Intelligent assembly method based on hundred-million-level point cloud

Abstract

This invention discloses an intelligent assembly method based on point clouds with hundreds of millions of data points. By constructing an efficient loading mechanism suitable for point cloud data with hundreds of millions of data points, a data management structure decoupling display and computation, and a hierarchical filtering and recursive pruning computation model based on BVH, this invention enables the virtual assembly process to achieve real-time response while ensuring geometric accuracy. Simultaneously, by combining multi-process cylindrical axial fitting and GPU parallel sealing ring intersection analysis technology, it achieves efficient computation for coaxiality assessment and compaction uniformity detection. This allows for accurate prediction of interference positions, optimization of assembly posture, and generation of reasonable assembly paths before assembly, significantly reducing the number of repeated trial and error attempts on-site, reducing reliance on manual labor, greatly improving assembly efficiency and accuracy, and enhancing the controllability and reliability of complex structure assembly processes.

Description

Technical Field

[0001] This invention belongs to the field of intelligent assembly, specifically, it relates to an intelligent assembly method based on a point cloud with hundreds of millions of points. Background Technology

[0002] As equipment manufacturing moves towards larger scale, precision, and complexity, the assembly process is playing an increasingly prominent role in the overall manufacturing process. This is especially true in the assembly of large electromechanical equipment, sealing structural components, and high-precision shaft components, where the fit between parts directly affects the overall operational stability and service life of the machine. However, traditional assembly methods rely heavily on manual experience for on-site positioning and adjustment. In practice, this often requires repeated alignment, disassembly, fine-tuning, and correction, representing a typical trial-and-error assembly process. This method not only demands high skill levels from operators but also struggles to accurately predict interference positions and assembly postures when dealing with complex curved surfaces or concealed fits, resulting in a large workload, long processing time, and low assembly efficiency. Furthermore, repeated contact can lead to surface wear or seal failure risks.

[0003] To improve assembly efficiency, virtual assembly technology based on 3D scanning and point cloud modeling has been gradually introduced in recent years. This involves collecting 3D point cloud data of the assembly, constructing a digital model in a computer environment, and performing assembly path analysis and collision simulation in virtual space. This method can identify potential interference areas in advance, optimize assembly posture, and output a theoretical assembly path before actual assembly, thereby reducing the number of repeated disassembly and assembly steps on-site. Typically, even with measurement or model errors, only two to three corrections to the assembly path are needed to meet assembly requirements. Compared to traditional manual trial-and-error methods, this significantly improves assembly efficiency and accuracy.

[0004] However, in practical engineering applications, with the improvement of scanning accuracy and the expansion of equipment scale, the point cloud data volume of a single assembly often reaches hundreds of millions. Such a large amount of data will generate huge computational pressure in the processes of loading, displaying, distance calculation, collision detection, and attitude analysis. If not handled properly, it can easily lead to problems such as system response lag, display stuttering, or even failure to provide real-time feedback. At the same time, when dealing with coaxial assemblies of cylindrical objects or rubber sealing ring clamping assemblies, axial fitting analysis and sealing clamping uniformity judgment are also required, which places higher demands on computational accuracy and real-time performance. Existing virtual assembly systems often struggle to balance display smoothness and computational accuracy when dealing with hundreds of millions of point cloud data, affecting the effectiveness of practical engineering applications.

[0005] In view of this, the present invention is proposed. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an intelligent assembly method based on a point cloud of hundreds of millions of points, which solves the problems mentioned in the background art.

[0007] To solve the above-mentioned technical problems, the basic concept of the technical solution adopted by the present invention is as follows: A smart assembly method based on a point cloud with hundreds of millions of points includes the following steps: S1: Read the point cloud data of the workpiece model, convert the text-formatted XYZ point cloud data into binary-formatted BXYZ point cloud data, and then load it to improve the reading efficiency of hundreds of millions of point cloud data. S2: Perform proportional sampling on the original point cloud data of hundreds of millions of points, use only the sampled point cloud for 3D visualization display, and establish a one-to-one mapping relationship between the sampled point cloud and the original point cloud, so that the display data and the calculation data are independent of each other, and improve the display smoothness while ensuring the calculation accuracy. S3: During the assembly process, BVH hierarchical bounding volume data structures are constructed for the source assembly point cloud and the target assembly point cloud respectively, and hierarchical filtering calculations are performed based on the tree recursive search algorithm to obtain the shortest distance between the two assemblies in real time, which is used to guide the assembly posture adjustment and avoid collisions. S4: When the assembly is a cylindrical structure, perform axial fitting calculation on the assembly point cloud, divide the hundreds of millions of point cloud data into multiple data blocks, and complete the local fitting calculation in multiple independent processes. Finally, merge the calculation results of each process to obtain the global axial fitting result, so as to improve CPU utilization and shorten the calculation time. S5: When the assembly includes a rubber sealing ring structure, the intersection area between the assembly point cloud and the sealing ring point cloud is calculated in real time during the assembly process. The intersection calculation task of the sealing ring point cloud is deployed to the GPU for parallel computing, and the distribution state of the assembly inside the sealing ring is displayed in color point cloud mode to evaluate the uniformity of compression.

[0008] Optionally, the BXYZ binary point cloud data is stored continuously in floating-point format for X, Y, and Z three-dimensional coordinate information to reduce the time consumption during character parsing and improve I / O reading efficiency.

[0009] Optionally, the proportional sampling adopts uniform sampling or sampling method based on spatial voxel division, and the sampling ratio is dynamically adjusted according to the display hardware performance.

[0010] Optionally, the one-to-one mapping relationship between the sampled point cloud and the original point cloud is stored through an index table structure. When performing distance, angle, or intersection calculations, the original point cloud data is called for calculation, and the steps are as follows: When completing the scaled sampling of the original point cloud, let the original point cloud set be... After sampling, a set of sampling point clouds is obtained. Each sampling point satisfies The This is a sampling mapping function; during the sampling process, the index number of each sampling point in the original point cloud array is recorded synchronously. And build an index mapping table. , where key value Indicates the sampling point number and the mapping value. This indicates the storage location of the corresponding original point in the original point cloud data, thereby establishing a one-to-one correspondence between the sampled point cloud and the original point cloud.

[0011] When performing distance calculations, candidate regions are first filtered based on sampled points, and then the original point cloud data is called through an index mapping function to participate in the calculation. The shortest distance is calculated according to...

[0012] Perform calculations, where and These represent the original point indices corresponding to the sampling points of the two assemblies, respectively, to ensure that the calculation accuracy is not affected by the sampling ratio.

[0013] During angle calculation, the corresponding original point coordinate data is extracted based on the index mapping relationship, a vector is constructed, and then... To solve for the included angle, where the vector... , All coordinates are composed of the original point cloud coordinates to ensure the accuracy of attitude assessment.

[0014] When performing intersection region determination, the original point cloud data and the target point cloud data are retrieved based on the index mapping, and a distance threshold is determined. If the condition is met... The point is determined to be an intersection point, where, To preset the spatial tolerance, this method enables collision and compaction area analysis based on the original precision data.

[0015] Optionally, the BVH hierarchical bounding volume data structure is constructed using a top-down recursive partitioning method, where each node corresponds to a bounding box, and the parent node's bounding box contains the bounding boxes of all its child nodes.

[0016] Optionally, the tree-based recursive search algorithm reduces the number of point-to-point calculations by prioritizing the determination of the minimum distance between bounding boxes and pruning the branch when the distance between bounding boxes is greater than the currently recorded minimum distance.

[0017] Optionally, the multi-process axial fitting calculation adopts a collaborative working mode of a main process and multiple sub-processes. The main process is responsible for data partitioning and result merging, while the sub-processes are responsible for performing least-squares cylindrical fitting calculations on the corresponding data blocks. The steps are as follows: The point cloud data is divided into KKK data sub-blocks according to spatial region division or equal number of sub-blocks. Each sub-block satisfies Furthermore, the sub-blocks do not overlap with each other, and then the main process allocates each data sub-block to different sub-processes to perform parallel computing tasks.

[0018] Secondly, each subprocess performs least-squares cylindrical fitting calculations on the allocated data sub-blocks, assuming the cylinder axis direction vector is a unit vector. A point on the axis is The radius of the cylinder is Then for any point within the sub-block Its distance from the axis is The subprocess minimizes the error function Solve for the local axial parameters of the corresponding sub-block. .

[0019] Finally, after all subprocesses have completed their calculations, the main process collects the fitting results for each sub-block and performs parameter fusion based on the number of data points or the reciprocal of the error as weights, while the axial direction vector is fused according to... Normalized synthesis is performed, where the weights are... , For the number of points in the sub-block, To address the fitting error, a global cylindrical axial fitting result is obtained, enabling efficient utilization of multi-core CPU resources and shortening the overall computation time.

[0020] Optionally, the GPU parallel computing improves real-time computing capabilities by loading the sealed circle point cloud data into the video memory and performing spatial inclusion judgment and intersection region filtering calculations on the GPU core.

[0021] Optionally, the shortest distance result displayed in real time during the assembly process is simultaneously shown in numerical and color gradient form, wherein a warning color is displayed when the distance is less than a preset safety threshold.

[0022] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art. Of course, any product implementing the present invention does not necessarily need to achieve all of the following advantages at the same time: This invention enables real-time response in the virtual assembly process while ensuring geometric accuracy by constructing an efficient loading mechanism suitable for point cloud data of hundreds of millions, a data management structure that decouples display and computation, and a hierarchical filtering and recursive pruning computation model based on BVH. At the same time, by combining multi-process cylindrical axial fitting and GPU parallel sealing ring intersection analysis technology, it achieves efficient calculation of coaxiality assessment and compression uniformity detection. Thus, interference positions can be accurately predicted, assembly posture optimized, and reasonable assembly paths generated before assembly, significantly reducing the number of repeated trial and error on site, reducing the degree of manual dependence, greatly improving assembly efficiency and assembly accuracy, and enhancing the controllability and reliability of the assembly process of complex structures.

[0023] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0024] The accompanying drawings described below are merely some embodiments. Those skilled in the art can obtain other drawings based on these drawings without any creative effort. In the drawings: Figure 1 This is a simulation diagram of a robotic arm.

[0025] It should be noted that these accompanying drawings and textual descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art by referring to specific embodiments. Detailed Implementation

[0026] The invention will now be described in further detail with reference to the accompanying drawings.

[0027] Please see Figure 1 As shown, this embodiment provides an intelligent assembly method based on a point cloud with hundreds of millions of points, including the following steps: S1: Read the point cloud data of the workpiece model, convert the text-formatted XYZ point cloud data into binary-formatted BXYZ point cloud data, and then load it to improve the reading efficiency of hundreds of millions of point cloud data. S2: Perform proportional sampling on the original point cloud data of hundreds of millions of points, use only the sampled point cloud for 3D visualization display, and establish a one-to-one mapping relationship between the sampled point cloud and the original point cloud, so that the display data and the calculation data are independent of each other, and improve the display smoothness while ensuring the calculation accuracy. S3: During the assembly process, BVH hierarchical bounding volume data structures are constructed for the source assembly point cloud and the target assembly point cloud respectively, and hierarchical filtering calculations are performed based on the tree recursive search algorithm to obtain the shortest distance between the two assemblies in real time, which is used to guide the assembly posture adjustment and avoid collisions. S4: When the assembly is a cylindrical structure, perform axial fitting calculation on the assembly point cloud, divide the hundreds of millions of point cloud data into multiple data blocks, and complete the local fitting calculation in multiple independent processes. Finally, merge the calculation results of each process to obtain the global axial fitting result, so as to improve CPU utilization and shorten the calculation time. S5: When the assembly includes a rubber sealing ring structure, the intersection area between the assembly point cloud and the sealing ring point cloud is calculated in real time during the assembly process. The intersection calculation task of the sealing ring point cloud is deployed to the GPU for parallel computing, and the distribution state of the assembly inside the sealing ring is displayed in color point cloud format to evaluate the uniformity of compression. This invention constructs a hierarchical processing architecture for hundreds of millions of point clouds, organically combining efficient data loading, display calculation decoupling, BVH hierarchical filtering, multi-process axial fitting, and GPU parallel sealing ring analysis to form a complete intelligent assembly calculation process. This enables real-time interaction and efficient analysis of hundreds of millions of point cloud data while ensuring geometric accuracy, improving assembly efficiency and enhancing collision warning and sealing evaluation capabilities, significantly improving the intelligence level of virtual assembly of complex equipment.

[0028] In this embodiment, the BXYZ binary point cloud data is stored continuously as floating-point numbers, providing X, Y, and Z three-dimensional coordinate information. This reduces the time consumption during character parsing and improves I / O read efficiency. By storing the BXYZ binary point cloud data continuously as floating-point numbers, the time consumption and CPU usage issues caused by parsing text data character by character are avoided. This significantly improves I / O read efficiency and memory access continuity, enabling the loading of hundreds of millions of point cloud data points in a short time, providing a stable data foundation for subsequent real-time assembly analysis.

[0029] In this embodiment, the proportional sampling adopts a uniform sampling method or a sampling method based on spatial voxel partitioning, and the sampling ratio is dynamically adjusted according to the display hardware performance. By adopting a uniform sampling method or a sampling method based on spatial voxel partitioning, and dynamically adjusting the sampling ratio according to the display hardware performance, adaptive control of the display layer data scale is achieved. This significantly reduces the rendering load while ensuring that the overall structural features of the model are not distorted, making the 3D display smoother. At the same time, it preserves complete original data for subsequent accurate calculations, improving the stability and adaptability of the system operation.

[0030] In this embodiment, the one-to-one mapping relationship between the sampled point cloud and the original point cloud is stored through an index table structure. When performing distance, angle, or intersection calculations, the original point cloud data is called for calculation. The steps are as follows: When completing the scaled sampling of the original point cloud, let the original point cloud set be... After sampling, a set of sampling point clouds is obtained. Each sampling point satisfies The This is a sampling mapping function; during the sampling process, the index number of each sampling point in the original point cloud array is recorded synchronously. And build an index mapping table. , where key value Indicates the sampling point number and the mapping value. This indicates the storage location of the corresponding original point in the original point cloud data, thereby establishing a one-to-one correspondence between the sampled point cloud and the original point cloud.

[0031] When performing distance calculations, candidate regions are first filtered based on sampled points, and then the original point cloud data is called through an index mapping function to participate in the calculation. The shortest distance is calculated according to...

[0032] Perform calculations, where and These represent the original point indices corresponding to the sampling points of the two assemblies, respectively, to ensure that the calculation accuracy is not affected by the sampling ratio.

[0033] During angle calculation, the corresponding original point coordinate data is extracted based on the index mapping relationship, a vector is constructed, and then... To solve for the included angle, where the vector... , All coordinates are composed of the original point cloud coordinates to ensure the accuracy of attitude assessment.

[0034] When performing intersection region determination, the original point cloud data and the target point cloud data are retrieved based on the index mapping, and a distance threshold is determined. If the condition is met... The point is determined to be an intersection point, where, To predetermine spatial tolerance, this method enables collision and compaction area analysis based on original precision data. By establishing a one-to-one index mapping relationship between the sampled point cloud and the original point cloud, the displayed data and the calculated data are made independent of each other. When performing distance, angle, or intersection calculations, the original high-precision point cloud data is always used in the calculation, thereby avoiding the problem of error accumulation caused by sampling. This achieves a balance between smooth visualization and geometric calculation accuracy, improving the reliability of assembly analysis results.

[0035] In this embodiment, the BVH hierarchical bounding volume data structure is constructed using a top-down recursive partitioning method. Each node corresponds to a bounding box, and the parent node's bounding box contains all its child node bounding boxes. By constructing the BVH hierarchical bounding volume structure using a top-down recursive partitioning method, the point cloud data forms a hierarchical organizational structure in space. Each parent node's bounding box contains all its child nodes, thereby enabling the rapid elimination of large-scale irrelevant regions during assembly calculations, significantly reducing the number of direct point-to-point calculations, and improving the efficiency of shortest distance calculations.

[0036] In this embodiment, the tree-structured recursive search algorithm prioritizes determining the minimum distance between bounding boxes. When the distance between bounding boxes exceeds the currently recorded minimum distance, the branch is pruned, thereby reducing the number of point pair calculations. By prioritizing the determination of the minimum distance between bounding boxes during the tree-structured recursive search process and performing pruning when the bounding box distance exceeds the currently recorded minimum distance, invalid node expansion can be effectively avoided, reducing the number of computational branches. Even with point clouds of hundreds of millions, real-time shortest distance updates can still be achieved, improving the response speed of collision detection and attitude adjustment.

[0037] In this embodiment, the multi-process axial fitting calculation adopts a collaborative working mode of a main process and multiple sub-processes. The main process is responsible for data partitioning and result merging, while the sub-processes are responsible for performing least-squares cylindrical fitting calculations on the corresponding data blocks. The steps are as follows: The point cloud data is divided into KKK data sub-blocks according to spatial region division or equal number of sub-blocks. Each sub-block satisfies Furthermore, the sub-blocks do not overlap with each other, and then the main process allocates each data sub-block to different sub-processes to perform parallel computing tasks.

[0038] Secondly, each subprocess performs least-squares cylindrical fitting calculations on the allocated data sub-blocks, assuming the cylinder axis direction vector is a unit vector. A point on the axis is The radius of the cylinder is Then for any point within the sub-block Its distance from the axis is The subprocess minimizes the error function Solve for the local axial parameters of the corresponding sub-block. .

[0039] Finally, after all subprocesses have completed their calculations, the main process collects the fitting results for each sub-block and performs parameter fusion based on the number of data points or the reciprocal of the error as weights, while the axial direction vector is fused according to... Normalized synthesis is performed, where the weights are... , For the number of points in the sub-block, To address fitting errors and obtain global cylindrical axial fitting results, efficient utilization of multi-core CPU resources and reduced overall computation time are achieved. A multi-process axial fitting mechanism, employing a main process and multiple sub-processes working collaboratively, divides large-scale point cloud data into multiple sub-blocks and performs least-squares cylindrical fitting operations in parallel. Finally, the results of each sub-block are fused, fully utilizing multi-core CPU computing resources. This significantly reduces computation time while maintaining fitting accuracy, improving the efficiency of coaxiality analysis for cylindrical assemblies.

[0040] In this embodiment, the GPU parallel computing improves real-time computing capabilities by loading the sealing ring point cloud data into the GPU memory and performing spatial inclusion judgment and intersection region filtering calculations on the GPU core. By loading the sealing ring point cloud data into the GPU memory and performing spatial inclusion judgment and intersection region filtering calculations on the GPU core, parallel processing of intersection operations is achieved, significantly improving the real-time performance of the sealing compression analysis. This allows for dynamic evaluation of sealing uniformity during assembly, enhancing the quality control capability of the sealing structure assembly.

[0041] In this embodiment, the shortest distance result calculated in real time during the assembly process is displayed synchronously in the form of numerical values ​​and color gradients, with a warning color displayed when the distance is less than a preset safety threshold. By synchronously displaying the shortest distance result calculated in real time during the assembly process in the form of numerical values ​​and color gradients, and using a warning color to indicate when the distance is less than the preset safety threshold, the changes in the gaps between the assembled parts can be intuitively reflected, enhancing the operator's safety perception ability, reducing the risk of misassembly and collision, and improving the visualization and safety of the assembly process.

[0042] based on Figure 1 As shown, the red and green point clouds represent two assemblies, respectively, while the blue point cloud represents the sealing ring. The assembly process involves the robot tool head carrying the red point cloud to the correct relative pose with the green point cloud. During the assembly process, it is necessary to display in real time the shortest distance between the red and green point clouds, the axial angle between them, and the distribution of the green point cloud within the blue point cloud.

[0043] This invention is not limited to the embodiments described above. Anyone should understand that structural changes made under the guidance of this invention, and any technical solutions that are the same as or similar to this invention, fall within the protection scope of this invention. Technical aspects, shapes, and structures not described in detail in this invention are all publicly known technologies.

Claims

1. An intelligent assembly method based on a point cloud with hundreds of millions of points, characterized in that, Includes the following steps: S1: Read the point cloud data of the workpiece model, convert the text-formatted XYZ point cloud data into binary-formatted BXYZ point cloud data, and then load it to improve the reading efficiency of hundreds of millions of point cloud data. S2: Perform proportional sampling on the original point cloud data of hundreds of millions of points, use only the sampled point cloud for 3D visualization display, and establish a one-to-one mapping relationship between the sampled point cloud and the original point cloud, so that the display data and the calculation data are independent of each other, and improve the display smoothness while ensuring the calculation accuracy. S3: During the assembly process, BVH hierarchical bounding volume data structures are constructed for the source assembly point cloud and the target assembly point cloud respectively, and hierarchical filtering calculations are performed based on the tree recursive search algorithm to obtain the shortest distance between the two assemblies in real time, which is used to guide the assembly posture adjustment and avoid collisions. S4: When the assembly is a cylindrical structure, perform axial fitting calculation on the assembly point cloud, divide the hundreds of millions of point cloud data into multiple data blocks, and complete the local fitting calculation in multiple independent processes. Finally, merge the calculation results of each process to obtain the global axial fitting result, so as to improve CPU utilization and shorten the calculation time. S5: When the assembly includes a rubber sealing ring structure, the intersection area between the assembly point cloud and the sealing ring point cloud is calculated in real time during the assembly process. The intersection calculation task of the sealing ring point cloud is deployed to the GPU for parallel computing, and the distribution state of the assembly inside the sealing ring is displayed in color point cloud mode to evaluate the uniformity of compression.

2. The intelligent assembly method based on a point cloud with hundreds of millions of points according to claim 1, characterized in that, The BXYZ binary point cloud data stores the X, Y, and Z three-dimensional coordinate information continuously in floating-point format to reduce the time consumption during character parsing and improve I / O reading efficiency.

3. The intelligent assembly method based on a point cloud with hundreds of millions of points according to claim 1, characterized in that, The proportional sampling adopts uniform sampling or sampling method based on spatial voxel division, and the sampling ratio is dynamically adjusted according to the display hardware performance.

4. The intelligent assembly method based on a point cloud with hundreds of millions of points according to claim 1, characterized in that, The one-to-one mapping relationship between the sampled point cloud and the original point cloud is stored through an index table structure. When performing distance, angle, or intersection calculations, the original point cloud data is called for calculation, and the steps are as follows: When completing the scaled sampling of the original point cloud, let the original point cloud set be... After sampling, a set of sampling point clouds is obtained. Each sampling point satisfies The This is a sampling mapping function; during the sampling process, the index number of each sampling point in the original point cloud array is recorded synchronously. And build an index mapping table. , where key value Indicates the sampling point number and the mapping value. This indicates the storage location of the corresponding original point in the original point cloud data, thereby establishing a one-to-one correspondence between the sampled point cloud and the original point cloud; When performing distance calculations, candidate regions are first filtered based on sampled points, and then the original point cloud data is called through an index mapping function to participate in the calculation. The shortest distance is calculated according to... Perform calculations, where and These represent the original point indices corresponding to the sampling points of the two assemblies, to ensure that the calculation accuracy is not affected by the sampling ratio; During angle calculation, the corresponding original point coordinate data is extracted based on the index mapping relationship, a vector is constructed, and then... To solve for the included angle, where the vector... , All coordinates are composed of the original point cloud coordinates to ensure the accuracy of attitude assessment; When performing intersection region determination, the original point cloud data and the target point cloud data are retrieved based on the index mapping, and a distance threshold is determined. If the condition is met... The point is determined to be an intersection point, where, To preset the spatial tolerance, this method enables collision and compaction area analysis based on the original precision data.

5. The intelligent assembly method based on a point cloud of hundreds of millions of points according to claim 1, characterized in that, The BVH hierarchical bounding volume data structure is constructed using a top-down recursive partitioning method. Each node corresponds to a bounding box, and the parent node's bounding box contains the bounding boxes of all its child nodes.

6. The intelligent assembly method based on a point cloud with hundreds of millions of points according to claim 1, characterized in that, The tree-based recursive search algorithm reduces the number of point-to-point calculations by prioritizing the minimum distance between bounding boxes and pruning branches when the distance between bounding boxes exceeds the currently recorded minimum distance.

7. The intelligent assembly method based on a point cloud of hundreds of millions of points according to claim 1, characterized in that, The multi-process axial fitting calculation adopts a collaborative working mode of a main process and multiple sub-processes. The main process is responsible for data partitioning and result merging, while the sub-processes are responsible for performing least-squares cylindrical fitting calculations on the corresponding data blocks. The steps are as follows: The point cloud data is divided into KKK data sub-blocks according to spatial region division or equal number of sub-blocks. Each sub-block satisfies Furthermore, each sub-block does not overlap with the others, and then the main process allocates each data sub-block to different sub-processes to perform parallel computing tasks. Secondly, each subprocess performs least-squares cylindrical fitting calculations on the allocated data sub-blocks, assuming the cylinder axis direction vector is a unit vector. A point on the axis is The radius of the cylinder is Then for any point within the sub-block Its distance from the axis is The subprocess minimizes the error function Solve for the local axial parameters of the corresponding sub-block. ; Finally, after all subprocesses have completed their calculations, the main process collects the fitting results for each sub-block and performs parameter fusion based on the number of data points or the reciprocal of the error as weights, while the axial direction vector is fused according to... Normalized synthesis is performed, where the weights are... , For the number of points in the sub-block, To address the fitting error, a global cylindrical axial fitting result is obtained, enabling efficient utilization of multi-core CPU resources and shortening the overall computation time.

8. The intelligent assembly method based on a point cloud with hundreds of millions of points according to claim 1, characterized in that, The GPU parallel computing improves real-time computing capabilities by loading the sealed circle point cloud data into the GPU memory and performing spatial inclusion judgment and intersection region filtering calculations on the GPU core.

9. The intelligent assembly method based on a point cloud of hundreds of millions of points according to claim 1, characterized in that, The shortest distance result displayed in real time during the assembly process is presented synchronously in the form of numerical values ​​and color gradients, with a warning color displayed when the distance is less than a preset safety threshold.

Images