Part feature plane distance calibration method based on pca-based point cloud filtering

Abstract

The application discloses a part feature plane distance calibration method based on PCA (Principal Component Analysis) point cloud filtering, which is used for calculating the distance between two feature planes on a part, performing principal component analysis on selected range point clouds on the two feature planes to obtain feature normal vectors of the feature planes, performing rotation transformation on all point clouds so that the feature normal vectors are aligned with any coordinate axis, filtering the range point clouds based on axis-aligned bounding boxes to obtain feature point clouds, performing least square plane fitting based on the feature point clouds to obtain two fitting planes, and calculating the distance between the two feature planes according to the number of point clouds contained in the two fitting planes; the application can effectively reduce the complexity of the point cloud extraction process, and then fit both of the two feature planes, thereby improving the accuracy of the final feature plane distance calculation result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of point cloud feature fitting, and particularly relates to a part feature plane distance calibration method based on PCA point cloud filtering. BACKGROUND

[0002] In order to calibrate the distance between two feature planes in a calibration part, the prior art usually adopts an attention mechanism-based point cloud data feature extraction method. The method first uses a pre-trained encoder to sample and expand the dimensions of the point cloud data on the feature planes for preprocessing, so as to obtain a point cloud feature pyramid. Secondly, a pre-trained decoder is used to obtain another feature pyramid of the point cloud data, and the feature pyramid is up-sampled to obtain up-sampled data. Finally, a cross-attention mechanism is used to process the two point cloud feature pyramids and the up-sampled data, so as to reduce the semantic information loss, and point cloud features are extracted through a backbone network. The main problem of the above prior art is that it only focuses on solving the segmentation problem of point cloud features, but does not pay attention to the fitting analysis problem of point cloud features, resulting in a long and complex point cloud extraction and calculation process, and reducing the efficiency of point cloud fitting.

[0003] Therefore, in view of the above problems in the prior art, the application discloses a part feature plane distance calibration method based on PCA point cloud filtering. SUMMARY

[0004] The part feature plane distance calibration method based on PCA point cloud filtering can effectively reduce the complexity of the point cloud extraction process, and then fit both of the two feature planes, thereby improving the accuracy of the distance calculation result between the final feature planes.

[0005] The application is implemented through the following technical scheme:

[0006] The part feature plane distance calibration method based on PCA point cloud filtering is used to calculate the distance between two feature planes on a part. Principal component analysis is performed on the selected range point clouds on the two feature planes to obtain the feature normals of the feature planes. The range point clouds are rotated and transformed so that the feature normals are aligned with any coordinate axis. The range point clouds are filtered based on the axis-aligned bounding box to obtain feature point clouds. Least square plane fitting is performed based on the feature point clouds to obtain two fitted planes. The distance between the two feature planes is calculated according to the number of point clouds contained in the two fitted planes.

[0007] In order to better implement the application, the method further comprises the following steps:

[0008] Step 1, selecting range point clouds on two feature planes, performing principal component analysis on the range point clouds to obtain feature normals respectively representing the two feature planes;

[0009] Step 2: Perform rotation transformation on the original point cloud and the range point cloud on the feature plane so that the feature normal of the feature plane is parallel and aligned with any coordinate axis;

[0010] Step 3: Based on the maximum and minimum coordinate points in the range point cloud after rotation transformation, establish an axis-aligned bounding box, set a bounding threshold to adjust the size of the axis-aligned bounding box, and filter the range point cloud through the axis-aligned bounding box to obtain the feature point cloud.

[0011] Step 4: Perform inverse transformation on the feature point cloud to obtain the mapped point cloud of the feature point cloud in the original point cloud of the feature plane. Based on the mapped point cloud, perform least squares plane fitting to obtain the first fitting plane representing the first feature plane and the second fitting plane representing the second feature plane.

[0012] Step 5: Downsample the mapped point cloud based on the number of point clouds. Calculate the distance between the two feature planes based on the difference in the number of mapped point clouds after downsampling.

[0013] To better realize the present invention, step 5 further includes:

[0014] Step 5.1: Downsample the mapped point cloud representing the two feature planes to obtain the first downsampled point cloud representing the first feature plane and the second downsampled point cloud representing the second feature plane.

[0015] Step 5.2: Compare the number of points in the downsampled feature point cloud. If the number of points in the first downsampled point cloud is less than or equal to half the number of points in the second downsampled point cloud, proceed to step 5.3; if the number of points in the first downsampled point cloud is greater than half the number of points in the second downsampled point cloud, proceed to step 5.4.

[0016] Step 5.3: Calculate only the first distance between points in the first downsampled point cloud and the second fitting plane as the distance between the two feature planes;

[0017] Step 5.4: Calculate the first distance between points in the first downsampled point cloud and the second fitting plane, and calculate the second distance between points in the second downsampled point cloud and the first fitting plane. Use the average of the sum of the first distance and the second distance as the distance between the two feature planes.

[0018] To better implement the present invention, further, in step 5.1, the number of points in the mapped point cloud is extracted. If the number of points is greater than or equal to 1000, the mapped point cloud is downsampled; if the number of points is less than 1000, the mapped point cloud is not downsampled.

[0019] To better realize the present invention, further, in step 5.1, the farthest point sampling algorithm is used to downsample the mapped point cloud.

[0020] To better realize the present invention, step 3 further includes:

[0021] Step 3.1: Traverse the coordinates of all points in the point cloud after the rotation transformation, extract the point with the largest coordinates on all three axes as the maximum coordinate point, and extract the point with the smallest coordinates on all three axes as the minimum coordinate point.

[0022] Step 3.2: Create an axis-aligned bounding box using the line segment connecting the maximum and minimum coordinate points as the diagonal.

[0023] Step 3.3: Set the bounding threshold and adjust the size of the axis-aligned bounding box. Select the points in the original point cloud of the feature plane that are located inside the axis-aligned bounding box to form the feature point cloud.

[0024] To better realize the present invention, further, after the size of the axis-aligned bounding box is adjusted, it is required that the axis-aligned bounding box can simultaneously intersect with the original point clouds of the two feature planes.

[0025] To better realize the present invention, step 1 further includes:

[0026] Step 1.1: Calculate the covariance matrix of the current feature plane based on the points in the range point cloud;

[0027] Step 1.2: Calculate the eigenvalues ​​and eigenvectors of the current feature plane based on the covariance matrix;

[0028] Step 1.3: Extract the direction of the eigenvector with the smallest eigenvalue as the feature normal of the current feature plane.

[0029] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0030] (1) The present invention determines the range of point cloud participating in plane fitting by manually selecting points, which can make the position of the extracted point cloud match the part to be detected more accurately;

[0031] (2) By performing principal component analysis on the selected points and using axis-aligned bounding boxes, the filtering direction of the extracted point cloud is limited to the coordinate axes of the point cloud coordinate system, which greatly reduces the time complexity of the point cloud extraction and filtering algorithm and improves efficiency.

[0032] (3) The present invention performs plane fitting on the point clouds of the two feature planes, which improves the accuracy of plane distance calculation. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0034] Example 1:

[0035] This embodiment of the part feature plane distance calibration method based on PCA point cloud filtering is used to calculate the distance between two feature planes on a part. Principal component analysis is performed on the selected range point cloud on the two feature planes to obtain the feature normal of the feature plane. All point clouds are rotated so that the feature normal is aligned with any coordinate axis. The range point cloud is filtered based on the axis-aligned bounding box to obtain the feature point cloud. Least squares plane fitting is performed based on the feature point cloud to obtain two fitting planes. The distance between the two feature planes is calculated based on the number of point clouds contained in the two fitting planes.

[0036] like Figure 1 As shown, the specific steps include:

[0037] Specifically, the following steps are included:

[0038] Step 1: Select a range of point clouds on the two feature planes, perform principal component analysis on the range of point clouds, and obtain the feature normals that represent the two feature planes respectively;

[0039] Step 2: Perform rotation transformation on the original point cloud and the range point cloud on the feature plane so that the feature normal of the feature plane is parallel and aligned with any coordinate axis;

[0040] Step 3: Based on the maximum and minimum coordinate points in the range point cloud after rotation transformation, establish an axis-aligned bounding box, set a bounding threshold to adjust the size of the axis-aligned bounding box, and filter the range point cloud through the axis-aligned bounding box to obtain the feature point cloud.

[0041] Step 4: Perform inverse transformation on the feature point cloud to obtain the mapped point cloud of the feature point cloud in the original point cloud of the feature plane. Based on the mapped point cloud, perform least squares plane fitting to obtain the first fitting plane representing the first feature plane and the second fitting plane representing the second feature plane.

[0042] Step 5: Downsample the mapped point cloud based on the number of point clouds. Calculate the distance between the two feature planes based on the difference in the number of mapped point clouds after downsampling.

[0043] Furthermore, step 1 specifically includes:

[0044] Step 1.1: Calculate the covariance matrix of the current feature plane based on the points in the range point cloud;

[0045] Step 1.2: Calculate the eigenvalues ​​and eigenvectors of the current feature plane based on the covariance matrix;

[0046] Step 1.3: Extract the direction of the eigenvector with the smallest eigenvalue as the feature normal of the current feature plane.

[0047] Furthermore, step 3 specifically includes:

[0048] Step 3.1: Traverse the coordinates of all points in the point cloud after the rotation transformation, extract the point with the largest coordinates on all three axes as the maximum coordinate point, and extract the point with the smallest coordinates on all three axes as the minimum coordinate point.

[0049] Step 3.2: Create an axis-aligned bounding box using the line segment connecting the maximum and minimum coordinate points as the diagonal.

[0050] Step 3.3: Set the bounding threshold and adjust the size of the axis-aligned bounding box. Select points in the original point cloud of the feature plane that are located inside the axis-aligned bounding box to form the feature point cloud. After the size of the axis-aligned bounding box is adjusted, it is required that the axis-aligned bounding box can simultaneously intersect with the original point clouds of two feature planes.

[0051] Furthermore, step 5 specifically includes:

[0052] Step 5.1: Use the farthest point sampling algorithm to downsample the mapping point cloud representing the two feature planes to obtain the first downsampled point cloud representing the first feature plane and the second downsampled point cloud representing the second feature plane.

[0053] Step 5.2: Compare the number of points in the downsampled feature point cloud. If the number of points in the first downsampled point cloud is less than or equal to half the number of points in the second downsampled point cloud, proceed to step 5.3; if the number of points in the first downsampled point cloud is greater than half the number of points in the second downsampled point cloud, proceed to step 5.4.

[0054] Step 5.3: Calculate only the first distance between points in the first downsampled point cloud and the second fitting plane as the distance between the two feature planes;

[0055] Step 5.4: Calculate the first distance between points in the first downsampled point cloud and the second fitting plane, and calculate the second distance between points in the second downsampled point cloud and the first fitting plane. Use the average of the sum of the first distance and the second distance as the distance between the two feature planes.

[0056] Furthermore, in step 5.1, the number of points in the mapped point cloud is extracted. If the number of points is greater than or equal to 1000, the mapped point cloud is downsampled; if the number of points is less than 1000, the mapped point cloud is not downsampled.

[0057] Example 2:

[0058] This embodiment discloses a method for calibrating the planar distance of part features based on PCA point cloud filtering, which is an improvement on Embodiment 1, specifically as follows:

[0059] Four points that can represent the point cloud region to be fitted in the plane are selected on each of the two feature planes of the part. The four points on the first feature plane are denoted as the range point cloud PSet1, and the four points on the second feature plane are denoted as the range point cloud PSet2.

[0060] Principal component analysis is performed on the four points in PSet1 to obtain the eigenvectors of the first characteristic plane. This involves calculating the covariance matrix of the four points in PSet1, then calculating the eigenvalues ​​and eigenvectors. The direction of the eigenvector corresponding to the smallest eigenvalue is considered as the eigenvector of the first characteristic plane.

[0061] Perform a rotation transformation on the original point cloud and the range point cloud PSet1 of the first feature plane so that the feature normal of the first feature plane is parallel to the X-axis in the reference coordinate system.

[0062] By finding the maximum coordinate point (X) in the rotated point cloud PSet1, the maximum coordinate point (X) is obtained. max Y max Z max ), minimum coordinate point (X) min Y min Z min An initial axis-aligned bounding box is defined using the line segment connecting the maximum and minimum coordinate points as the diagonal. A bounding threshold is set based on the point cloud's accuracy and point spacing to adjust the size of the axis-aligned bounding box, ensuring it intersects with the original point clouds representing two feature planes simultaneously. Points within the axis-aligned bounding box from the original point cloud are extracted to form a feature point cloud, thus filtering the original point cloud. The feature point cloud is inversely transformed to obtain its mapping point cloud in the original point cloud of the feature plane. This mapping point cloud, denoted as cloud1, is fitted using the global least squares method to obtain the first fitting plane, plane1. Similarly, the mapping point cloud, denoted as cloud2, is obtained by inversely transforming the feature point cloud representing the second feature plane, and fitted using the global least squares method to obtain the second fitting plane, plane2.

[0063] Since the fitted planes are not necessarily parallel, only an approximate value can be calculated. When the point clouds of mapped point clouds cloud1 or cloud2 are too dense, calculating the plane distance will consume a lot of resources. To make the sampling points more evenly distributed, when the number of points in the mapped point cloud is greater than or equal to 1000, the farthest point sampling algorithm is used to downsample the mapped point clouds cloud1 or cloud2.

[0064] After downsampling, the difference in the number of points in the mapped point cloud cloud1 and the mapped point cloud cloud2 is determined as follows:

[0065] When the number of points in the mapped point cloud cloud1 is less than or equal to half the number of points in the mapped point cloud cloud2, only the distance from the points in the mapped point cloud cloud cloud1 to the second fitting plane plane2 is calculated. That is, the distance between the two feature planes is calculated as follows:

[0066] ;

[0067] in: d represents the distance from the i-th point in cloud1 to plane2; d represents the distance between the two feature planes; num1 represents the number of points in cloud1.

[0068] When the number of points in the mapped point cloud cloud1 is greater than half the number of points in the mapped point cloud cloud2, it is necessary to calculate the distances from points in cloud1 to plane plane2 and from points in cloud2 to plane plane1. In this case, the calculated planar distances are...

[0069] ;

[0070] in: Let be the distance from the i-th point in cloud1 to plane2; num1 represents the distance from the i-th point in cloud2 to plane1; num2 represents the number of points in cloud1; num2 represents the number of points in cloud2; and d represents the distance between the two feature planes.

[0071] The rest of this embodiment is the same as that of Embodiment 1, so it will not be repeated here.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for calibrating the distance between feature planes of a part based on PCA-based point cloud filtering, used to calculate the distance between two feature planes on a part, characterized in that... Principal component analysis is performed on the selected range point cloud on the two feature planes to obtain the feature normal of the feature plane. All point clouds are rotated and transformed so that the feature normal is aligned with any coordinate axis. The range point cloud is filtered based on the axis-aligned bounding box to obtain the feature point cloud. Least square plane fitting is performed based on the feature point cloud to obtain two fitting planes. The distance between the two feature planes is calculated based on the number of point clouds contained in the two fitting planes. Specifically, the following steps are included: Step 1: Select a range of point clouds on the two feature planes, perform principal component analysis on the range of point clouds, and obtain the feature normals that represent the two feature planes respectively; Step 2: Perform rotation transformation on the original point cloud and the range point cloud on the feature plane so that the feature normal of the feature plane is parallel and aligned with any coordinate axis; Step 3: Based on the maximum and minimum coordinate points in the range point cloud after rotation transformation, establish an axis-aligned bounding box, set a bounding threshold to adjust the size of the axis-aligned bounding box, and filter the range point cloud through the axis-aligned bounding box to obtain the feature point cloud. Step 4: Perform inverse transformation on the feature point cloud to obtain the mapped point cloud of the feature point cloud in the original point cloud of the feature plane. Based on the mapped point cloud, perform least squares plane fitting to obtain the first fitting plane representing the first feature plane and the second fitting plane representing the second feature plane. Step 5: Downsample the mapped point cloud based on the number of point clouds, and calculate the distance between the two feature planes based on the difference in the number of mapped point clouds after downsampling. Step 5 specifically includes: Step 5.1: Downsample the mapped point cloud representing the two feature planes to obtain the first downsampled point cloud representing the first feature plane and the second downsampled point cloud representing the second feature plane. Step 5.2: Compare the number of points in the downsampled feature point cloud. If the number of points in the first downsampled point cloud is less than or equal to half the number of points in the second downsampled point cloud, proceed to step 5.3; if the number of points in the first downsampled point cloud is greater than half the number of points in the second downsampled point cloud, proceed to step 5.

4. Step 5.3: Calculate only the first distance between points in the first downsampled point cloud and the second fitting plane as the distance between the two feature planes; Step 5.4: Calculate the first distance between points in the first downsampled point cloud and the second fitting plane, and calculate the second distance between points in the second downsampled point cloud and the first fitting plane. Use the average of the sum of the first distance and the second distance as the distance between the two feature planes.

2. The method for calibrating the distance between feature planes of a part based on PCA point cloud filtering according to claim 1, characterized in that, In step 5.1, the number of points in the mapped point cloud is extracted. If the number of points is greater than or equal to 1000, the mapped point cloud is downsampled; if the number of points is less than 1000, the mapped point cloud is not downsampled.

3. The part feature plane distance calibration method based on PCA point cloud filtering according to claim 2, characterized in that, In step 5.1, the farthest point sampling algorithm is used to downsample the mapped point cloud.

4. The method for calibrating the planar distance of part features based on PCA point cloud filtering according to any one of claims 1-3, characterized in that, Step 3 specifically includes: Step 3.1: Traverse the coordinates of all points in the point cloud after the rotation transformation, extract the point with the largest coordinates on all three axes as the maximum coordinate point, and extract the point with the smallest coordinates on all three axes as the minimum coordinate point. Step 3.2: Create an axis-aligned bounding box using the line segment connecting the maximum and minimum coordinate points as the diagonal. Step 3.3: Set the bounding threshold and adjust the size of the axis-aligned bounding box. Select the points in the original point cloud of the feature plane that are located inside the axis-aligned bounding box to form the feature point cloud.

5. The part feature plane distance calibration method based on PCA point cloud filtering according to claim 4, characterized in that, After the size of the axis-aligned bounding box is adjusted, it is required that the axis-aligned bounding box can simultaneously intersect with the original point clouds of the two feature planes.

6. The method for calibrating the distance between feature planes of a part based on PCA point cloud filtering according to any one of claims 1-3, characterized in that, Step 1 specifically includes: Step 1.1: Calculate the covariance matrix of the current feature plane based on the points in the range point cloud; Step 1.2: Calculate the eigenvalues ​​and eigenvectors of the current feature plane based on the covariance matrix; Step 1.3: Extract the direction of the eigenvector with the smallest eigenvalue as the feature normal of the current feature plane.

Images