A cluster batch optimization and adjustment system for real-life 3D model processing

Abstract

The present invention belongs to the technical field of three-dimensional modeling. The present invention discloses a cluster batch optimization adjustment system for real-scene three-dimensional model processing, including a cluster architecture construction module, which is used to construct a node cluster including a management node, a storage node and n computing nodes; a data acquisition module, which is used to acquire point cloud data, texture image data and batch optimization task data; a data processing module, which is used to pre-process the point cloud data and texture image data; divide the pre-processed point cloud data and texture image data into m regional blocks according to space, and distribute them to n computing nodes of the node cluster for parallel processing; a task scheduling module, which is used to decompose the batch optimization task data into γ independent subtasks; and adopt an adaptive load balancing algorithm to dynamically distribute the γ independent subtasks to the n computing nodes of the node cluster. This provides an efficient, stable and intelligent solution for large-scale real-scene three-dimensional model processing.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional modeling, and more specifically, to a cluster batch optimization adjustment system for real-scene three-dimensional model processing. Background Art

[0002] The patent with patent publication number CN117392328A discloses a three-dimensional real scene modeling method and system based on drone clusters, involving the field of three-dimensional modeling technology, obtaining communication permissions of several drones, and then building a drone cluster, and performing drone cluster collaborative operations; drone cluster collaborative operations include cluster data interaction, cluster path collaboration and cluster task collaboration; obtaining several oblique photography images through the drone cluster, performing image feature extraction, image registration fusion and image correction on the oblique photography images, and then generating modeling image data; obtaining modeling image data to generate external point cloud data and internal mapping data, building a three-dimensional real scene external model according to the external point cloud data, generating a three-dimensional real scene internal model according to the internal mapping data, and then merging the three-dimensional real scene external model and the three-dimensional real scene internal model to generate a three-dimensional real scene model; and monitoring the modeling area, updating the three-dimensional real scene model in real time, thereby ensuring the safety of the drone cluster and the accuracy of modeling.

[0003] Existing real-life 3D model processing technologies have computing bottlenecks that lead to low processing efficiency. When processing large-scale real-life 3D models, they usually rely on single-machine computing or traditional parallel computing methods. These methods cannot fully utilize computing resources when faced with massive point cloud data and complex geometric structures, resulting in slow processing speeds. In tasks that require a lot of computing, such as model refinement and texture repair, single-machine computing is prone to problems such as insufficient memory and long computing time.

[0004] Traditional texture mapping methods rely on geometric feature matching or projection algorithms, which make it difficult to achieve high-precision texture fitting and repair in large-scale models. With the improvement of texture image resolution, traditional methods often cannot meet the needs of high-precision texture synthesis, resulting in obvious misalignment, blur or distortion on the model surface. In the process of large-scale model generation, the details of the model are often missing or the topology structure is wrong due to excessive data volume or insufficient algorithm processing capabilities. These problems are difficult to solve in traditional point cloud processing or mesh optimization algorithms, which often rely on manual repair or rule-based optimization methods, making it difficult to achieve automation and efficiency.

[0005] Current 3D modeling technologies usually lack real-time processing capabilities and cannot meet the needs of modeling and updating models in dynamic environments. The processing flow of existing technologies is generally simple and non-scalable. When the amount of data increases further, the computing power of the system is likely to become a bottleneck, affecting the real-time and accuracy of model processing.

[0006] In view of this, the present invention proposes a cluster batch optimization adjustment system for real-scene three-dimensional model processing to solve the above problems. Summary of the invention

[0007] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solution: a cluster batch optimization adjustment system for real-scene three-dimensional model processing, comprising:

[0008] Cluster architecture building module, used to build a node cluster including management nodes, storage nodes and n computing nodes;

[0009] Data acquisition module, used to collect point cloud data, texture image data and batch optimization task data;

[0010] The data processing module is used to pre-process the point cloud data and texture image data; the pre-processed point cloud data and texture image data are divided into m regional blocks according to space, and are distributed to n computing nodes of the node cluster for parallel processing;

[0011] The task scheduling module is used to decompose the batch optimization task data into γ independent subtasks; the adaptive load balancing algorithm is used to dynamically allocate the γ independent subtasks to the n computing nodes of the node cluster; each computing node executes the allocated independent subtask to obtain the optimized regional blocks;

[0012] The model optimization module is used to submit the optimized regional blocks to the management node for integration processing to generate a complete real-scene 3D model; perform global optimization on the real-scene 3D model to generate a multi-resolution hierarchical model;

[0013] The output module is used to convert the format of the multi-resolution hierarchical model and save it to the storage node; each module is connected by wired and / or wireless means.

[0014] Furthermore, the method for constructing a node cluster including a management node, a storage node and n computing nodes includes:

[0015] Define the types and number of nodes in the node cluster, including one management node, one storage node and n computing nodes; divide the node functions according to the node types, define the functions of the management node as task scheduling and resource allocation, the function of the storage node as data storage, and the function of the n computing nodes as executing parallel computing tasks; configure performance parameters for the n computing nodes, including computing power, memory capacity, network bandwidth and current load.

[0016] Furthermore, the point cloud data consists of v discrete points with attributes; the attributes include point coordinates, reflection intensity, color value, normal vector and density distribution; the texture image data includes texture images and texture mapping coordinates corresponding to the point cloud data; the batch optimization task data is the goal of optimizing the real-life three-dimensional model, including mesh refinement and smoothing, topology repair, texture mapping and texture super-resolution optimization.

[0017] Furthermore, the method for preprocessing point cloud data and texture image data includes:

[0018] The deep learning model PointNet++ is used to perform denoising on the point cloud data, and the noise points and abnormal points in the point cloud data are identified and removed; cluster sampling is performed on the denoised point cloud data, and a clustering algorithm is used to identify and retain the cluster center points in the point cloud data; the coordinates of the retained cluster center points are aligned to obtain the preprocessed point cloud data; the texture image data is formatted in a unified manner, and the bilateral filtering algorithm is used to remove noise interference in the texture image data to obtain the preprocessed texture image data.

[0019] Furthermore, the method of dividing the pre-processed point cloud data and texture image data into m regional blocks according to space and distributing them to n computing nodes of the node cluster for parallel processing includes:

[0020] The preprocessed point cloud data is reconstructed using the Delaunay triangulation algorithm to generate a three-dimensional mesh model; the preprocessed texture image data is mapped to the three-dimensional mesh model through texture mapping coordinates to obtain a three-dimensional mesh model with texture; the three-dimensional mesh model with texture is spatially segmented using the octree algorithm and divided into m regional blocks. The specific method steps are as follows:

[0021] S51. For a 3D mesh model with texture, define an initial bounding box B 0 ; The range of the initial bounding box is determined by the extreme values ​​of the grid vertex coordinates of the 3D grid model with texture: B 0 =[x min ,x max ]×[y min ,y max ]×[z min ,z max ]; among them, B 0 is the initial bounding box; x min and x max are the minimum and maximum coordinate values ​​of all mesh vertices of the 3D mesh model with texture on the x-axis; y min and maxare the minimum and maximum coordinate values ​​of all mesh vertices of the 3D mesh model with texture on the y-axis; min and z max They are the minimum and maximum coordinate values ​​of all mesh vertices of the 3D mesh model with texture on the z-axis;

[0022] S52, initial bounding box B 0 Perform recursive division, each time dividing into 8 sub-areas, each sub-area corresponds to a sub-cube; the current bounding box is preset to be divided into 8 sub-areas B 1 ,B 2 ,...,B 8 ,The spatial range and center point of each sub-region are calculated by the following formula:

[0023]

[0024] Among them, B i,j is the bounding box of the sub-region with index j at the recursive level i; The x-axis coordinate of the center point of the space currently enclosing the box area; The y-axis coordinate of the center point of the space currently enclosing the box area; is the z-axis coordinate of the spatial center point of the current bounding box area; i is the index of the current recursive layer of the bounding box; j is the sub-area index in the current recursive layer of the bounding box;

[0025] S53, repeat step S52 until the number of points contained in the sub-area is less than the preset minimum point threshold N min Stop when , and use the sub-region as a leaf node of the octree;

[0026] The preset minimum point threshold is dynamically adjusted through the point threshold adjustment formula. The point threshold adjustment formula is: Among them, N′ min is the minimum point threshold after dynamic adjustment; N min is the preset minimum point threshold; b is the number of discrete points in the three-dimensional grid model; β is the proportional factor that controls the increase in the preset minimum point threshold; k is the baseline adjustment factor that controls the basic value of the minimum point threshold; α is the logarithmic adjustment factor used to adjust the sensitivity of the change with the number of points;

[0027] S54. According to the coordinate attribution of the mesh vertices, the meshes in the three-dimensional mesh model with texture are allocated to corresponding sub-regions, and then m regional blocks are obtained and evenly allocated to n computing nodes of the node cluster for parallel processing.

[0028] Furthermore, the method of dynamically allocating γ independent subtasks to n computing nodes of a node cluster using an adaptive load balancing algorithm includes:

[0029] According to the performance parameters of n computing nodes, the node performance of n computing nodes is evaluated by the node performance comprehensive scoring formula; the node performance comprehensive scoring formula is: n′ =ω 1 ·C n′ +ω 2 ·M n′ +ω 3 ·B n′ -ω 4 ·L n′ ; Among them, H n′ is the comprehensive performance score of the n′th computing node; C n′ is the computing power of the n′th computing node; M n′ is the memory capacity of the n′th computing node; B n′ is the network bandwidth of the n′th computing node; L n′ is the current load of the n′th computing node; n′ is the index of the computing node; ω 1 is the weight factor of computing power; ω 2 is the weight factor of memory capacity, ω 3 is the weight factor of network bandwidth; ω 4 is the weight factor of the current load;

[0030] The priority of γ independent subtasks is calculated by the priority calculation formula; the priority calculation formula is: γ′ =μ 1 ·DA γ′ +μ 2 ·CO γ′ -μ 3 ·DE γ′ ; Among them, T γ′ is the priority of the γ′th independent subtask; DA γ′ The amount of data required to be processed for the γ′th independent subtask; CO γ′ The task complexity of the γ′th independent subtask is calculated; DE γ′ is the computation time required to complete the γ′th independent subtask; μ 1 is the weight of the impact of data volume on priority; μ 2 Calculate the weight of the task complexity on the priority; μ 3 To calculate the weight of the impact of time on priority;

[0031] Based on the adaptive load balancing algorithm, the task allocation rule is constructed to dynamically allocate γ independent subtasks to the n computing nodes in the node cluster; the task allocation rule is: Assign(γ′,n′) represents the rule for assigning the γ′th independent subtask to the n′th computing node; H n′ -L n′ is the remaining capacity of the n′th computing node; λ is the sensitivity coefficient; the sensitivity coefficient λ is adjusted by the coefficient adjustment formula; the coefficient adjustment formula is: Among them, λ′ is the adjusted sensitivity coefficient; λ 0 is the initial sensitivity coefficient; α′ is the adjustment factor; R missed is the number of unfinished independent subtasks of the current computing node; R total The total number of independent subtasks assigned to the current computing node;

[0032] When the remaining capacity of the n′th computing node is greater than or equal to the priority of the γ′th independent subtask multiplied by the sensitivity coefficient λ, the γ′th independent subtask is assigned to the n′th computing node, otherwise it is not assigned;

[0033] Continuously monitor the operating status of the node cluster. When the current load of a computing node exceeds the maximum load threshold, dynamically adjust the independent subtasks assigned to the computing node, prioritize the migration of independent subtasks with low priority, and select the computing node with the lowest current load and the highest comprehensive node performance score among the remaining computing nodes as the newly assigned computing node.

[0034] Furthermore, each computing node executes the assigned independent subtask to obtain the optimized regional block method, which includes:

[0035] After each computing node receives an independent subtask, it executes the independent subtask on the assigned regional block. Through the parallel processing of node clusters, the generative adversarial network is used to perform mesh refinement, texture mapping and texture super-resolution optimization on the assigned regional blocks. The smoothing algorithm is applied to smooth the transition area of ​​the regional blocks, and the AI ​​algorithm is used to perform topological repair to automatically fill the topological defects in the regional blocks. After all computing nodes have completed all the received independent subtasks, the regional blocks on which each computing node has completed the independent subtask are collected to obtain the optimized regional blocks.

[0036] Furthermore, the method of submitting the optimized regional blocks to the management node for integration processing to generate a complete real-scene three-dimensional model includes:

[0037] Each computing node uploads all optimized regional blocks respectively and submits them to the management node in a unified manner; the management node integrates the topological structures of all optimized regional blocks and eliminates boundary gaps to merge them into a complete real-life 3D model.

[0038] Furthermore, the method of globally optimizing the real-scene three-dimensional model to generate a multi-resolution hierarchical model includes:

[0039] The real-scene 3D model is calibrated with texture mapping, scale calibration and brightness calibration to obtain a globally optimized real-scene 3D model; the globally optimized real-scene 3D model is imported into the multi-resolution level generation tool Simplifiedgon to automatically generate multi-resolution levels and simplified parameters; the automatic generation process of Simplifiedgon is started, and the globally optimized real-scene 3D model is simplified and rendered according to the generated simplified parameters to obtain a multi-resolution level model.

[0040] Furthermore, the method of converting the format of the multi-resolution hierarchical model and saving it to the storage node includes:

[0041] According to the requirements of the storage node, the multi-resolution hierarchical model is converted into a model file in the target format, and the model file is compressed using lossless compression technology; the compressed model file is backed up by the system and uploaded to the storage node using a high-performance transmission protocol.

[0042] The technical effects and advantages of the cluster batch optimization adjustment system for real-scene three-dimensional model processing of the present invention are as follows:

[0043] The present invention builds a multi-node cluster and makes full use of distributed computing resources. It can split large-scale data and complex computing tasks into multiple parallel tasks, significantly improving processing efficiency. Each node can focus on local processing of data and finally merge the results, effectively breaking through the bottleneck of single-machine computing and significantly accelerating the model generation, optimization and repair process. It combines the generative adversarial network in deep learning to perform texture synthesis and repair, and achieves higher-precision texture mapping and automatic repair of missing parts through cluster parallel training and reasoning. The AI ​​model can automatically identify and repair errors in texture mapping, avoid manual intervention in traditional methods, and improve processing accuracy.

[0044] The preset minimum point threshold is dynamically adjusted through the point threshold adjustment formula to optimize the termination condition, prevent the computing and storage overhead caused by excessive subdivision, and improve the resource utilization efficiency of the system; dynamically adjust the sensitivity coefficient to quickly adapt to changes in task and node status and reduce delays caused by unreasonable allocation; use an adaptive load balancing algorithm to adjust task allocation according to the performance, load conditions and characteristics of each computing node and the computing task, maximize the use of computing resources, and comprehensively improve the utilization of cluster resources, task execution efficiency and system stability; through dynamic allocation and adjustment strategies, it can adapt to complex and changeable computing scenarios, providing an efficient, stable and intelligent solution for large-scale data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic diagram of the structure of a cluster batch optimization and adjustment system for real-scene three-dimensional model processing according to the present invention;

[0046] Figure 2 A schematic flow chart of a cluster batch optimization and adjustment method for real-scene three-dimensional model processing according to the present invention; DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] Embodiment 1

[0049] See also Figure 1 As shown, the cluster batch optimization and adjustment system for real-scene three-dimensional model processing described in this embodiment includes:

[0050] Cluster architecture building module, used to build a node cluster including management nodes, storage nodes and n computing nodes;

[0051] Data acquisition module, used to collect point cloud data, texture image data and batch optimization task data;

[0052] The data processing module is used to pre-process the point cloud data and texture image data; the pre-processed point cloud data and texture image data are divided into m regional blocks according to space, and are distributed to n computing nodes of the node cluster for parallel processing;

[0053] The task scheduling module is used to decompose the batch optimization task data into γ independent subtasks; the adaptive load balancing algorithm is used to dynamically allocate the γ independent subtasks to the n computing nodes of the node cluster; each computing node executes the allocated independent subtask to obtain the optimized regional blocks;

[0054] The model optimization module is used to submit the optimized regional blocks to the management node for integration processing to generate a complete real-scene 3D model; perform global optimization on the real-scene 3D model to generate a multi-resolution hierarchical model;

[0055] The output module is used to convert the format of the multi-resolution hierarchical model and save it to the storage node; each module is connected by wired and / or wireless means.

[0056] The method of building a node cluster including a management node, a storage node, and n computing nodes includes:

[0057] Define the types and number of nodes in the node cluster, including one management node, one storage node and n computing nodes; divide the node functions according to the node types, define the functions of the management node as task scheduling and resource allocation, the function of the storage node as data storage, and the function of the n computing nodes as executing parallel computing tasks; configure performance parameters for the n computing nodes, including computing power, memory capacity, network bandwidth and current load.

[0058] Computing power is the core indicator of the performance parameters of n computing nodes, including the number of CPU cores and GPU performance; memory capacity determines the amount of data that a single computing node can accommodate during task execution, which is especially important for processing large-scale point cloud data or texture optimization; network bandwidth determines the speed of data transmission between n computing nodes, which directly affects the efficiency of distributed computing tasks; current load is used to evaluate the real-time operating status of n computing nodes and optimize task scheduling, including CPU usage, memory usage, task queue length and network occupancy; the following is an example configuration of a small node cluster containing 4 computing nodes:

[0059]

[0060] The point cloud data consists of v discrete points with attributes; the attributes include point coordinates, reflection intensity, color value, normal vector and density distribution; the texture image data includes texture images and texture mapping coordinates corresponding to the point cloud data; the batch optimization task data is the goal of optimizing the real-life 3D model, including mesh refinement and smoothing, topology repair, texture mapping and texture super-resolution optimization.

[0061] Methods for preprocessing point cloud data and texture image data include:

[0062] The deep learning model PointNet++ is used to perform denoising on the point cloud data, and the noise points and abnormal points in the point cloud data are identified and removed; cluster sampling is performed on the denoised point cloud data, and a clustering algorithm is used to identify and retain the cluster center points in the point cloud data; the coordinates of the retained cluster center points are aligned to obtain the preprocessed point cloud data; the texture image data is formatted in a unified manner, and a bilateral filtering algorithm is used to remove noise interference in the texture image data to obtain the preprocessed texture image data.

[0063] The method of dividing the preprocessed point cloud data and texture image data into m regional blocks according to space and distributing them to n computing nodes of the node cluster for parallel processing includes:

[0064] The preprocessed point cloud data is reconstructed using the Delaunay triangulation algorithm to generate a three-dimensional mesh model; the preprocessed texture image data is mapped to the three-dimensional mesh model through texture mapping coordinates to obtain a three-dimensional mesh model with texture; the three-dimensional mesh model with texture is spatially segmented using the octree algorithm and divided into m regional blocks. The specific method steps are as follows:

[0065] S51. For a 3D mesh model with texture, define an initial bounding box B 0 ; The range of the initial bounding box is determined by the extreme values ​​of the grid vertex coordinates of the 3D grid model with texture: B 0 =[x min ,x max ]×[y min ,y max ]×[z min ,z max ]; among them, B 0 is the initial bounding box; x min and x max are the minimum and maximum coordinate values ​​of all mesh vertices of the 3D mesh model with texture on the x-axis; y min and max are the minimum and maximum coordinate values ​​of all mesh vertices of the 3D mesh model with texture on the y-axis; min and z max They are the minimum and maximum coordinate values ​​of all mesh vertices of the 3D mesh model with texture on the z-axis;

[0066] S52, initial bounding box B 0 Perform recursive division, each time dividing into 8 sub-areas, each sub-area corresponds to a sub-cube; the current bounding box is preset to be divided into 8 sub-areas B 1 ,B 2 ,...,B 8,The spatial range and center point of each sub-region are calculated by the following formula:

[0067]

[0068] Among them, B i,j is the bounding box of the sub-region with index j at the recursive level i; The x-axis coordinate of the center point of the space currently enclosing the box area; The y-axis coordinate of the center point of the space currently enclosing the box area; is the z-axis coordinate of the spatial center point of the current bounding box area; i is the index of the current recursive layer of the bounding box; j is the sub-area index in the current recursive layer of the bounding box;

[0069] S53, repeat step S52 until the number of points contained in the sub-area is less than the preset minimum point threshold N min Stop when , and use the sub-region as a leaf node of the octree;

[0070] The preset minimum point threshold is dynamically adjusted through the point threshold adjustment formula. The point threshold adjustment formula is: Among them, N′ min is the minimum point threshold after dynamic adjustment; N min is the preset minimum point threshold; b is the number of discrete points in the three-dimensional grid model; β is the proportional factor that controls the increase in the preset minimum point threshold; k is the baseline adjustment factor that controls the basic value of the minimum point threshold; α is the logarithmic adjustment factor used to adjust the sensitivity of the change with the number of points;

[0071] For example, suppose there is a 3D mesh model with 10,000 vertices, and the coordinate range of the mesh vertices is: x∈[0,100], y∈[0,200], z∈[0,150]; the proportional factor β that controls the increase in the preset minimum point threshold is 0.01; the baseline adjustment factor k is 50; the logarithmic adjustment factor α is 2; N min The preset minimum point count threshold is 60; the range of the initial bounding box is determined by the extreme values ​​of the vertex coordinates: B 0 =[0,100]×[0,200]×[0,150]; the center point of each sub-area of ​​the initial bounding box is: After division, the range of each sub-area of ​​the initial bounding box is:

[0072] Subregion 1: x∈[0,50],y∈[0,100],z∈[0,75];

[0073] Subregion 2: x∈[0,50],y∈[0,100],z∈[75,150];

[0074] Subregion 3: x∈[0,50],y∈[100,200],z∈[0,75];

[0075] Sub-area 4: x∈[0,50], y∈[100,200], z∈[75,150]; the other 4 areas are calculated in the same way; the formula adjusted according to the point threshold is: That is, stop until the number of points contained in the sub-area is less than 163.

[0076] S54. According to the coordinate attribution of the mesh vertices, the meshes in the three-dimensional mesh model with texture are allocated to corresponding sub-regions, and then m regional blocks are obtained and evenly allocated to n computing nodes of the node cluster for parallel processing.

[0077] The method of dynamically allocating γ independent subtasks to n computing nodes of a node cluster using an adaptive load balancing algorithm includes:

[0078] According to the performance parameters of n computing nodes, the node performance of n computing nodes is evaluated by the node performance comprehensive scoring formula; the node performance comprehensive scoring formula is: n′ =ω 1 ·C n′ +ω 2 ·M n′ +ω 3 ·B n′ -ω 4 ·L n′ ; Among them, H n′ is the comprehensive performance score of the n′th computing node; C n′ is the computing power of the n′th computing node; M n′ is the memory capacity of the n′th computing node; B n′ is the network bandwidth of the n′th computing node; L n′ is the current load of the n′th computing node; n′ is the index of the computing node; ω 1 is the weight factor of computing power; ω 2 is the weight factor of memory capacity, ω 3 is the weight factor of network bandwidth; ω 4 is the weight factor of the current load;

[0079] For example, suppose there are three computing nodes with the following performance parameters:

[0080] Node Number Computing power Memory capacity Network bandwidth Current load Node 1 200 32 1000 50 Node 2 150 64 800 30 Node 3 300 16 1200 70

[0081] Assume ω 1 is 0.4; 2 =0.3,ω 3 is 0.2; 4is 0.1; according to the node performance comprehensive scoring formula, it can be concluded that: H 1 =0.4×200+0.3×32+0.2×1000-0.1×50=284.6;

[0082] H 2 =0.4×150+0.3×64+0.2×800-0.1×30=236.2;

[0083] H 3 =0.4×300+0.3×16+0.2×1200-0.1×70=357.8;

[0084] The priority of γ independent subtasks is calculated by the priority calculation formula; the priority calculation formula is: γ′ =μ 1 ·DA γ′ +μ 2 ·CO γ′ -μ 3 ·DE γ′ ; Among them, T γ′ is the priority of the γ′th independent subtask; DA γ′ The amount of data required to be processed for the γ′th independent subtask; CO γ′ The task complexity of the γ′th independent subtask is calculated; DE γ′ is the computation time required to complete the γ′th independent subtask; μ 1 is the weight of the impact of data volume on priority; μ 2 Calculate the weight of the task complexity on the priority; μ 3 To calculate the weight of the impact of time on priority;

[0085] Based on the adaptive load balancing algorithm, the task allocation rule is constructed to dynamically allocate γ independent subtasks to the n computing nodes in the node cluster; the task allocation rule is: Assign(γ′,n′) represents the rule for assigning the γ′th independent subtask to the n′th computing node; H n′ -L n′ is the remaining capacity of the n′th computing node; λ is the sensitivity coefficient, which is used to adjust the balance between task priority and node load; the sensitivity coefficient λ is adjusted by the coefficient adjustment formula; the coefficient adjustment formula is: Among them, λ′ is the adjusted sensitivity coefficient; λ 0 is the initial sensitivity coefficient; α′ is the adjustment factor used to control the sensitivity of response task delay; R missed is the number of unfinished independent subtasks of the current computing node; R total The total number of independent subtasks assigned to the current computing node;

[0086] Assume that the initial sensitivity coefficient is 1; the adjustment factor is 0.5; the number of unfinished independent subtasks of the current computing node is 10; the total number of independent subtasks assigned to the current computing node is 20; adjust the sensitivity coefficient λ through the coefficient adjustment formula, and the adjusted sensitivity coefficient can be obtained as follows:

[0087]

[0088] When the remaining capacity of the n′th computing node is greater than or equal to the priority of the γ′th independent subtask multiplied by the sensitivity coefficient λ, the γ′th independent subtask is assigned to the n′th computing node, otherwise it is not assigned;

[0089] Continuously monitor the operating status of the node cluster. When the current load of a computing node exceeds the maximum load threshold, dynamically adjust the independent subtasks assigned to the computing node, prioritize the migration of independent subtasks with low priority, and select the computing node with the lowest current load and the highest comprehensive node performance score among the remaining computing nodes as the newly assigned computing node.

[0090] Each computing node executes the assigned independent subtask, and the method of obtaining the optimized regional block includes:

[0091] After each computing node receives an independent subtask, it executes the independent subtask on the assigned regional block. Through the parallel processing of node clusters, the generative adversarial network is used to perform mesh refinement, texture mapping and texture super-resolution optimization on the assigned regional blocks. The smoothing algorithm is applied to smooth the transition area of ​​the regional blocks, and the AI ​​algorithm is used to perform topological repair to automatically fill the topological defects in the regional blocks. After all computing nodes have completed all the received independent subtasks, the regional blocks on which each computing node has completed the independent subtask are collected to obtain the optimized regional blocks.

[0092] Submitting the optimized regional blocks to the management node for integration processing to generate a complete real-scene 3D model includes:

[0093] Each computing node uploads all optimized regional blocks respectively and submits them to the management node in a unified manner; the management node integrates the topological structures of all optimized regional blocks and eliminates boundary gaps to merge them into a complete real-life 3D model.

[0094] The method of globally optimizing the real-scene 3D model and generating a multi-resolution hierarchical model includes:

[0095] The real-scene 3D model is calibrated with texture mapping, scale calibration and brightness calibration to obtain a globally optimized real-scene 3D model; the globally optimized real-scene 3D model is imported into the multi-resolution level generation tool Simplifiedgon to automatically generate multi-resolution levels and simplified parameters; the automatic generation process of Simplifiedgon is started, and the globally optimized real-scene 3D model is simplified and rendered according to the generated simplified parameters to obtain a multi-resolution level model.

[0096] Methods for converting the format of a multi-resolution hierarchical model and saving it to a storage node include:

[0097] According to the requirements of the storage node, the multi-resolution hierarchical model is converted into a model file in the target format, and the model file is compressed using lossless compression technology; the compressed model file is backed up by the system and uploaded to the storage node using a high-performance transmission protocol.

[0098] This embodiment builds a multi-node cluster and makes full use of distributed computing resources. It can split large-scale data and complex computing tasks into multiple parallel tasks, significantly improving processing efficiency; each node can focus on local processing of data and finally merge the results, effectively breaking through the bottleneck of single-machine computing and significantly accelerating the model generation, optimization and repair process; combined with the generative adversarial network in deep learning, texture synthesis and repair are performed, and higher-precision texture mapping and automatic repair of missing parts are achieved through cluster parallel training and reasoning; the AI ​​model can automatically identify and repair errors in texture mapping, avoid manual intervention in traditional methods, and improve processing accuracy;

[0099] The preset minimum point threshold is dynamically adjusted through the point threshold adjustment formula to optimize the termination condition, prevent the computing and storage overhead caused by excessive subdivision, and improve the resource utilization efficiency of the system; dynamically adjust the sensitivity coefficient to quickly adapt to changes in task and node status and reduce delays caused by unreasonable allocation; use an adaptive load balancing algorithm to adjust task allocation according to the performance, load conditions and characteristics of each computing node and the computing task, maximize the use of computing resources, and comprehensively improve the utilization of cluster resources, task execution efficiency and system stability; through dynamic allocation and adjustment strategies, it can adapt to complex and changeable computing scenarios, providing an efficient, stable and intelligent solution for large-scale data processing.

[0100] Embodiment 2

[0101] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides a cluster batch optimization adjustment method for real scene 3D model processing, including:

[0102] S1. Build a node cluster including a management node, a storage node, and n computing nodes;

[0103] S2, collect point cloud data, texture image data and batch optimization task data;

[0104] S3, preprocessing the point cloud data and texture image data; dividing the preprocessed point cloud data and texture image data into m regional blocks according to space, and distributing them to n computing nodes of the node cluster for parallel processing;

[0105] S4, decompose the batch optimization task data into γ independent subtasks; use the adaptive load balancing algorithm to dynamically allocate the γ independent subtasks to the n computing nodes of the node cluster; each computing node executes the allocated independent subtask to obtain the optimized regional blocks;

[0106] S5, submitting the optimized regional blocks to the management node for integration processing to generate a complete real-scene 3D model; performing global optimization on the real-scene 3D model to generate a multi-resolution hierarchical model;

[0107] S6. Convert the multi-resolution hierarchical model into a new format and save it to a storage node; each module is connected to another module via wired and / or wireless means.

[0108] Since the electronic device introduced in this embodiment is an electronic device used to implement a cluster batch optimization and adjustment system for real-scene three-dimensional model processing in the embodiment of this application, based on the cluster batch optimization and adjustment system for real-scene three-dimensional model processing introduced in the embodiment of this application, the technical personnel of this field can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application is not described in detail here. As long as the technical personnel of this field implement the electronic device used in the cluster batch optimization and adjustment system for real-scene three-dimensional model processing in the embodiment of this application, it belongs to the scope of protection of this application.

[0109] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0110] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A cluster batch optimization and adjustment system for real-scene three-dimensional model processing, characterized in that: include: Cluster architecture building module, used to build a node cluster including management nodes, storage nodes and n computing nodes; Data acquisition module, used to collect point cloud data, texture image data and batch optimization task data; A data processing module, used for preprocessing point cloud data and texture image data; The pre-processed point cloud data and texture image data are divided into m regional blocks according to space, and are distributed to n computing nodes of the node cluster for parallel processing; Task scheduling module, used to decompose batch optimization task data into γ independent subtasks; Adopting adaptive load balancing algorithm, γ independent subtasks are dynamically allocated to n computing nodes in the node cluster; each computing node executes the allocated independent subtask to obtain optimized regional blocks; The model optimization module is used to submit the optimized regional blocks to the management node for integration processing to generate a complete real-life 3D model; Perform global optimization on the real-scene 3D model to generate a multi-resolution hierarchical model; The output module is used to convert the format of the multi-resolution hierarchical model and save it to the storage node; each module is connected by wired and / or wireless means.

2. The cluster batch optimization and adjustment system for real-scene three-dimensional model processing according to claim 1, characterized in that: The method for constructing a node cluster including a management node, a storage node and n computing nodes includes: Define the types and number of nodes in the node cluster, including one management node, one storage node and n computing nodes; divide the node functions according to the node types, define the functions of the management node as task scheduling and resource allocation, the function of the storage node as data storage, and the function of the n computing nodes as executing parallel computing tasks; configure performance parameters for the n computing nodes, including computing power, memory capacity, network bandwidth and current load.

3. The cluster batch optimization and adjustment system for real-scene 3D model processing according to claim 2, characterized in that: The point cloud data consists of v discrete points with attributes; the attributes include point coordinates, reflection intensity, color value, normal vector and density distribution; the texture image data includes texture images and texture mapping coordinates corresponding to the point cloud data; the batch optimization task data is the goal of optimizing the real-life three-dimensional model, including mesh refinement and smoothing, topology repair, texture mapping and texture super-resolution optimization.

4. The cluster batch optimization and adjustment system for real-scene 3D model processing according to claim 3, characterized in that: The method for preprocessing point cloud data and texture image data comprises: The deep learning model PointNet++ is used to perform denoising on the point cloud data, and the noise points and abnormal points in the point cloud data are identified and removed; cluster sampling is performed on the denoised point cloud data, and a clustering algorithm is used to identify and retain the cluster center points in the point cloud data; the coordinates of the retained cluster center points are aligned to obtain the preprocessed point cloud data; the texture image data is formatted in a unified manner, and a bilateral filtering algorithm is used to remove noise interference in the texture image data to obtain the preprocessed texture image data.

5. The cluster batch optimization and adjustment system for real-scene three-dimensional model processing according to claim 4, characterized in that: The method of dividing the pre-processed point cloud data and texture image data into m regional blocks according to space and distributing them to n computing nodes of a node cluster for parallel processing includes: The preprocessed point cloud data is reconstructed using the Delaunay triangulation algorithm to generate a three-dimensional mesh model; the preprocessed texture image data is mapped to the three-dimensional mesh model through texture mapping coordinates to obtain a three-dimensional mesh model with texture; the three-dimensional mesh model with texture is spatially segmented using the octree algorithm and divided into m regional blocks. The specific method steps are as follows: S51. For the three-dimensional mesh model with texture, define an initial bounding box B0; the range of the initial bounding box is determined by the extreme values ​​of the mesh vertex coordinates of the three-dimensional mesh model with texture: B0=[x min ,x max ]×[y min ,y max ]×[z min ,z max ]; where B0 is the initial bounding box; x min and x max are the minimum and maximum coordinate values ​​of all mesh vertices of the 3D mesh model with texture on the x-axis; y min and max are the minimum and maximum coordinate values ​​of all mesh vertices of the 3D mesh model with texture on the y-axis; min and z max They are the minimum and maximum coordinate values ​​of all mesh vertices of the 3D mesh model with texture on the z-axis; S52, recursively divide the initial bounding box B0 into 8 sub-areas each time, each sub-area corresponds to a sub-cube; it is assumed that the current bounding box is divided into 8 sub-areas B1, B2, ..., B8, and the spatial range and center point of each sub-area are calculated by the following formula: Among them, B i,j is the bounding box of the sub-region with index j at the recursive level i; The x-axis coordinate of the center point of the space currently enclosing the box area; The y-axis coordinate of the center point of the space currently enclosing the box area; is the z-axis coordinate of the spatial center point of the current bounding box area; i is the index of the current recursive layer of the bounding box; j is the sub-area index in the current recursive layer of the bounding box; S53, repeat step S52 until the number of points contained in the sub-area is less than the preset minimum point threshold N min Stop when , and use the sub-region as a leaf node of the octree; The preset minimum point threshold is dynamically adjusted through the point threshold adjustment formula. The point threshold adjustment formula is: Among them, N′ min is the minimum point threshold after dynamic adjustment; N min is the preset minimum point threshold; b is the number of discrete points in the three-dimensional grid model; β is the proportional factor that controls the increase in the preset minimum point threshold; k is the baseline adjustment factor that controls the basic value of the minimum point threshold; α is the logarithmic adjustment factor used to adjust the sensitivity of the change with the number of points; S54. According to the coordinate attribution of the mesh vertices, the meshes in the three-dimensional mesh model with texture are allocated to corresponding sub-regions, and then m regional blocks are obtained and evenly allocated to n computing nodes of the node cluster for parallel processing.

6. The cluster batch optimization and adjustment system for real-scene 3D model processing according to claim 5, characterized in that: The method of dynamically allocating γ independent subtasks to n computing nodes of a node cluster using an adaptive load balancing algorithm includes: According to the performance parameters of n computing nodes, the node performance of n computing nodes is evaluated by the node performance comprehensive scoring formula; the node performance comprehensive scoring formula is: n′ =ω1·C n′ +ω2·M n′ +ω3·B n′ -ω4·L n′ ; Among them, H n′ is the comprehensive performance score of the n′th computing node; C n′ is the computing power of the n′th computing node; M n′ is the memory capacity of the n′th computing node; B n′ is the network bandwidth of the n′th computing node; L n′ is the current load of the n′th computing node; n′ is the index of the computing node; ω1 is the weight factor of computing power; ω2 is the weight factor of memory capacity; ω3 is the weight factor of network bandwidth; ω4 is the weight factor of current load; The priority of γ independent subtasks is calculated by the priority calculation formula; the priority calculation formula is: γ′ =μ1·DA γ′ +μ2·CO γ′ -μ3·DE γ′ ; Among them, T γ′ is the priority of the γ′th independent subtask; DA γ′ The amount of data required to be processed for the γ′th independent subtask; CO γ′ The task complexity of the γ′th independent subtask is calculated; DE γ′ is the computing time required to complete the γ′th independent subtask; μ1 is the weight of the impact of data volume on priority; μ2 is the weight of the impact of task computational complexity on priority; μ3 is the weight of the impact of computational time on priority; Based on the adaptive load balancing algorithm, the task allocation rule is constructed to dynamically allocate γ independent subtasks to the n computing nodes in the node cluster; the task allocation rule is: Assign(γ′,n′) represents the rule for assigning the γ′th independent subtask to the n′th computing node; H n′ -L n′ is the remaining capacity of the n′th computing node; λ is the sensitivity coefficient; the sensitivity coefficient λ is adjusted by the coefficient adjustment formula; the coefficient adjustment formula is: Among them, λ′ is the adjusted sensitivity coefficient; λ0 is the initial sensitivity coefficient; α′ is the adjustment factor; R missed is the number of unfinished independent subtasks of the current computing node; R total The total number of independent subtasks assigned to the current computing node; When the remaining capacity of the n′th computing node is greater than or equal to the priority of the γ′th independent subtask multiplied by the sensitivity coefficient λ, the γ′th independent subtask is assigned to the n′th computing node, otherwise it is not assigned; Continuously monitor the operating status of the node cluster. When the current load of a computing node exceeds the maximum load threshold, dynamically adjust the independent subtasks assigned to the computing node, prioritize the migration of independent subtasks with low priority, and select the computing node with the lowest current load and the highest comprehensive node performance score among the remaining computing nodes as the newly assigned computing node.

7. The cluster batch optimization and adjustment system for real-scene 3D model processing according to claim 6, characterized in that: The method for each computing node to execute the assigned independent subtask to obtain optimized regional blocks includes: After each computing node receives an independent subtask, it executes the independent subtask on the assigned regional block. Through the parallel processing of node clusters, the generative adversarial network is used to perform mesh refinement, texture mapping and texture super-resolution optimization on the assigned regional blocks. The smoothing algorithm is applied to smooth the transition area of ​​the regional blocks, and the AI ​​algorithm is used to perform topological repair to automatically fill the topological defects in the regional blocks. After all computing nodes have completed all the received independent subtasks, the regional blocks on which each computing node has completed the independent subtask are collected to obtain the optimized regional blocks.

8. The cluster batch optimization and adjustment system for real-scene 3D model processing according to claim 7, characterized in that: The method of submitting the optimized regional blocks to the management node for integration processing to generate a complete real-scene three-dimensional model includes: Each computing node uploads all optimized regional blocks respectively and submits them to the management node in a unified manner; the management node integrates the topological structures of all optimized regional blocks and eliminates boundary gaps to merge them into a complete real-life 3D model.

9. The cluster batch optimization and adjustment system for real-scene 3D model processing according to claim 8, characterized in that: The method of globally optimizing the real-scene three-dimensional model to generate a multi-resolution hierarchical model includes: The real-life 3D model is calibrated for texture mapping, scale and brightness to obtain a globally optimized real-life 3D model; the globally optimized real-life 3D model is imported into the multi-resolution layer generation tool Simplygon to automatically generate multi-resolution layers and simplified parameters; the automatic generation process of Simplygon is started, and the globally optimized real-life 3D model is simplified and rendered according to the generated simplified parameters to obtain a multi-resolution layer model.

10. The cluster batch optimization and adjustment system for real-scene three-dimensional model processing according to claim 9, characterized in that: The method of converting the format of the multi-resolution hierarchical model and saving it to the storage node includes: According to the requirements of the storage node, the multi-resolution hierarchical model is converted into a model file in the target format, and the model file is compressed using lossless compression technology; the compressed model file is backed up by the system and uploaded to the storage node using a high-performance transmission protocol.

Images