An automated welding system and working method based on vision recognition packaging line technology

Abstract

The present invention provides an automated welding system and working method based on vision recognition of a packing line, which is used for welding and fixing the packing lines around the inner and outer sides of a steel coil. The welding system includes a pair of vertical circular ring structures connected by slide rails, and also includes a pair of manipulators for welding the packing lines. The inner side of the circular ring structure is adjacent to the outer side of the steel coil and the packing lines. A plurality of 3D cameras for collecting point cloud data of the steel coil and the packing lines are arranged on the inner side of the circular ring structure. The 3D cameras are connected to the control modules of the manipulators. The manipulators are located on the left and right sides of the steel coil. The present invention can effectively improve the packing efficiency and packing effect of the steel coil, reduce production costs, save resources, and ensure the personal safety of workers.

Description

Technical Field

[0001] The present invention relates to the technical field of steel production, and in particular to an automated welding system and working method based on a vision recognition packing line technology. Background Art

[0002] In the steel industry, due to the need for transportation, steel coils need to be packed. In the current packing tasks, the packing work of the packing line is mainly carried out manually. The manual packing work has low efficiency and the packing effect is not good enough. Some packing lines will break or even pop out due to insufficient fastening, which brings great trouble and danger to the railway transportation process. Therefore, in order to improve the packing fastening effect and ensure the personal safety of workers, it is possible to consider using robotic vision-based automatic welding to carry out the fastening work of the packing line. In recent years, with the continuous update of image processing technology and the continuous decline of hardware costs, some image detection methods have been applied to the identification of steel coils. For example, the identification of the head end of the steel coil and the identification of the steel coil generally use two-dimensional images for scene matching and frame difference for head or coil detection. In addition, using machine learning methods to detect and classify targets is a current research hotspot. In recent years, three-dimensional data acquisition technology has gradually developed and combined with neural networks to form a new three-dimensional feature detection branch. The three-dimensional data processing and detection technology based on deep learning has developed rapidly and become increasingly mature, providing technical support for target detection.

[0003] In this case, introducing an intelligent packing line fastening system based on machine vision technology for automatic welding can improve work efficiency, improve work effects, and reduce the negative impact of the working environment, ensure the packing effect, ensure the personal safety of workers, and ensure that the packing line will not break or pop out during transportation, which can effectively promote the development of the industry of fastening the packing line of steel coils. Summary of the Invention

[0004] The present invention provides an automated welding system and working method based on a vision recognition packing line technology, which can effectively improve the packing efficiency and packing effect of steel coils, reduce production costs, save resources, and ensure the personal safety of workers.

[0005] The present invention adopts the following technical solutions.

[0006] An automated welding system based on a vision recognition packing line technology is used to weld and fix the packing lines around the inner and outer sides of a steel coil. The welding system includes a pair of vertical circular ring structures connected by slide rails, and also includes a pair of manipulators for welding the packing lines; the inner sides of the circular ring structures are adjacent to the outer sides of the steel coil and the packing lines; a plurality of 3D cameras for collecting the point cloud data of the steel coil and the point cloud data of the packing lines are arranged on the inner sides of the circular ring structures; the 3D cameras are connected to the control module of the manipulator; the manipulators are located on the left and right sides of the steel coil.

[0007] Each steel coil is attached with four packing wires; four 3D cameras are provided inside the inner side of each ring structure; a groove for laying cables is provided inside the inner side of the ring structure; the cables laid in the groove are connected to the control modules of the 3D cameras and the manipulator.

[0008] The inner circumferential surface of the ring structure can be moved within a small range by the 3D cameras, so that the shooting directions of the four 3D cameras are respectively aligned with the four packing wires.

[0009] The slide rail is connected to the upper part of the ring structure; after the steel coil is attached with the packing wires, it is carried by a horizontal U-shaped hook to the welding operation area of the manipulator, and the U-shaped hook is used to maintain the shape of the steel coil when the 3D cameras of the ring structure collect data and the manipulator performs the welding operation.

[0010] A working method of an automatic welding system based on a vision recognition packing wire technology, adopting the automatic welding system based on the vision recognition packing wire technology described above, the 3D camera is an industrial 3D camera; the control module of the manipulator is a PC; the coordinate-related acquisition data of the industrial camera and the motion control data of the manipulator are unified in the same world coordinate system, and the method includes the following steps;

[0011] Step S1: The industrial 3D camera collects data. Specifically, in the workshop, the steel coil carried by the U-shaped hook reaches the designated position, the ring structure slides on the slide rail into the working position, the manipulators on both sides of the steel coil enter the welding working position, and the industrial camera inside the ring structure translates along with the ring structure on the slide rail to obtain the point cloud data within the required range of the steel coil.

[0012] Step S2: The original point cloud data collected by the industrial 3D camera is transmitted to the PC side for processing to identify the position information of the packing wire in space, and then determine the positions of the welding points on the packing wire.

[0013] Step S3: The PC transmits the specific coordinate values of the welding point positions obtained in Step S2 to the manipulator, and the manipulator pulls the welding wire to perform welding connection on the welding points. After completion, the robotic arm withdraws from the working position, and the steel coil and the packing wire are carried by the U-shaped hook to withdraw from the working position, and the work process is completed.

[0014] The said Step S2 includes the following steps;

[0015] Step A1: Filter and denoise the point cloud data, and perform filtering operation on the point cloud using a voxel filtering algorithm;

[0016] Step A2: Sample and compress the filtered point cloud data, using a uniform sampling and compression method;

[0017] Step A3: Preliminarily process the point cloud data, identify the steel coil point cloud and the packing line point cloud through the PointNet and PointNet++ neural network frameworks, and mark and display them with points of different colors.

[0018] Step A4: Project the point cloud onto the XOY plane to obtain the vertical top-down images of the steel coil and the packing line point clouds.

[0019] Step A5: Grayscale the two-dimensional packing line image in the vertical top-down image of the packing line point cloud obtained in Step A4, use binarization to extract the image contour edges, then use the Sobel operator to identify and extract specific lines in the image to obtain the position information of the packing line. Combine the depth information of the packing line position measured by the 3D camera to obtain the specific coordinate values of the solder joints in space.

[0020] In the point cloud filtering in Step A1, the point cloud sampling in Step A2, and the point cloud projection in Step A4, the VisualStudio2017 software is used.

[0021] In Step A3, a neural network algorithm is adopted, with Pycharm as the neural network architecture platform, on which the point cloud data set is retrieved, and the point clouds of the steel coil and the packing line are classified and segmented.

[0022] In the PointNet neural network framework, first, the features of the original point cloud data points are dimensionally increased through a fully connected layer, from the original 3 dimensions or 6 dimensions to 1024 dimensions. After passing through the max pooling layer, the 1024-dimensional features of each point are reduced to one-dimensional features, and then the features of each point are classified by category through a fully connected layer.

[0023] In the PointNet++ neural network, the extraction of local features is added. First, the key points of the point cloud features are extracted through farthest point sampling. Then, with the farthest point sampling point as the center, a circle with a specific radius is selected to extract the features of the circle, which is the extraction of the local features of the point cloud. After that, the features of the circles with different radii are spliced. Finally, the extraction of local features and global features is obtained through a fully connected layer to obtain more specific and representative features of the point cloud. When performing farthest point sampling, the FPS principle is adopted, specifically as follows:

[0024] Step B1: The input point cloud has N points. Select a point P0 from the point cloud as the starting point to obtain the sampling point set S = {P0}.

[0025] Step B2: Calculate the distances from all points to P0 to form an M-dimensional array L, and select the point corresponding to the maximum value in it as P1, and update the sampling point set S = {P0, P1}.

[0026] Step B3: Calculate the distances from all points to P1. For each point Pi, if its distance to P1 is less than L[i], then update L[i] = d(Pi, P1). Thus, the array L stores the nearest distances from each point to the sampling point set S.

[0027] Step B4: Select the point corresponding to the maximum value in L as P2, and update the sampling point set S = {P0, P1, P2}.

[0028] Step B5: Repeat steps B2 - B4 until the required number of target sampling points M is reached.

[0029] The voxel filtering in step A1 creates a three-dimensional voxel grid from the input point cloud data. Inside each voxel, the center of gravity of all points in the voxel is used to approximate other points in the voxel, so that all points in the voxel are finally represented by a single center of gravity point to achieve the purpose of downsampling.

[0030] The uniform sampling compression algorithm in step A2 creates a 3D voxel grid on the input point cloud data. Then, in each voxel, the point closest to the center of the voxel is used to replace all points in the voxel.

[0031] The point cloud projection in step A4 is to compress the point cloud projection from the direction of the Z-axis, compress all the point cloud into the XOY plane to form a two-dimensional image for the subsequent determination of the point cloud position recognition. The method for determining the point cloud position recognition is as follows:

[0032] Step B1: First, convert the two-dimensional image into a grayscale image, binarize the grayscale image to show the contour edges of the steel coil and the packing line. Then, use the Sobel operator to process the binarized image to identify the horizontal lines in the image, that is, the shape features of the packing line.

[0033] Step B2: By determining the position of the horizontal lines in the image and combining with the depth information of the packing line measured by the industrial camera, obtain the position information of the packing line in space.

[0034] Step B3: Determine the positions of the welding points on the packing line, and finally the manipulator performs automatic welding.

[0035] In step B2, in the aspect of horizontal line recognition in the image, the sobel operator in the y direction is used for calculation, that is, the Gy operator of sobel is used to eliminate the edge gradient in the horizontal direction to obtain the horizontal lines in the image.

[0036] In step B3, based on the packing line, walk a preset distance from the edge of the steel coil inward to obtain the positions of the points to be welded to determine the positions of the welding points on the packing line.

[0037] The system described in the present invention is based on machine vision processing technology. The 3D camera scans the point cloud data of the steel coil and the packing line, and then transmits it to the computer for data processing to obtain the shape and pose information of the packing line. The pose information is used for the welding work of the manipulator, making the entire work process automated.

[0038] The present invention aims to realize the automation of the welding of the steel coil packing line. Through innovation in structure and communication connection with the manipulator, an automated welding system for the packing line is realized.

[0039] In terms of structure, the present invention has two circular ring structures connected by sliding rails at both ends. The industrial 3D camera installed inside the circular ring collects the original data of the packing line. The circular ring translates on the sliding rail so that the industrial 3D camera can collect the point cloud data of the target part of the packing line. Then, wiring is carried out through the groove inside the circular ring to realize the connection between the industrial camera and the PC side. After the point cloud is processed at the PC side, the result is sent to the manipulator. Based on the unified coordinate system, the manipulator can automatically weld the packing line to complete the work process. In terms of the neural network structure, based on the PointNet and PointNet++ neural network models, the classification and segmentation parameters and the input value parameters of the data points are optimized for the subject, improving the classification and segmentation effect of PointNet++ for the packing line point cloud.

[0040] The automated welding system for the packing line realized based on the point cloud target recognition and positioning technology of the present invention can effectively improve the packing efficiency and packing effect of the steel coil, reduce production costs, save resources, ensure the personal safety of workers, and further improve the production efficiency of the enterprise on this basis, promoting the development of the industry. Economically, since it liberates human labor, it can save labor costs and the mechanical automation improves the productivity level, which can promote the economic development of the steel industry and achieve the maximization of benefits. Brief Description of the Drawings

[0041] The following further elaborates on the present invention in conjunction with the drawings and specific implementation manners:

[0042] Attached Figure 1 is a schematic diagram of the device of the solution described in the present invention;

[0043] Attached Figure 2 is a schematic photo of the steel coil with an additional packing line;

[0044] Attached Figure 3 is a schematic diagram of the working process of the system described in the present invention;

[0045] Attached Figure 4 is a schematic diagram of the PointNet neural network framework;

[0046] Attached Figure 5It is a schematic diagram of calculating using the Sobel operator in the y direction for horizontal line recognition in an image;

[0047] Appendix Figure 6 It is a schematic diagram of using the Gy operator of Sobel to eliminate the edge gradients in the horizontal direction to obtain the horizontal lines (horizontal packing lines) in the image;

[0048] In the figure: 1 - 3D camera; 2 - slide rail; 3 - packing line; 4 - circular ring structure; 5 - U-shaped hook; 6 - steel wire coil; 7 - manipulator. Specific implementation manner

[0049] As shown in the figure, an automated welding system based on vision recognition of packing line technology is used to weld and fix the packing lines 3 around the inner and outer sides of the steel wire coil 6. The welding system includes a pair of vertical circular ring structures 4 connected by a slide rail 2, and also includes a pair of manipulators for welding the packing lines; the inner side of the circular ring structure is adjacent to the outer side of the steel wire coil and the packing line; there are multiple 3D cameras 1 for collecting point cloud data of the steel wire coil and the packing line on the inner side of the circular ring structure; the 3D cameras are connected to the control module of the manipulator; the manipulators are located on the left and right sides of the steel wire coil.

[0050] Each steel wire coil is attached with four packing lines; there are four 3D cameras on the inner side of each circular ring structure; there are grooves for laying cables on the inner side of the circular ring structure; the cables laid in the grooves connect the 3D cameras and the control module of the manipulator.

[0051] The inner ring surface of the circular ring structure allows the 3D cameras to move within a small range, so that the shooting directions of the four 3D cameras are respectively aligned with the four packing lines.

[0052] The slide rail is connected to the upper part of the circular ring structure; after the steel wire coil is attached with the packing line, it is transported by a horizontal U-shaped hook to the welding operation area of the manipulator, and the U-shaped hook is used to maintain the shape of the steel wire coil when the 3D cameras of the circular ring structure collect data and the manipulator performs welding operations.

[0053] The working method of the automated welding system based on vision recognition of packing line technology uses the above-mentioned automated welding system based on vision recognition of packing line technology. The 3D cameras are industrial 3D cameras; the control module of the manipulator is a PC; the coordinate-related acquisition data of the industrial cameras and the motion control data of the manipulator are unified in the same world coordinate system. The method includes the following steps;

[0054] Step S1, the industrial 3D camera collects data, specifically: in the workshop, the steel coil carried by the U-shaped hook 5 arrives at the designated position, the circular ring structure slides on the slide rail to enter the working position, the manipulators on both sides of the steel coil enter the welding working position, and the industrial camera inside the circular ring structure translates on the slide rail with the circular ring structure to obtain point cloud data within the required range of the steel coil;

[0055] Step S2, the original point cloud data collected by the industrial 3D camera is transmitted to the PC for processing, the position information of the baling line in space is identified, and then the position of the welding point on the baling line is determined;

[0056] Step S3, the PC transmits the specific coordinate value of the welding point position obtained in step S2 to the robot arm, and the robot arm pulls the welding line to weld the welding point. After completion, the robot arm withdraws from the working position, and the steel coil and the baling line are carried out of the working position by the U-shaped hook, and the work process is completed.

[0057] The step S2 comprises the following steps:

[0058] Step A1, filtering and denoising the point cloud data, using a voxel filtering algorithm to perform a filtering operation on the point cloud;

[0059] Step A2, sampling and compressing the filtered point cloud data, using a uniform sampling compression method;

[0060] Step A3: preliminarily process the point cloud data, identify the steel coil point cloud and the baling wire point cloud through the PointNet and PointNet++ neural network frameworks, and mark and display them with different color points;

[0061] Step A4, projecting the point cloud onto the XOY plane to obtain a vertical top view image of the point cloud of the steel coil and the baling wire;

[0062] Step A5, graying the two-dimensional baling line image in the vertical overhead image of the baling line point cloud obtained in step A4, using binarization to extract the image contour edge, and then using the Sobel operator to identify and extract specific straight lines from the image to obtain the position information of the baling line, combined with the baling line position depth information measured by the 3D camera, to obtain the specific coordinate value of the welding point in space.

[0063] In the point cloud filtering of step A1, the point cloud sampling of step A2, and the point cloud projection of step A4, Visual Studio 2017 software is used.

[0064] In step A3, a neural network algorithm is used, and Pycharm is used as a neural network architecture platform to retrieve the point cloud data set, and classify and segment the point clouds of the steel coil and the baling wire;

[0065] In the PointNet neural network framework, first, the features of the original point cloud data points are dimensionally increased through a fully connected layer, from the original 3 or 6 dimensions to 1024 dimensions. After passing through the max pooling layer, the 1024-dimensional features of each point are reduced to a one-dimensional feature. Then, the features of each point are classified into categories through a fully connected layer.

[0066] In PointNet++, the extraction of local features is added. First, the feature key points of the point cloud are extracted through farthest point sampling. Then, with the farthest point sampling point as the center, a circle with a specific radius is selected to extract the features of the circle, which is the extraction of the local features of the point cloud. After that, the features of circles with different radii are spliced. Finally, the extraction of local features and global features is obtained through a fully connected layer to obtain more specific and representative features of the point cloud. When performing farthest point sampling, the FPS principle is adopted, specifically as follows:

[0067] Step B1: The input point cloud has N points. Select a point P0 from the point cloud as the starting point, and obtain the sampling point set S = {P0}.

[0068] Step B2: Calculate the distances from all points to P0, form an M-dimensional array L, and select the point corresponding to the maximum value in it as P1. Update the sampling point set S = {P0, P1}.

[0069] Step B3: Calculate the distances from all points to P1. For each point Pi, if its distance to P1 is less than L[i], then update L[i] = d(Pi, P1). Therefore, the array L stores the closest distance from each point to the sampling point set S.

[0070] Step B4: Select the point corresponding to the maximum value in L as P2, and update the sampling point set S = {P0, P1, P2}.

[0071] Step B5: Repeat steps B2 - B4 until the required number of target sampling points M is reached.

[0072] For the voxel filtering in step A1, a three-dimensional voxel grid is created from the input point cloud data. Inside each voxel, the center of gravity of all points in the voxel is used to approximately represent other points in the voxel. In this way, all points in the voxel are finally represented by a center of gravity point to achieve the purpose of downsampling.

[0073] For the uniform sampling compression algorithm in step A2, a 3D voxel grid is created on the input point cloud data. Then, in each voxel, the point closest to the center of the voxel is used to replace all points in the voxel.

[0074] The point cloud projection in step A4 is to perform point cloud compression projection from the direction of the Z-axis, compress all the point clouds into the XOY plane to form a two-dimensional image for determining the position recognition of the point cloud later. The method is as follows:

[0075] In step B1, first convert the two-dimensional image into a grayscale image, binarize the grayscale image to show the contour edges of the steel coil and the packing line. Then use the Sobel operator to process the binarized image to identify the horizontal lines in the image, which are the shape features of the packing line;

[0076] In step B2, by determining the position of the horizontal line in the image and combining with the depth information of the packing line measured by the industrial camera, obtain the position information of the packing line in space;

[0077] In step B3, determine the positions of the welding points on the packing line, and finally the manipulator performs automatic welding.

[0078] In step B2, in terms of horizontal line recognition in the image, calculate using the Sobel operator in the y direction, that is, use the Gy operator of Sobel to eliminate the edge gradients in the horizontal direction to obtain the horizontal lines in the image;

[0079] In step B3, with the packing line as the reference, walk a preset distance from the edge of the steel coil inward to obtain the positions of the points to be welded to determine the positions of the welding points on the packing line.

Claims

1. An automated welding system based on vision recognition packing line technology, which is used for welding and fixing the packing lines around the inner and outer sides of a steel coil, and is characterized in that: The welding system includes a pair of vertical circular ring structures connected by slide rails, and also includes a pair of manipulators for welding the packing wires; the inner sides of the circular ring structures are adjacent to the outer sides of the steel coils and the packing wires; multiple 3D cameras for collecting the point cloud data of the steel coils and the packing wires are arranged on the inner sides of the circular ring structures; the 3D cameras are connected to the control module of the manipulator; the manipulators are located on the left and right sides of the steel coils.

2. The automated welding system based on the visual recognition packing line technology according to claim 1, characterized in that: Four packing wires are attached to each steel coil; four 3D cameras are arranged on the inner side of each circular ring structure; grooves for laying cables are arranged on the inner side of the circular ring structure; the cables laid in the grooves connect the 3D cameras and the control module of the manipulator.

3. The automated welding system based on the visual recognition packing line technology according to claim 2, wherein: The inner ring surface of the circular ring structure allows the 3D cameras to move within a small range, so that the shooting directions of the four 3D cameras are respectively aligned with the four packing wires.

4. An automated welding system based on a vision recognition packing line technology according to claim 1, characterized in that: The slide rails are connected to the upper parts of the circular ring structures; after the packing wires are attached to the steel coils, the steel coils are transported to the welding operation area of the manipulator by a horizontal U-shaped hook, and the U-shaped hook is used to maintain the shape of the steel coils when the 3D cameras of the circular ring structures collect data and the manipulators perform welding operations.

5. The working method of the automated welding system based on the visual recognition packing line technology, using the automated welding system based on the visual recognition packing line technology described in claim 4, is characterized in that: The 3D cameras are industrial 3D cameras; the control module of the manipulator is a PC; the coordinate-related acquisition data of the industrial cameras and the motion control data of the manipulators are unified in the same world coordinate system. The method includes the following steps; Step S1: The industrial 3D cameras collect data. Specifically, in the workshop, the steel coils transported by the U-shaped hook reach the designated position, the circular ring structures slide on the slide rails into the working position, the manipulators on both sides of the steel coils enter the welding working position, and the industrial cameras on the inner sides of the circular ring structures translate along with the circular ring structures on the slide rails to obtain the point cloud data within the required range of the steel coils. Step S2: The original point cloud data collected by the industrial 3D cameras is transmitted to the PC side for processing to identify the position information of the packing wires in space, and then the positions of the welding points on the packing wires are determined. Step S3: The PC transmits the specific coordinate values of the welding point positions obtained in Step S2 to the manipulator. The manipulator pulls the welding wire to perform welding connection on the welding points. After completion, the robotic arm withdraws from the working position, and the steel coils and the packing wires are transported out of the working position by the U-shaped hook, and the work process is completed.

6. The working method of the automated welding system based on the visual recognition packing line technology according to claim 5, characterized in that: The said Step S2 includes the following steps; Step A1: Filter and denoise the point cloud data, and perform filtering operation on the point cloud using the voxel filtering algorithm. Step A2: Sample and compress the filtered point cloud data using the uniform sampling compression method. Step A3: Preliminarily process the point cloud data, identify the point clouds of the steel coils and the packing wires through the PointNet and PointNet++ neural network frameworks, and mark and display them with points of different colors. Step A4: Project the point cloud onto the XOY plane to obtain the vertical top view images of the point clouds of the steel coils and the packing wires. Step A5: Grayscale the two-dimensional packing line image in the vertical top-down image of the packed line point cloud obtained in Step A4, use binarization to extract the image contour edges, then use the Sobel operator to identify and extract specific lines in the image to obtain the position information of the packing line, and combine the depth information of the packing line position measured by the 3D camera to obtain the specific coordinate values of the solder joints in space.

7. The working method of the automated welding system based on the vision recognition packing line technology according to claim 6, characterized in that: In the point cloud filtering in Step A1, the point cloud sampling in Step A2, and the point cloud projection in Step A4, the VisualStudio 2017 software is used.

8. The working method of the automated welding system based on the visual recognition packaging line technology according to claim 6, characterized in that: In Step A3, a neural network algorithm is adopted, with Pycharm as the neural network architecture platform, where the point cloud data set is retrieved, and the point clouds of the steel coil and the packing line are classified and segmented. In the PointNet neural network framework, first, the features of the original point cloud data points are dimensionally increased through a fully connected layer, from the original 3 dimensions or 6 dimensions to 1024 dimensions. After passing through the max pooling layer, the 1024-dimensional features of each point are reduced to one-dimensional features, and then the features of each point are classified by a fully connected layer. In the PointNet++ neural network, the extraction of local features is increased. First, the key points of the point cloud features are extracted through farthest point sampling. Then, with the farthest point sampling point as the center, a circle with a specific radius is selected to extract the features of the circle, which is the extraction of the local features of the point cloud. After that, the features of circles with different radii are spliced, and finally, the extraction of local features and global features is obtained through a fully connected layer to obtain more specific and representative features of the point cloud. When performing farthest point sampling, the FPS principle is adopted. Specifically: Step B1: The input point cloud has N points. Select a point P0 from the point cloud as the starting point to obtain the sampling point set S = {P0}. Step B2: Calculate the distances from all points to P0 to form an M-dimensional array L, and select the point corresponding to the maximum value in it as P1, and update the sampling point set S = {P0, P1}. Step B3: Calculate the distances from all points to P1. For each point P i , if its distance to P1 is less than L[i], then update L[i] = d(P i , P1). Therefore, the array L stores the shortest distance from each point to the sampling point set S; Step B4: Select the point corresponding to the maximum value in L as P2, and update the sampling point set S = {P0, P1, P2}. Step B5: Repeat Steps B2 - B4 until the required number of target sampling points M is reached. The voxel filtering in Step A1 creates a three-dimensional voxel grid through the input point cloud data. Inside each voxel, the center of gravity of all points in the voxel is used to approximately represent other points in the voxel, so that all points in the voxel are finally represented by a center of gravity point to achieve the purpose of downsampling. The uniform sampling compression algorithm in Step A2 creates a 3D voxel grid on the input point cloud data. Then, in each voxel, the point closest to the center of the voxel is used to replace all points in the voxel.

9. The working method of the automated welding system based on the visual recognition packing line technology according to claim 8, characterized in that: The point cloud projection in Step A4 is to compress and project the point cloud from the direction of the Z axis, compress all the point clouds into the XOY plane to form a two-dimensional image for the subsequent determination of the point cloud position identification. The method for determining the point cloud position identification is as follows: Step B1: First, convert the two-dimensional image into a grayscale image, binarize the grayscale image to reveal the contour edges of the steel coil and the packing line. Then, use the Sobel operator to process the binarized image to identify the horizontal lines in the image, which are the shape features of the packing line. Step B2: By determining the position of the horizontal lines in the image and combining with the depth information of the packing line measured by the industrial camera, obtain the position information of the packing line in space. Step B3: Determine the positions of the solder joints on the packing line, and finally, the manipulator performs automatic welding.

10. The working method of the automated welding system based on the visual recognition packing line technology according to claim 9, characterized in that: In Step B2, in terms of identifying the horizontal lines in the image, the sobel operator in the y direction is used for calculation, that is, the Gy operator of sobel is used to eliminate the edge gradients in the horizontal direction to obtain the horizontal lines in the image. In Step B3, taking the packing line as the reference, walk a preset distance from the edge of the steel coil inward to obtain the positions of the points to be welded, so as to determine the positions of the solder joints on the packing line.

Images