Geometric shape and position dimension measurement method based on three-dimensional point cloud

Abstract

The invention provides a three-dimensional point cloud-based geometric shape and position dimension measurement method and application thereof. The method utilizes a three-dimensional scanning technology to obtain three-dimensional point cloud data of an industrial part, and processes and measures the three-dimensional point cloud data through the following steps: 1, preprocessing: removing isolated points, outliers and Gaussian noise in the point cloud through filtering; 2, geometric feature extraction: extracting geometric features with linear and arc features, and dividing the geometric features into single-cluster point clouds; 3, straightness and roundness measurement: fitting the point clouds with geometric characteristics into a straight line or a circle so as to measure straightness and roundness; and 4, arc length measurement: providing a method for sequentially searching adjacent points on the linear point cloud, calculating the sum of the distances between the adjacent points, and further measuring the arc length. The application field of the invention comprises the measurement and the quality control of the part size in the manufacturing industry. The method performs non-contact measurement based on the three-dimensional point cloud data, and has the characteristics of high efficiency, high precision and strong anti-interference performance.

Description

Technical Field

[0001] The invention relates to the field of measurement technology, and in particular to a method for measuring the geometric shape and size of a large part based on a three-dimensional point cloud. Background Art

[0002] In traditional manufacturing, the commonly used method for measuring part dimensions is based on direct contact measuring tools, such as three-dimensional coordinate measuring machines, vernier calipers, etc. However, this traditional method has problems such as being time-consuming and labor-intensive, having large limitations on the objects that can be measured, and being easily interfered by human factors. It cannot meet the needs of modern industry for fast and accurate measurement of large parts. In order to solve these problems, a variety of non-contact measurement technologies have emerged in recent years, such as ultrasonic ranging and binocular vision measurement, which have improved the accuracy and efficiency of measuring complex objects, but there are still challenges in measuring the geometric shape and size of large parts.

[0003] With the development of 3D scanning technology, the acquisition of 3D point cloud data has become simpler and more convenient. The application prospects of 3D point clouds in the fields of cultural relics protection, industrial production, intelligent robots and human-computer interaction are becoming increasingly broad. The measurement technology based on 3D point clouds has the advantages of non-contact, high efficiency and high precision, and has become an important development direction in the field of measurement. However, there are currently few point cloud processing methods for the size measurement of large parts in the industry, and most methods only project 3D features into 2D for size measurement, which makes it difficult to achieve high-precision and high-efficiency measurement. Summary of the invention

[0004] The present invention aims to provide a method for measuring the geometric shape and size of large parts based on three-dimensional point clouds to overcome the shortcomings of the prior art. Through the steps of point cloud preprocessing, geometric feature extraction, feature fitting and size measurement, the method can accurately and efficiently measure the geometric shape and size of large parts such as roundness, straightness, arc length, etc., and realize the evaluation and optimization of part processing accuracy.

[0005] In order to achieve the above objectives, the present invention proposes a method for measuring the geometric shape and size of large parts, comprising the following steps:

[0006] (1) Perform point cloud preprocessing and use statistical filtering and bilateral filtering to remove noise.

[0007] Statistical filtering is used to remove isolated points and outliers in the original point cloud, and the sampling points p i Perform statistical analysis on the k neighboring points of the point and calculate the average distance d from all neighboring points to the point i , all d i It should be Gaussian distributed, with its mean μ and standard deviation σ known. Given a standard deviation threshold coefficient K, it is assumed that d i Points outside the range of [μ-Kσ, μ+Kσ] are noise and are deleted.

[0008] Bilateral filtering based on normal vector estimation is used to remove Gaussian noise in the original point cloud. The k neighborhood points of the sampling point are weighted averaged, and the mean of the filter convolution kernel is 0. Given a neighborhood point distance standard deviation σ d and the standard deviation of the distance in the normal direction σ n , if the distance from the neighborhood point to the sampling point is d d , the distance in the normal direction is d n , then the convolution kernel is:

[0009]

[0010] (2) Extract the straight line and arc features of the part and divide them into single cluster point clouds respectively.

[0011] The RANSAC plane fitting algorithm is used to extract the key plane. After intercepting the point cloud near the plane as the element to be fitted, the initial element is three randomly selected non-collinear points in the point cloud. The mathematical model to be calculated is:

[0012] Ax+By+Cz+D=0,A 2 +B 2 +C 2 ≠0

[0013] Set the distance threshold T. If the distance from the point to be checked to the fitting plane is less than T, it is an internal point and can be considered on the plane. Otherwise, it is an external point. After multiple iterations, the plane model containing the most internal points is selected. At this time, the internal points are the extracted plane.

[0014] The edge extraction algorithm based on normal vector estimation is used to extract the boundary points of each plane. First, the k nearest neighbor points p of any point P in the plane are searched. i (i=1,2,...,k), and P form a reference point set X, then fit the micro-cut plane of X, and project it onto the micro-cut plane. The projection point set is X′={(x i ′,y i ′,z i ′)|i=0,1,...,k}, taking the projection point P′ of the sampling point as the starting point, p i ′(i=1, 2, ..., k) is the end point, and the vector set is obtained For any vector Calculate the angle α between it and other vectors j And sort them, and then get the angle δ between adjacent vectors j , when the maximum angle δ between adjacent vectors maxWhen it is greater than the angle threshold ε, the sampling point P is judged to be a boundary point. By performing edge extraction on each plane, a closed edge contour line containing straight line or arc features can be obtained.

[0015] The method of rotation interception is used to extract the generatrix of the part. A plane Q passing through the axis l is selected, and the point cloud is continuously rotated around the axis l at a certain angle. After each rotation, all points on the plane Q are extracted. Since the distribution of the point cloud is scattered, a distance threshold needs to be set. Points whose distance to the plane is less than the threshold can be regarded as on the plane. At this time, what is obtained is the intersection of plane Q and the cylindrical surface of the part, which is a closed contour containing the inner and outer generatrixes. The rotation angle determines the position of the extracted generatrix on the part.

[0016] For closed contour point clouds that need to measure straightness and roundness, the method of directly cutting off 1% of the two ends is adopted, while retaining almost all the points on the straight line or arc feature; for point clouds that need to measure arc length, the plane where the two ends to be removed are located is extracted, and the part of the point cloud that overlaps with the plane is deleted, so that the two ends and the main part of the arc can be accurately segmented. Each group of point clouds obtained contains two straight line or arc features, which are divided into two clusters using Euclidean clustering. In each group of point clouds, any point P is selected as a seed point, and k neighborhood points are obtained by the k-nearest neighbor search algorithm. The points whose Euclidean distance to the seed point is less than the set threshold t are classified into the same cluster Q with P, and new seed points are selected in Q to continue searching. If the number of points in Q does not increase, the clustering process of Q ends. If the number of points in Q is within the threshold range, the clustering result is saved, and new seed points are selected in the remaining point clouds. Repeat the above steps until all point clouds are traversed and clustering is completed.

[0017] (3) Perform feature fitting and measure straightness and roundness based on the fitting results.

[0018] The RANSAC algorithm is used to fit the single cluster point cloud after clustering to a straight line and a circle. The straight line fitting requires randomly selecting two points from the point cloud to be fitted as the initial elements. The mathematical model for calculation is the spatial straight line equation, which contains 6 parameters, namely the three-dimensional coordinates of a point on the straight line and the three components of the direction vector of the straight line. The circle fitting requires randomly selecting three non-collinear points from the point cloud to be fitted as the initial elements. The mathematical model for calculation is the spatial circle equation, which contains 7 parameters, namely the three-dimensional coordinates of the center of the circle, the three components of the normal vector of the plane where the circle is located, and the radius of the circle. Set the distance threshold T. If the distance from the point to be checked to the fitted straight line or circle is less than T, it is an internal point and can be considered on the straight line or circle. Otherwise, it is an external point. After multiple iterations, the straight line model or circle model containing the most internal points is selected as the fitting result.

[0019] Calculate the offset Δ of each point in the point cloud relative to the fitted line or fitted circle i, the offset of the straight line point cloud relative to the fitted line is the distance from the point to the spatial line, and the offset of the arc point cloud relative to the fitted circle is:

[0020]

[0021] Where (x0, y0, z0) is the coordinate of the center of the fitting circle, r is the fitting radius, (x i ,y i , z i ) is any point P in the arc point cloud i The coordinates of i Twice the maximum value is the corresponding straightness or roundness measurement result.

[0022] (4) For the extracted arc features, search all points in adjacent order and calculate the arc length.

[0023] Select any point in the arc feature as the current point, extract the point closest to it, that is, the next point adjacent to it as the new current point, extract the point closest to it again and determine whether it has been extracted. If this point has been extracted, select the nearest point from the remaining points and determine again until the point closest to the current point and not extracted is found as the next point in the remaining points. Repeat the above process. The new current point must be the adjacent point on the same side of all the points currently extracted. Therefore, the extracted points must continue to extend in the same direction in the adjacent order. Set a distance threshold. If the distance between the next point and the current point is greater than the threshold, or all points have been extracted, it means that the current point is already the endpoint of the arc. At this time, stop extracting, take the current point as the starting point, and search the point cloud in sequence. You can search all points on the arc in the order of extending in one direction along the arc, and calculate the Euclidean distance between every two adjacent points. The sum of the distances is the arc length measurement result.

[0024] The large-scale part geometric shape and size measurement method of the present invention has the following advantages: adopting a non-contact measurement method avoids the limitations of traditional contact measurement and interference of human factors; based on three-dimensional point cloud data processing, the range of measurable objects is expanded, and it is suitable for complex large parts that are not easy to measure directly; through geometric feature extraction and fitting, the measurement of geometric shape and size such as straightness, roundness, arc length, etc. is realized, and then the evaluation of part processing accuracy is realized; the arc length is calculated by the sum of the distances between adjacent points of the arc point cloud, and the calculation result is more accurate, and part defects can be accurately detected. Therefore, the present invention has broad application prospects and commercial value in the field of industrial manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flow chart of a method for measuring the size of a part geometry based on a three-dimensional point cloud is provided in an embodiment.

[0026] Figure 2 A schematic diagram of rocket tank wall panel parts provided for an embodiment.

[0027] Figure 3 The preprocessed wall panel point cloud provided in the embodiment.

[0028] Figure 4 Schematic diagram of features and key planes to be extracted from the wall panel parts provided in the embodiment.

[0029] Figure 5 The key plane extraction result of the wall panel point cloud provided in the embodiment.

[0030] Figure 6 A schematic diagram of the generatrix extraction of a wall panel part provided in an embodiment.

[0031] Figure 7 This is the Euclidean clustering result of the straight line point cloud provided in the embodiment.

[0032] Figure 8 This is the Euclidean clustering result of the arc point cloud provided in the embodiment. DETAILED DESCRIPTION

[0033] The present invention will now be described in further detail with reference to the accompanying drawings.

[0034] In order to solve the problem that the geometric shape and size of rocket tank wall panel parts are difficult to measure in the prior art, the present invention provides a wall panel part geometric size detection method based on three-dimensional point cloud, the process is as follows Figure 1 shown. Figure 2 Shown is a schematic diagram of a rocket tank wall panel part in an embodiment of the present invention. The part is a 1 / 4 cylindrical surface with an outer radius of 1150mm, an inner radius of 1147mm, a wall panel thickness of 3mm and a height of 500m.

[0035] This embodiment includes the following steps:

[0036] (1) Preprocess the original point cloud and use statistical filtering and bilateral filtering to remove noise.

[0037] Statistical filtering is used to remove isolated points and outliers in the original point cloud, and the sampling points p i Statistical analysis is performed on the 50 neighboring points of the point, and the average distance d from all neighboring points to the point is calculated. i , all d i It should be Gaussian distributed, with its mean μ and standard deviation σ known. i Points outside the range of [μ-σ, μ+σ] are noise and are deleted.

[0038] Bilateral filtering based on normal vector estimation is used to remove Gaussian noise in the original point cloud, and the k neighborhood points of the sampling point are weighted averaged. If the distance from the neighborhood point to the sampling point is d d , the distance in the normal direction is d n , then the convolution kernel is:

[0039]

[0040] After the original point cloud is filtered by bilateral filtering and statistical filtering, isolated points, outliers, Gaussian noise, etc. are effectively removed, and the edge features are well maintained. The preprocessed point cloud is as follows Figure 3 shown.

[0041] (2) Extract the straight line features and arc features of the side wall, generatrix, etc. of the wall panel and divide them into single cluster point clouds respectively.

[0042] like Figure 4 As shown, for the wall panel parts, the geometric features that need to be extracted include arc features and straight line features such as side edges and generatrixes, among which the arcs and side edges are the edges where the inner wall, outer wall, bottom surface and side surface intersect, respectively. Therefore, it is necessary to first extract and fit the four key planes of the wall panel from the point cloud, including two side surfaces and two bottom surfaces.

[0043] The RANSAC plane fitting algorithm is used to extract the key planes. The points with a distance less than 0.1 mm from the fitting plane are taken as internal points. A single layer and a complete plane can be extracted. The four key planes extracted are as follows: Figure 5 shown.

[0044] The edge extraction algorithm based on normal vector estimation is used to extract the boundary points of each plane respectively. The 100 neighborhood points of any point P in the plane are searched, and the 100 neighborhood points are combined with P to form a reference point set and projected onto the micro-cut plane. The vector set is obtained with the projection point of the sampling point and the projection point of the neighboring point as the end point, and the angle δ between adjacent vectors is calculated. j , when the maximum angle δ between adjacent vectors max When it is greater than the angle threshold π / 1.5, the point P is judged to be a boundary point. By performing edge extraction on the four key planes, a closed edge contour line containing straight line or arc features can be obtained.

[0045] The inner and outer generatrixes to be extracted are the parts where the plane Q passing through the axis l intersects with the inner and outer arc surfaces of the wall panel, such as Figure 6 As shown in the figure, this feature can be used to extract the generatrix. Select a plane Q passing through the axis l, rotate the point cloud around the axis l by 30° and 60° respectively, and extract the points with a distance less than 0.1 mm from the plane Q after each rotation, and you can get two sets of closed contours containing the inner and outer generatrixes.

[0046] For closed contour point clouds that need to measure straightness and roundness, the method of directly cutting off 1% of the two ends is adopted, while retaining almost all the points on the straight line or arc feature. For the bottom edge contour containing arc features, the part of the point cloud that overlaps with the two side faces is deleted, and the two ends can be accurately removed and the main part of the arc can be segmented.

[0047] Each point cloud obtained contains two straight line or arc features, which are divided into two clusters using Euclidean clustering. In each point cloud, select any point P as the seed point, and obtain 50 neighborhood points through the k-nearest neighbor search algorithm. The points with a Euclidean distance less than 2 mm from the seed point are classified into the same cluster Q with P until the number of points in Q no longer increases. Select a new seed point from the remaining point clouds, and repeat the above steps until all point clouds are traversed and clustering is completed. Each side, busbar or arc point cloud is divided into two clusters of point clouds with straight line features or arc features, such as Figure 7 and Figure 8 shown.

[0048] (3) Perform feature fitting and measure straightness and roundness based on the fitting results.

[0049] The RANSAC algorithm is used to fit the point clouds of the side and the main line into straight lines, and the point cloud corresponding to the arc part into a circle. The straight line fitting requires randomly selecting two points from the point cloud to be fitted as initial elements, and the mathematical model for calculation is the spatial straight line equation; the circle fitting requires randomly selecting three non-collinear points from the point cloud to be fitted as initial elements, and the mathematical model for calculation is the spatial circle equation, which contains 7 parameters. The fitting result is used as a benchmark to calculate the offset Δ of each point in the point cloud relative to the fitted straight line or fitted circle i , the offset of the straight line point cloud relative to the fitted line is the distance from the point to the spatial line, and the offset of the arc point cloud relative to the fitted circle is:

[0050]

[0051] Where (x0, y0, z0) is the coordinate of the center of the fitting circle, r is the fitting radius, (x i ,y i , z i ) is any point P in the arc point cloud i The coordinates of i Twice the maximum value is the corresponding straightness or roundness measurement result.

[0052] (4) For the extracted arc features, search all points in adjacent order and calculate the arc length.

[0053] Select any point in the arc feature as the current point, extract the closest point to it, that is, the next point adjacent to it as the new current point, and extract the closest point to it again. If this point has been extracted, select the nearest point from the remaining points and judge again until the nearest point to the current point that has not been extracted is found among the remaining points as the next point. Repeat the above process continuously. The new current point must be the adjacent point on the same side of all the points currently extracted. Therefore, the extracted points must continue to extend in the same direction in the adjacent order until the endpoint of the arc is extracted. At this time, stop the extraction, take the power off as the starting point, and search the point cloud in sequence. You can search all points on the arc in the order of extending in one direction along the arc, and calculate the Euclidean distance between every two adjacent points. The sum of the distances is the arc length measurement result.

Claims

1. A method for measuring geometric shape and size based on three-dimensional point cloud, characterized in that The following steps are involved: (1) Preprocess the acquired point cloud data, use statistical filtering to remove isolated points and outliers in the original point cloud, and use bilateral filtering based on normal vector estimation to remove Gaussian noise in the original point cloud; (2) Extracting the geometric features of the parts from the preprocessed point cloud data, including edge contours and generatrix lines, and dividing them into single cluster point clouds with different geometric features; (3) The extracted geometric features are fitted into straight lines or circles using the RANSAC algorithm, and the straightness and roundness are measured based on the fitting results; (4) For the extracted arc features, search all points in order of adjacent points and calculate the Euclidean distance between adjacent points. The sum of the distances between adjacent points is the arc length measurement result.

2. The method according to claim 1, wherein in the preprocessing step in (1), the statistical filtering step comprises: i Perform statistical analysis on the k neighboring points of the point and calculate the average distance d from all neighboring points to the point i , all d i It should be Gaussian distributed, with its mean μ and standard deviation σ known. Given a standard deviation threshold coefficient K, it is assumed that d i Points outside the range of [μ-Kσ, μ+Kσ] are noise and are deleted; the bilateral filtering step includes weighted averaging of the k neighborhood points of the sampling point, the mean of the filter convolution kernel is 0, and a neighborhood point distance standard deviation σ is given. d and the standard deviation of the distance in the normal direction σ n , if the distance from the neighborhood point to the sampling point is d d , the distance in the normal direction is d n , then the convolution kernel is:

3. According to the method of claim 1, in the step of extracting edge contour lines in (2), the key plane is first extracted by using the RANSAC plane fitting algorithm, and the point cloud near the plane is intercepted as the element to be fitted. The initial elements are three non-collinear points randomly selected from the point cloud, and the mathematical model to be calculated is: Ax+By+Cz+D=0,A 2 +B 2 +C 2 ≠0 Set the distance threshold T. If the distance between the point to be checked and the fitting plane is less than T, it is an internal point and can be considered on the plane. Otherwise, it is an external point. After multiple iterations, the plane model containing the most internal points is selected. At this time, the internal points are the extracted planes. Then, the edge extraction algorithm based on normal vector estimation is used to extract the boundary points of each plane, which is the edge contour line of the part. First, the k nearest neighbor points p of the sampling point P are searched. i (i=1,2,...,k), and P form a reference point set X, then fit the micro-cut plane of X, and project it onto the micro-cut plane. The projection point set is X′={(x i ′,y i ′,z i ′)|i=0,1,...,k}, taking the projection point P′ of the sampling point as the starting point, p i ′(i=1, 2, ..., k) is the end point, and the vector set is obtained For any of the vectors Calculate the angle α between it and other vectors j And sort them, and then get the angle δ between adjacent vectors j , when the maximum angle δ between adjacent vectors max When it is greater than the angle threshold ε, the sampling point P is judged to be a boundary point.

4. According to the method described in claim 1, in the step of extracting the generatrix in (2), a plane Q passing through the axis line l is selected, and the point cloud is continuously rotated around the axis line l at a certain angle. After each rotation, the points on the plane Q are extracted from all the points. Since the distribution of the point cloud is scattered, it is necessary to set a distance threshold. Points whose distance to the plane is less than the threshold can be regarded as being on the plane. At this time, what is obtained is the intersection of the plane Q and the cylindrical surface of the part, which is a closed contour containing the inner and outer generatrixes. The rotation angle determines the position of the extracted generatrix on the part.

5. According to the method of claim 1, in the step of dividing the point cloud into single cluster point clouds with different geometric features in (2), firstly, the two ends of the extracted closed contour line are removed by using geometric relationship. For point clouds that need to measure straightness and roundness, a method of directly cutting off 1% of the two ends is adopted, while retaining almost all points on the straight line or arc feature. For point clouds that need to measure arc length, the plane where the two ends to be removed are located is extracted, and the part of the point cloud that overlaps with the plane is deleted, so that the two ends and the main part of the arc can be accurately segmented; then, each group of point clouds containing two geometric features is divided into two clusters by using Euclidean clustering, and any point P is selected as a seed point. K neighborhood points are obtained by using a k-nearest neighbor search algorithm. Points whose Euclidean distance to the seed point is less than a set threshold t are classified into the same cluster Q with P, and new seed points are selected in Q to continue searching. If the number of points in Q does not increase, the clustering process of Q is terminated. If the number of points in Q is within the threshold range, the clustering result is saved, and new seed points are selected in the remaining point clouds. The above steps are repeated until all point clouds are traversed and clustering is completed.

6. According to the method of claim 1, in the step of fitting a straight line and a circle using the RANSAC algorithm in (3), the fitting of the straight line requires randomly selecting two points from the point cloud to be fitted as initial elements, and the mathematical model to be calculated is a spatial straight line equation, which includes 6 parameters, namely the three-dimensional coordinates of a point on the straight line and the three components of the direction vector of the straight line. The fitting of the circle requires randomly selecting three non-collinear points from the point cloud to be fitted as initial elements, and the mathematical model to be calculated is a spatial circle equation, which includes 7 parameters, namely the three-dimensional coordinates of the center of the circle, the three components of the normal vector of the plane where the circle is located, and the radius of the circle. A distance threshold T is set. If the distance from the point to be checked to the fitted straight line or circle is less than T, it is an internal point and can be considered to be on the straight line or circle. Otherwise, it is an external point. After multiple iterations, the straight line model or circle model containing the most internal points is selected as the fitting result.

7. According to the method of claim 1, in the step of calculating the straightness and roundness in (3), the offset Δ of each point in the point cloud relative to the fitted line or the fitted circle is calculated. i , the offset of the straight line point cloud relative to the fitted line is the distance from the point to the spatial line, and the offset of the arc point cloud relative to the fitted circle is: Where (x0, y0, z0) is the coordinate of the center of the fitting circle, r is the fitting radius, (x i ,y i , z i ) is any point P in the arc point cloud i The coordinates of i Twice the maximum value is the corresponding straightness or roundness measurement result.

8. According to the method described in claim 1, in the step of searching the point cloud in sequence in adjacent order in said (4), select any point as the current point, extract the point closest to it, that is, the next point adjacent to it as the new current point, extract the point closest to it again and judge whether it has been extracted. If this point has been extracted, select the nearest point from the remaining points and judge again until the point closest to the current point and not extracted is found among the remaining points as the next point. Repeat the above process continuously. The new current point must be the adjacent point on the same side of all the points currently extracted. Therefore, the extracted points must extend continuously in the same direction in adjacent order.

9. According to the method described in claim 1, in the step of searching all points on the arc point cloud in (4), select any point as the starting point, search points extending in the same direction in sequence, set a distance threshold, and if the distance between the next point and the current point is greater than the threshold, or all points have been extracted, it means that the current point is already the endpoint of the arc. At this time, stop the extraction, take the current point as the starting point, and search the point cloud in sequence, so that all points on the arc can be searched in the order of extending in one direction along the arc.

Images