A high-precision registration method for multi-view point clouds of welding workpieces

Abstract

The invention discloses a high-precision registration method for multi-view point clouds of welding workpieces. The method is characterized in that spot welding is first performed on the welding grooves of two workpieces to obtain workpieces to be welded; a structured light camera is used to collect multiple frames of original point clouds of the workpieces to be welded at different viewpoints, and background point clouds are removed to obtain multiple frames of workpiece point clouds at different viewpoints; then, the workpiece upper surface point cloud containing the welding groove is extracted, and each frame of the workpiece upper surface point cloud is roughly registered; the quality weight of each point in the workpiece upper surface point cloud is calculated, and then the point clouds of adjacent frames are precisely registered by a weighted ICP algorithm to optimize the adjacent frame poses, so as to achieve precise splicing between multiple frames of workpiece upper surface point clouds at different viewpoints, and obtain a complete workpiece upper surface point cloud containing the welding groove; the method has the advantage that the influence of point cloud distortion on point cloud precise registration is well reduced, and the accuracy of point cloud registration is improved, thereby ensuring the accuracy of subsequent three-dimensional model reconstruction and the accuracy of automatic welding.

Description

Technical Field

[0001] The present invention relates to the field of intelligent welding technology, and in particular to a high-precision registration method for multi-view point clouds of welding workpieces. Background Art

[0002] Traditional workpiece welding is usually done manually, but manual welding has problems such as poor consistency and low efficiency, and cannot meet the current mass production welding industry needs. In order to improve the welding quality and efficiency of workpieces, robot automated welding is gradually replacing manual welding.

[0003] Before automatically welding the workpiece, it is necessary to first collect the morphological features of the workpiece to be welded. Due to the limitations of the field of view of the current 3D scanning equipment, it is impossible to directly obtain the complete point cloud data of the target object. Therefore, it is necessary to collect point cloud data of different perspectives of the target object at different angles, and then perform registration processing (i.e. point cloud registration) on the obtained point cloud data of different perspectives. The accuracy of point cloud registration will directly affect the accuracy of subsequent 3D reconstruction. At present, when using a structured light camera to collect the point cloud of the surface of the target object, there will inevitably be a certain degree of distortion, and the point cloud distortion will affect the subsequent point cloud precise registration, that is, it will affect the accuracy of the point cloud registration, thereby affecting the accuracy of subsequent automatic welding. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a high-precision registration method for multi-view point clouds of welding workpieces, which can reduce the influence of point cloud distortion on subsequent point cloud precise registration, so as to improve the accuracy of point cloud registration.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: a high-precision registration method for multi-view point clouds of welding workpieces, comprising the following specific steps:

[0006] (1) Spot welding is first performed on the welding grooves of two workpieces to connect them together to obtain the workpieces to be welded;

[0007] (2) Installing a structured light camera on the end arm of the robot, using the structured light camera to collect multiple frames of original point clouds of the workpiece to be welded at different perspectives, and removing the background point clouds of the workpiece to be welded from the multiple frames of original point clouds to obtain multiple frames of workpiece point clouds at different perspectives;

[0008] (3) dividing the workpiece side point cloud in the workpiece point cloud into independent clusters, and extracting the workpiece upper surface point cloud containing the welding groove from the clusters;

[0009] (4) Unify the upper surface point clouds of the workpiece at different perspectives in each frame into the robot base coordinate system, and perform coarse registration on the upper surface point clouds of the workpiece at different perspectives in each frame;

[0010] (5) Calculate the quality weight of each point in the workpiece upper surface point cloud, and then use the weighted ICP algorithm to accurately align the point clouds of adjacent frames to optimize the poses of adjacent frames, so as to achieve accurate stitching of the workpiece upper surface point clouds of multiple frames with different perspectives and obtain a complete workpiece upper surface point cloud including the welding groove.

[0011] Furthermore, in the step (2), the background point cloud in the original point cloud is removed by a random sampling consistency algorithm.

[0012] Furthermore, in the step (3), the workpiece side point cloud in the workpiece point cloud is divided into independent clusters by using the NV-DBSCAN clustering algorithm.

[0013] Furthermore, the specific process of the NV-DBSCAN clustering algorithm is as follows: Calculate the point cloud to be clustered P = {p1, p2, p33p n The normal vector of each point in} constitutes the normal vector set N = {n1,n2,n33n n}, for point p in point cloud P i 、p j , if its corresponding normal vector n i 、n j Satisfies the relationship: Where: γ is the threshold for determining whether two points satisfy the normal vector consistency constraint, then the point p is defined i and p j Satisfy the normal vector consistency constraint and move point p i and p j Grouped into the same cluster of point clouds.

[0014] Furthermore, in the step (5), the quality weight of each point in the workpiece surface point cloud is calculated as follows:

[0015] For a point P in the point cloud, the off-axis weight function f(α) of the point cloud quality of the point P is defined by calculating the angle α between the straight line connecting the optical center of the structured light camera and the point P and the optical axis of the camera:

[0016] Among them: α1 is the maximum value of angle α, k α is the decay exponent of the off-axis weight function f(α), k α The larger the value is defined, the faster the off-axis weight function f(α) decays as α increases;

[0017] At the same time, the direction vector of the camera optical axis is z, and the local normal vector of point P in the point cloud is n. Then the angle between the camera optical axis and the normal vector n is β, and the point cloud quality tilt weight function f(β) of the point P is defined by the angle β: Where: β1 is the maximum value of angle β, kβ is the decay exponent of the tilt weight function f(β), k β The larger the value is defined, the faster the tilt weight function f(β) decays as β increases.

[0018] Furthermore, in step (5), the specific steps of the weighted ICP algorithm are:

[0019] (5-1) The rough registration result of step (4) is defined as the initial transformation matrix T0;

[0020] (5-2) Corresponding point matching: The point clouds collected from each frame constitute a multi-frame point cloud set Q = {P1, P2, ..., P k ,…,P m-1 ,P m}, m is the number of point clouds, which is equal to the number of frames of the original point cloud of the workpiece to be welded collected by the structured light camera. Take one of the point clouds P in Q k is the source point cloud, k is initialized to 2, for the source point cloud P k Each point in Find its nearest neighbor in the target point cloud P1

[0021] (5-3) Use Euclidean distance to measure the point pair error e i : T is the transformation matrix during the optimization process;

[0022] (5-4) Calculate the weight value ω of the matching point pair i : Where: f i (α) is the source point cloud P k Points in The off-axis weight function value of the point cloud quality, f i (β) is the source point cloud P k Points in The point cloud quality tilt weight function value;

[0023] (5-5) By minimizing the weighted error, the transformation matrix T is iteratively optimized and the source point cloud P is transformed using the weighted least squares method. k The weighted error sum of the Euclidean distances between all points in and the corresponding nearest neighbor points in the target point cloud P1 is used as the optimization target:

[0024] Where: T * It represents the optimal solution of the transformation matrix T obtained by iterative optimization, that is, the optimal transformation matrix, and m represents the number of point clouds;

[0025] (5-6) When the number of iterations reaches the set maximum number of iterations or when the difference between the optimization variables is less than the threshold, the iteration is stopped and convergence is completed;

[0026] (5-7), using the obtained optimal transformation matrix T * The source point cloud P k Perform pose transformation and transform the source point cloud P k Combine with the target point cloud P1 to form a new point cloud, and perform point cloud downsampling to obtain a new point cloud Complete the update of the target point cloud P1; set k=k+1 to increase in sequence until k=m, and stop updating; the point cloud finally updated is the spliced ​​point cloud obtained by precise registration.

[0027] Compared with the prior art, the advantage of the present invention is that the method calculates the quality weight of each point in the point cloud, that is, by establishing a point cloud quality weight model to evaluate the accuracy of the point cloud collected by the structured light camera, and in the subsequent point cloud registration, provides weight parameters for the traditional ICP algorithm, and performs point cloud precise registration through the weighted ICP algorithm to more accurately match adjacent point clouds, thereby greatly reducing the impact of point cloud distortion on point cloud precise registration and improving the accuracy of point cloud registration, thereby ensuring the accuracy of subsequent three-dimensional model reconstruction and automatic welding. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of the present invention.

[0029] Figure 2 Schematic diagram of point cloud quality weight of the present invention. DETAILED DESCRIPTION

[0030] The present invention is further described in detail below with reference to the accompanying drawings.

[0031] like Figure 1 As shown, a high-precision registration method for multi-view point clouds of welding workpieces includes the following specific steps:

[0032] (1) Spot welding is first performed on the welding grooves of two workpieces to connect them together to obtain the workpieces to be welded;

[0033] (2) Install a structured light camera on the end arm of the robot, use the structured light camera to collect multiple frames of original point clouds of the workpiece to be welded at different perspectives (i.e., different local parts), and use a random sampling consistency algorithm (i.e., RANSAC algorithm) to remove the background point clouds of the workpiece to be welded in the multiple frames of original point clouds to obtain multiple frames of workpiece point clouds at different perspectives;

[0034] (3) The side point cloud of the workpiece in the workpiece point cloud is divided into independent clusters by using the NV-DBSCAN clustering algorithm, and the upper surface point cloud of the workpiece containing the welding groove is extracted from the clusters;

[0035] The specific process of NV-DBSCAN clustering algorithm is as follows: Calculate the point cloud to be clustered P = {p1, p2, p33p n The normal vector of each point in} constitutes the normal vector set N = {n1,n2,n33n n}, for point p in point cloud P i 、p j , if its corresponding normal vector n i 、n j Satisfies the relationship: Where: γ is the threshold for determining whether two points satisfy the normal vector consistency constraint, then the point p is defined i and p j Satisfy the normal vector consistency constraint and move point p i and p j Classify into the same cluster of point clouds;

[0036] (4) Unifying the upper surface point clouds of the workpiece at different viewing angles in each frame into the robot base coordinate system, and roughly registering the upper surface point clouds of the workpiece at different viewing angles in each frame; the specific method of roughly registering the point clouds is well known in the art and will not be elaborated here;

[0037] (5) Calculate the mass weight of each point in the workpiece surface point cloud, such as Figure 2 As shown, specifically:

[0038] For a point P in the point cloud, the off-axis weight function f(α) of the point cloud quality of the point P is defined by calculating the angle α between the straight line connecting the optical center of the structured light camera and the point P and the optical axis of the camera:

[0039] Among them: α1 is the maximum value of angle α, k α is the decay exponent of the off-axis weight function f(α), k α The larger the value is defined, the faster the off-axis weight function f(α) decays as α increases;

[0040] At the same time, the direction vector of the camera optical axis is z, and the local normal vector of point P in the point cloud is n. Then the angle between the camera optical axis and the normal vector n is β, and the point cloud quality tilt weight function f(β) of the point P is defined by the angle β: Where: β1 is the maximum value of angle β, k β is the decay exponent of the tilt weight function f(β), k β The larger the value is defined, the faster the tilt weight function f(β) decays as β increases;

[0041] Then, the weighted ICP algorithm is used to precisely align the point clouds of adjacent frames to optimize the poses of adjacent frames, and to achieve accurate splicing of the workpiece upper surface point clouds of multiple frames with different perspectives, so as to obtain a complete workpiece upper surface point cloud including the welding groove; the specific steps of the weighted ICP algorithm are as follows:

[0042] (5-1) The rough registration result of step (4) is defined as the initial transformation matrix T0;

[0043] (5-2) Corresponding point matching: The point clouds collected from each frame constitute a multi-frame point cloud set Q = {P1, P2, ..., P k ,…,P m-1 ,P m}, m is the number of point clouds, which is equal to the number of frames of the original point cloud of the workpiece to be welded collected by the structured light camera. Take one of the point clouds P in Q k is the source point cloud, k is initialized to 2, for the source point cloud P k Each point in Find its nearest neighbor in the target point cloud P1

[0044] (5-3) Use Euclidean distance to measure the point pair error e i : T is the transformation matrix during the optimization process;

[0045] (5-4) Calculate the weight value ω of the matching point pair i : Where: f i (α) is the source point cloud P k Points in The off-axis weight function value of the point cloud quality, f i (β) is the source point cloud P k Points in The point cloud quality tilt weight function value;

[0046] (5-5) By minimizing the weighted error, the transformation matrix T is iteratively optimized and the source point cloud P is transformed using the weighted least squares method. k The weighted error sum of the Euclidean distances between all points in and the corresponding nearest neighbor points in the target point cloud P1 is used as the optimization target:

[0047] Where: T * It represents the optimal solution of the transformation matrix T obtained by iterative optimization, that is, the optimal transformation matrix, and m represents the number of point clouds;

[0048] (5-6) When the number of iterations reaches the set maximum number of iterations or when the difference between the optimization variables is less than the threshold, the iteration is stopped and convergence is completed;

[0049] (5-7), using the obtained optimal transformation matrix T * The source point cloud P k Perform pose transformation and transform the source point cloud P k Combine with the target point cloud P1 to form a new point cloud, and perform point cloud downsampling to obtain a new point cloud Complete the update of the target point cloud P1; set k=k+1 to increase in sequence until k=m, and stop updating; the point cloud finally updated is the spliced ​​point cloud obtained by precise registration.

[0050] The protection scope of the present invention includes but is not limited to the above embodiments, and its protection scope is subject to the claims. Any replacement, deformation, and improvement of the technology that can be easily thought of by technicians in this field shall fall within the protection scope of the present invention.

Claims

1. A high-precision registration method for multi-view point clouds of welding workpieces, characterized in that The specific steps include: (1) Spot welding is first performed on the welding grooves of two workpieces to connect them together to obtain the workpieces to be welded; (2) Installing a structured light camera on the end arm of the robot, using the structured light camera to collect multiple frames of original point clouds of the workpiece to be welded at different perspectives, and removing the background point clouds of the workpiece to be welded from the multiple frames of original point clouds to obtain multiple frames of workpiece point clouds at different perspectives; (3) The side point cloud of the workpiece in the workpiece point cloud is divided into independent clusters by the NV-DBSCAN clustering algorithm, and the upper surface point cloud of the workpiece containing the welding groove is extracted from the clusters; wherein: The specific process of the NV-DBSCAN clustering algorithm is as follows: Calculate the point cloud to be clustered P = {p1, p2, p33p n The normal vector of each point in} constitutes the normal vector set N = {n1,n2,n33n n }, for point p in point cloud P i 、p j , if its corresponding normal vector n i 、n j Satisfies the relationship: Where: γ is the threshold for determining whether two points satisfy the normal vector consistency constraint, then the point p is defined i and p j Satisfy the normal vector consistency constraint and move point p i and p j Classify into the same cluster of point clouds; (4) Unify the upper surface point clouds of the workpiece at different perspectives in each frame into the robot base coordinate system, and perform coarse registration on the upper surface point clouds of the workpiece at different perspectives in each frame; (5) Calculate the mass weight of each point in the point cloud on the workpiece surface. The calculation method is: For a point P in the point cloud, the off-axis weight function f(α) of the point cloud quality of the point P is defined by calculating the angle α between the straight line connecting the optical center of the structured light camera and the point P and the optical axis of the camera: Among them: α1 is the maximum value of angle α, k α is the decay exponent of the off-axis weight function f(α), k α The larger the value is defined, the faster the off-axis weight function f(α) decays as α increases; At the same time, the direction vector of the camera optical axis is z, and the local normal vector of point P in the point cloud is n. Then the angle between the camera optical axis and the normal vector n is β, and the point cloud quality tilt weight function f(β) of the point P is defined by the angle β: Where: β1 is the maximum value of angle β, k β is the decay exponent of the tilt weight function f(β), k β The larger the value is defined, the faster the tilt weight function f(β) decays as β increases; Then, the point clouds of adjacent frames are precisely aligned through the weighted ICP algorithm to optimize the poses of adjacent frames, so as to achieve accurate stitching of the workpiece surface point clouds of multiple frames with different perspectives, and obtain a complete workpiece surface point cloud including the welding groove; wherein: the weight value of the matching point pair calculated in the weighted ICP algorithm is obtained by the off-axis weight function value of the point cloud quality and the tilt weight function value of the point cloud quality.

2. A high-precision registration method for multi-view point clouds of welding workpieces as claimed in claim 1, characterized in that: In the step (2), the background point cloud in the original point cloud is removed by a random sampling consistency algorithm.

3. A high-precision registration method for multi-view point clouds of welding workpieces as claimed in claim 1, characterized in that: In step (5), the specific steps of the weighted ICP algorithm are: (5-1) The rough registration result of step (4) is defined as the initial transformation matrix T0; (5-2) Corresponding point matching: The point clouds collected from each frame constitute a multi-frame point cloud set Q = {P1, P2, ..., P k ,…,P m-1 ,P m }, m is the number of point clouds, which is equal to the number of frames of the original point cloud of the workpiece to be welded collected by the structured light camera. Take one frame of point cloud P in Q k is the source point cloud, k is initialized to 2, for the source point cloud P k Each point in Find its nearest neighbor in the target point cloud P1 (5-3) Use Euclidean distance to measure the point pair error e i : T is the transformation matrix during the optimization process; (5-4) Calculate the weight value ω of the matching point pair i : Where: f i (α) is the source point cloud P k Points in The off-axis weight function value of the point cloud quality, f i (β) is the source point cloud P k Points in The point cloud quality tilt weight function value; (5-5) By minimizing the weighted error, the transformation matrix T is iteratively optimized and the source point cloud P is transformed using the weighted least squares method. k The weighted error sum of the Euclidean distances between all points in and the corresponding nearest neighbor points in the target point cloud P1 is used as the optimization target: Where: T * It represents the optimal solution of the transformation matrix T obtained by iterative optimization, that is, the optimal transformation matrix, and m represents the number of point clouds; (5-6) When the number of iterations reaches the set maximum number of iterations or when the difference between the optimization variables is less than the threshold, the iteration is stopped and convergence is completed; (5-7), using the obtained optimal transformation matrix T * The source point cloud P k Perform pose transformation and transform the source point cloud P k Combine with the target point cloud P1 to form a new point cloud, and perform point cloud downsampling to obtain a new point cloud Complete the update of target point cloud P1; Let k=k+1 and increase in sequence until k=m, then stop updating; the point cloud finally updated is the spliced ​​point cloud obtained by precise registration.

Images