Mechanical arm-based automatic installation method for curtain wall

Abstract

The application discloses a curtain wall automatic installation method based on a mechanical arm, and relates to the technical field of automatic installation.The method comprises the following steps: step S1, scanning a building facade and constructing a dense point cloud model of an installation surface through an RGB-D camera; step S2, extracting the center point coordinates and normal vector of a curtain wall keel supporting surface; step S3, calculating the target pose of a curtain wall panel to be installed according to the supporting surface parameters; step S4, describing the joint movement of the mechanical arm by using a six-degree polynomial, and optimizing the trajectory parameters by using a genetic algorithm; and step S5, calculating the pose deviation in real time during the movement of the mechanical arm, and dynamically adjusting the pose of an end effector.The method constructs an automatic installation method through mechanical arm path planning and modeling of the installation surface, so that the installation precision is high, the construction period is shortened, the cost is reduced, and the danger of manual installation is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automated installation technology, in particular to a curtain wall automated installation method based on a mechanical arm. BACKGROUND

[0002] For aesthetic and energy-saving reasons, most high-rise buildings are installed with curtain walls, of which glass curtain walls are the most common. The size of the curtain wall of a high-rise building is mostly between 0.6m x 0.6m and 0.9m x 1.2m, and the size of some specific curtain walls can be even larger, and the weight can also be larger. It is very dangerous and difficult to install curtain walls with large size and weight on high-rise buildings. Due to the lack of special curtain wall installation equipment at present, the installation of curtain walls mainly relies on simple hoisting equipment and the construction of scaffolding for installation, and a large number of manual labor is often required during installation. As an important part of modern buildings, the installation quality and efficiency of building curtain walls have an important influence on the entire construction project. Intelligent construction technology, with its characteristics of high efficiency, precision, safety, etc., is gradually changing the traditional installation method of building curtain walls. By introducing advanced technical means and equipment, the construction unit realizes the automation, informatization and intelligentization of the curtain wall installation process, so as to improve the engineering quality, shorten the construction period and reduce the construction cost.

[0003] Intelligent positioning and assembly for on-site installation—For the on-site installation of building curtain walls, the importance of intelligent positioning and assembly is self-evident. Using high-precision positioning systems and advanced laser scanning technology, curtain wall panels are precisely positioned at the predetermined position, with the error controlled within a very small range. Intelligent hoisting equipment automatically adjusts the force according to the positioning information, smoothly and accurately hoists the curtain wall panels and installs them in place.

[0004] A curtain wall installation operation monitoring and guiding system and method are disclosed in Chinese Patent No. CN119494466A. Real-time position installation data, actual material installation data and real-time temperature data of the curtain wall installation process are collected; the preset position installation data, material installation data and preset temperature data are input into a material position monitoring model to obtain preset force interaction simulation data of the curtain wall installation operation material position; the real-time position installation data, actual material installation data and real-time temperature data are input into the material position monitoring model to obtain real-time force interaction simulation data of the curtain wall installation operation material position; the preset force interaction simulation data and real-time force interaction simulation data are analyzed to obtain an analysis result of the curtain wall installation and issue a warning signal, and the workers adjust the curtain wall installation according to the analysis result, which can monitor the changes in force and installation direction in real time during the curtain wall installation process, and can avoid curtain wall installation errors that do not meet the curtain wall installation specifications in advance.

[0005] A curtain wall installation control system and method based on visual positioning of a mechanical arm are disclosed in Chinese Patent No. CN118752472B. The system includes a seven-degree-of-freedom mechanical arm, a distance measuring system, an AGV cart, and a visual control system. The distance measuring system is used to obtain a first distance and a second distance and send them to the visual control system. The visual control system is used to determine mechanical arm adjustment information based on the first distance and control the seven-degree-of-freedom mechanical arm to adjust its pose based on the mechanical arm adjustment information. After the pose adjustment of the seven-degree-of-freedom mechanical arm is completed, the second distance is used to determine cart adjustment information, which is sent to the AGV cart. The AGV cart is used to adjust the pose of the AGV cart based on the cart adjustment information. The visual control system is used to control the seven-degree-of-freedom mechanical arm to install the glass curtain wall into the installation frame of the glass curtain wall after the pose adjustment of the AGV cart is completed. This application can improve work efficiency, ensure the safety of construction personnel, and ensure the stability of construction equipment.

[0006] The above patents all have the problems raised in the background art: positioning relies on manual calibration, control lines need to be marked on the building structure in advance, and is greatly affected by measurement errors and the environment; the safety of mechanical arm trajectory planning is not considered, and collision with curtain wall keels or installed panels is easy; the installation process lacks dynamic compensation of the pose, leading to misalignment of the panel and the keel. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a curtain wall automatic installation method based on a mechanical arm to overcome the shortcomings of the prior art.

[0008] To achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0009] The curtain wall automatic installation method based on a mechanical arm comprises the following steps:

[0010] Step S1, scanning the building facade by an RGB-D camera and constructing a dense point cloud model of the installation surface;

[0011] Step S2, extracting the center point coordinates and normal vector of the curtain wall keel support surface;

[0012] Step S3, calculating the target pose of the curtain wall panel to be installed according to the support surface parameters;

[0013] Step S4, using a six-degree polynomial to describe the joint motion of the mechanical arm, and optimizing the trajectory parameters by a genetic algorithm;

[0014] Step S5, calculating the pose deviation in real time during the motion of the mechanical arm and dynamically adjusting the pose of the end effector.

[0015] Further, in step S1, the dense point cloud model of the installation surface is constructed, which specifically comprises the following steps:

[0016] Step S1.1, initialize the voxel grid and pre-process the image captured by the RGB-D camera;

[0017] Step S1.2, convert the pixel coordinates and depth values into 3D points in the camera coordinate system through the camera intrinsic matrix;

[0018] Step S1.3, dynamically adjust the voxel based on the curvature;

[0019] Step S1.4, calculate the signed distance function value of each voxel x along the camera optical center ray to the surface;

[0020] Step S1.5, introduce a truncation distance to calculate the truncated signed distance function value of each voxel x;

[0021] Step S1.6, calculate the weight value of each voxel x;

[0022] Step S1.7, fuse the current frame data into the global truncated signed distance function field;

[0023] Step S1.8, generate a triangular mesh when all depth maps are fused to complete the construction of the dense point cloud model of the installation surface.

[0024] Further, in the step S1.3, the voxel is dynamically adjusted based on the curvature, which specifically includes the following steps:

[0025] Calculate the covariance matrix of the neighborhood point set of the point p, which is equal to the sum of the product of the difference between the coordinates of each point in the neighborhood point set and the center coordinates of the neighborhood point set, divided by the number of points in the neighborhood point set;

[0026] Calculate the curvature based on the covariance matrix, which is equal to the minimum eigenvalue of the covariance matrix divided by the sum of all eigenvalues;

[0027] Classify the curvature to update the voxel size, when the curvature is greater than 0.02, the voxel size is set to 1mm, otherwise the voxel size is set to 5mm.

[0028] Further, in the step S1.5, the TSDF value of each voxel x is calculated, and the specific formula is:

[0029]

[0030] wherein, represents the truncated signed distance function value of the voxel x, represents the signed distance function value of the voxel x, represents the truncation distance.

[0031] Further, in the step S1.6, a normal vector direction constraint is introduced to calculate the weight value of each voxel x, which is equal to the cosine value of the angle between the surface normal vector and the line of sight direction divided by the square of the Euclidean norm of the camera coordinate system coordinate of the voxel, wherein the angle between the surface normal vector and the line of sight direction refers to the angle between the normal vector at the image point and the direction vector from the camera optical center to the center point of the voxel.

[0032] Further, the step S2 specifically comprises the following steps:

[0033] Step S2.1, fitting an initial plane equation using a RANSAC coarse segmentation algorithm to segment the curtain keel support surface;

[0034] Step S2.2, calculating the center point coordinate of the support surface after RANSAC coarse segmentation, which is equal to the weighted average value of the three-dimensional coordinates of all points in the support surface point cloud, wherein the weight coefficient is the contribution degree of each point;

[0035] Step S2.3, calculating the spatial distribution covariance of the point cloud relative to the center point, which is equal to the sum of the products of each point and the center point difference;

[0036] Step S2.4, solving the eigenvector corresponding to the minimum eigenvalue of the covariance matrix, which is the support surface normal vector.

[0037] Further, in the step S3, according to the support surface parameters, the target pose of the to-be-installed curtain panel is solved, specifically comprising:

[0038] Calculating the to-be-installed curtain panel coordinate, which is equal to the support surface center point coordinate plus the product of the installation gap and the normal vector, and then plus the offset of the panel on the support surface tangent plane;

[0039] Calculating the to-be-installed curtain panel attitude rotation matrix, which is composed of a unit matrix and a rotation matrix around the rotation axis, wherein the rotation axis is obtained by the cross product of the z-axis and the target normal vector, and the rotation angle is determined by the angle between the z-axis and the target normal vector.

[0040] Further, in the step S4, a sixth order polynomial is used to describe the joint motion of the mechanical arm, and the specific formula is:

[0041]

[0042] wherein, represents the joint motion trajectory of the mechanical arm, represents time, , , , , , and respectively represent constant term coefficient, first order term coefficient, second order term coefficient, third order term coefficient, fourth order term coefficient, fifth order term coefficient and sixth order term coefficient;

[0043] The angles, angular velocities and angular accelerations of each joint of the mechanical arm at the initial time and the terminal time are taken as boundary conditions into Solving obtains , , , , and ;

[0044] The sixth order term coefficient is solved by genetic algorithm for optimization;

[0045] According to the solved , , , , , and , the motion trajectory with continuous joint angle, angular velocity and angular acceleration is planned.

[0046] Further, the sixth order term coefficient is solved by genetic algorithm for optimization, specifically including the following steps:

[0047] An optimization function is designed, and the function value is the trajectory length weight multiplied by the total length of the path of the end effector, plus the joint rotation weight multiplied by the sum of the absolute values of the changes of all joint angles, plus the safety threshold weight multiplied by the maximum deviation of the actual distance from the safety threshold;

[0048] Randomly generate 200 sets of sixth order term coefficients;

[0049] Calculate the optimization function value of the trajectory corresponding to each sixth order term coefficient, and select the top twenty individuals in the inverse function value order as parents;

[0050] Generate a new population by crossing and mutating, and iterate until convergence;

[0051] The optimal sixth order term coefficient is obtained after optimization, and the optimal trajectory is generated by substituting the sixth order polynomial.

[0052] Further, the step S5 specifically includes the following steps:

[0053] According to the actual pose and the target pose of the end of the mechanical arm, the pose compensation amount is obtained, which is equal to the target pose minus the actual pose;

[0054] The rotation matrix in the pose compensation amount is converted into an axis-angle representation to obtain a rotation vector;

[0055] The angle compensation quantity of the mechanical arm joint is solved by a Jacobian pseudo-inverse method, and the compensation quantity is equal to the pseudo-inverse of the Jacobian matrix multiplied by the pose compensation quantity.

[0056] Compared with the prior art, the beneficial effects of the present application are as follows:

[0057] 1、The method realizes the automatic curtain wall installation based on the mechanical arm through the vision-planning-control closed loop, and improves the curtain wall installation precision and safety.

[0058] 2、The present application constructs the building facade dense point cloud by RGB-D point cloud and plane segmentation algorithm, accurately segments the curtain wall support surface, introduces dynamic voxel downsampling, adaptively adjusts the voxel size according to the point cloud curvature, and reduces 70% of the calculation amount; when calculating the voxel weight, the normal vector direction constraint is introduced, which strengthens the contribution of front observation and weakens the contribution of side observation.

[0059] 3、The present application plans the joint space of the mechanical arm trajectory by a six-order polynomial, obtains the continuous trajectory of the angular velocity and angular acceleration of each joint of the mechanical arm, ensures the continuous acceleration (jerk without sudden change), avoids the shaking of the mechanical arm, and provides an additional optimization degree of freedom compared with a quintic polynomial.

[0060] 4、The present application designs an optimization function, minimizes the optimization function by a genetic algorithm, finds the optimal parameters that minimize the optimization function value, and finally brings the optimal parameters into the six-order polynomial to obtain the optimal motion trajectory. BRIEF DESCRIPTION OF DRAWINGS

[0061] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments made with reference to the accompanying drawings:

[0062] Figure 1 The flowchart of the embodiment of the present application is shown in the figure;

[0063] Figure 2 The flowchart of the construction of the point cloud model of the embodiment of the present application is shown in the figure;

[0064] Figure 3 The six-order polynomial trajectory function graph of the embodiment of the present application is shown in the figure;

[0065] Figure 4 The six-order polynomial coefficient evolution graph of the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is described in detail below with reference to the drawings and specific embodiments.

[0067] As Figure 1As shown, the mechanical arm-based curtain wall automatic installation method comprises the following steps:

[0068] Step S1, scanning the building facade through an RGB-D camera and constructing a dense point cloud model of the installation surface;

[0069] Step S2, extracting the center point coordinates and normal vector of the curtain wall keel support surface;

[0070] Step S3, calculating the target pose of the curtain wall panel to be installed according to the support surface parameters;

[0071] Step S4, describing the joint motion of the mechanical arm with a sixth-order polynomial and optimizing the trajectory parameters through a genetic algorithm;

[0072] Step S5, calculating the pose deviation in real time during the motion of the mechanical arm and dynamically adjusting the pose of the end effector.

[0073] In the step S1, the dense point cloud model of the installation surface is constructed, specifically comprising the following steps:

[0074] Step S1.1, initializing the voxel grid and pre-processing the image collected by the RGB-D camera;

[0075] Step S1.2, converting the pixel coordinates (u, v) and the depth value d into 3D points in the camera coordinate system through the camera intrinsic matrix K;

[0076] Step S1.3, dynamically adjusting the voxel based on curvature calculation;

[0077] Step S1.4, calculating the SDF value of each voxel x along the camera optical center ray to the surface;

[0078] Step S1.5, introducing a truncation distance to calculate the TSDF value of each voxel x;

[0079] Step S1.6, calculating the weight value of each voxel x;

[0080] Step S1.7, fusing the current frame data into the global TSDF field;

[0081] Step S1.8, generating a triangular mesh after all depth maps are fused to complete the construction of the dense point cloud model of the installation surface.

[0082] TSDF stands for Truncated Signed Distance Function, which is translated as truncated signed distance function. Usually, a three-dimensional space to be modeled is first selected, and then the three-dimensional space is divided into many small blocks, with a resolution of 256x256x256 or 128x128x128. Each small block is called a voxel.

[0083] Each voxel in the TSDF model stores the distance of the small block to the nearest object surface. If the small block is in front of the object surface, it stores a positive value; if the small block is behind the object surface, it stores a negative value. Further, it is generally considered that the object surface has a thickness, so the values that are too large and too small are set to 1 or -1, so that the distance after truncation is obtained, that is, the so-called TSDF model. Finally, according to the definition, the place where the TSDF is 0 is the place where the reconstructed surface is located.

[0084] In the step S1.3, the voxel is dynamically adjusted based on the curvature calculation, and specifically includes the following steps:

[0085] The covariance matrix of the neighborhood point set of the point p is calculated, and the calculation formula is:

[0086]

[0087] wherein, the covariance matrix is represented by, the number of points in the neighborhood point set is represented by, the neighborhood point coordinates are represented by, the center coordinates of the neighborhood point set are represented by;

[0088] The curvature is calculated based on the covariance matrix, and the calculation formula is:

[0089]

[0090] wherein, the curvature is represented by, , and the eigenvalues of the covariance matrix are represented by;

[0091] The curvature classification is updated to the voxel size, and the calculation formula is:

[0092]

[0093] wherein, the voxel size is represented by, and the unit is represented by.

[0094] In the step S1.5, the TSDF value of each voxel x is calculated, and the specific formula is:

[0095]

[0096] wherein, the TSDF value of the voxel x is represented by, the SDF value of the voxel x is represented by, the truncation distance is represented by, ;

[0097] In the step S1.6, the normal vector direction constraint is introduced to calculate the weight value of each voxel x, and the specific formula is:

[0098]

[0099] wherein, represents the weight value of each voxel x, represents the included angle between the surface normal vector and the line-of-sight direction, represents the camera coordinate system coordinate of the voxel, represents the 3D coordinates of the voxel center point in the world coordinate system, represents the position of the camera optical center in the world coordinate system, represents the normal vector at the image point (u, v), represents the included angle between two vectors, represents the square of the L2 norm.

[0100] The step S2 specifically includes the following steps:

[0101] Step S2.1, randomly sampling a point set to fit an initial plane equation, and using a RANSAC coarse segmentation algorithm to segment the curtain keel support surface;

[0102] RANSAC is mainly used to process data sets containing noise and outliers. It finds the best fitting model through random sampling and iteration, ignoring the influence of outliers, so it performs well in the case of a large amount of noise or a small number of outliers in the data. Steps: RANSAC selects an inner point sample randomly to fit the model, and judges the outliers by calculating the error between the inner points and the model. It iteratively selects the model with the maximum number of inner points as the final fitting model.

[0103] Step S2.2, calculating the center point coordinates of the support surface after RANSAC coarse segmentation, and the calculation formula is:

[0104]

[0105] wherein, represents the center point coordinates of the support surface, represents the weight coefficient of the kth point, represents the three-dimensional coordinates of the kth point, represents the number of support surface point clouds;

[0106] Weight determined by point cloud density and curvature, the weight increases in high density / high curvature area:

[0107] Point density: the number of point clouds per unit volume, reflecting the scanning quality, and high density areas usually correspond to clear structure surfaces;

[0108] Curvature effect: When the curvature is large (such as the edge of a bolt hole). Increase, ensure It should be close to the actual geometric center of the structure to avoid deflecting the centroid in planar areas.

[0109] Step S2.3: Calculate the spatial distribution covariance of the point cloud relative to the center point. The calculation formula is as follows:

[0110]

[0111] in, This represents the spatial distribution covariance matrix of the point cloud relative to its center point.

[0112] Step S2.4: Solve for the covariance matrix The eigenvector corresponding to the smallest eigenvalue is the normal vector of the support surface. .

[0113] make sure The direction pointing to the curtain wall panel installation direction is reversed when the angle with the direction of gravity is less than 85°.

[0114] In step S3, the target pose of the curtain wall panel to be installed is calculated based on the support surface parameters, specifically including:

[0115] The specific formula for calculating the coordinates of the curtain wall panel to be installed is as follows:

[0116]

[0117] in, Indicates the coordinates of the curtain wall panel to be installed. Represents the normal vector of the support surface. This indicates the installation gap between the curtain wall panel to be installed and the keel;

[0118] The formula for calculating the rotation matrix of the curtain wall panel to be installed is as follows:

[0119]

[0120] in, This represents the rotation matrix indicating the attitude of the curtain wall panel to be installed. Represents the identity matrix. Indicates the z-axis and The included angle, This represents the cross product matrix of the rotation axes.

[0121] In step S4, a sixth-order polynomial is used to describe the joint motion of the robotic arm. The specific formula is as follows:

[0122]

[0123] wherein, represents the trajectory of the joint movement of the mechanical arm, represents time, 、 、 、 、 、 and represent constant term coefficient, first term coefficient, second term coefficient, third term coefficient, fourth term coefficient, fifth term coefficient and sixth term coefficient, respectively;

[0124] The angles, angular velocities and angular accelerations of each joint of the mechanical arm at the initial time and the terminal time are taken as boundary conditions and brought into solving to obtain 、 、 、 、 and ;

[0125] The sixth term coefficient is solved by genetic algorithm for optimization;

[0126] According to the solved 、 、 、 、 、 and , the motion trajectory with continuous joint angles, angular velocities and angular accelerations is planned.

[0127] The sixth term coefficient is solved by genetic algorithm for optimization, specifically including the following steps:

[0128] An optimization function is designed, and the specific formula is:

[0129]

[0130] wherein, represents the optimization function value, represents the total length of the path of the end effector, represents the sum of the absolute values of the changes of all joint angles, represents the maximum deviation of the actual distance from the safety threshold , 、 and represent the weight coefficients of the trajectory length, joint rotation amount and safety threshold, respectively;

[0131] As shown in Table 1, the weight coefficient setting basis is shown:

[0132] Table 1:

[0133]

[0134] Safety threshold Need to be greater than the mechanical arm positioning error (usually 2~5mm), each physical quantity dimension is different, need to be normalized to the same order of magnitude: the normalized term = the maximum value - the minimum value original value - the minimum value, avoid a certain item factor value is too large and dominant function:

[0135] Trajectory length weight: shorten the path → reduce the motion time and energy consumption;

[0136] Joint rotation weight: reduce the total sum of each joint rotation → reduce motor load and mechanical wear;

[0137] Safety threshold weight: maximize the minimum gap → avoid collision risk.

[0138] Randomly generate 200 sets of six-term coefficients;

[0139] Calculate the F value of the trajectory corresponding to each six-term coefficient, and select the top 20 individuals in the inverse order as the parent;

[0140] Generate a new population by crossover and mutation, and iterate until convergence, with the convergence condition being that the F value changes by less than 10 for 20 consecutive generations -5 .

[0141] The optimal six-term coefficient is obtained after optimization, and the optimal trajectory is generated by substituting the six-term polynomial.

[0142] The step S5 specifically includes the following steps:

[0143] According to the actual pose of the end of the mechanical arm and the target pose, the pose compensation amount is obtained, and the calculation formula is:

[0144]

[0145] Among them, represents the pose compensation amount, represents the actual pose of the end of the mechanical arm, represents the target pose;

[0146] Convert the rotation matrix in the pose compensation amount to the axis angle representation to obtain ;

[0147] The angle compensation amount of the joints of the mechanical arm is solved by the pseudo-inverse method of Jacobian, and the calculation formula is:

[0148]

[0149] Among them, represents an angle compensation amount of a joint of the robot arm, represents a pseudo-inverse of the Jacobian matrix.

[0150] The joint angle compensation amount is directly input into a robot arm controller to realize closed-loop control.

[0151] As shown in Figure 3 The high-order polynomial ensures the continuity of acceleration and controllability of jerk in trajectory planning: the continuity of acceleration avoids mechanical vibration; the position, velocity, acceleration and jerk of the start and end points can be specified simultaneously.

[0152] As shown in Figure 4 The high-order coefficients are usually used to control the convergence characteristics at the end of the trajectory, and the low-order coefficients mainly affect the initial motion state; the oscillation amplitude reflects the sensitivity of the parameters to the boundary conditions, and a sudden change in the initial acceleration will cause a dramatic change in the quadratic term.

[0153] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc.

[0154] The examples described in the present application are only used to describe the preferred embodiments of the present application, and do not limit the concept and scope of the present application. Without departing from the design idea of the present application, various modifications and improvements of the technical solutions of the present application made by the engineering technicians in the art should fall within the protection scope of the present application.

Claims

1. A method for automated installation of curtain walls based on a robot arm, characterized in that, The method comprises the following steps: Step S1, scanning the building facade through an RGB-D camera and constructing a dense point cloud model of the installation surface; Step S2, extracting the center point coordinates and normal vector of the curtain wall keel support surface; Step S3, calculating the target pose of the curtain wall panel to be installed according to the support surface parameters; Step S4, describing the joint movement of the mechanical arm by using a sixth order polynomial, and optimizing the trajectory parameters by using a genetic algorithm; Step S5, calculating the pose deviation in real time during the movement of the mechanical arm, and dynamically adjusting the pose of the end effector.

2. The method of claim 1, wherein, In the step S1, the dense point cloud model of the installation surface is constructed, and the construction specifically comprises the following steps: Step S1.1, initializing a voxel grid and pre-processing the image collected by the RGB-D camera; Step S1.2, converting the pixel coordinates and depth values into 3D points in the camera coordinate system through the camera intrinsic matrix; Step S1.3, dynamically adjusting the voxel based on the curvature; Step S1.4, calculating the signed distance function value of each voxel x along the camera optical center ray to the surface; Step S1.5, introducing a truncation distance to calculate the truncated signed distance function value of each voxel x; Step S1.6, calculating the weight value of each voxel x; Step S1.7, fusing the current frame data into the global truncated signed distance function field; Step S1.8, generating a triangular mesh after all depth maps are fused to complete the construction of the dense point cloud model of the installation surface.

3. The method of claim 2, wherein, In the step S1.3, the voxel is dynamically adjusted based on the curvature, and the adjustment specifically comprises the following steps: calculating the covariance matrix of the neighborhood point set of the point p, which is equal to the sum of the products of the difference between the coordinates of each point in the neighborhood point set and the center coordinate of the neighborhood point set, divided by the number of points in the neighborhood point set; calculating the curvature based on the covariance matrix, which is equal to the minimum eigenvalue of the covariance matrix divided by the sum of all eigenvalues; classifying the curvature to update the voxel size, and setting the voxel size to 1mm when the curvature is greater than 0.02, otherwise setting the voxel size to 5mm.

4. The method of claim 3, wherein, In the step S1.5, the TSDF value of each voxel x is calculated, and the specific formula is: wherein, denotes the truncated signed distance function value of a voxel x, denotes the signed distance function value of a voxel x, denotes the truncated distance.

5. The method of claim 4, wherein, In the step S1.6, the weight value of each voxel x is calculated by introducing a normal vector direction constraint, and the weight value is equal to the cosine value of the angle between the surface normal vector and the line of sight direction divided by the square of the Euclidean norm of the camera coordinate system coordinate of the voxel, wherein the angle between the surface normal vector and the line of sight direction refers to the angle between the normal vector at the image pixel and the direction vector from the camera optical center to the center point of the voxel.

6. The method of claim 5, wherein, The step S2 specifically comprises the following steps: Step S2.1, fitting an initial plane equation using a random sample point set, and using a RANSAC coarse segmentation algorithm to segment the curtain wall keel support surface; Step S2.2, calculating the center point coordinates of the support surface after RANSAC coarse segmentation, which is equal to the weighted average value of the three-dimensional coordinates of all points in the support surface point cloud, wherein the weight coefficient is the contribution degree of each point; Step S2.3, calculating the spatial distribution covariance of the point cloud relative to the center point, which is equal to the sum of the products of the difference between each point and the center point; Step S2.4, solving the eigenvector corresponding to the minimum eigenvalue of the covariance matrix, that is, the support surface normal vector.

7. The method of claim 6, wherein, In the step S3, the target pose of the curtain panel to be installed is calculated according to the support surface parameters, specifically including: Calculate the coordinate of the curtain panel to be installed, which is equal to the coordinate of the center point of the support surface plus the product of the installation gap and the normal vector, and then add the offset of the panel on the tangent plane of the support surface; Calculate the attitude rotation matrix of the curtain panel to be installed, which is composed of the unit matrix and the rotation matrix around the rotation axis, wherein the rotation axis is obtained by the cross product of the z-axis and the target normal vector, and the rotation angle is determined by the angle between the z-axis and the target normal vector.

8. The method of claim 7, wherein, In the step S4, the joint motion of the robot arm is described by a sixth order polynomial, and the specific formula is: wherein, represents a mechanical arm joint motion trajectory, represents time, , , , , , and represent constant term coefficients, first term coefficients, second term coefficients, third term coefficients, fourth term coefficients, fifth term coefficients, and sixth term coefficients, respectively. The angles, angular velocities and angular accelerations of each joint of the mechanical arm at the initial time and the terminal time are taken as boundary conditions Solving obtains , , , , and ; The coefficient of the sixth order term Optimization is performed by genetic algorithm; According to the solving result , , , , , and planning the motion trajectory with continuous joint angle, angular velocity and angular acceleration.

9. The method of claim 8, wherein, said coefficient of the sixth order term The optimization solution by genetic algorithm, specifically comprising the following steps: Design an optimization function, which is the product of the trajectory length weight and the total length of the path of the end effector, plus the product of the joint rotation weight and the sum of the absolute values of all joint angle changes, plus the product of the safety threshold weight and the maximum deviation of the actual distance from the safety threshold; Randomly generate 200 sets of sixth order coefficients; Calculate the optimization function value of the trajectory corresponding to each sixth order coefficient, and select the top twenty individuals in the inverse function value order as the parents; Generate a new population through crossover and mutation, and iterate until convergence; After optimization, the optimal sixth order coefficient is obtained, which is substituted into the sixth order polynomial to generate the optimal trajectory.

10. The method of claim 9, wherein, The step S5 specifically includes the following steps: According to the actual pose and the target pose of the end of the robot arm, the pose compensation amount is obtained, which is equal to the target pose minus the actual pose; Convert the rotation matrix in the pose compensation amount to axis-angle representation to obtain the rotation vector; Solve the angle compensation amount of the joints of the robot arm by the pseudo-inverse method of Jacobian, which is equal to the pseudo-inverse of the Jacobian matrix multiplied by the pose compensation amount.

Images