A method, system and medium for creating a three-dimensional feature descriptor based on local surface change information

Abstract

The present invention discloses a method, system and medium for creating a three-dimensional feature descriptor based on local surface change information. The method is used to realize the joint encoding of local surface geometry and spatial information, including local feature descriptor calculation, encoding the geometry and spatial information of the neighborhood of key points in the point cloud into a high-dimensional vector; correspondence estimation, finding matching point pairs on different point clouds based on the similarity of feature descriptors, and filtering out erroneous matching points using certain constraint screening conditions. The present invention establishes an LRA on the local surface of the point cloud, and then simultaneously encodes the spatial and geometric information of the local surface of the point cloud. The spatial information is encoded by radially dividing the local space on the LRA; the geometric information is encoded by counting five geometric attributes with strong robustness, so that the LSVSH feature descriptor shows the best performance in various evaluation indicators, while achieving a good balance in efficiency, descriptiveness and robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud coding for image processing, and in particular to a method, system and medium for creating a three-dimensional feature descriptor based on local surface change information. Background Art

[0002] In recent years, with the development of 3D computer vision equipment and applications, the processing of 3D point cloud data has attracted increasing research and attention. Local feature description of point clouds, a key technology in 3D computer vision, is widely used in object registration, recognition, and reconstruction. However, because the acquired raw point cloud data often contains interference such as partial overlap, occlusion, and noise, highly discriminative and robust description of local feature in point clouds remains a challenging task.

[0003] In the past two decades, researchers have proposed many local feature descriptors in an attempt to solve this problem. These local feature descriptors are divided into three categories: normal vector-based, local reference frame (LRF)-based, and local reference axis (LRA)-based. The normal vector-based local feature descriptor uses a normal vector with low repeatability as the LRA, which results in low robustness of this type of algorithm and is sensitive to noise and point cloud resolution changes. Although the LRF-based feature descriptor can obtain more dimensional information of the support area, due to the low repeatability of the x and y axes in the LRF, it is unable to accurately describe the local information and exhibits weak robustness. Compared with the previous two algorithms, the LRA-based feature descriptor has superior descriptiveness and robustness.

[0004] In practical applications, a local feature descriptor needs to meet several performance requirements (such as descriptiveness, robustness, compactness, efficiency, etc.). However, existing local feature descriptions find it difficult to strike a balance between these performance requirements. Moreover, since most of them only encode the spatial information or geometric features of the local surface of the point cloud, it is difficult to achieve a comprehensive and robust description of the local surface spatial and geometric information. This will inevitably lead to limited descriptiveness and robustness of the local feature descriptors. For example, the designed local feature descriptors can only be applied to specific scenarios, have poor generalization capabilities, and have low description accuracy when faced with interference such as noise, point cloud resolution changes, and occlusion. Although some feature descriptors encode local spatial information and geometric information simultaneously, the feature attributes used are weakly discriminable and robust, which is insufficient to describe all the feature information of the local surface. Therefore, their descriptiveness and robustness need to be further improved. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a three-dimensional feature descriptor based on local surface variation information, introduce a new geometric attribute - curvature attribute, to further improve the performance of the algorithm, and propose a new feature descriptor called Local Surface Variation based StatisticsHistogram (LSVSH for short), which realizes the joint encoding of local surface geometry and spatial information. Compared with some existing feature attributes, the proposed curvature attribute has stronger discrimination power and is more robust to interference such as noise. With the help of this curvature attribute, the LSVSH feature descriptor not only has higher descriptiveness and efficiency, but also has stronger robustness to interference such as noise, different grid resolutions and clutter.

[0006] To solve the above technical problems, the present invention adopts a technical solution: providing a method for creating a three-dimensional feature descriptor based on local surface change information, which is used to realize the joint encoding of local surface geometry and spatial information. The creation method includes the following steps:

[0007] Step 1: Calculate the local descriptor to encode the geometric and spatial information of the neighborhood of the key point of the point cloud into a high-dimensional vector. The specific steps are as follows:

[0008] Step 1.1: Given a point cloud or surface, obtain the points of the key point p within the spherical neighborhood of the support radius using the kd_tree technique;

[0009] Step 1.2: Establish LRA for the spherical neighborhood of key point p;

[0010] Step 1.3: Divide the local space into several subspaces evenly along the radial direction;

[0011] Step 1.4: For each point in the subspace, encode five geometric attributes, including four angle attributes and one curvature attribute;

[0012] Step 1.5: Generate statistical histograms corresponding to the five geometric attributes in each radially divided subspace, and normalize them to 1.

[0013] Step 1.6: Use different weights to weight the five subhistograms together to form the final LSVSH histogram.

[0014] Step 2: Estimation of correspondence relationships: Find matching point pairs on different point clouds based on the similarity of feature descriptors, and use certain constraint filtering conditions to filter out incorrect matching points. The specific steps are as follows:

[0015] Step 2.1: Extract a certain number of key points from the model point cloud and scene point cloud respectively, and generate corresponding feature descriptors;

[0016] Step 2.2: Find the two scene features with the closest and second closest Euclidean distance to each model feature in the scene features. If the ratio of the closest distance to the second closest distance is less than a certain set threshold, the model feature and the closest scene feature are considered to form a matching pair;

[0017] Step 2.3: If the distance between the corresponding feature points of a certain set of feature matching pairs is small enough, then this set of feature matching pairs is considered to be a correct matching pair; otherwise, it is considered to be an incorrect matching pair.

[0018] Furthermore, in step 1.2, the specific steps for establishing LRA for the spherical neighborhood of key point p are:

[0019] Step 1.2.1, given key points p and R LRA , and get R LRA The point set within the determined spherical area is defined as Q LRA ={q1,q2,q3,...,q n}, where n is the number of neighboring points of p;

[0020] Step 1.2.2, set Q LRA The mean coordinates of all neighborhood points in , based on Q LRA Perform covariance analysis, and the covariance matrix cov(p) can be calculated as:

[0021]

[0022] Where q j represents the jth neighboring point, and n is the number of neighboring points of p;

[0023] Step 1.2.3. Calculate the eigenvector LRA(p) corresponding to the minimum eigenvalue of cov(p), and correct the direction of LRA(p) using the following formula:

[0024]

[0025] Where LRA(p) is the direction of LRA at point p, “·” represents the dot product between two vectors, and n is Q LRA The number of midpoints, n(q i ) is the normal vector of the j-th neighborhood point.

[0026] Furthermore, in step 1.3, the specific steps of evenly dividing the local space into several subspaces along the radial direction are as follows:

[0027] Step 1.3.1: Given a point cloud or surface area, support radius R descriptor, extract the point set in the spherical neighborhood of the key point p, defined as Q descriptor ={q1,q2,q3,...,q k}, where k is the number of neighborhood points;

[0028] Step 1.3.2: Generate the local LRA at the key point p using the method in step 1.2.

[0029] Step 1.3.3: After generating the local space of p, divide the local space into N uniformly along the radial direction. r subspaces;

[0030] Step 1.3.4: For each point in the subspace, encode five geometric attributes Among them, α, β, η and δ are four angle attributes, and their value range is [0,180°]. It is the curvature property, and its value range is [0,1].

[0031] Furthermore, in step 1.3.4, for each neighborhood point q in the subspace i , four angle attributes (α(q i ),β(q i ),η(q i ),δ(q i )) means as follows:

[0032]

[0033] In the formula pq i Indicates p to q i The vector, n(q i ) represents q i where LRA(p) is the normal vector at p, LRA(p) represents the LRA direction at p, “×” represents the cross product between two vectors, “·” represents the dot product between two vectors, i is an integer and i∈[1,k].

[0034] For a neighborhood point q i , its curvature properties The approximate solution is as follows:

[0035]

[0036] Where λ1(q i ),λ2(q i ) and λ3(q i ) represents q i The eigenvalues ​​of the three principal axis directions of the surface, and λ1(q i )≥λ2(q i )≥λ3(q i )>0.

[0037] Furthermore, in order to improve Characterize the effect of local surface geometric features, further execute the following formula, Scaled to a normal distribution with center at 0 and standard deviation at 1:

[0038]

[0039] In the formula It is q i The mean of the curvature values ​​of the surface, It is q i The standard deviation of the surface curvature values.

[0040] Furthermore, in step 1.5, the method for generating the statistical histograms corresponding to the five geometric attributes is:

[0041] The statistical histograms of the five geometric attributes are recorded as H1, H2, H3, H4 and H5 respectively, and the number of blocks used to count the five geometric attributes is recorded as N. α 、N β 、N η 、N η and Then the lengths of H1, H2, H3, H4 and H5 are N respectively. r ×N α 、N r ×N β 、N r ×N η 、N r ×N δ and

[0042] Furthermore, in step 1.6, the five sub-histograms are weightedly connected to form the final LSVSH histogram represented as: f{γ1H1,γ2H2,γ3H3,γ4H4,γ5H5}, where γ1, γ2, γ3, γ4 and γ5 are the weights corresponding to the five geometric attributes.

[0043] To achieve the above-mentioned objectives, the present invention also provides a system for creating a three-dimensional feature descriptor based on local surface change information. The system includes a memory, a processor, and a program for creating a three-dimensional feature descriptor based on local surface change information stored on the processor. When the program for creating a three-dimensional feature descriptor based on local surface change information is run by the processor, all the steps of the creation method described above are executed.

[0044] To achieve the above-mentioned purpose, a computer-readable storage medium is provided, comprising a program for creating a three-dimensional feature descriptor based on local surface change information stored on the computer-readable storage medium, wherein the program for creating a three-dimensional feature descriptor based on local surface change information executes all the steps of the creation method described above when the program is run by a processor.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] The present invention discloses a three-dimensional feature descriptor based on local surface change information. The descriptor first establishes an LRA on the local surface of the point cloud, and then simultaneously encodes the spatial and geometric information of the local surface of the point cloud. Spatial information is encoded by radially dividing the local space on the LRA; geometric information is encoded by counting five highly robust geometric attributes (including the introduced curvature attribute). Finally, the performance of the LSVSH feature descriptor is comprehensively evaluated through a large number of experiments. The experimental results show that compared with other feature descriptors, the LSVSH feature descriptor exhibits the best performance in various evaluation indicators, while achieving a good balance in efficiency, descriptiveness and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of two commonly used local coordinate systems: LRF and LRA;

[0048] Figure 2 Schematic diagram of the process of generating LSVSH feature descriptors for the present invention;

[0049] Figure 3 Schematic diagram of parameter selection for the LSVSH feature descriptor in the embodiment;

[0050] Figure 4 Two example models and two corresponding example scenarios in the four data sets in the embodiment;

[0051] Figure 5 The RPC performance evaluation results of the nine feature descriptors in the embodiment on the B3R dataset are shown;

[0052] Figure 6 The RPC performance evaluation results of the nine feature descriptors in the embodiment on the U3M, U3OR and QuLD datasets are shown;

[0053] Figure 7 The time required to calculate nine descriptors at different support radii in the embodiment. DETAILED DESCRIPTION

[0054] The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more precise definition of the protection scope of the present invention.

[0055] See also Figure 1 , Local Reference Axis (LRA) and Local Reference Frame (LRF) are two commonly used local coordinate systems for constructing local feature descriptors. LRF is composed of three mutually orthogonal coordinate axes, which can represent complete local spatial information (i.e. information in radial, azimuth and elevation directions). Compared with LRF, LRA has only one guide axis and lacks spatial information in the azimuth direction. Although LRF can represent more information in the local space, the repeatability of its x and y axes is significantly lower than that of LRA, especially under interference such as noise, point cloud resolution changes and occlusion. Using axes with less repeatability to encode spatial information in the local space may have a negative impact on the performance of the descriptor. In order to more accurately segment the space, the present invention is constructed based on LRA.

[0056] In order to improve the robustness of LRA to interference such as noise, and also to improve the time efficiency of calculating LRA, through experiments, we choose to calculate the radius of LRA (defined as R LRA ) is set to 15mr (mr is the average distance between all nearest point pairs in the point cloud). Given key points p and R LRA , we can get R LRA The point set within the determined spherical area is defined as Q LRA ={q1,q2,q3,...,q n}(where n is the number of neighboring points of p). Q LRA The mean coordinates of all neighborhood points in , based on Q LRA Perform covariance analysis, and the covariance matrix cov(p) can be calculated as:

[0057]

[0058] Where q j represents the jth neighborhood point, and n is the number of neighborhood points of p.

[0059] Calculate the eigenvector LRA(p) corresponding to the minimum eigenvalue of cov(p). The normal vector determined by LRA(p) is usually ambiguous, that is, the direction of the normal vector cannot be determined. Here, the following formula is used to correct the direction of LRA(p).

[0060]

[0061] Where LRA(p) is the direction of LRA at point p, “·” represents the dot product between two vectors, and n is Q LRA The number of midpoints, n(q i ) is the normal vector of the j-th neighborhood point.

[0062] After constructing LRA, the next task is to design a suitable feature representation method to encode local surface information. Given a point cloud or surface area, support radius (defined as R descriptor ), extract the point set in the spherical neighborhood of the key point p, defined as Q descriptor ={q1,q2,q3,...,q k} (where k is the number of neighborhood points), and use the above method to generate the local LRA at p, such as Figure 2 (a) and 2(b). After generating the local space of p, the local space is evenly divided into N r subspaces, such as Figure 2 (c) For each point in the subspace, five geometric attributes are encoded like Figure 2 (d)-2(h). α, β, η and δ are four angle attributes, with a value range of [0,180°]. It is the curvature property, and its value range is [0,1].

[0063] For a neighborhood point q i , four angle attributes (α(q i ),β(q i ),η(q i ),δ(q i )) means as follows:

[0064]

[0065] In the formula pq i Indicates p to q i The vector, n(q i ) represents q i where LRA(p) is the normal vector at p, LRA(p) represents the LRA direction at p, “×” represents the cross product between two vectors, “·” represents the dot product between two vectors, i is an integer and i∈[1,k].

[0066] For a neighborhood point q i , its curvature properties The approximate solution is as follows:

[0067]

[0068] Where λ1(q i ),λ2(q i) and λ3(q i ) represents q i The eigenvalues ​​of the three principal axis directions of the surface, and λ1(q i )≥λ2(q i )≥λ3(q i )>0.

[0069] To improve Characterize the effect of local surface geometric features, further execute the following formula, Normal distribution scaled to have a center of 0 and a standard deviation of 1.

[0070]

[0071] In the formula It is q i The mean of the curvature values ​​of the surface, It is q i The standard deviation of the surface curvature values.

[0072] After calculating the five geometric attributes of all local points, their distribution is statistically analyzed in each subspace divided along the radial direction. The statistical histograms of the five geometric attributes are denoted as H1, H2, H3, H4 and H5, respectively. Figure 2 (i)-2(m). Assume that the number of blocks used to count the five geometric attributes is N α 、N β 、N η 、N η and Then the lengths of H1, H2, H3, H4 and H5 are N respectively. r ×N α 、N r ×N β 、N r ×N η 、N r ×N δ and In order to enhance the robustness to different grid resolutions, the five sub-histograms are normalized to 1. Since each sub-histogram contributes differently to the performance of the descriptor, the five sub-histograms are weighted and connected to form the final LSVSH histogram: f{γ1H1,γ2H2,γ3H3,γ4H4,γ5H5}, where γ1, γ2, γ3, γ4 and γ5 are the weights corresponding to the five geometric attributes, as shown in Figure 2 (n) and 2(o).

[0073] The LSVSH feature descriptor has twelve key parameters, namely the number of spatial subdivisions N in the radial direction r , Count the number of blocks used for the five geometric attributes (N α 、Nβ 、N η 、N η and ), the weight parameters of the five statistical histograms (γ1, γ2, γ3, γ4 and γ5) and the support radius R descriptor , refer to the relevant paper settings (such as Yang JQ, Cao ZG, Zhang QA fast and robust local descriptor for 3D point cloud registration[J]. Information Sciences, 2016, 346-347: 163-179) to set R descriptor The parameter is set to 20 μr. To obtain the values ​​of the remaining eleven parameters, a scene with 0.3 μr standard deviation Gaussian noise and 1 / 4 grid decimation from the University of Bologna 3D Retrieval (B3R) dataset was selected for testing. Descriptor performance is typically evaluated using the recall and 1-precision curves (RPC curves). To present the performance of descriptors compactly and quantitatively, the area enclosed by the RPC curve and the coordinate axes (defined as AUCpr) was used to measure the performance of the LSVSH descriptor under different parameter settings.

[0074] The following first introduces the process of drawing the RPC curve:

[0075] Step 1: Extract key points from the source point cloud P and the target point cloud Q, and generate the corresponding feature descriptor set, denoted as (i is the number of key points in the source point cloud P) and (j is the number of key points in the target point cloud Q).

[0076] Step 2: In the source point cloud feature descriptor set F P Search for each target point cloud feature descriptor set F P The feature descriptors with the closest and second closest Euclidean distance are represented as σ1 and σ2 respectively, and σ is used to represent the ratio of σ1 and σ2: σ = σ1 / σ2. If σ is less than a set threshold, it is considered that the source point cloud feature and the nearest target point cloud feature constitute a matching pair.

[0077] Step 3: If the distance between the corresponding feature points of a set of feature matching pairs is small enough (for example, less than the support radius R descriptor half of the feature matching pair), then this set of feature matching pairs is considered to be a correct matching pair, otherwise it is considered to be an incorrect matching pair. Finally, the recall rate (recall) and precision (precision) are defined as follows:

[0078]

[0079] While using RPC to evaluate the performance of 3D feature descriptors, in order to present the performance of the descriptor more compactly and quantitatively, the area enclosed by the RPC curve and the coordinate axes (AUCpr) is used to measure the performance of the descriptor. The ideal AUCpr value should be close to 1. A higher AUCpr value indicates a stronger performance of the feature descriptor, and vice versa.

[0080] The following introduces the twelve key parameters of LSVSH feature descriptor (R descriptor The setting process of the fixed setting is 20mr:

[0081] During parameter setting, the current parameter value is selected based on the fixed values ​​of other parameters. Specifically, when measuring a parameter value, other parameters are kept fixed, allowing the value of this parameter to vary within a defined range. Based on the test results, the optimal value of this parameter is selected, fixed, and the values ​​of other parameters are continued to be measured using this parameter value.

[0082] N r 、N α 、N β 、N η 、N η and The initial values ​​of are set to 5, 15, 15, 15, 15 and 12 respectively, and the range of variation is set to 2 to 20 with a step size of 1; the initial values ​​of weight parameters γ1, γ2, γ3, γ4 and γ5 are all set to 1, and the range of variation is set to 0.1 to 1 with a step size of 0.1, such as Figure 3 As shown in Table 1, the solid marks represent the selected parameter values. Table 1 shows the process of LSVSH feature descriptor parameters.

[0083] Table 1 Parameter setting process of LSVSH feature descriptor

[0084]

[0085] Based on the test results, the descriptiveness and compactness of the descriptor are weighed, N r 、N α 、N β 、N η 、N η 、 The values ​​of γ1, γ2, γ3, γ4, and γ5 are set to 8, 13, 12, 12, 13, 13, 0.9, 1.3, 1.2, 0.6, and 1, respectively.

[0086] To comprehensively evaluate the performance of LSVSH feature description, four popular benchmark datasets were selected. These are B3R, the University of Western Australia 3D Object Reconstruction Dataset (UWA3M), the University of Western Australia 3D Object Recognition Dataset (UWAOR), and the Queen's University LiDAR Dataset (QuLD). Detailed features of these datasets are listed in Table 2.

[0087] Table 2 Detailed characteristics of the four benchmark datasets

[0088]

[0089] Figure 4 Some example models and corresponding scenes from four datasets are shown. These datasets were chosen because they contain point cloud data of different quality and interference types (such as noise and occlusion), ensuring a comprehensive and adequate evaluation of the descriptors.

[0090] Figure 5 The RPC performance evaluation of the present invention and other eight mainstream local feature descriptors on the B3R dataset is shown. The numbers in brackets are the AUCpr values ​​calculated on the curve. The B3R dataset can evaluate the robustness of the descriptor to noise and different grid resolutions. It can be observed that under different interference type point cloud conditions (including Gaussian noise of different intensities, grid extraction, and a combination of Gaussian noise and grid extraction), the performance of the present invention is higher than that of other feature descriptors. In particular, when facing the harsh test scenario of 0.5mr standard deviation Gaussian noise combined with 1 / 8 grid extraction, the performance of the present invention is still stable enough, such as Figure 5 (f) This shows that the present invention is sufficiently robust to noise and varying grid resolution.

[0091] Figure 6 The RPC performance evaluation of the proposed method and eight other mainstream local feature descriptors on the U3M, U3OR, and QuLD datasets is shown. The numbers in parentheses are the AUCpr values ​​calculated on the curves. The U3M and U3OR datasets can be used to evaluate the robustness of feature descriptors to clutter, occlusion, and mesh boundaries. Figure 7 The time required to calculate the nine descriptors under different support radii is shown, and the numbers in brackets represent the corresponding average time consumption. From Figure 6(a) and Figure 6 (b) It can be seen that the proposed method achieves the best performance on both datasets and is significantly ahead of other feature descriptors. This verifies that the proposed method is highly robust to clutter, occlusion, and grid boundaries. The QuLD dataset incorporates interference such as clutter, occlusion, noise, and varying resolution. Figure 6(c) It can be observed that the performance of all feature descriptors, including the proposed method, on the QuLD dataset degrades significantly. The proposed method achieves the best performance, significantly outperforming the other feature descriptors. This demonstrates that the proposed method has high descriptive power and robustness to various common interferences, and also performs well on low-quality datasets.

[0092] To verify the time efficiency of our present invention, we compared and evaluated the time efficiency of nine local feature descriptors, including ours. Since computational efficiency is only related to the number of points within the support radius, we only use the B3R dataset to evaluate the efficiency of different descriptors. Specifically:

[0093] Step 1: First, randomly extract 1000 key points from each model in the B3R dataset, and extract a total of 6000 key points.

[0094] Step 2: Increase the support radius from 5mr to 40mr in steps of 5mr. Generate features for the nine local feature descriptors at keypoints at different radii. For a specific support radius, record the total time it takes for each descriptor to generate features at 6000 keypoints.

[0095] Figure 7 The time statistics of all nine local feature descriptors are shown, and the numbers in brackets represent the corresponding average time consumption. It can be observed from the figure that with the increase of the support radius, the calculation time of all nine feature descriptors gradually becomes longer. Specifically, the SI feature descriptor shows the highest time efficiency, followed by the SDASS feature descriptor. This is because the SI feature descriptor only considers the spatial information of the local surface and is more efficient. The LSVSH and DLFS feature descriptors also achieved relatively superior efficiency performance. RoPS and TriSI are the two most time-consuming feature descriptors, and the time spent is significantly longer than the other seven feature descriptors. This is mainly because the LRF constructed in these two descriptors is more time-consuming.

[0096] The purpose of the present invention is to disclose a three-dimensional feature descriptor based on local surface change information. The descriptor first establishes an LRA on the local surface of the point cloud, and then simultaneously encodes the spatial and geometric information of the local surface of the point cloud. Spatial information is encoded by radially dividing the local space on the LRA; geometric information is encoded by counting five highly robust geometric attributes (including the introduced curvature attribute). Finally, the performance of the LSVSH feature descriptor is comprehensively evaluated through a large number of experiments. The experimental results show that compared with other feature descriptors, the LSVSH feature descriptor shows the best performance in various evaluation indicators, while achieving a good balance in efficiency, descriptiveness and robustness.

[0097] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention's description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for creating a three-dimensional feature descriptor based on local surface change information, characterized by: The method for realizing the joint encoding of local surface geometry and spatial information includes the following steps: Step 1: Calculate the local descriptor to encode the geometric and spatial information of the neighborhood of the key point of the point cloud into a high-dimensional vector. The specific steps are as follows: Step 1.1: Given a point cloud or surface, obtain the points of the key point p within the spherical neighborhood of the support radius using the kd_tree technique; Step 1.2: Establish LRA for the spherical neighborhood of key point p; Step 1.3: Divide the local space into several subspaces evenly along the radial direction; Step 1.4: For each point in the subspace, encode five geometric attributes, including four angle attributes and one curvature attribute; Step 1.5: Generate statistical histograms corresponding to the five geometric attributes in each radially divided subspace, and normalize them to 1. Step 1.6: Use different weights to weight the five subhistograms together to form the final LSVSH histogram. Step 2: Estimation of correspondence relationships: Find matching point pairs on different point clouds based on the similarity of feature descriptors, and use certain constraint filtering conditions to filter out incorrect matching points. The specific steps are as follows: Step 2.1: Extract a certain number of key points from the model point cloud and scene point cloud respectively, and generate corresponding feature descriptors; Step 2.2: Find the two scene features with the closest and second closest Euclidean distance to each model feature in the scene features. If the ratio of the closest distance to the second closest distance is less than a certain set threshold, the model feature and the closest scene feature are considered to form a matching pair; Step 2.3: If the distance between the corresponding feature points of a certain set of feature matching pairs is small enough, then this set of feature matching pairs is considered to be a correct matching pair; otherwise, it is considered to be an incorrect matching pair.

2. The method for creating a three-dimensional feature descriptor based on local surface change information according to claim 1, characterized in that: In step 1.2, the specific steps to establish the LRA for the spherical neighborhood of the key point p are: Step 1.2.1, given key points p and ,get The point set within the determined spherical area is defined as , where n is the number of neighboring points of p; Step 1.2.2, set for The mean coordinates of all neighborhood points in , based on right Perform covariance analysis, and its covariance matrix It can be calculated as: In the formula Indicates the Neighborhood points, is the number of p neighborhood points; Step 1.2.3, calculation The eigenvector corresponding to the minimum eigenvalue and corrected by the following formula Directions: Where is the direction of the LRA at point p, "." represents the dot product between two vectors, for The number of midpoints, For the Normal vectors of neighboring points.

3. The method for creating a three-dimensional feature descriptor based on local surface change information according to claim 2, characterized in that: In step 1.3, the specific steps for evenly dividing the local space into several subspaces along the radial direction are: Step 1.3.1: Given a point cloud or surface area, support radius , extract the point set in the spherical neighborhood of the key point p, defined as , where k is the number of neighborhood points; Step 1.3.2: Generate the local LRA at the key point p using the method in step 1.2; Step 1.3.3: After generating the local space of p, divide the local space evenly along the radial direction into subspace; Step 1.3.4: For each point in the subspace, encode five geometric attributes ,in 、 、 and There are four angle attributes, with a value range of [0, 180°], is the curvature property, and its value range is [0, 1].

4. The method for creating a three-dimensional feature descriptor based on local surface change information according to claim 3, characterized in that: In step 1.3.4, for each neighborhood point in the subspace , four angle attributes It is expressed as follows: In the formula Indicates p to vector, express The normal vector at represents the LRA direction at p, " " represents the cross product between two vectors, " represents the dot product between two vectors, is an integer and ; For a neighborhood point , its curvature properties The approximate solution is as follows: In the formula 、 and express The eigenvalues ​​of the three principal axis directions of the surface, and .

5. The method for creating a three-dimensional feature descriptor based on local surface change information according to claim 4, characterized in that: To improve Characterize the effect of local surface geometric features, further execute the following formula, Scaled to a normal distribution with center at 0 and standard deviation at 1: In the formula yes The mean of the curvature values ​​of the surface, yes The standard deviation of the surface curvature values.

6. The method for creating a three-dimensional feature descriptor based on local surface change information according to claim 5, characterized in that: In step 1.5, the method for generating the statistical histograms corresponding to the five geometric attributes is: The statistical histograms of the five geometric attributes are recorded as , the number of blocks used to count the five geometric attributes are recorded as ,but The lengths are .

7. The method for creating a three-dimensional feature descriptor based on local surface change information according to claim 6, characterized in that: In step 1.6, the five subhistograms are weightedly connected to form the final LSVSH histogram represented as: ,in are the weights corresponding to the five geometric attributes.

8. A system for creating three-dimensional feature descriptors based on local surface change information, characterized by: The system includes a memory, a processor, and a creation program for a three-dimensional feature descriptor based on local surface change information stored on the processor. When the creation program for a three-dimensional feature descriptor based on local surface change information is run by the processor, all steps of the creation method according to any one of claims 1 to 7 are executed.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a program for creating a three-dimensional feature descriptor based on local surface change information, and the program for creating a three-dimensional feature descriptor based on local surface change information executes all steps of the creation method according to any one of claims 1 to 7 when the program is run by the processor.

Images