A soybean three-dimensional point cloud processing method based on a cubic reference

Abstract

The application discloses a soybean three-dimensional point cloud processing method based on a cubic reference object and belongs to the three-dimensional point cloud processing field. The method comprises the following steps: acquiring multi-view video data of a whole target scene through a video recording device and converting the multi-view video data into picture data sets; acquiring the pose of the video recording device and scene sparse point clouds based on the picture data sets, reconstructing scene dense point clouds; filtering the scene dense point clouds to obtain reference object point clouds; restoring scale information by using the reference object point clouds; and correcting the point cloud coordinate system by using the reference object point clouds after the scale is restored. The application optimizes the multi-view three-dimensional reconstruction process, and the point clouds after the scale is restored can be used for phenotype analysis, which is beneficial to the development of the soybean industry.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional point cloud processing, and more specifically, to a method for processing three-dimensional point clouds of soybeans based on a cube reference. Background Technology

[0002] In recent years, with the continuous development of the field of 3D reconstruction, its application in agriculture has been extensively explored, such as obtaining crop phenotypic data in the field and 3D reconstruction of crops at various stages. The application of 3D reconstruction in agriculture can provide breeding researchers with rich phenotypic information and track crop growth, thereby enabling timely understanding of the environmental impact on crops. As an important part of my country's agriculture, the soybean industry has an increasing demand for 3D reconstruction of soybeans. Currently, common methods for 3D reconstruction of crops include LiDAR-based 3D reconstruction, Kinect-based 3D reconstruction, and multi-view-based 3D reconstruction. LiDAR-based methods have the disadvantages of high cost and long processing time, but offer the highest reconstruction accuracy. Kinect-based 3D reconstruction is inexpensive and can simultaneously acquire depth and color maps, but it has relatively high noise levels. Multi-view-based 3D reconstruction methods are low-cost, have good reconstruction results, and are less affected by environmental limitations, but the reconstructed point cloud lacks scale information and cannot correct the coordinate system. Summary of the Invention

[0003] To address the issues of multi-view-based 3D reconstruction lacking scale information and unable to correct coordinate systems, this invention proposes a soybean 3D point cloud processing method based on a cubic reference object. This method involves capturing soybean videos and converting them into image datasets. Based on the image datasets, the pose and sparse point cloud of the recording device are obtained, and a dense point cloud of the overall scene is further reconstructed. The overall point cloud is then filtered to obtain a reference object point cloud, which is then used for scale recovery and coordinate system correction.

[0004] The specific technical solution is as follows:

[0005] A method for processing soybean 3D point clouds based on a cube reference includes the following steps:

[0006] The video data of the entire target scene from multiple perspectives is acquired using recording equipment and converted into an image dataset.

[0007] Based on the image dataset, the pose of the recording device and the sparse point cloud of the scene are obtained, and the dense point cloud of the scene is reconstructed.

[0008] Filter dense point clouds in the scene to obtain reference point clouds;

[0009] Recovering scale information using reference point clouds;

[0010] The point cloud coordinate system is corrected using the restored scaled reference point cloud.

[0011] Furthermore, the process of acquiring multi-view video data of the entire target scene through a recording device and converting it into an image dataset includes:

[0012] A cube reference object is placed within the scene. The color of the cube reference object is different from that of the scene, and the color of the top face of the cube reference object is different from the colors of the other faces.

[0013] Video data containing a cube reference object and soybean plants from multiple perspectives was acquired using a recording device and converted into an image dataset; a color threshold was set to generate a masked image dataset.

[0014] Furthermore, the step of acquiring the pose of the recording device and the sparse point cloud of the scene based on the aforementioned image dataset, and reconstructing the dense point cloud of the scene, includes:

[0015] Based on the image dataset and mask image dataset, obtain the pose of the recording device and the sparse point cloud of the scene;

[0016] The scene is reconstructed using the pose of the video recording device and the sparse point cloud of the scene, generating a dense point cloud of the entire scene.

[0017] Furthermore, the filtered scene dense point cloud to obtain the reference point cloud includes:

[0018] Set a color threshold to filter out the cube reference point cloud from the dense point cloud of the scene, which are the top face point cloud and other face point clouds of the cube reference.

[0019] A filtering algorithm is used to filter discrete points in the top point cloud and other surface point clouds of a cube reference object.

[0020] Clustering algorithms are used to further filter the filtered top surface point cloud of the reference object to obtain a refined top surface point cloud of the reference object.

[0021] Furthermore, the method of recovering scale information using reference point clouds includes:

[0022] Based on the detailed top surface point cloud, calculate its diagonal length, and obtain its side length using the geometric relationship of the reference object;

[0023] Calculate the ratio of the point cloud edge length to the true edge length to recover the point cloud scale information. The calculation formula is as follows:

[0024]

[0025]

[0026] Where: scale represents the scale ratio, Len pointLen represents the side length of the calculated top surface point cloud. truth V represents the actual side length. i V represents the coordinates of the i-th 3D point within the scene point cloud. i ′ represents the coordinates of the i-th 3D point in the scene point cloud after the point cloud scale information is restored, and n is the number of 3D points in the scene.

[0027] Furthermore, the calculation of the diagonal length of the top surface point cloud, using the geometric relationships of the reference object to obtain the side length of the top surface point cloud, includes:

[0028] Calculate the convex hull points of the top surface point cloud;

[0029] Calculate the distance between the convex hull points, and select the pair of convex hull points with the farthest distance as the diagonal;

[0030] Using the fact that the diagonal of a square is the same as the side length of the square Using the geometric relationship of multiples, the side length of the top surface point cloud is obtained, and the calculation formula is as follows:

[0031] hullpoints = hull(plane) b )

[0032] point dist ={||hullpoints i -hullpoints j ||2},i=1,…,N; j=1,…,N

[0033]

[0034] Where hull(·) represents the Hull algorithm, used to find the convex hull points of the top plane; hullpoints represents the set of convex hull points of the top point cloud; N represents the total number of convex hull points in the convex hull point set. i hullpoints j Let point represent the i-th and j-th convex hull points of the top point cloud. dist The set of Euclidean distances for each pair of convex hull points within the top point cloud; percentile(point dist ,99) indicates taking the point dist The 99th percentile value is used as the diagonal.

[0035] Furthermore, the step of correcting the point cloud coordinate system using the reference point cloud after scaling includes:

[0036] For the point cloud of other faces of the reference object, the multi-plane RANSAC algorithm is used to calculate the equations of all planes in the point cloud of other faces of the reference object and the three-dimensional points contained in each plane, forming a set of equations of other faces. The equation of a single plane in the point cloud of other faces can be expressed as ax + by + cz + d = 0.

[0037] For a detailed top surface point cloud of a reference object, the planar equation and the contained three-dimensional points of the top surface point cloud of the reference object are calculated using the single-plane RANSAC algorithm.

[0038] Calculate the dot product of the normal vectors of the other face equations and the top face equation to obtain three approximately mutually perpendicular planes;

[0039] The rotation matrix is ​​obtained by using the normal vectors of three approximately mutually perpendicular planes calculated using Schmidt orthogonalization.

[0040] Calculate the translation matrix using the center of the top point cloud;

[0041] Based on the rotation and translation matrices, the original scene coordinate system is rotated and translated to a coordinate system with the center of the top point cloud as the origin, resulting in the scene point cloud after the coordinate system is corrected.

[0042] Furthermore, the single-plane RANSAC algorithm and the multi-plane RANSAC algorithm include:

[0043] Randomly select n points from the restored reference point cloud and calculate the plane formed by these points and its plane equation.

[0044] Calculate the distance from all points to the plane, and classify the points into inner and outer points based on a distance threshold;

[0045] Store n randomly selected points and all found interior points, which form a subset of the set of planes of the plane to be found;

[0046] Repeat the above process until the maximum number of repetitions is set;

[0047] Save the subset of planes containing the most interior points in the set of planes as a found plane; this is the single-plane RANSAC algorithm.

[0048] Repeat the process of calculating the single-plane RANSAC result from the remaining outliers and removing the resulting plane point set from the outliers until the number of outliers is less than a set threshold. This is the multi-plane RANSAC algorithm.

[0049] Furthermore, the calculation of the normal vector dot product of the other face equation set and the top face equation includes:

[0050] Calculate the dot product of the normal vector of the top surface point cloud and the normal vectors of the equation set of other surfaces, and find the two normal vectors with the smallest dot product. The calculation formula is as follows:

[0051]

[0052] r1 = min|Normals·Normal|

[0053] r2=min|Normals·Normal+Normals·r1|

[0054] Where: Normals is the set of normal vectors of other surface equation sets, Normal is the normal vector of the top surface point cloud, r1 is the normal vector with the smallest absolute value of the dot product of the normal vector sets of other surfaces and the top surface normal vector, and r2 is the normal vector with the smallest absolute value of the sum of the dot products of r1 and the top surface normal vector with the normal vector sets of other surfaces respectively; a b b b c b The coefficients of the equation of the top plane; a m b m c m This represents the coefficients of the plane equation for the m-th other face.

[0055] Furthermore, the rotation matrix is ​​obtained by taking the normal vectors of the three approximately mutually perpendicular planes obtained through the Schmidt orthogonalization, including:

[0056] Let the orthogonal basis z = Normal;

[0057] The difference between r1 and its projection onto the orthogonal basis z can be used to obtain the orthogonal basis. Where <.,.> represents the inner product;

[0058] Then r2 and its projections onto the orthogonal basis z and x are: Then, subtracting the sum of the two projections from r2 yields the orthogonal basis.

[0059] Finally, unitization is performed.

[0060] Obtain the rotation matrix in the point cloud world The superscript T indicates transpose.

[0061] Furthermore, the calculation of the translation matrix using the center of the top surface point cloud includes:

[0062]

[0063] T = -R·center

[0064] in: Let be the i-th point in the top plane, N represent the number of three-dimensional points in the top plane, center is the center point of the top point cloud, and T is the translation matrix.

[0065] The beneficial effects of this invention are: This invention proposes a method for processing soybean 3D point clouds based on a cubic reference, which can achieve multi-view processing... Figure 3 The point cloud scale information is recovered during 3D reconstruction, and the point cloud coordinate system is corrected. The resulting soybean point cloud has good quality and high precision, and can be used for phenotypic analysis. Attached Figure Description

[0066] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or approximate component shown in the various figures can be represented by the same symbol. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings, wherein:

[0067] Figure 1 This is a flowchart of the soybean 3D point cloud processing method based on a cube reference object according to the present invention;

[0068] Figure 2 This is a schematic diagram of the mask image of the present invention;

[0069] Figure 3 The sparse point cloud generated by COLMAP in this invention;

[0070] Figure 4 This refers to the blue planar point cloud of the present invention;

[0071] Figure 5 This is a flowchart of the recovery scale of the present invention;

[0072] Figure 6 This is a flowchart of the correction coordinate system of the present invention;

[0073] Figure 7 This is the result of measuring the side length of a cube using meshlab in this invention. Detailed Implementation

[0074] To make the above-mentioned objectives, features and advantages of the present invention more apparent and understandable, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] like Figure 1 As shown, the overall process of the soybean 3D point cloud processing method based on a cube reference object of the present invention includes the following steps:

[0076] Step 1: Obtain a soybean multi-view dataset;

[0077] Place the cube reference object in the soybean plant scene, use a mobile phone as a recording device, and use the mobile phone to shoot the entire target scene from a distance and from top to bottom to capture a multi-angle video.

[0078] After converting the video into an image dataset using ffmpeg, the RGB values ​​are converted to HSV, a mask is set, and a masked image dataset is generated. Figure 2 This is a schematic diagram of a masked image. The background was removed by masking, while the soybean plant and the cube reference were preserved.

[0079] Step 2: Obtain the pose and overall sparse point cloud;

[0080] Input image dataset and mask image set; use COLMAP software to call the feature extraction interface, using OpenCV as the camera model; use COLMAP software to call the feature matching interface to find the correspondence between feature points in different images; use COLMAP software to call the incremental SFM interface to obtain the phone pose and sparse scene point cloud, such as... Figure 3 As shown.

[0081] Step 3: Reconstruct the dense point cloud of the scene;

[0082] Based on the pose and sparse point cloud of the recording device, the COLMAP software is used to call the dense point cloud interface to generate a scene dense point cloud, which includes soybean plants and cubes.

[0083] Step 4: Filter the point cloud to obtain the reference point cloud;

[0084] S4.1 converts the dense point cloud of the scene from RGB to HSV. By setting the HSV threshold, it segments the point cloud with a coarse red color (other face point cloud) and the point cloud with a blue color (top face point cloud). Statistical filtering is used to perform statistical analysis on the neighborhood of each point to remove points that do not meet the standards.

[0085] S4.2 For the blue plane, the DBSCAN point cloud clustering method is used to select the point cloud with the highest density to obtain the final refined blue plane, i.e., the refined top surface point cloud, as shown below. Figure 4 As shown.

[0086] Step 5: Restore the point cloud scale, the process is as follows Figure 5 As shown;

[0087] S5.1 uses the Qhull algorithm to obtain all convex hull points of the blue plane:

[0088] hullpoints = Qhull(plane) b )

[0089] S5.2 Calculate the Euclidean distance between each pair of convex hull points:

[0090] point dist ={||hullpoints i -hullpoints j ||2},i=1,…,N; j=1,…,N

[0091] Where n is the number of 3D points in the cube point cloud.

[0092] S5.3 Take the 99th percentile value as the farthest distance between two points:

[0093] point sorted =sort(point) dist )

[0094] idx = n × 99%

[0095]

[0096] S5.4 divided by Get the side length Len of the blue plane cube :

[0097]

[0098] S5.5 calculates the ratio of the point cloud edge length to the true edge length, and then recovers the point cloud scale:

[0099]

[0100]

[0101] Where the actual side length is 1 decimeter, V i A 3D point within the scene's point cloud.

[0102] Step 6: Correct the point cloud coordinate system, the process is as follows: Figure 6 As shown;

[0103] S6.1 The equation of the blue plane is obtained using the single-plane RANSAC algorithm. The specific process is as follows:

[0104] Randomly select three points within a finely detailed blue point cloud and calculate the equation of the plane formed by these points;

[0105] Calculate the distance from all points to the plane based on the plane; points with a distance less than 0.02 are considered interior points.

[0106] Store the plane equation and all selected points obtained in this iteration;

[0107] Repeat the above process until it has been repeated 2000 times;

[0108] The plane containing the most interior points is selected as the blue plane, and its equation is obtained. The plane equation can be expressed as:

[0109] ax + by + cz + d = 0

[0110] S6.2 The plane equations of other surfaces of the reference object are calculated using the multi-plane RANSAC algorithm. The specific process is as follows:

[0111] The initial plane is obtained using the single-plane RANSAC algorithm;

[0112] Remove the points contained in the initial plane from the reference point cloud, and then use the single-plane RANSAC algorithm to obtain the next plane;

[0113] The loop stops when the number of reference points in the point cloud is less than 500, and the resulting planar points and plane equations are recorded. Theoretically, five planes with reference points can be obtained, but in practice, stopping the loop when the number of reference points in the point cloud is less than 500 will result in at least five planes.

[0114] S6.3 Calculate the dot product of the normal vector of the blue plane and the normal vectors of the other planes, and find the two normal vectors with the smallest dot product. The calculation formula is as follows:

[0115]

[0116] r1 = min|Normals r ·Normal b |

[0117] r2=min|Normals r ·Normal b +Normals r ·r1|

[0118] Among them: Normals r For all the normal vectors of the red plane, similarly for Normal... b Let r1 be the normal vector matrix of the blue plane, r2 be the normal vector whose dot product with the normal vector of the blue plane is the smallest, and r3 be the normal vector whose sum of the dot products of r1 and the normal vector of the blue plane with the normal vector of the red plane is the smallest. b b b c b The coefficients of the equation for the blue plane; a m b m c m Let represent the coefficients of the equation for the m-th red plane.

[0119] S6.4 The rotation matrix R is obtained using Schmidt orthogonalization, and the specific process is as follows:

[0120] Let the orthogonal basis z = Normal;

[0121] The difference between r1 and its projection onto the orthogonal basis z can be used to obtain the orthogonal basis. Where <.,.> represents the inner product;

[0122] Then r2 and its projections onto the orthogonal basis z and x are: Then, subtracting the sum of the two projections from r2 yields the orthogonal basis.

[0123] Finally, unitization is performed.

[0124] Obtain the rotation matrix in the point cloud world The superscript T indicates transpose.

[0125] S6.5 Calculate the mean of the blue plane, which is the center point: in Let N be the i-th point in the blue plane, and let N represent the number of three-dimensional points in the blue plane. Calculate the translation matrix T = -R·center using the center of the blue plane.

[0126] S6.6 Correcting the coordinate system using rotation and translation matrices:

[0127]

[0128] Where n is the number of 3D points in the scene.

[0129] Example 1

[0130] Multi-view video data of soybeans was captured, converted into an image dataset using ffmpeg, and the images were downsampled and color thresholded to generate a masked image dataset. COLMAP software was used to generate the phone's pose and sparse point cloud. COLMAP software was also used to generate a dense point cloud. Cube point cloud data was used to recover the point cloud scale information and correct the coordinate system. The scale information of the rescaled point cloud was then measured using MeshLab. Figure 7 As shown, the side length of the cube is 9.7 centimeters, with an error of only 3 millimeters.

[0131] It should be noted that, for the sake of simplicity, the aforementioned method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously.

[0132] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0133] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for processing three-dimensional soybean point clouds based on a cubic reference, characterized in that: Includes the following steps: The video data of the entire target scene from multiple perspectives is acquired using recording equipment and converted into an image dataset. Based on the image dataset, the pose of the recording device and the sparse point cloud of the scene are obtained, and the dense point cloud of the scene is reconstructed. Filter dense point clouds in the scene to obtain reference point clouds; Recovering scale information using reference point clouds; The point cloud coordinate system is corrected using the restored scaled reference point cloud.

2. The soybean 3D point cloud processing method based on a cube reference as described in claim 1, characterized in that: The process of acquiring multi-view video data of the entire target scene through a recording device and converting it into an image dataset includes: A cube reference object is placed within the scene. The color of the cube reference object is different from that of the scene, and the color of the top face of the cube reference object is different from the colors of the other faces. Video data containing a cube reference object and soybean plants from multiple perspectives was acquired using a recording device and converted into an image dataset; a color threshold was set to generate a masked image dataset.

3. The soybean 3D point cloud processing method based on a cube reference as described in claim 1, characterized in that: The process of acquiring the pose of the recording device and the sparse point cloud of the scene based on the aforementioned image dataset, and reconstructing the dense point cloud of the scene, includes: Based on the image dataset and mask image dataset, obtain the pose of the recording device and the sparse point cloud of the scene; The scene is reconstructed using the pose of the video recording device and the sparse point cloud of the scene, generating a dense point cloud of the entire scene.

4. The soybean 3D point cloud processing method based on a cube reference as described in claim 1, characterized in that: The filtering of dense point clouds in the scene to obtain reference point clouds includes: Set a color threshold to filter out the cube reference point cloud from the dense point cloud of the scene, which are the top face point cloud and other face point clouds of the cube reference. A filtering algorithm is used to filter discrete points in the top point cloud and other surface point clouds of a cube reference object. Clustering algorithms are used to further filter the filtered top surface point cloud of the reference object to obtain a refined top surface point cloud of the reference object.

5. The soybean 3D point cloud processing method based on a cube reference as described in claim 4, characterized in that: The method of recovering scale information using reference point clouds includes: Based on the detailed top surface point cloud, calculate its diagonal length, and obtain its side length using the geometric relationship of the reference object; Calculate the ratio of the point cloud edge length to the true edge length to recover the point cloud scale information. The calculation formula is as follows: Where: scale represents the scale ratio, Len point Len represents the side length of the calculated top surface point cloud. truth V represents the actual side length. i V represents the coordinates of the i-th 3D point within the scene point cloud. i ′ represents the coordinates of the i-th 3D point in the scene point cloud after the point cloud scale information is restored, and n is the number of 3D points in the scene.

6. The soybean 3D point cloud processing method based on a cube reference as described in claim 5, characterized in that: The calculation of the diagonal length of the top surface point cloud, which utilizes the geometric relationships of a reference object to obtain the side length of the top surface point cloud, includes: Calculate the convex hull points of the top surface point cloud; Calculate the distance between the convex hull points, and select the pair of convex hull points with the farthest distance as the diagonal; Using the fact that the diagonal of a square is the same as the side length of the square Using the geometric relationship of multiples, the side length of the top surface point cloud is obtained, and the calculation formula is as follows: hullpoints=hull(plane b ) point dist ={||hullpoints i -hullpoints j ||2},i=1,...,N;j=1,...,N Where hull(·) represents the Hull algorithm, used to find the convex hull points of the top plane; hullpoints represents the set of convex hull points of the top point cloud; N represents the total number of convex hull points in the convex hull point set. i hullpoints j Let point represent the i-th and j-th convex hull points of the top point cloud. dist Percentile is the set of Euclidean distances between each pair of convex hull points within the top point cloud; dist ,99) indicates taking the point dist The 99th percentile value is used as the diagonal.

7. The soybean 3D point cloud processing method based on a cube reference as described in claim 4, characterized in that: The process of correcting the point cloud coordinate system using the restored reference point cloud includes: For the point cloud of other faces of the reference object, the multi-plane RANSAC algorithm is used to calculate the equations of all planes in the point cloud of other faces of the reference object and the three-dimensional points contained in each plane, forming a set of equations of other faces. The equation of a single plane in the point cloud of other faces can be expressed as ax + by + cz + d = 0. For a detailed top surface point cloud of a reference object, the planar equation and the contained three-dimensional points of the top surface point cloud of the reference object are calculated using the single-plane RANSAC algorithm. Calculate the dot product of the normal vectors of the other face equations and the top face equation to obtain three approximately mutually perpendicular planes; The rotation matrix is ​​obtained by using the normal vectors of three approximately mutually perpendicular planes calculated using Schmidt orthogonalization. Calculate the translation matrix using the center of the top point cloud; Based on the rotation and translation matrices, the original scene coordinate system is rotated and translated to a coordinate system with the center of the top point cloud as the origin, resulting in the scene point cloud after the coordinate system is corrected.

8. The soybean 3D point cloud processing method based on a cube reference as described in claim 7, characterized in that: The single-plane RANSAC algorithm and the multi-plane RANSAC algorithm include: Randomly select n points from the restored reference point cloud and calculate the plane formed by these points and its plane equation. Calculate the distance from all points to the plane, and classify the points into inner and outer points based on a distance threshold; Store n randomly selected points and all found interior points, which form a subset of the set of planes of the plane to be found; Repeat the above process until the maximum number of repetitions is set; Save the subset of planes containing the most interior points in the set of planes as a found plane; this is the single-plane RANSAC algorithm. Repeat the process of calculating the single-plane RANSAC result from the remaining outliers and removing the resulting plane point set from the outliers until the number of outliers is less than a set threshold. This is the multi-plane RANSAC algorithm.

9. The soybean 3D point cloud processing method based on a cube reference as described in claim 7, characterized in that: The calculation of the normal vector dot product of the equations of other surfaces and the equation of the top surface includes: Calculate the dot product of the normal vector of the top surface point cloud and the normal vectors of the equation set of other surfaces, and find the two normal vectors with the smallest dot product. The calculation formula is as follows: r1 = min|Normals·Normal| r2=min|Normals·Normal+Normals·r1| Where: Normals is the set of normal vectors of other surface equation sets, Normal is the normal vector of the top surface point cloud, r1 is the normal vector with the smallest absolute value of the dot product of the normal vector sets of other surfaces and the top surface normal vector, and r2 is the normal vector with the smallest absolute value of the sum of the dot products of r1 and the top surface normal vector with the normal vector sets of other surfaces respectively; a b b b c b The coefficients of the equation of the top plane; a m b m c m This represents the coefficients of the plane equation for the m-th other face.

10. The soybean 3D point cloud processing method based on a cubic reference as described in claim 9, characterized in that: The rotation matrix is ​​obtained by taking the normal vectors of the three approximately mutually perpendicular planes obtained through the Schmidt orthogonalization, including: Let the orthogonal basis z = Normal; The difference between r1 and its projection onto the orthogonal basis z can be used to obtain the orthogonal basis. Where <., .> denotes the inner product; Then r2 and its projections onto the orthogonal basis z and x are: Then, subtracting the sum of the two projections from r2 yields the orthogonal basis. Finally, unitization is performed. Obtain the rotation matrix in the point cloud world The superscript T indicates transpose.

Images