A three-dimensional local surface feature description method based on deviation angle statistics

Abstract

The application is a three-dimensional local surface feature description method based on deviation angle statistics, which overcomes the problem of poor distinction of similar local curved surfaces in the prior art. The method comprises the following steps: 1: obtaining a three-dimensional point cloud of a scene and extracting feature points of the point cloud data of the scene; 2: calculating a local reference axis of a neighborhood within a spherical radius range of the feature points and transforming all points in the neighborhood to the local reference axis; 3: encoding the spatial information of the spherical neighborhood, dividing the spherical neighborhood into a plurality of sub-space neighborhoods; 4: calculating the normal of each neighborhood point in the spherical neighborhood; 5: calculating the deviation angle between the normal of the neighborhood point and the local reference axis, counting the deviation angles in the sub-space and forming a sub-histogram; 6: finding the nearest three feature points of the feature point, calculating the deviation angles of the local reference axes between the feature point and the nearest three feature points, and forming a sub-histogram; and 7: connecting all the sub-histograms and normalizing to form a final histogram.

Description

Technical fields:

[0001] This invention belongs to the field of 3D computer vision technology and relates to a 3D local surface feature description method based on deviation angle statistics, that is, a method for characterizing the local surface features of a 3D object by using the deviation angle statistics between key points and neighboring points, and between key points and adjacent key points. Background technology:

[0002] With the rise of low-cost sensors and high-speed computing systems such as Microsoft Kinect, 3D data is readily available, and 3D vision has wide applications in robotics, reverse engineering, autonomous driving, and biometric systems. Local feature descriptors play a crucial role in these applications, and extracting high-performance local feature descriptors from point clouds remains a challenging task and a hot research area in 3D technology.

[0003] Currently, to improve robustness against noise interference, point cloud resolution variations, and occlusion, many 3D local feature descriptors have been proposed in existing research. Local reference frames and feature encoding are the two main components determining their performance. Local frames can be divided into local reference frames (LRFs) and local reference axes (LRAs). The repeatability of the x / y axes of an LRF is more susceptible to various disturbances (e.g., noise, varying mesh resolution, and symmetric surfaces) than that of the z-axis, and constructing an LRF requires more time than constructing an LRA. Therefore, local descriptors constructed on LRAs are more robust to various disturbances. However, in practical computation, LRAs constructed at different neighborhood scales yield different performance results, making it impossible to simultaneously acquire robustness against multiple disturbances. Feature encoding typically only encodes the geometric and spatial information of the local surfaces of keypoints. However, many similar or symmetric local surfaces exist on the same object; encoding only the information of the local surfaces of keypoints cannot distinguish these similar or symmetric local surfaces. Summary of the Invention:

[0004] The purpose of this invention is to provide a three-dimensional local surface feature description method based on deviation angle statistics, which overcomes the problems of existing technologies that cannot simultaneously obtain robustness to multiple disturbances and cannot distinguish similar or symmetrical surfaces of the same object in actual calculations. The local feature descriptor of this invention has stronger descriptiveness and robustness.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] 1. A method for describing three-dimensional local surface features based on deviation angle statistics, characterized by the following steps:

[0007] Step 1: Obtain the 3D point cloud of the scene and extract the feature points P from the point cloud data of the scene;

[0008] Step 2: Calculate the local reference axis of the neighborhood within the spherical radius R of the feature point p. The origin of the local reference axis coincides with the origin of the feature point. Transform all points in the neighborhood to the local reference axis to achieve alignment with the local reference axis and attitude normalization, thereby obtaining rigid transformation invariance to the feature description.

[0009] Step 3: Encode the spatial information of the spherical neighborhood and divide the spherical neighborhood into multiple sub-spatial neighborhoods;

[0010] Step 4: Calculate the normal to each neighboring point within the spherical neighborhood;

[0011] Step 5: Calculate the deviation angle between the normal of the neighboring point and the local reference axis, count the deviation angles in the subspace, and form a sub-histogram;

[0012] Step 6: Find the three nearest feature points of the feature point p, calculate the deviation angle of the local reference axis between the feature point p and the three nearest feature points, and form a sub-histogram;

[0013] Step 7: Connect all the sub-histograms and normalize them to form the final histogram, which is the local feature descriptor.

[0014] Step 2, establishing a local reference axis for the feature point p includes:

[0015] Determine the radius R of the local reference axis for calculation. z Based on the calculated radius R z Calculate the covariance matrix of the neighborhood points and generate the local reference axis;

[0016] To eliminate the ambiguity of the local reference axes, the computational radius R of multiple local reference axes is adopted. s Based on the calculated radius R s Determine the direction of the local reference axis.

[0017] In step 3, the spherical neighborhood is divided based on the distance between the local reference axis and the projected radial distance. The division ranges of the local reference axis and the projected radial distance are [0, 2R] and [0, R], respectively, where R is the support radius.

[0018] In step 4, the normal direction of the neighboring points is determined by the local correspondence method, and the Poisson disk sampling method is used to reduce computational redundancy.

[0019] In step 6, set the search radius R for finding the nearest feature point. f In the search radius R fWithin a spherical neighborhood, retrieve the nearest feature point and gradually expand R. f The range is calculated until the three nearest feature points are found.

[0020] Compared with the prior art, the advantages and effects of the present invention are as follows:

[0021] (1) Local reference axes of different scales correspond to different performances, and the computational domain R s A local reference axis smaller than the support radius R can improve the robustness of the local reference axis to clutter and occlusion, but performs poorly in low-resolution data. Using points in the neighborhood of all radii to determine the direction and sign of the LRA can achieve high robustness to noise and varying point cloud resolution, but has poor robustness to grid boundaries. Therefore, the multi-scale local reference axis in this invention can cope with scenarios under different interference conditions.

[0022] (2) In this invention, the spherical neighborhood is divided based on the local reference axis and the radial distance of the projection, and the geometric information in each subspace is encoded to obtain a high descriptiveness of the local surface;

[0023] (3) The calculation of the normal is extremely time-consuming. Points adjacent to each other on the local surface have similar characteristics. Therefore, in this invention, the normal direction is redirected by local correspondence and sampling is performed using the Poisson disk method, which greatly reduces computational redundancy.

[0024] (4) Traditional local feature descriptors have the defect of poor discrimination of similar curved surfaces. In this invention, the geometric information of feature points and neighboring feature points is combined to improve the ability to distinguish similar surfaces. Moreover, this technology can be combined with any similar three-dimensional local feature description technology to improve the ability to distinguish similar surfaces. Attached image description:

[0025] Figure 1 This is a flowchart illustrating the three-dimensional local surface feature description method based on deviation angle statistics according to an embodiment of this application.

[0026] Figure 2 This is a schematic diagram of the process of establishing a local reference axis in the spherical neighborhood of the feature point p according to an embodiment of this application.

[0027] Figure 3 This is a schematic diagram of spatial information encoded based on a local reference axis according to an embodiment of the present application.

[0028] Figure 4 This is a schematic diagram illustrating the calculation of the deviation angle between the local reference axis and the normal according to an embodiment of this application.

[0029] Figure 5 This is a schematic diagram of the statistical feature points and nearest neighbor feature geometric information according to an embodiment of this application.

[0030] Figure 6 This is a schematic flowchart of feature matching according to an embodiment of the present application. Detailed implementation method:

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] This invention discloses a 3D local surface feature description method based on deviation angle statistics. The method utilizes key points from point cloud data, selects points on the local surface within a certain range of the key points, establishes a local reference axis for the local surface of the key point, divides the local surface of the key point into multiple subspaces based on the local reference axis, calculates the deviation angle between the local reference axis and the normals of all points within the subspace, and combines the geometric information of the key point and its neighboring key points to finally form a 3D local feature descriptor for the local surface of the key point. In this field, after describing a 3D object using the local surface feature description method of this application, subsequent 3D feature matching or 3D target recognition can be performed.

[0033] Example:

[0034] The present application will now be described in detail with reference to specific embodiments and accompanying drawings.

[0035] like Figure 1 As shown, this application proposes a three-dimensional local surface feature description method based on deviation angle statistics, the method specifically including:

[0036] Obtain the 3D point cloud of the scene;

[0037] Extract the feature point set P = {p1, p2, p3, ..., p} from the scene point cloud. n};

[0038] In this embodiment, the scene point cloud is the surface of the scene represented by a set of three-dimensional coordinate points after the point cloud acquisition device describes the scene. The three-dimensional point cloud can be acquired in real time or stored in advance.

[0039] In this embodiment, feature points are also referred to as key points, which represent important shape information of the original point cloud. The method for obtaining feature points is not limited in this application.

[0040] Establish a local reference axis for the local surface of the feature point, and combine it with Figure 2Given a feature point p and a support radius R in the feature point set, all points within the radius range of the sphere are defined as neighborhood points of point p. These neighborhood points form a local surface Q = {q1, q2, q3, ..., q...} n};

[0041] The subset of the local surface Q is defined as The direction of the z-axis is used to calculate the local surface subset Q. z The covariance matrix Cov(Q) z ) is defined as:

[0042]

[0043] Where k is the local surface subset Q z Size, It is the local surface subset Q z The centroid will correspond to the covariance matrix Cov(Q) z The eigenvector v(p) of the smallest eigenvalue of ) is set as the z-axis.

[0044] The direction of the eigenvector is random. To eliminate the ambiguity of the eigenvector v(p), the computational domain R is used. s To generate a local reference axis.

[0045]

[0046] Where n is the computational domain R s of points, pq i For the key point p to the neighboring point q i The vectors are denoted by "·", and the dot product is represented between the vectors.

[0047] The computational domain R s Defined as: R s =λR.

[0048] λ is a scaling factor that adjusts the size of the computational domain, and R is the support radius. The computational domain R used to estimate the z-axis is... s Using a radius smaller than the support radius R can improve the robustness of the LRF to clutter and occlusion, but it performs poorly in low-resolution data. On the other hand, using points within the neighborhood of all radii to determine the direction and sign of the LRA achieves high robustness to noise and varying point cloud resolution, but poor robustness to grid boundaries.

[0049] The number of neighborhood points involved in calculating the z-axis varies even within the same scale neighborhood subset of different data. For example, when LiDAR collects data from distant scenes, the point cloud acquired is relatively sparse. That is, within the same radius, the number of points in the scene is much smaller than the number of points in the model. This is essentially related to the average grid resolution of the point cloud.

[0050] Obtain the average point cloud resolution s.pr of the scene and the average point cloud resolution m.pr of the model.

[0051] A scaling factor λ is introduced to adjust the computational domain R. s To improve robustness to resolution changes:

[0052]

[0053] Where c is a constant, based on experience with previous methods, it is considered that when the average point cloud resolution of the scene and the model are roughly the same, the scaling factor λ is taken as 1 / 3, that is, the constant c = 3 is the preferred implementation. The support radius R is the computational domain R. s Upper limit of threshold.

[0054] The local reference axis is aligned with the feature point p and aligned with the z-axis of the local surface Q.

[0055] Based on the local reference axis, the spatial information of the local space Q is encoded, combined with... Figure 3 The local space is encoded by using projected radial distance and height distance, and the local space is divided into multiple subspaces.

[0056] In a preferred embodiment of this invention, the number of divisions along the LRA axis and the projected radial direction are 15 and 5, respectively.

[0057] The geometric information of the local space Q is encoded, which consists of the normal information of all points within the local space. A local correspondence method is used to reorient the normals of each surface, ensuring that the point's direction aligns with the direction of most normals within its radius neighborhood. For each point p... i and its original normal n i and k nearest neighbors {p i1 ,p i2 ,p i3 ,…,p ik},calculate That is, p i The centroid of the neighborhood is shown below:

[0058]

[0059] Eliminate the ambiguity of the normal and determine the direction of the normal.

[0060]

[0061] The normal is sampled using the Poisson disk sampling algorithm to retain necessary descriptive information and eliminate redundancy.

[0062] Furthermore, such as Figure 4 As shown, in each subspace, the deviation angle between the local reference axis and the point normal is calculated. The deviation angle between the normal and the LRA is calculated as follows:

[0063] nd i (q k =arccos(LRA(p i ),n i (q k ))

[0064] Where LRA(p) represents the local reference axis at the feature point p; n i Represents the local point q k The normal at the location; nd i Representing LRA(p) and n i The deviation angle between them, nd i The range is [0, π]. Geometric information is encoded into a partition, generating a sub-histogram h of the statistical offset angles within that partition. i .

[0065] Based on the generation of sub-histograms in the aforementioned subspace, combined with Figure 5 Based on the feature point p, find the three nearest feature points P of the feature point p. k1 P k2 and P k3 The distance between a feature point and other feature points is calculated using the following formula:

[0066]

[0067] To improve search efficiency, set the search radius R. f Within a spherical region of the search radius, the three nearest feature points are retrieved, and the search radius R is gradually expanded. f The range is calculated until the three nearest feature points P are found. k1 P k2 and P k3 .

[0068] Step 3.7: After obtaining the three nearest feature points, calculate the deviation angle between the local reference axes at the feature points to represent the geometric relationship between the feature points.

[0069] ld(P i ,P k )=cos -1 (P i ,P k )

[0070] Where ld(P) i ,P kThe deviation angle between the local reference axes of the feature points is represented by ), where the deviation angle ranges from [0, π]. A sub-histogram H is generated based on this deviation angle. kp The sub-histogram h of each subspace i With the sub-histogram H kp The data is then normalized to fully describe the features of the local surface, forming the final three-dimensional local feature descriptor.

[0071] Based on the above embodiments, such as Figure 6 As shown, in real-time matching, the average point cloud resolution of the scene is unknown. Multiple scaling factors are predetermined, and the local reference axis and three-dimensional local feature descriptor of the model determined by the scaling factors are calculated and stored. After determining the average resolution of the scene, the actual scaling factor is calculated, and the three-dimensional local feature descriptor obtained closest to the actual scaling factor is found.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. All equivalent structural changes made based on the description and drawings of the present invention should be included within the scope of patent protection of the invention.

Claims

1. A three-dimensional local surface feature description method based on deviation angle statistics, characterized in that: The method comprises the following steps: Step 1: obtaining a three-dimensional point cloud of a scene, and extracting a feature point P of point cloud data of the scene; Step 2: calculating a local reference axis of a neighborhood within a spherical radius R of the feature point p, the origin of the local reference axis coinciding with the origin of the feature point, transforming all points of the neighborhood to the local reference axis, realizing alignment with the local reference axis and attitude normalization, and obtaining rigidity transformation invariance of a feature description; Step 3: encoding spatial information of a spherical neighborhood, and dividing the spherical neighborhood into a plurality of subspace neighborhoods; Step 4: calculating a normal of each neighborhood point in the spherical neighborhood; Step 5: calculating a deviation angle of the normal of the neighborhood point and the local reference axis, counting the deviation angle in the subspace, and forming a sub-histogram 1; Step 6: finding three nearest feature points of the feature point p, calculating a deviation angle of the local reference axis between the feature point p and the three nearest feature points, and forming a sub-histogram 2; Step 7: connecting all sub-histograms and normalizing to form a final histogram, that is, a local feature descriptor.

2. The method of claim 1, wherein the method is based on a deviation angle statistic. In step 2, the local reference axis is established for the feature point p, and the local reference axis comprises: determining a radius R of a local reference axis z , based on the neighborhood points of the computed radius R z , computing a covariance matrix, generating the local reference axis; eliminating the ambiguity of the local reference axis, calculating a radius R of a plurality of the local reference axes s determining the direction of the local reference axis based on the calculated radius R s ​ 3. The method of claim 1, wherein the method is based on a deviation angle statistic. In step 3, the spherical neighborhood is divided based on the local reference axis and a projection radial distance, the division ranges of the local reference axis and the projection radial distance being [0, 2R] and [0, R] respectively, and R being a support radius.

4. The method of claim 1, wherein the method is based on a deviation angle statistic. In step 4, a local corresponding method is used to determine the normal direction of the neighborhood point, and a Poisson disc sampling method is used to reduce calculation redundancy.

5. The method of claim 1, wherein the method is based on a deviation angle statistic. In step 6, set the search radius R for finding the nearest feature point. f In the search radius R f Within a spherical neighborhood, retrieve the nearest feature point and gradually expand R. f The range is calculated until the three nearest feature points are found.

Images