Three-dimensional model automatic generation method and system based on point cloud data

Abstract

The invention relates to the technical field of three-dimensional modeling, in particular to a three-dimensional model automatic generation method and system based on point cloud data, and the method comprises the steps: obtaining consumption-level point cloud data, employing a multi-scale feature recognition model to recognize multi-scale features in the consumption-level point cloud data, and dividing regions; setting a hierarchical processing mechanism to perform hierarchical processing on the consumption-level point cloud data to obtain production-level point cloud data; using a point cloud semantic segmentation model to identify functional regions and structural features of an object, segmenting production-level point cloud data into different semantic regions, and allocating reconstruction strategies and parameters for each semantic region to perform three-dimensional reconstruction; geometric parameters of a three-dimensional grid generated in the three-dimensional reconstruction process are detected in real time; when the geometric parameters do not meet the requirements of the target production process, automatically triggering the local grid correction model to carry out geometric adjustment; and establishing a multi-target evaluation model, adaptively adjusting each target weight according to the user-customized demand type and the selected production process, and generating a final three-dimensional model.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional modeling, and in particular to a method and system for automatically generating a three-dimensional model based on point cloud data. Background Art

[0002] In recent years, point cloud processing technologies based on deep learning have developed rapidly. Network structures such as PointNet, PointNet++, and DGCNN have made significant progress in tasks such as point cloud classification, segmentation, and feature extraction. However, existing deep learning methods mainly focus on the semantic understanding of point clouds and are still insufficient in high-precision reconstruction for manufacturing. Parametric modeling technology describes geometric shapes through mathematical parameters and supports flexible shape editing and modification, but most existing parametric methods require manual design of parameter structures and are difficult to automatically adapt to point cloud data of arbitrary shapes. In the field of CAD / CAM, multi-constraint optimization technology is widely used in design optimization and process planning, but how to effectively integrate these constraints into the point cloud reconstruction process to achieve the integration of design and manufacturing remains an open problem.

[0003] A comprehensive analysis of existing technologies shows that the following major technical problems exist in point cloud 3D reconstruction in C2M personalized customization scenarios: there is a lack of specialized processing methods for the characteristics of point cloud data from consumer-grade devices, making it difficult to convert low-quality inputs into production-level precision models; the model structure generated by existing methods is not conducive to personalized editing and lacks semantic partitioning and parametric representation; the reconstruction process lacks consideration of manufacturing constraints, and the generated models often do not meet production requirements; there is a lack of a unified framework to balance multiple goals such as reconstruction accuracy, editing flexibility, and production feasibility.

[0004] Therefore, in the C2M personalized customization scenario, how to automatically convert the low-quality point cloud data obtained by ordinary users through consumer-grade devices into a high-quality three-dimensional model that meets the constraints of industrial production processes and supports personalized editing operations is an urgent problem to be solved.

[0005] Therefore, a method and system for automatically generating three-dimensional models based on point cloud data are proposed. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and system for automatically generating three-dimensional models based on point cloud data, so as to automatically convert low-quality point cloud data obtained by ordinary users through consumer-grade devices into high-quality three-dimensional models that meet industrial production process constraints and support personalized editing operations.

[0007] To achieve the above object, the present invention provides the following technical solutions: A method for automatically generating a three-dimensional model based on point cloud data includes:

[0008] The consumer-level point cloud data acquired by the receiving consumer device is identified using a multi-scale feature recognition model, and the consumer-level point cloud data is divided into regions; A hierarchical processing mechanism is set to process the consumer-level point cloud data hierarchically to obtain production-level point cloud data; the hierarchical processing mechanism includes a precision-first processing layer, an adaptive balance processing layer, and an efficiency-first processing layer; The production-level point cloud data is segmented into different semantic regions using a point cloud semantic segmentation model to identify the functional regions and structural features of the object, and a reconstruction strategy and parameters are assigned to each semantic region according to the attributes of the different semantic regions for three-dimensional reconstruction; Based on a pre-set multi-process rule library, the geometric parameters of the three-dimensional mesh generated during the three-dimensional reconstruction process are detected in real time; when the geometric parameters are detected to be inconsistent with the requirements of the target production process, a local mesh correction model is automatically triggered for geometric adjustment to obtain a corrected mesh model; A multi-objective evaluation model is established to adaptively adjust the weights of each target according to the user's customized demand type and selected production process, and the corrected mesh model is optimized to generate a final three-dimensional model.

[0009] Preferably, the multi-scale feature recognition model includes: a multi-scale feature extraction unit that extracts geometric features of the consumer-level point cloud data at different scales; a feature classification unit that classifies the extracted geometric features, including edge features, corner features, curved surface change features, and flat region features; a feature importance evaluation unit that calculates the importance weight of each geometric feature based on feature type, local geometric complexity, and manufacturing criticality; and a key region positioning unit that regionally divides the consumer-level point cloud data according to the importance weight and spatial continuity constraints, including key regions, transition regions, and flat regions.

[0010] Preferably, the precision-first processing layer performs precision-first processing on the identified key regions; the precision-first processing uses densification sampling and multi-view registration strategies to increase the point cloud density of the consumer-level point cloud data in the key regions to N times the original density while preserving the tiny geometric features, to obtain high-precision point cloud data; The adaptive balance processing layer performs adaptive balance processing on the identified transition regions; the adaptive balance processing uses medium-density sampling and adaptive parameter adjustment strategies to dynamically adjust the processing precision according to the distance from adjacent key regions and the local geometric complexity, to obtain balanced point cloud data; The efficiency-first processing layer performs efficiency-first processing on the identified flat regions; the efficiency-first processing uses sparse sampling, fast interpolation, and batch processing strategies to reduce the point cloud density of the consumer-level point cloud data in the flat regions to M% of the original density while ensuring basic accuracy, to obtain high-efficiency point cloud data; The hierarchical processing mechanism further comprises an adaptive fusion unit; the adaptive fusion unit adaptively fuses the high-precision point cloud data, the balanced point cloud data and the high-efficiency point cloud data to obtain production-level point cloud data.

[0011] Preferably, the point cloud semantic segmentation model comprises: a structure feature extraction unit for extracting structure features in the production-level point cloud data, including local geometric features and context features; a functional area identification unit for identifying functional areas of an object according to the structure features by using a pre-trained semantic classifier, including load-bearing structure areas, decorative areas, connection interface areas and functional opening areas; a region attribute analysis unit for analyzing attributes of each semantic area, including geometric parameters, functional attributes and manufacturing constraints; a reconstruction strategy allocation unit for allocating corresponding reconstruction strategies and parameters to each functional area according to the attributes of each semantic area; the parameters include grid density parameters and error tolerance parameters.

[0012] Preferably, the local grid correction model comprises: a three-dimensional reconstruction unit for performing three-dimensional reconstruction based on the reconstruction strategies and parameters allocated to each functional area by the point cloud semantic segmentation model; a geometric parameter real-time detection unit for real-time calculation of geometric parameters of a three-dimensional grid generated in the three-dimensional reconstruction process, including wall thickness, inclination angle, fillet radius and aperture size of the three-dimensional grid; a local grid correction triggering unit for comparing the detected geometric parameters with 3D printing rules, numerical control machining rules and injection molding rules in a preset multi-process rule library to identify rule violation areas that do not meet the requirements of a target production process; a geometric adjustment unit for automatically selecting a local grid correction strategy for geometric adjustment according to the rule violation type and severity of the rule violation areas to obtain a corrected grid model.

[0013] Preferably, the process of generating the final three-dimensional model comprises: establishing a multi-objective evaluation model based on geometric accuracy, manufacturing feasibility, material usage, processing cost and functional integrity; dynamically adjusting weight coefficients of each evaluation objective according to a user customization requirement type and a selected process type to generate a personalized weight configuration; weighting each evaluation objective based on the personalized weight configuration to obtain a multi-objective function; solving an optimal solution set of the multi-objective function by using a multi-objective optimization algorithm, and selecting a final three-dimensional model scheme from the optimal solution set in combination with user preferences and engineering constraints to optimize and correct the grid model and generate the final three-dimensional model.

[0014] Preferably, a three-dimensional model automatic generation system based on point cloud data comprises: A consumer-level feature identification module is configured to receive consumer-level point cloud data acquired by a consumer-level device, identify multi-scale features in the consumer-level point cloud data by using a multi-scale feature identification model, and divide regions. The production-level point cloud generation module is configured to set a hierarchical processing mechanism to process the consumer-level point cloud data hierarchically to obtain production-level point cloud data; the hierarchical processing mechanism comprises a precision-first processing layer, an adaptive balance processing layer and an efficiency-first processing layer; The semantic modeling module is configured to identify the functional areas and structural features of the object by using a point cloud semantic segmentation model, segment the production-level point cloud data into different semantic areas, assign a reconstruction strategy and parameters to each semantic area according to the attributes of the different semantic areas, and perform three-dimensional reconstruction; The local grid correction module is configured to detect the geometric parameters of the three-dimensional grid generated in the three-dimensional reconstruction process in real time based on a preset multi-process rule library; when it is detected that the geometric parameters do not meet the requirements of the target production process, a local grid correction model is automatically triggered to perform geometric adjustment to obtain a corrected grid model; The three-dimensional model generation module is configured to establish a multi-target evaluation model, adaptively adjust the weights of each target according to the user's customized demand type and selected production process, optimize the corrected grid model, and generate a final three-dimensional model.

[0015] Compared with the prior art, the present application has the following advantages: 1、The hierarchical processing mechanism is used to realize intelligent balance optimization of processing efficiency and geometric precision. The multi-scale feature recognition model can automatically identify multi-scale features and important areas in the point cloud data, and then the three-layer processing mechanism of precision-first, adaptive balance and efficiency-first is used for differential processing, which can shorten the overall processing time on the premise of ensuring the accuracy of the key area, and significantly improve the conversion efficiency from consumer-level data to production-level model.

[0016] 2、The local grid correction model integrates a multi-process rule library and a real-time detection and correction mechanism, which can detect the geometric parameters of the generated grid in real time during the three-dimensional reconstruction process. When it is found that the generated grid does not meet the requirements of the 3D printing, numerical control machining or injection molding process, the corresponding correction strategy is automatically triggered to perform geometric adjustment, which can solve the key problem that the generated three-dimensional model in the prior art cannot be directly used for manufacturing, improve the success rate of manufacturing the generated three-dimensional model, and shorten the cycle time from design to production.

[0017] 3、The application realizes personalized automatic optimization for different user needs and production processes through a multi-objective evaluation model and an adaptive weight adjustment mechanism, greatly improving the intelligent level and user friendliness of the system. The multi-objective evaluation model of the application can automatically identify the type of user's customized needs and the type of selected production process, and then intelligently adjust the weight coefficient of the evaluation dimension. By using a multi-objective optimization algorithm to solve the optimal solution set, one-key automatic generation of a three-dimensional model is realized, which can generate an optimal three-dimensional model that meets the user's individual needs and adapts to the specific production process, improving user satisfaction and the comprehensive performance of the generated model in multiple evaluation dimensions. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A flowchart of a three-dimensional model automatic generation method based on point cloud data provided by an embodiment of the application is shown in the figure. Figure 2 A flowchart of generating production-level point cloud data provided by an embodiment of the application is shown in the figure. Figure 3 A structure diagram of a local grid correction model provided by an embodiment of the application is shown in the figure. Figure 4 A structure diagram of a three-dimensional model automatic generation system based on point cloud data provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0020] The application proposes a three-dimensional model automatic generation method and system based on point cloud data, which can automatically convert low-quality point cloud data obtained by ordinary users through consumer-level devices into high-quality three-dimensional models that meet industrial production process constraints and support individualized editing operations. To illustrate that the method of the application can reduce the quality of point cloud data to high-quality three-dimensional models that meet industrial production process constraints and support individualized editing operations, the effectiveness of the application will be illustrated from two embodiments below.

[0021] Embodiment one: In this embodiment, the method proposed by the application is used to automatically generate a three-dimensional model from point cloud data obtained by a user through a mobile phone scan. The user scanned a customized tea cup using the LiDAR sensor of an iPhone 14 Pro and obtained consumer-level point cloud data containing about 150,000 points. Figure 1This is a specific flow chart of the system of the present invention, including: receiving consumer-grade point cloud data acquired by consumer-grade devices, using a multi-scale feature recognition model to identify multi-scale features in the consumer-grade point cloud data, and dividing the area; setting a hierarchical processing mechanism to perform hierarchical processing on the consumer-grade point cloud data to obtain production-grade point cloud data; using a point cloud semantic segmentation model to identify the functional areas and structural features of the object, dividing the production-grade point cloud data into different semantic areas, assigning reconstruction strategies and parameters to each semantic area according to the attributes of different semantic areas, and performing three-dimensional reconstruction; based on a preset multi-process rule library, real-time detection of the geometric parameters of the three-dimensional grid generated during the three-dimensional reconstruction process; when it is detected that the geometric parameters do not meet the target production process requirements, automatically triggering the local grid correction model to perform geometric adjustments to obtain a corrected grid model; establishing a multi-objective evaluation model, adaptively adjusting the weights of each target according to the user's customized demand type and the selected production process, optimizing the corrected grid model, and generating a final three-dimensional model. The following is based on Figure 1 The following content is described: Receive consumer-grade point cloud data acquired by consumer-grade devices, use a multi-scale feature recognition model to identify multi-scale features in the consumer-grade point cloud data, and divide the area; The multi-scale feature recognition model includes: a multi-scale feature extraction unit, which extracts geometric features of consumer-grade point cloud data at different scales; a feature classification unit, which classifies the extracted geometric features, including edge features, corner features, surface change features, and flat area features; a feature importance evaluation unit, which calculates the importance weight of each geometric feature based on feature type, local geometric complexity, and manufacturing criticality; and a key area positioning unit, which divides the consumer-grade point cloud data into regions according to the importance weight and spatial continuity constraints, including key regions, transition regions, and flat regions.

[0022] Specifically, the consumer-grade point cloud data obtained by the user through the mobile phone LiDAR sensor is received, and the point cloud density is about 800-1200 points per square centimeter.

[0023] The multi-scale feature extraction unit is based on a multi-scale neighborhood analysis method with radius search. Three different search radii r1=2mm, r2=5mm, and r3=10mm are set to extract local, medium, and global scale geometric features, respectively. For the tea cup model, the local curvature variation feature of the handle connection is extracted at the r1 scale, the medium smooth transition feature of the cup surface is identified at the r2 scale, and the global geometric shape feature of the overall profile is captured at the r3 scale. By calculating the maximum principal curvature κ1, the minimum principal curvature κ2, and the Gaussian curvature K=κ1×κ2 of each point in the consumer-level point cloud data, a feature vector F=[κ1, κ2, K, ∇n] is generated; wherein the maximum principal curvature κ1 and the minimum principal curvature κ2 are obtained by local surface fitting and eigenvalue decomposition; ∇n is the normal vector change rate, which is obtained by calculating the average angle difference between the normal vector of the current point and the normal vectors of other points in the neighborhood.

[0024] The feature classification unit adopts a classification strategy based on threshold and machine learning. The geometric feature threshold is set: the edge feature judgment condition is |∇n|>0.3 and K>0.1; the corner feature judgment condition is κ1>0.5 and κ2>0.5; the surface variation feature judgment condition is 0.05<K<0.1; the flat area feature judgment condition is K<0.05 and |∇n|<0.1. Using the geometric feature threshold, 1847 edge feature points (mainly concentrated in the tea cup rim and handle profile), 423 corner feature points (tea cup handle connection and bottom corner), 8736 surface variation feature points (tea cup body surface), and about 130,000 flat area feature points are successfully identified.

[0025] The feature importance evaluation unit is based on feature type, local geometric complexity, and manufacturing criticality. The geometric complexity is obtained by calculating the local curvature variation rate and the standard deviation of the surface normal vector gradient, reflecting the degree of change in geometric shape; the manufacturing criticality is determined by analyzing the influence of the region on the functionality, structural strength, and assembly accuracy of the final product, and is obtained by using expert knowledge base and historical data statistics. The importance weight of each geometric feature is calculated using a weighted scoring formula, wherein the weighted weights of the weighted scoring formula are obtained according to expert experience. After calculation, the tea cup handle connection area obtains the highest importance weight of 0.95, the rim edge area weight is 0.88, the cup surface area weight is 0.45, and the bottom flat area weight is 0.15.

[0026] The key region positioning unit divides the region according to the importance weight and the spatial continuity constraint. A three-level division strategy based on a weight threshold is adopted: a key region (importance weight ≥ 0.7), a transition region (0.3 ≤ importance weight < 0.7), and a flat region (importance weight < 0.3). At the same time, a region growing algorithm is adopted to ensure spatial continuity, that is, the divided region remains connected in three-dimensional space, avoiding isolated points or broken regions. Finally, the tea cup point cloud is divided into: a key region (handle connection region and cup edge, about 21,000 points), a transition region (cup surface curved surface region, about 87,000 points), and a flat region (most of the bottom and inner wall region, about 42,000 points).

[0027] By adopting a multi-scale feature recognition model, geometric features can be fully extracted at different scales, and the importance can be evaluated by combining the type, complexity and manufacturing criticality, so as to accurately divide the key, transition and flat regions, and provide accurate basis for subsequent differentiated processing and semantic modeling. The model significantly improves the accuracy and robustness of region recognition, ensures that the key region is prioritized in the subsequent reconstruction process, and helps to generate a high-quality three-dimensional model that meets the structural and functional requirements and process adaptability.

[0028] Further, a hierarchical processing mechanism is set to process the consumer-level point cloud data to obtain production-level point cloud data; the hierarchical processing mechanism includes a precision-first processing layer, an adaptive balance processing layer and an efficiency-first processing layer; with reference to Figure 2 ; The precision-first processing layer performs precision-first processing on the identified key region; the precision-first processing adopts a densification sampling and multi-view registration strategy to increase the point cloud density of the consumer-level point cloud data of the key region to N times the original density, and retains the micro-geometric features, to obtain high-precision point cloud data; The adaptive balance processing layer performs adaptive balance processing on the identified transition region; the adaptive balance processing adopts a medium-density sampling and adaptive parameter adjustment strategy to dynamically adjust the processing precision according to the distance from the adjacent key region and the local geometric complexity, to obtain balanced point cloud data; The efficiency-first processing layer performs efficiency-first processing on the identified flat region; the efficiency-first processing adopts a sparse sampling, fast interpolation and batch processing strategy to reduce the point cloud density of the consumer-level point cloud data of the flat region to M% of the original density on the premise of ensuring basic accuracy, to obtain high-efficiency point cloud data; The hierarchical processing mechanism further includes an adaptive fusion unit; the adaptive fusion unit adaptively fuses the high-precision point cloud data, the balanced point cloud data and the high-efficiency point cloud data to obtain production-level point cloud data.

[0029] Specifically, a hierarchical processing mechanism is set to differentiate the identified different regions: The precision-first processing layer processes the key regions (handle connection region and cup mouth edge) with precision-first processing. A densification sampling strategy is adopted. First, the missing data points are completed by cubic spline interpolation based on adjacent points. Then, the ICP algorithm is used for multi-view accurate registration, with registration error controlled within 0.1 mm. Through multi-view registration and interpolation subdivision, the point cloud density of the key region is increased from the original 900 points / cm 2 to 4500 points / cm 2 (N=5 times). For the complex curved surface of the handle connection, a subdivision algorithm based on normal vector constraint is used to preserve the tiny geometric features, ensuring that the key manufacturing parameters such as the fillet radius of the connection are accurately reconstructed.

[0030] The adaptive balance processing layer processes the transition region (cup body surface) with adaptive balance processing. A medium-density sampling strategy is adopted, and the processing precision is dynamically adjusted according to the distance to the adjacent key region and the local geometric complexity . The processing precision adjustment formula is: wherein, is the basic processing precision, is the distance influence coefficient, is the distance attenuation parameter, is the complexity influence coefficient. According to the calculation, the processing precision of the cup body surface near the handle connection is increased to 3200 points / cm 2 , while the middle part of the cup body far from the key region maintains a moderate accuracy of 1800 points / cm 2 .

[0031] The efficiency-first processing layer processes the flat regions (bottom and inner wall) with efficiency-first processing. A sparse sampling, fast interpolation, and batch processing strategy is adopted to reduce the point cloud density to 35% of the original density (M=35). For large-area flat regions, a plane fitting algorithm is used to represent the region as a mathematical plane equation, with fitting error controlled within 0.05 mm. A fast bilinear interpolation algorithm is used to generate regularly distributed sparse point clouds on the plane, significantly improving processing efficiency while ensuring basic geometric accuracy to meet manufacturing requirements.

[0032] The adaptive fusion unit uses a weighted average and transition smoothing algorithm to perform gradual transition at the boundaries of different processing precision regions, avoiding density jumps and geometric discontinuities. At the same time, a 5mm transition zone is set in the boundary region, and a Gaussian weight function is used to achieve smooth transition. The final production-level point cloud data has a total of about 185,000 points, an increase of 23% compared to the consumer-level point cloud data, a 5-fold increase in precision in the key region, and a 65% reduction in overall processing time compared to uniform high-precision processing.

[0033] Through the three-layer processing mechanism of precision priority, adaptive balance and efficiency priority, combined with dense sampling, multi-view registration and sparse interpolation strategy, the accuracy and processing efficiency of point cloud data can be effectively balanced according to the regional characteristics. Especially in the key area, the accuracy is guaranteed while the overall computational overhead is significantly reduced, and the seamless integration of different levels of data is realized through adaptive fusion, finally obtaining production-level point cloud data with high precision, high efficiency and manufacturing feasibility.

[0034] Further, the functional areas and structural features of the object are identified by using a point cloud semantic segmentation model, the production-level point cloud data is segmented into different semantic areas, and the reconstruction strategy and parameters are allocated to each semantic area according to the attributes of different semantic areas for three-dimensional reconstruction. The point cloud semantic segmentation model comprises: a structural feature extraction unit for extracting structural features in the production-level point cloud data, including local geometric features and context features; a functional area identification unit for identifying functional areas of the object according to the structural features by using a pre-trained semantic classifier, including load-bearing structural areas, decorative areas, connection interface areas and functional opening areas; a region attribute analysis unit for analyzing the attributes of each semantic area, including geometric parameters, functional attributes and manufacturing constraints; a reconstruction strategy allocation unit for allocating corresponding reconstruction strategies and parameters to each functional area according to the attributes of each semantic area; the parameters include grid density parameters and error tolerance parameters.

[0035] Specifically, the structural feature extraction unit adopts a deep learning model based on the existing PointNet++ network architecture to extract multi-level structural features of the point cloud. The network contains 4 Set Abstraction layers, which extract features of different resolutions with a down-sampling rate of 0.1, 0.2, 0.4 and 0.8, respectively. For the cup model, the first layer extracts local surface normal and curvature features, the second layer identifies edge and connection line features, the third layer captures overall shape contour, and the fourth layer understands global structural relationship. At the same time, an attention mechanism is introduced to allocate higher weights to important feature areas, improving the feature representation ability of key areas.

[0036] The functional area recognition unit divides the tea cup into functional areas using a pre-trained semantic classifier. The semantic classifier is trained on a dataset containing 50,000 different tea cup models. The specific training process is as follows: first, the training samples are expanded to 200,000 through data augmentation techniques (including rotation, scaling, noise addition, etc.), then the cross-entropy loss function and Adam optimizer are used for end-to-end training, the learning rate is set to 0.001, the batch size is 32, and the training period is 200 epochs. The network extracts global and local features of the point cloud through multiple layers of perception and maximum pooling operations, and combines spatial context information for semantic segmentation prediction. To improve the generalization ability of the model, Dropout regularization (probability 0.5) and early stopping mechanism are introduced during training. When the validation set accuracy does not improve for 10 consecutive epochs, training is stopped. The final trained semantic classifier has an overall segmentation accuracy of 94.2% on the test set, which can effectively identify the various functional areas of the tea cup and provide accurate semantic information for subsequent reconstruction strategy allocation.

[0037] Four main functional areas are identified: load-bearing structure area (cup body main body, accounting for 68% of the total point cloud), decorative area (cup surface texture area, accounting for 15%), connection interface area (handle connection, accounting for 12%), and functional opening area (cup opening area, accounting for 5%). Each area is assigned a corresponding semantic label and functional attribute description.

[0038] The area attribute analysis unit conducts in-depth attribute analysis of each semantic area. For the load-bearing structure area, the wall thickness distribution (2.5-3.2mm), structural strength requirements, and material usage are analyzed; for the decorative area, the surface quality requirements and detail retention level are analyzed; for the connection interface area, the stress concentration points, fillet radius, and connection strength are analyzed; for the functional opening area, the edge smoothness and dimensional accuracy requirements are analyzed. Based on the results of the area attribute analysis, a complete attribute profile is established for each area, including geometric parameters, functional attributes, and manufacturing constraints.

[0039] The reconstruction strategy allocation unit allocates differentiated reconstruction strategies and parameters to different areas based on the attributes of each semantic area. For example, the load-bearing structure area adopts a high-strength reconstruction strategy, with a grid density of 3000 faces / cm² and an error tolerance of ±0.1mm; the decorative area adopts a high-surface quality strategy, with a grid density of 4500 faces / cm² and an error tolerance of ±0.05mm; the connection interface area adopts an ultra-high precision strategy, with a grid density of 6000 faces / cm² and an error tolerance of ±0.03mm; the functional opening area adopts an edge optimization strategy, focusing on optimizing edge smoothness and roundness.

[0040] The point cloud semantic segmentation model realizes the accurate identification of the functional areas of the object by structural feature extraction and semantic classification, and assigns individualized reconstruction strategies according to the geometric and manufacturing properties of different areas, thereby significantly improving the correspondence between the model structure and function. The model makes the generated three-dimensional model have higher semantic integrity and functional adaptability, and provides a basis for subsequent geometric adjustment and manufacturing matching, thereby enhancing the intelligent perception and structural understanding ability of the automatic modeling system for complex objects.

[0041] Further, the geometric parameters of the three-dimensional mesh generated in the three-dimensional reconstruction process are detected in real time based on a preset multi-process rule library; when it is detected that the geometric parameters do not meet the requirements of the target production process, a local mesh correction model is automatically triggered to perform geometric adjustment to obtain a corrected mesh model; the corrected mesh model is referenced Figure 3 ; The local mesh correction model comprises: a three-dimensional reconstruction unit that performs three-dimensional reconstruction based on the reconstruction strategies and parameters assigned by the point cloud semantic segmentation model for each functional area; a geometric parameter real-time detection unit that calculates the geometric parameters of the three-dimensional mesh generated in the three-dimensional reconstruction process in real time, including the wall thickness, inclination angle, fillet radius and aperture size of the three-dimensional mesh; a local mesh correction triggering unit that compares the detected geometric parameters with the 3D printing rules, numerical control machining rules and injection molding rules in the preset multi-process rule library to identify rule violation areas that do not meet the requirements of the target production process; and a geometric adjustment unit that automatically selects a local mesh correction strategy for geometric adjustment according to the rule violation type and severity of the rule violation area to obtain a corrected mesh model.

[0042] Specifically, the three-dimensional reconstruction unit performs initial three-dimensional reconstruction based on the aforementioned assigned reconstruction strategies. The existing Marching Cubes algorithm is used to generate a basic mesh, and then Loop subdivision and Laplace smoothing are performed for mesh optimization.

[0043] The geometric parameter real-time detection unit calculates the geometric parameters of the three-dimensional mesh in real time during the reconstruction process. A parallel computing architecture is adopted to immediately perform geometric detection on each newly generated mesh patch. The main detection parameters include: the wall thickness calculation adopts the ray casting method with a detection accuracy of 0.01 mm; the inclination angle is calculated by the included angle between the patch normal vector and the vertical direction; the fillet radius is obtained by the local curvature fitting algorithm; and the aperture size is obtained by boundary recognition and distance measurement. The detection found that there is a 0.8 mm thin wall (which does not meet the 3D printing minimum wall thickness requirement of 1.0 mm) at the connection of the handle, the inclination angle of the cup body side wall reaches 85° (which exceeds the injection molding limit of 80°), and there is a 0.3 mm sharp fillet (which does not meet the minimum fillet requirement of 0.5 mm for numerical control machining) at the bottom.

[0044] The local grid correction triggering unit compares the detection result with a preset multi-process rule library in real time. The preset multi-process rule library includes three types of process rules: 3D printing rules (minimum wall thickness 1.0 mm, maximum overhang angle 45°, minimum feature size 0.4 mm); numerical control machining rules (minimum fillet radius 0.5 mm, minimum slot width 2.0 mm, maximum depth-to-diameter ratio 5:1); injection molding rules (draft slope ≥ 1°, wall thickness uniformity deviation ≤ 20%, minimum draft angle 1.5°). The system detects 7 violation areas, and marks the violation type, severity (high / medium / low) and impact range.

[0045] The geometry adjustment unit automatically selects a correction strategy according to the violation type. For the 0.8 mm thin wall thickness problem at the handle connection, a bidirectional offset algorithm is used to adjust the wall thickness of the handle connection to 1.2 mm; for the 85° large inclination angle problem of the cup body side wall, a gradual adjustment strategy is used to correct the cup body side wall angle to 78°; for the 0.3 mm sharp corner of the bottom, a curve fitting algorithm is used to expand the bottom corner radius to 0.6 mm.

[0046] By integrating the real-time detection of geometric parameters and the multi-process rule comparison mechanism, violation areas that do not meet manufacturing requirements can be identified and located in time during the three-dimensional reconstruction process, and then local geometry adjustment is automatically triggered to ensure that the generated model meets the precision and structural specifications of various manufacturing processes such as 3D printing, numerical control machining and injection molding. The comparison mechanism effectively improves the manufacturing adaptability and one-time molding success rate of the three-dimensional model, reduces rework and manual intervention, and shortens the overall cycle from product design to production.

[0047] Further, a multi-objective evaluation model is established, and the target weights are adjusted adaptively according to the user's customized demand type and selected production process to optimize the corrected grid model and generate the final three-dimensional model.

[0048] The process of generating the final three-dimensional model includes: establishing a multi-objective evaluation model based on geometric precision, manufacturing feasibility, material usage, processing cost and functional integrity; dynamically adjusting the weight coefficients of each evaluation target according to the user's customized demand type and selected process type to generate personalized weight configuration; weighting each evaluation target based on the personalized weight configuration to obtain a multi-objective function; using a multi-objective optimization algorithm to solve the optimal solution set of the multi-objective function, and selecting the final three-dimensional model scheme from the optimal solution set in combination with user preferences and engineering constraints to optimize the corrected grid model and generate the final three-dimensional model.

[0049] Specifically, the multi-objective evaluation model is established: the geometric accuracy is calculated by the average deviation from the original point cloud; the manufacturing feasibility is evaluated by the process rule compliance rate; the material usage is calculated by the volume; the processing cost is evaluated by the process complexity and time; and the functional integrity is calculated by the key feature retention rate.

[0050] User demand identification and weight adjustment: the user selects the demand type of "personal customization + 3D printing", and the system automatically adjusts the weight configuration. For the personal customization demand, the geometric accuracy weight is increased to 0.35, and the functional integrity weight is increased to 0.25; for the 3D printing process, the manufacturing feasibility weight is increased to 0.25, the processing cost weight is appropriately reduced to 0.1, and the material usage weight remains 0.05. The final weight configuration is [0.35, 0.25, 0.05, 0.10, 0.25].

[0051] Multi-objective optimization solution: the existing improved NSGA-II genetic algorithm is used to solve the multi-objective optimization problem. The population size is set to 200, the evolution number is set to 150, the crossover probability is set to 0.9, and the mutation probability is set to 0.1. After optimization calculation, an optimal solution set containing 47 non-dominated solutions is generated. Combined with user preferences (giving priority to accuracy and functional integrity), the scheme with the highest comprehensive score is selected from the optimal solution set to optimize the modified grid model, and the final three-dimensional model is obtained.

[0052] Through the multi-objective evaluation model combined with user customization demand and process type dynamic weight configuration adjustment, and through the multi-objective optimization algorithm to quickly solve the optimal solution set, the customized automatic generation of three-dimensional model is realized. The multi-objective evaluation model not only considers multi-dimensional indexes such as geometric accuracy, manufacturing cost and functional integrity, but also integrates user preferences and engineering constraints, improves the comprehensive performance of the model between accuracy, function and manufacturing, and significantly enhances the individual response ability and intelligent optimization level of the system.

[0053] The present application realizes the whole-process intelligent conversion from consumer-level point cloud data to high-quality three-dimensional model by constructing a complete automatic process covering multi-scale feature recognition, hierarchical point cloud processing, semantic region reconstruction, process rule detection and multi-objective optimization. The method not only improves the point cloud processing efficiency, but also significantly enhances the manufacturing adaptability and individualized reconstruction ability of the model, effectively solves the technical bottleneck that low-quality data is difficult to support industrial production in traditional technology, realizes the automatic conversion of low-quality point cloud data obtained by ordinary users through consumer-level equipment into high-quality three-dimensional model which meets the constraints of industrial production process and supports individualized editing operation, has strong universality and engineering practical value.

[0054] Example two: In embodiment one, the method of the present application successfully realizes the automatic conversion of low-quality point cloud data obtained by ordinary users through consumer-grade devices into high-quality three-dimensional models that meet the constraints of industrial production processes and support personalized editing operations. To further verify the effectiveness of the present application, a three-dimensional model automatic generation system based on point cloud data is proposed in the embodiments of the present application to automatically generate three-dimensional models from point cloud data obtained by another user.

[0055] A three-dimensional model automatic generation system based on point cloud data, the specific structure diagram is shown in Figure 4 A consumer-grade feature recognition module is configured to receive consumer-grade point cloud data obtained by a consumer-grade device, identify multi-scale features in the consumer-grade point cloud data using a multi-scale feature recognition model, and divide regions. The multi-scale feature recognition model includes a multi-scale feature extraction unit that extracts geometric features of the consumer-grade point cloud data at different scales, a feature classification unit that classifies the extracted geometric features, including edge features, corner features, curved surface change features, and flat region features, a feature importance evaluation unit that calculates the importance weight of each geometric feature based on feature type, local geometric complexity, and manufacturing criticality, and a key region positioning unit that divides regions of the consumer-grade point cloud data according to the importance weight and spatial continuity constraints, including key regions, transition regions, and flat regions.

[0056] Further, after receiving the consumer-grade point cloud data, a quality evaluation model is established based on point cloud completeness, noise level, density uniformity, and key feature recognizability. When the quality evaluation score is lower than a preset threshold, an automatic data preprocessing enhancement module is triggered to use existing deep learning-based point cloud denoising algorithms, intelligent missing data completion algorithms, and density adaptive balancing algorithms to preprocess and optimize the consumer-grade point cloud data, and then the data quality is improved before entering the multi-scale feature recognition process.

[0057] By introducing a quality evaluation model and an automatic preprocessing enhancement mechanism after receiving the consumer-grade point cloud data, the problems of insufficient completeness, noise interference, density non-uniformity, and missing key structures in the original point cloud can be effectively identified and processed, thereby significantly improving the stability and accuracy of the subsequent modeling process. When the quality evaluation score is lower than the threshold, the system can automatically call deep learning-driven denoising, completion, and density balancing algorithms to adaptively enhance the point cloud, effectively avoiding feature recognition bias and reconstruction errors caused by low-quality input. This mechanism realizes active protection of modeling quality from the data source, enhances the adaptability of the system to diverse data environments, and improves the robustness of the overall model reconstruction and the reliability of the final three-dimensional model.

[0058] ​Further, the production-level point cloud generation module is configured to set a hierarchical processing mechanism to process the consumer-level point cloud data hierarchically to obtain production-level point cloud data; the hierarchical processing mechanism includes a precision-first processing layer, an adaptive balance processing layer, and an efficiency-first processing layer; The precision-first processing layer is configured to perform precision-first processing on the identified key region; the precision-first processing adopts a densification sampling and multi-view registration strategy to increase the point cloud density of the consumer-level point cloud data of the key region to N times of the original density, and retain the micro geometric features, to obtain high-precision point cloud data; The adaptive balance processing layer is configured to perform adaptive balance processing on the identified transition region; the adaptive balance processing adopts a medium-density sampling and adaptive parameter adjustment strategy to dynamically adjust the processing precision according to the distance from the adjacent key region and the local geometric complexity, to obtain balanced point cloud data; The efficiency-first processing layer is configured to perform efficiency-first processing on the identified flat region; the efficiency-first processing adopts a sparse sampling, fast interpolation, and batch processing strategy to reduce the point cloud density of the consumer-level point cloud data of the flat region to M% of the original density on the premise of ensuring basic precision, to obtain high-efficiency point cloud data; The hierarchical processing mechanism further includes an adaptive fusion unit; the adaptive fusion unit is configured to adaptively fuse the high-precision point cloud data, the balanced point cloud data, and the high-efficiency point cloud data to obtain production-level point cloud data.

[0059] Further, the semantic modeling module is configured to identify the functional regions and structural features of the object by using a point cloud semantic segmentation model, segment the production-level point cloud data into different semantic regions, assign a reconstruction strategy and parameters to each semantic region according to the attributes of the different semantic regions, and perform three-dimensional reconstruction; The point cloud semantic segmentation model includes: a structural feature extraction unit configured to extract structural features in the production-level point cloud data, including local geometric features and context features; a functional region identification unit configured to identify the functional regions of the object according to the structural features by using a pre-trained semantic classifier, including load-bearing structural regions, decorative regions, connection interface regions, and functional opening regions; a region attribute analysis unit configured to analyze the attributes of each semantic region, including geometric parameters, functional attributes, and manufacturing constraints; a reconstruction strategy allocation unit configured to assign a corresponding reconstruction strategy and parameters to each functional region according to the attributes of each semantic region; the parameters include grid density parameters and error tolerance parameters.

[0060] Further, the local grid correction module is configured to detect the geometric parameters of the three-dimensional grid generated in the three-dimensional reconstruction process in real time based on a pre-set multi-process rule library; when it is detected that the geometric parameters do not meet the requirements of the target production process, a local grid correction model is automatically triggered to perform geometric adjustment to obtain a corrected grid model; The local mesh correction model comprises: a three-dimensional reconstruction unit, which performs three-dimensional reconstruction based on the reconstruction strategy and parameters assigned to each functional region by the point cloud semantic segmentation model; a geometric parameter real-time detection unit, which calculates geometric parameters of a three-dimensional mesh generated in the three-dimensional reconstruction process in real time, including wall thickness, inclination angle, fillet radius, and aperture size of the three-dimensional mesh; a local mesh correction triggering unit, which compares the detected geometric parameters with 3D printing rules, numerical control machining rules, and injection molding rules in a preset multi-process rule library, and identifies a rule violation region that does not meet the requirements of a target production process; and a geometric adjustment unit, which automatically selects a local mesh correction strategy for geometric adjustment according to the rule violation type and severity of the rule violation region, and obtains a corrected mesh model.

[0061] Further, the correction process adopts a local optimization strategy, and only the meshes within a 5mm range around the rule violation region are adjusted, and the geometry of other regions remains unchanged.

[0062] By limiting the geometric correction range to within 5mm around the rule violation region, the overall structure and geometric features of the original three-dimensional model can be maintained to the greatest extent while ensuring that the production process constraints are met, effectively avoiding error propagation and unnecessary performance loss caused by global adjustment. The local optimization strategy not only improves the correction efficiency and reduces the computational overhead, but also preserves the user's individualized design intent in non-rule violation regions, which helps to maintain the consistency and stability of the model design on the basis of ensuring manufacturing feasibility, and improves the processing precision and engineering practicability of the system for local rule violation problems.

[0063] Further, the three-dimensional model generation module is configured to establish a multi-objective evaluation model, dynamically adjust the weights of each objective according to the user's customized demand type and selected production process, optimize the corrected mesh model, and generate a final three-dimensional model.

[0064] The process of generating the final three-dimensional model comprises: establishing a multi-objective evaluation model based on geometric precision, manufacturing feasibility, material usage, processing cost, and functional integrity; dynamically adjusting the weight coefficients of each evaluation objective according to the user's customized demand type and selected process type, and generating a personalized weight configuration; weighting each evaluation objective based on the personalized weight configuration to obtain a multi-objective function; solving the optimal solution set of the multi-objective function using a multi-objective optimization algorithm, and selecting a final three-dimensional model scheme from the optimal solution set in combination with user preferences and engineering constraints, optimizing the corrected mesh model, and obtaining the final three-dimensional model.

[0065] Further, a user interaction feedback mechanism is introduced in the optimization process, allowing the user to adjust the preference weight in real time during the optimization iteration process, and the system dynamically updates the fitness function according to the user feedback, and automatically adjusts the inertia weight and acceleration coefficient according to the convergence state by using the adaptive parameter adjustment strategy, ensuring that the algorithm maintains global search ability while improving convergence speed and solution quality.

[0066] By introducing the user interaction feedback mechanism, the user can adjust the weight preference in real time during the optimization iteration process, realizing the personalized dynamic guidance of the optimization direction, and significantly improving the matching degree of the optimization result to the user demand. At the same time, the system updates the fitness function in real time according to the user feedback, and automatically adjusts the inertia weight and acceleration coefficient in different convergence stages by combining the adaptive parameter adjustment strategy, which not only guarantees the global search ability and prevents falling into local optimum, but also improves the convergence speed and solution quality. This mechanism realizes the organic integration of human-computer collaborative optimization and algorithm adaptive control, enhances the flexibility, intelligence and convergence stability of the three-dimensional model generation process, and is suitable for high-quality model optimization tasks under multi-objective complex constraints.

[0067] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for automatically generating a three-dimensional model based on point cloud data, characterized in that: include: Receive consumer-grade point cloud data acquired by consumer-grade devices, use a multi-scale feature recognition model to identify multi-scale features in the consumer-grade point cloud data, and divide the area; A hierarchical processing mechanism is set up to process consumer-grade point cloud data in layers to obtain production-grade point cloud data; the hierarchical processing mechanism includes an accuracy priority processing layer, an adaptive balance processing layer, and an efficiency priority processing layer; Utilize the point cloud semantic segmentation model to identify the functional areas and structural features of objects, segment production-level point cloud data into different semantic regions, and assign reconstruction strategies and parameters to each semantic region based on its attributes for 3D reconstruction. The geometric parameters of the 3D mesh generated during the 3D reconstruction process are detected in real time based on a preset multi-process rule library. When it is detected that the geometric parameters do not meet the target production process requirements, the local mesh correction model is automatically triggered to make geometric adjustments to obtain a corrected mesh model. A multi-objective evaluation model is established to adaptively adjust the weight of each objective according to the user's customized demand type and the selected production process, optimize the modified grid model, and generate the final three-dimensional model.

2. The method for automatically generating a three-dimensional model based on point cloud data according to claim 1, characterized in that: The multi-scale feature recognition model includes: a multi-scale feature extraction unit, which extracts geometric features of consumer-grade point cloud data at different scales; a feature classification unit, which classifies the extracted geometric features, including edge features, corner features, surface change features, and flat area features; a feature importance evaluation unit, which calculates the importance weight of each geometric feature based on feature type, local geometric complexity, and manufacturing criticality; and a key area positioning unit, which divides the consumer-grade point cloud data into regions according to the importance weight and spatial continuity constraints, including key regions, transition regions, and flat regions.

3. The method for automatically generating a three-dimensional model based on point cloud data according to claim 2, characterized in that: The precision priority processing layer performs precision priority processing on the identified key areas; the precision priority processing uses dense sampling and multi-view registration strategies to increase the point cloud density of the consumer-grade point cloud data of the key areas to N times the original density, while retaining tiny geometric features to obtain high-precision point cloud data; The adaptive balancing processing layer performs adaptive balancing processing on the identified transition area; the adaptive balancing processing adopts a medium-density sampling and adaptive parameter adjustment strategy, dynamically adjusts the processing accuracy according to the distance to the adjacent key area and the local geometric complexity, and obtains balanced point cloud data; The efficiency-prioritized processing layer performs efficiency-prioritized processing on the identified flat area; the efficiency-prioritized processing adopts sparse sampling, fast interpolation, and batch processing strategies to reduce the point cloud density of the consumer-grade point cloud data of the flat area to M% of the original density while ensuring basic accuracy, thereby obtaining high-efficiency point cloud data; The hierarchical processing mechanism also includes an adaptive fusion unit; the adaptive fusion unit adaptively fuses the high-precision point cloud data, the balanced point cloud data, and the high-efficiency point cloud data to obtain production-level point cloud data.

4. The method for automatically generating a three-dimensional model based on point cloud data according to claim 1, wherein: The point cloud semantic segmentation model includes: a structural feature extraction unit, which extracts structural features from production-level point cloud data, including local geometric features and contextual features; a functional area identification unit, which uses a pre-trained semantic classifier to identify the functional areas of an object based on the structural features, including load-bearing structure areas, decorative areas, connection interface areas, and functional opening areas; a regional attribute analysis unit, which analyzes the attributes of each semantic area, including geometric parameters, functional attributes, and manufacturing constraints; and a reconstruction strategy allocation unit, which allocates corresponding reconstruction strategies and parameters to each functional area based on the attributes of each semantic area; the parameters include grid density parameters and error tolerance parameters.

5. The method for automatically generating a three-dimensional model based on point cloud data according to claim 1, wherein: The local mesh correction model includes: a three-dimensional reconstruction unit, which performs three-dimensional reconstruction based on the reconstruction strategy and parameters assigned to each functional area by the point cloud semantic segmentation model; a real-time geometric parameter detection unit, which calculates in real time the geometric parameters of the three-dimensional mesh generated during the three-dimensional reconstruction process, including the wall thickness, tilt angle, fillet radius and aperture size of the three-dimensional mesh; a local mesh correction trigger unit, which compares the detected geometric parameters with the 3D printing rules, CNC machining rules and injection molding rules in the preset multi-process rule library to identify illegal areas that do not meet the target production process requirements; and a geometric adjustment unit, which automatically selects a local mesh correction strategy for geometric adjustment based on the violation type and severity of the illegal area to obtain a corrected mesh model.

6. The method for automatically generating a three-dimensional model based on point cloud data according to claim 1, characterized in that: The process of generating the final 3D model includes: establishing a multi-objective evaluation model based on geometric accuracy, manufacturing feasibility, material usage, processing cost and functional integrity; dynamically adjusting the weight coefficient of each evaluation objective according to the user's customized demand type and selected process type to generate a personalized weight configuration; weighting each evaluation objective based on the personalized weight configuration to obtain a multi-objective function; using a multi-objective optimization algorithm to solve the optimal solution set of the multi-objective function, and combining user preferences and engineering constraints to select the final 3D model solution from the optimal solution set, optimize and correct the mesh model, and obtain the final 3D model.

7. A three-dimensional model automatic generation system based on point cloud data, characterized in that: include: The consumer-grade feature recognition module is used to receive consumer-grade point cloud data acquired by consumer-grade devices, identify multi-scale features in the consumer-grade point cloud data using a multi-scale feature recognition model, and divide the area. The production-grade point cloud generation module is used to set a hierarchical processing mechanism to perform hierarchical processing on the consumer-grade point cloud data to obtain production-grade point cloud data. The hierarchical processing mechanism includes an accuracy priority processing layer, an adaptive balance processing layer, and an efficiency priority processing layer. The semantic modeling module is used to use the point cloud semantic segmentation model to identify the functional areas and structural features of objects, segment the production-level point cloud data into different semantic areas, and assign reconstruction strategies and parameters to each semantic area according to the attributes of different semantic areas to perform three-dimensional reconstruction; the local mesh correction module is used to detect the geometric parameters of the three-dimensional mesh generated during the three-dimensional reconstruction process in real time based on the preset multi-process rule library; when it is detected that the geometric parameters do not meet the target production process requirements, the local mesh correction model is automatically triggered to perform geometric adjustments to obtain a corrected mesh model; the three-dimensional model generation module is used to establish a multi-objective evaluation model, adaptively adjust the weights of each objective according to the user's customized demand type and the selected production process, optimize the corrected mesh model, and generate the final three-dimensional model.

Images