data.

[0016] Optionally, determining the local feature descriptor of each query point and the local feature descriptor of its corresponding neighboring points in the 3D point cloud data specifically includes:

[0017] Determine the local density of the point cloud from the points within the neighborhood of each query point to the query point;

[0018] Based on the fitted plane of the point cloud within the neighborhood range of each query point;

[0019] Determine the distance from the nearest points within the neighborhood of each query point to the fitted plane;

[0020] The variance of the point cloud fitting plane distance is determined by the variance of the distances from the nearest points in the neighborhood of each query point to the fitting plane.

[0021] Optionally, the step of statistically weighting the local feature descriptor of each query point and the local feature descriptors of its corresponding neighboring points to obtain the statistical local feature descriptor of the corresponding query point specifically includes:

[0022] Using formula Determine the statistical local feature descriptor for the query point;

[0023] in, query point Statistical local feature descriptors, For query point Local feature descriptors, Neighboring points Local feature descriptors, The number of points within the neighborhood. For query point To nearby points The distance.

[0024] Optionally, the ICP algorithm is improved based on the feature differences and average matching distance between point pairs in the source and target point clouds to perform point cloud registration, specifically including:

[0025] Select from the target point cloud using the farthest point sampling method 1 sample point, and according to Each sample point determines the sample point set;

[0026] Perform corresponding point matching on the feature differences between point pairs in 3D point cloud data and sample point sets, and remove erroneous point pairs;

[0027] The relationship between corresponding points is determined by singular value decomposition, and the transformation model from 3D point cloud data to target point cloud is determined based on the relationship between corresponding points.

[0028] The ICP algorithm is improved by using the average matching distance as a metric to determine the target transformation model; the target transformation model is the transformation model when the average matching distance reaches the minimum value.

[0029] Point cloud registration is performed using a target transformation model.

[0030] Optionally, corresponding point matching is performed on the feature differences between corresponding point pairs in the 3D point cloud data and the sample point set, specifically including:

[0031] Using formula Determine the feature differences between point pairs in the source point cloud and the target point cloud;

[0032] in, Point pairs between source and target point clouds The characteristic differences between them For points in the source point cloud, For the corresponding point in the target point cloud, For the first Item parameters.

[0033] Optionally, the step of using the average matching distance as a metric to improve the ICP algorithm and determine the target transformation model specifically includes:

[0034] Determine the current average matching distance between points in the current 3D point cloud data and corresponding points in the target point cloud in the current transformation model;

[0035] Determine if the current average matching distance has reached the minimum value;

[0036] If the target transformation model is achieved, then the current transformation model will be determined as the target transformation model.

[0037] If the target is not reached, then return to the method of selecting from the target point cloud using the farthest point sampling. 1 sample point, and according to The steps for determining the sample point set are: update the transformation model for each sample point until the average matching distance reaches the minimum value.

[0038] Optionally, using the formula Determine the average matching distance;

[0039] in, for and The average matching distance For sample points in the target point cloud; Statistical local feature descriptors and Corresponding points, This represents the number of sampled matching point pairs.

[0040] A point cloud registration system based on statistical local feature description and matching includes:

[0041] The local feature descriptor determination module is used to determine the local feature descriptor of each query point and the local feature descriptor of its corresponding neighboring points in the 3D point cloud data; the local feature descriptor includes: point cloud local density, point cloud fitting plane distance variance, Gaussian curvature, and mean curvature; the 3D point cloud data is the source point cloud;

[0042] The statistical local feature descriptor determination module is used to statistically weight the local feature descriptor of each query point and the local feature descriptor of the corresponding neighboring points to obtain the statistical local feature descriptor of the corresponding query point.

[0043] The module for obtaining statistical local feature descriptors of source and target point clouds is used to obtain statistical local feature descriptors of source and target point clouds.

[0044] The point cloud registration module is used to improve the ICP algorithm based on the feature differences and average matching distance between point pairs in the source and target point clouds, and to register the point clouds.

[0045] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0046] This invention provides a point cloud registration method and system based on statistical local feature description and matching. It constructs a 4-dimensional feature descriptor, called a local feature descriptor (LFD), using the local density of the point cloud, the variance of the distance between the fitted planes, the Gaussian curvature, and the average curvature. Through statistical weighting, a statistical local feature descriptor (SLFD) for the query point is obtained, accurately describing the local features of the query point. Then, corresponding points are matched using the feature differences (FD) between point pairs to eliminate erroneous point pairs, addressing the problems of missing data and noise interference in the point cloud. Finally, the average matching distance (MMD) is used as a metric to improve the ICP algorithm for point cloud registration, solving the problem of low registration accuracy when the initial pose is poor. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 The flowchart of a point cloud registration method based on statistical local feature description and matching provided by the present invention is shown below.

[0049] Figure 2 This is a schematic diagram of the local density of the point cloud;

[0050] Figure 3 This is a schematic diagram of the fitted plane (where (a) represents the curved portion and (b) represents the flat portion).

[0051] Figure 4 The neighborhood range of SLFD;

[0052] Figure 5 Point cloud registration flowchart;

[0053] Figure 6 A schematic diagram showing the registration of the Bunny model under different transformation states;

[0054] Figure 7 A schematic diagram illustrating the registration of the Bunny model under different levels of data loss.

[0055] Figure 8 A schematic diagram showing the registration of the China Dragon model under different noise conditions;

[0056] Figure 9A diagram illustrating the registration process for the actual Bottle item. Detailed Implementation

[0057] 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, and 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.

[0058] The purpose of this invention is to provide a point cloud registration method and system based on statistical local feature description and matching, which can improve the accuracy and robustness of point cloud registration.

[0059] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] like Figure 1 As shown, the point cloud registration method based on statistical local feature description and matching provided by the present invention includes:

[0061] S101, determine the local feature descriptor of each query point and the local feature descriptor of the corresponding neighboring points in the three-dimensional point cloud data respectively; the local feature descriptor includes: point cloud local density, point cloud fitting plane distance variance, Gaussian curvature and mean curvature.

[0062] S101 also includes:

[0063] Three-dimensional point cloud data is acquired using a 3D scanner; the 3D point cloud data serves as the source point cloud.

[0064] The KD-tree method is used to determine the neighboring points within the neighborhood range corresponding to each query point in the three-dimensional point cloud data.

[0065] S101 specifically includes:

[0066] S1001, Determine the local density of the point cloud from the point in the neighborhood of each query point to the query point;

[0067] S1001 describes the density of the point cloud within a given range using local point cloud density. The local density is estimated by the average distance between points in the point cloud. The distance between points is represented by the distance between a given point and its nearest neighbor. A smaller average distance indicates a denser point cloud distribution; a larger average distance indicates a more dispersed point cloud distribution. For example... Figure 2 As shown, the distribution is more dispersed in flat areas and more dense in areas with greater curvature.

[0068] Assuming the query point Within the neighborhood The query point is selected using the KD-tree method. The KD-tree algorithm is more efficient than the nearest Euclidean distance method, using the nearest neighbor points. Represents points within the neighborhood. Distance between other points Represents points within the neighborhood. The minimum distance between it and other points is:

[0069] (1)

[0070] Within its neighborhood Point to The local density of the point cloud can be expressed as:

[0071] (2)

[0072] S1002, based on the fitting plane of the point cloud within the neighborhood range of each query point;

[0073] In S1002, the distribution of the point cloud within a neighborhood is effectively described by the variance of the distance from neighboring points to the fitted plane of the neighborhood. First, the fitted plane L of the point cloud within the neighborhood is obtained using the points in the neighborhood; then, the distance between each point in the neighborhood and L is calculated. Finally, the planar distance is fitted within the neighborhood range. The variance describes the distribution of the point cloud within this neighborhood. Figure 3 It can be seen that when the variance of the fitting plane distance is small, the fitting plane conforms to the distribution of the point cloud, and the point cloud distribution is relatively flat; when the variance of the fitting plane distance is large, the fitting plane deviates from the distribution of the point cloud, and the point cloud is more curved.

[0074] Through the neighborhood radius Given points, fit a quadratic surface using the least squares method. According to the least squares principle, we can obtain:

[0075]

[0076] Differentiate the coefficients in the above equation and set them to 0 to obtain the coefficients of the quadratic surface. Rewrite the curve equation in the form of the L-parameter equation of the surface:

[0077]

[0078] At this point, the variance of the fitted plane distance is:

[0079]

[0080] in, Neighboring points Distance to the fitted plane; This is the average value of the fitted plane distance.

[0081] Curvature is classified into principal curvature, Gaussian curvature, and mean curvature. Principal curvature is further divided into maximum principal curvature and minimum principal curvature, representing the curvature values ​​perpendicular to the surfaces of minimum and maximum curvature, respectively. Gaussian curvature Gaussian curvature is the product of the two principal curvatures. Its magnitude depends on the distance on the surface but is independent of how the surface is embedded in space; therefore, Gaussian curvature represents the internal geometry of a point cloud. The mean curvature of the surface... The curvature that describes the embedding of a surface in the surrounding space is used to represent the degree of external curvature of the surface. Therefore, combining Gaussian curvature and mean curvature can comprehensively depict the local curvature at a point.

[0082] The first fundamental formula for a surface can be expressed as:

[0083]

[0084] The second fundamental formula for a surface can be expressed as:

[0085]

[0086] Gaussian curvature:

[0087] (8)

[0088] Mean curvature:

[0089]

[0090] in, This is the first fundamental formula for surfaces. , , , This is the second fundamental formula for surfaces. It is the unit normal vector.

[0091] S102, statistically weight the local feature descriptor of each query point and the local feature descriptors of its corresponding neighboring points to obtain the statistical local feature descriptor (LFD) for the corresponding query point; define the four descriptors in the local features, with the point cloud local density as a parameter. Local distribution characteristics as parameters Gaussian curvature is a parameter The mean curvature is a parameter. The LFD representation is as follows .

[0092] Statistically weighting the feature descriptors of the query point and its neighbors can reduce the impact of outliers and noise within the neighborhood. Using the reciprocal of the distance between the query point and its neighbors as a weight can weaken the influence of distant points on the neighborhood features of the query point. Through statistical weighting, a Statistical Local Feature Descriptor (SLFD) can be obtained for the query point.

[0093] S102 specifically includes:

[0094] Using formula Determine the statistical local feature descriptor for the query point;

[0095] in, query point Statistical local feature descriptors, For query point Local feature descriptors, Neighboring points Local feature descriptors, The number of points within the neighborhood. For query point To nearby points The distance.

[0096] The calculation steps for SLFD are as follows:

[0097] Step 1: Select query point using KD-tree method The nearest point.

[0098] Step 2: Calculate the query point The local density of the point cloud, the variance of the distance between the fitted plane of the point cloud, the Gaussian curvature, and the mean curvature are used to obtain the local feature descriptor (LFD) of the point.

[0099] Step 3: As Figure 4 As shown, neighboring points LFD and query point The SLFD of the query point is obtained by statistically weighting the LFD.

[0100] Point cloud registration is the process of establishing a correspondence between corresponding point pairs in two point clouds, thereby creating a transformation model. Traditional geometric feature-based registration algorithms match point pairs with consistent features in two point clouds and then establish a transformation model to achieve point cloud registration. This paper determines the correspondence by calculating the feature differences between point pairs and uses the Mean Match Distance (MMD) instead of the Root Mean Squared Error (RMSE) as a metric to calculate the deviation between two point clouds. This algorithm is called the Feature Consistency Iterative Closest Point (FC-ICP) algorithm.

[0101] S103, Obtain the statistical local feature descriptors of the source point cloud and the target point cloud;

[0102] S104, Improve the ICP algorithm based on the feature differences and average matching distance between point pairs in the source point cloud and the target point cloud, and register the point cloud.

[0103] S104 specifically includes:

[0104] Select from the target point cloud using the farthest point sampling method 1 sample point, and according to Each sample point determines the sample point set;

[0105] The feature differences between corresponding point pairs in 3D point cloud data and sample point sets are matched to eliminate erroneous point pairs.

[0106] After describing the local features of a point cloud, finding corresponding matching point pairs is a challenge in point cloud registration. The correspondence is determined by the feature difference between two points. The feature difference (FD) between two different point clouds is shown below:

[0107]

[0108] in, For points in the source point cloud, For the corresponding point in the target point cloud, For the first Item parameters.

[0109] The smaller the FD (Feature Determination), the more similar the features between two points. When FD is 0, the features between two points are identical. To further improve the matching accuracy of points with the same features, a threshold of 0.002 is selected. When FD exceeds the threshold, the features of the two points are significantly different.

[0110] The relationship between corresponding points is determined by singular value decomposition, and the transformation model from 3D point cloud data to target point cloud is determined based on the relationship between corresponding points.

[0111] The ICP algorithm is improved by using the average matching distance as a metric to determine the target transformation model; the target transformation model is the transformation model when the average matching distance reaches the minimum value.

[0112] Point cloud registration is performed using a target transformation model.

[0113] Using formula Determine the average matching distance;

[0114] in, for and The average matching distance For sample points in the target point cloud; Statistical local feature descriptors and Corresponding points, This represents the number of sampled matching point pairs.

[0115] The specific steps for point cloud registration are as follows, and the flowchart of the registration algorithm is shown below. Figure 5 As shown:

[0116] Step 1: Calculate the SLFD of the source point cloud and the target point cloud;

[0117] Step 2: Select from the target point cloud T using the farthest point sampling (FPS) method. Let there be n sample points, and define these points as the sample point set Q.

[0118] Step 3: Find the point in the source point cloud P that has the same SLFD as the sample point set Q. Randomly select a point from the point set Q as the corresponding point in the point cloud P, minimizing its FD. If the FD is greater than the threshold, it means that there is no corresponding point, indicating that there is noise and outliers in the neighborhood of the point. In this case, remove the point from the point set Q.

[0119] Step 4: Calculate the relationship between corresponding points in the two point clouds using Singular Value Decomposition (SVD) to obtain the transformation model from the source point cloud to the target point cloud;

[0120]

[0121] in, The number of transformations.

[0122] Step 5: Calculate the average matching distance Measure the degree of matching between two point clouds;

[0123]

[0124] in, For sample points in the target point cloud; SLFD features in source point cloud and Corresponding points; This represents the number of sampled matching point pairs.

[0125] Step 6: Iterate through the above steps, updating the transformation relationship between the source point cloud and the target point cloud. Stop iterating when the minimum value is reached.

[0126] To verify the applicability and feasibility of the algorithm, this paper conducts verification on a computer with an Intel Core i5-6200 2.4 GHz CPU and 2GB RAM, using C++ in the Visual Studio 2019 environment and the PCL 1.11.0 point cloud public library. This paper uses the Bunny (35,947 points), China Dragon (437,645 points) point cloud models from the Stanford University point cloud dataset, and the Bottle point cloud (21,469 points) obtained by the Creaform Handy SCAN 700 high-precision industrial-grade handheld 3D laser scanner as experimental objects, setting up four sets of experiments.

[0127] The first set of experiments considered the impact of different initial poses on point cloud registration, using Bunny models with different initial poses as source point clouds for registration. The second set of experiments considered point cloud registration with varying degrees of data loss, selecting Bunny models with 10% (32,865 points), 30% (26,127 points), and 50% (17,325 points) data loss after random rotation and translation transformations as source point clouds for registration. The third set of experiments considered point cloud registration under different noise conditions, using the China Dragon model and adding Gaussian noise interference with standard deviations of 1mm, 2mm, and 3mm to the rotated original point cloud data as source point clouds for registration. Finally, the algorithm was validated in practical applications by performing random rotation and translation transformations on the actual Bottle point cloud data, partially truncating the data, and adding Gaussian noise interference with a standard deviation of 2mm (17,837 points).

[0128] This paper compares FC-ICP with the ICP algorithm, an ICP algorithm based on bidirectional distance scaling, an improved ICP point cloud registration algorithm based on fast point feature histograms, and a point cloud registration algorithm based on improved 3D shape context to verify the algorithm's advantages. Registration time is calculated from the start of reading point cloud data until registration is complete. The root mean square error of the distance between the registered point cloud and the target point cloud is used as the evaluation index for point cloud registration accuracy. Defined as:

[0129]

[0130] in, This represents the Euclidean distance between corresponding point pairs after registration. Therefore, the larger the root mean square error, the greater the distance between corresponding points in the two point clouds, meaning the lower the point cloud registration accuracy.

[0131] During the scanning process of 3D scanners, the scanned object will undergo rotation and displacement changes under different environments. ICP and its variants are prone to getting trapped in local optima when the initial pose is poor. To analyze the registration accuracy of the algorithm under non-ideal conditions, random rotation and translation of the point cloud were performed experimentally. The ICP algorithm based on bidirectional distance scaling is abbreviated as Bidirectional ICP. An improved ICP point cloud registration algorithm based on fast point feature histograms is denoted as Paper1, and a point cloud registration algorithm based on improved 3D shape context is denoted as Paper2. The registration results are as follows: Figure 6 As shown.

[0132] Table 1. Registration data of the Bunny model under different transformation states

[0133]

[0134] Table 1 shows the registration data of the Bunny model under different transformation states. As can be seen from the table, the ICP and bidirectional ICP algorithms perform registration using the minimum distance between point pairs, resulting in poor registration accuracy for point clouds with poor initial poses. Compared to these two algorithms, the FC-ICP algorithm improves accuracy by two orders of magnitude. Compared to the ICP algorithm, the FC-ICP algorithm improves speed by 25%, and compared to the bidirectional ICP algorithm, the improvement is even greater. The Paper1 and Paper2 algorithms perform point cloud registration by extracting feature points and then matching them using the local features of these feature points. Feature point extraction has limitations; therefore, although both algorithms still achieve good results with poor initial poses, their overall registration accuracy is poor. Compared to the Paper1 algorithm, the FC-ICP algorithm improves registration accuracy by over 93.45% and saves over 24.29% of time. Compared to the Paper2 algorithm, the FC-ICP algorithm improves registration time by over 24.7% and registration accuracy by over 95.29%. FC-ICP performs feature description on the overall point cloud and finds corresponding point pairs. As can be seen, the FC-ICP algorithm has good registration results for point clouds under any transformation state.

[0135] During the scanning process of 3D scanners, environmental factors such as occlusion and missing data can cause incomplete point cloud data. To analyze the registration accuracy of the algorithm under non-ideal conditions, the original point cloud data was subjected to rotation and translation transformations. Then, Bunny models with 10%, 30%, and 50% data missing values ​​were selected as source point clouds for registration to verify the algorithm. Some point cloud registration results are shown below. Figure 7 As shown.

[0136] Table 2. Registration data of the Bunny model under different transformation states

[0137]

[0138] Table 2 shows the registration data of the Bunny model under different levels of data loss. As can be seen from the table, partially missing point clouds have a significant impact on ICP and its variants. The ICP and bidirectional ICP algorithms have poor registration accuracy for partially missing point clouds. FC-ICP registration accuracy is more than 99.7% higher than the ICP algorithm, saving more than 21.41% of time, and 99.25% higher than the bidirectional ICP algorithm, saving even more time. Partially missing point clouds have a smaller impact on point cloud registration algorithms based on traditional geometric features, such as Paper1 and Paper2. Compared with the Paper1 algorithm, the point cloud registration accuracy is improved by 67.33%, and the speed is improved by 26.1%. Compared with the Paper1 algorithm, the point cloud registration accuracy is improved by 89.11%, and the speed is improved by 16.3%. The FC-ICP algorithm finds the same SLFD to establish correct matching point pairs and removes non-matching sampling points. Therefore, this algorithm can obtain better registration results even with partially missing point cloud data, effectively improving registration accuracy.

[0139] During the scanning process of 3D scanners, noise interference can cause anomalies in point cloud data. Using ChinaDragon as the target point cloud, random rotation transformations were applied to the original point cloud data, followed by the addition of random Gaussian noise with standard deviations of 1mm, 2mm, and 3mm. The registration accuracy and time of the algorithm in noisy environments were compared to verify its robustness. The data processing time was measured from the start of point cloud reading, using the root mean square error before and after registration as the metric. Due to the large amount of data in the ChinaDragon model, this part of the experiment took a considerable amount of time. The registration results of the ChinaDragon model under different noise conditions are shown below. Figure 8 As shown.

[0140] Table 3. Bunny model registration data under different levels of data loss.

[0141]

[0142] Table 3 shows the registration data of the China Dragon model under different noise conditions. According to Table 3, under high noise conditions, FC-ICP is affected by SLFD, resulting in inaccurate point-to-point matching, increased number of point cloud model transformation iterations, and increased registration time. The ICP algorithm and the bidirectional ICP algorithm get stuck in the optimal solution, resulting in lower registration accuracy. Since both Paper1 and Paper2 algorithms extract feature points through local point cloud features, noise significantly impacts feature point extraction. Compared to Paper1, the registration accuracy is improved by 48.95%, and the speed is improved by 5.24%. The Paper2 algorithm is significantly affected by noise and has poor robustness. FC-ICP improves registration accuracy by 22.75% and registration time by 6.89%. Therefore, the FC-ICP algorithm can effectively complete registration when noise is low. However, when noise is high, estimating the point cloud curvature requires planar fitting of the neighborhood, which introduces bias, leading to significant errors in curvature estimation. The robustness of FC-ICP still needs improvement.

[0143] To verify the effectiveness of FC-ICP in practical applications, point clouds of shower gel bottles obtained by the Creaform Handy SCAN 700 high-precision industrial-grade handheld 3D laser scanner were used as experimental data. To simulate the registration process of point clouds in actual situations, random rotation and translation transformations were performed on the original data, a portion of the data was cut off, and Gaussian noise with a mean square error of 2 mm was added. Figure 9 Table 4 shows the registration results for the Bottle object, including the registration error and registration time.

[0144] Table 4 Registration data for the actual Bottle

[0145]

[0146] As shown in Table 4, during the physical registration process, the ICP algorithm and the bidirectional ICP algorithm get stuck in local optima. The Paper1 and Paper2 algorithms are affected by noise and have low accuracy. Compared with other algorithms, the FC-ICP algorithm improves accuracy by more than 70.52% and registration rate by more than 9.92%.

[0147] To address the challenges of point cloud registration under conditions of initial pose difference transformation, missing data, and noise interference, this paper proposes a point cloud registration algorithm based on statistical local feature description and matching. A 4-dimensional statistical local feature descriptor (SLFD) is constructed using point cloud local density, variance of the fitted plane distance, Gaussian curvature, and mean curvature, providing a more accurate and comprehensive description of the local features of the point cloud. Corresponding points are found by identifying feature differences between point pairs, establishing the correspondence, and an improved FC-ICP algorithm is used for point cloud model registration. Experimental results using the Stanford University point cloud dataset and a real-world dataset show that the registration accuracy is improved by over 67.33% in the absence of noise and by over 22.75% in the presence of noise. The registration speed is significantly improved when the point cloud data is limited, saving 16.3% of the time; when the point cloud data is large, the speed is improved by 5.24%. Therefore, under different environments, the algorithm presented in this paper has higher registration accuracy and registration speed, and stronger robustness compared with the ICP algorithm, bidirectional ICP algorithm, improved ICP point cloud registration algorithm based on fast point feature histogram, and point cloud registration based on improved 3D shape context.

[0148] Corresponding to the above method, the present invention also provides a point cloud registration system based on statistical local feature description and matching, comprising:

[0149] The local feature descriptor determination module is used to determine the local feature descriptor of each query point and the local feature descriptor of the corresponding neighboring points in the three-dimensional point cloud data; the local feature descriptor includes: point cloud local density, point cloud fitting plane distance variance, Gaussian curvature and mean curvature.

[0150] The statistical local feature descriptor determination module is used to statistically weight the local feature descriptor of each query point and the local feature descriptor of the corresponding neighboring points to obtain the statistical local feature descriptor of the corresponding query point.

[0151] The module for obtaining statistical local feature descriptors of source and target point clouds is used to obtain statistical local feature descriptors of source and target point clouds.

[0152] The point cloud registration module is used to improve the ICP algorithm based on the feature differences and average matching distance between point pairs in the source and target point clouds, and to register the point clouds.

[0153] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0154] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A point cloud registration method based on statistical local feature description and matching, characterized in that, include: Determine the local feature descriptor of each query point and the local feature descriptor of its corresponding neighboring points in the 3D point cloud data respectively; The local feature descriptors include: local density of the point cloud, variance of the distance between the point cloud fitting plane, Gaussian curvature, and average curvature; the three-dimensional point cloud data is the source point cloud; The local feature descriptor of each query point and the local feature descriptor of its corresponding neighboring points are statistically weighted to obtain the statistical local feature descriptor of the corresponding query point. Obtain statistical local feature descriptors for the source and target point clouds; The ICP algorithm is improved based on the feature differences and average matching distance between point pairs in the source point cloud and the target point cloud to perform point cloud registration. The step of statistically weighting the local feature descriptor of each query point and the local feature descriptors of its corresponding neighboring points to obtain the statistical local feature descriptor of the corresponding query point specifically includes: Using formula Determine the statistical local feature descriptor for the query point; in, query point Statistical local feature descriptors, For query point Local feature descriptors, Neighboring points Local feature descriptors, The number of points within the neighborhood. For query point To nearby points The distance; The ICP algorithm is improved based on the feature differences and average matching distance between point pairs in the source and target point clouds to perform point cloud registration, specifically including: Select from the target point cloud using the farthest point sampling method 1 sample point, and according to Each sample point determines the sample point set; Perform corresponding point matching on the feature differences between point pairs in 3D point cloud data and sample point sets, and remove erroneous point pairs; The relationship between corresponding points is determined by singular value decomposition, and the transformation model from 3D point cloud data to target point cloud is determined based on the relationship between corresponding points. The ICP algorithm is improved by using the average matching distance as a metric to determine the target transformation model; the target transformation model is the transformation model when the average matching distance reaches the minimum value. Point cloud registration is performed using a target transformation model.

2. The point cloud registration method based on statistical local feature description and matching according to claim 1, characterized in that, Before determining the local feature descriptor of each query point and the corresponding local feature descriptor of its neighboring points in the 3D point cloud data, the process also includes: Three-dimensional point cloud data is acquired using a 3D scanner; the 3D point cloud data serves as the source point cloud. The KD-tree method is used to determine the neighboring points within the neighborhood range corresponding to each query point in the three-dimensional point cloud data.

3. The point cloud registration method based on statistical local feature description and matching according to claim 2, characterized in that, The step of determining the local feature descriptor of each query point and the local feature descriptor of its corresponding neighboring points in the 3D point cloud data specifically includes: Determine the local density of the point cloud from the points within the neighborhood of each query point to the query point; Based on the fitted plane of the point cloud within the neighborhood range of each query point; Determine the distance from the nearest points within the neighborhood of each query point to the fitted plane; The variance of the point cloud fitting plane distance is determined by the variance of the distances from the nearest points in the neighborhood of each query point to the fitting plane.

4. The point cloud registration method based on statistical local feature description and matching according to claim 1, characterized in that, Matching corresponding points based on feature differences between point pairs in 3D point cloud data and sample point sets, specifically including: Using formula Determine the feature differences between point pairs in the source point cloud and the target point cloud; in, Point pairs between source and target point clouds The characteristic differences between them For points in the source point cloud, For the corresponding point in the target point cloud, For the first Item parameters.

5. The point cloud registration method based on statistical local feature description and matching according to claim 1, characterized in that, The improvement of the ICP algorithm using average matching distance as a metric to determine the target transformation model specifically includes: Determine the current average matching distance between points in the current 3D point cloud data and corresponding points in the target point cloud in the current transformation model; Determine if the current average matching distance has reached the minimum value; If the target transformation model is achieved, then the current transformation model will be determined as the target transformation model. If the target is not reached, then return to the method of selecting from the target point cloud using the farthest point sampling. 1 sample point, and according to The steps for determining the sample point set are: update the transformation model for each sample point until the average matching distance reaches the minimum value.

6. The point cloud registration method based on statistical local feature description and matching according to claim 5, characterized in that, Using formula Determine the average matching distance; in, for and The average matching distance For sample points in the target point cloud; Statistical local feature descriptors and Corresponding points, This represents the number of sampled matching point pairs.

7. A point cloud registration system based on statistical local feature description and matching, used to implement the point cloud registration method based on statistical local feature description and matching as described in any one of claims 1-6, characterized in that, include: The local feature descriptor determination module is used to determine the local feature descriptor of each query point and the local feature descriptor of its corresponding neighboring points in the 3D point cloud data; the local feature descriptor includes: point cloud local density, point cloud fitting plane distance variance, Gaussian curvature, and mean curvature; the 3D point cloud data is the source point cloud; The statistical local feature descriptor determination module is used to statistically weight the local feature descriptor of each query point and the local feature descriptor of the corresponding neighboring points to obtain the statistical local feature descriptor of the corresponding query point. The module for obtaining statistical local feature descriptors of source and target point clouds is used to obtain statistical local feature descriptors of source and target point clouds. The point cloud registration module is used to improve the ICP algorithm based on the feature differences and average matching distance between point pairs in the source and target point clouds, and to register the point clouds.

Images