Battlefield target three-dimensional model automatic alignment method, system and electronic equipment

Abstract

The application discloses a battlefield target three-dimensional model automatic alignment method and system and electronic equipment, relates to the technical field of computer graphics and three-dimensional model processing, and comprises the following steps: performing pretreatment on an input battlefield target three-dimensional model to be corrected, calculating an axis alignment bounding box, normalizing a model geometric center to a coordinate origin, and generating a bounding box six-face orthogonal projection rendering image; inputting the six-face rendering image and a structured text prompt into a multi-modal large model, identifying a bounding box direction surface through independent semantic reasoning combined with a majority voting mechanism, extracting a corresponding directed normal vector, and constructing a three-axis orthogonal semantic local coordinate system; taking the semantic local coordinate system as a source coordinate system and a predefined target world coordinate system as a target, constructing a rigid rotation matrix, and applying the rigid rotation matrix to all geometric data of the three-dimensional model to complete automatic correction and alignment output of the coordinate axis orientation. High-precision automatic semantic-driven normalization of the three-dimensional model coordinate axis orientation is realized.

Description

Technical Field

[0001] This invention relates to the fields of computer graphics and 3D model processing technology, and in particular to a method, system and electronic device for automatic alignment of 3D battlefield target models. Background Technology

[0002] Three-dimensional digital models are core data asset carriers in fields such as digital twins, virtual reality, intelligent manufacturing, and computer graphics. With the rapid popularization of AI-powered 3D reconstruction technologies, such as diffusion models and neural radiation fields, the sources of 3D model generation have expanded from traditional modeling software or scanning equipment to automated generation pipelines driven by multimodal data such as text and images. However, the local coordinate system of 3D models generated by these technologies is often arbitrarily determined by the initial algorithm settings or the initial pose of the acquisition device, lacking a necessary connection with the direction of gravity in the real world or the semantic orientation of the object itself. This leads to serious inconsistencies in the coordinate axis orientation of models from different sources. 3D model coordinate axis orientation normalization technology is widely used in downstream tasks such as 3D data preprocessing, automatic scene layout, physical simulation initialization, and 3D content generation, and is an indispensable key link in the 3D model standardization processing pipeline.

[0003] Existing methods for 3D model coordinate axis alignment mainly fall into two categories. The first category is a 3D model pose normalization method based on Principal Component Analysis (PCA). This method extracts the 3D coordinates of all vertices from the input 3D mesh model to form a point cloud dataset. After zero-mean processing of the point cloud data, it calculates the 3D spatial covariance matrix, then performs eigenvalue decomposition on the covariance matrix to extract three mutually orthogonal eigenvectors as the estimated coordinate axis directions of the model. Finally, it constructs a rotation matrix according to predefined axial mapping rules and applies a rigid body rotation transformation to the original model to complete the coordinate axis alignment. The second category is a 3D model "up-axis" estimation method based on physical stability analysis. This method extracts candidate bottom planes from the 3D mesh, calculates the centroid projection position of the model on each candidate bottom plane, determines whether it is located inside the supporting polygon, and evaluates the physical stability of the model in that pose. It selects the bottom normal vector corresponding to the most physically stable placement pose as the model's "up" direction and constructs a coordinate transformation matrix accordingly to apply rotation alignment.

[0004] However, all of the aforementioned existing methods have significant technical shortcomings. Regarding principal component analysis (PCA)-based methods, the core deficiencies are: First, they cannot correctly handle geometrically asymmetric models, resulting in unreliable orientation results. Since PCA is essentially a purely statistical analysis method, the eigenvectors of the covariance matrix only reflect the statistical regularity of the point cloud's geometric distribution, not the model's inherent semantic structure. Therefore, they will inevitably fail for models such as vehicles and weapons where the geometric distribution and semantic orientation are inconsistent. Second, there is a problem of symbolic ambiguity. The eigenvectors obtained from PCA eigenvalue decomposition only define the direction lines of the axes, not the directed axes themselves, making it impossible to distinguish whether the "up axis" points to the top or bottom, or the "front axis" points to the front or back. Third, they are highly sensitive to noise and attached geometry. When the 3D model contains attached geometry or reconstructs noisy points, even a small number of large noise geometry objects can have a decisive impact on the final result. Regarding the physical stability analysis-based method, its main drawbacks are: First, substituting physical stability for semantic orientation contains a fundamental logical fallacy; models with the most physically stable postures, such as those of aircraft and humans, that are inconsistent with the semantic standard orientation cannot obtain correct orientation results. Second, it can only estimate a single "up axis," and cannot provide effective estimates for the "front axis" and "side axis," resulting in incomplete coordinate axis normalization. Third, it is highly dependent on the geometric complexity and topological quality of the model, has poor generalization ability, and its applicability to models with non-mesh representations such as non-manifold meshes, normal vector flips, or point clouds is severely limited.

[0005] The common fundamental flaw of both types of existing technologies is that they both treat the orientation normalization problem of 3D models as a purely geometric or purely mechanical problem, while ignoring the fact that the problem is essentially a semantic problem that is highly related to human cognition and common sense about the physical world. They completely lack the ability to understand the category knowledge, functional attributes, usage scenarios and human cognitive conventions of the objects represented by the model, making it difficult to meet the engineering needs of automated preprocessing of large-scale 3D model libraries. Summary of the Invention

[0006] The present invention aims to provide an automatic alignment method and system for 3D models of battlefield targets based on the semantics of the azimuth plane of reconnaissance images, so as to overcome the shortcomings of the existing technology. The technical problem to be solved by the present invention is achieved through the following technical solution.

[0007] According to a first aspect of this application, a method for automatic alignment of three-dimensional battlefield target models is provided, comprising the following steps: S1 parses the input 3D model file of the battlefield target, extracts the vertex coordinate set, calculates the axis aligned with the bounding box, and normalizes the geometric center of the model to the origin of the coordinates by translation transformation. Based on the normal vector directions of the six azimuth planes of the bounding box, an orthogonal projection virtual camera is set up, and orthogonal projection rendering is performed on the model from the six directions of positive X, negative X, positive Y, negative Y, positive Z, and negative Z respectively to generate six azimuth rendering images. S2, construct a multimodal large model input by combining the six directional rendering images with text prompts containing task description, directional semantic definition and structured output format requirements, perform K independent semantic reasoning on the same set of inputs, perform directional majority voting on the K reasoning results, and adopt the recognition result when the recognition results of the front, top and side directions pass the vote with more than half of the support rate respectively. S3. Based on the orientation plane recognition results output in step S2 and the bounding box orientation plane correspondence mapping table, extract the front semantic direction vector, the top semantic direction vector and the side semantic direction vector, verify that the dot product of each pair of the three vectors is approximately zero to confirm the orthogonality, verify the cross product relationship of the three vectors to confirm the right-hand coordinate system, and construct a three-axis orthogonal semantic local coordinate system with the geometric center of the model as the reference point. S4. Construct the source coordinate system basis vector matrix by column using the three unit basis vectors of the semantic local coordinate system, and construct the target coordinate system basis vector matrix by column using the corresponding standard axis of the target world coordinate system. Solve the rotation matrix by multiplying the target basis vector matrix by the transpose of the source coordinate system basis vector matrix, and perform orthogonality verification and determinant verification on the rotation matrix. S5 applies the validated rotation matrix to all vertex coordinates, vertex normals, and face normals of the 3D model, completing a rigid body rotation transformation with the geometric center of the model as the reference point, and outputting a standardized 3D model whose coordinate axes are completely aligned with the target world coordinate system.

[0008] Preferably, the orthographic projection rendering in step S1 specifically includes: The projection plane size of each orthogonal projection virtual camera is set according to the side length of the corresponding direction of the bounding box to ensure that the model is fully presented in the rendered image without pruning; The rendering process preserves the model's texture maps, material colors, and basic lighting effects. After rendering is complete, six orientation rendering images with uniform resolution and a mapping table showing the correspondence between the six images and the bounding box orientation plane are output.

[0009] Preferably, step S1 further includes: performing size normalization processing on the three-dimensional model, calculating the maximum side length of the bounding box in the three-axis direction, performing a proportional scaling transformation on the coordinates of all vertices of the model using the reciprocal of the maximum side length as a scaling factor, and normalizing the model to the range of a unit bounding box.

[0010] Preferably, the construction of the text prompt in step S2 includes: The text prompts indicate that the input image is a six-view orthographic projection of a 3D model, and each image is labeled with a number corresponding to its bounding box face. The text prompts include explicit definitions of the semantics of front, top, and side; The requirement is to output the correspondence between azimuth numbers and semantic categories in a predefined JSON format for the multimodal large model; For symmetrical object categories where semantic orientation judgment is ambiguous, guide the multimodal large model to make judgments based on functional features or texture distribution.

[0011] Preferably, step S3 further includes: when the orthogonality verification of the three extracted semantic orientation vectors fails, triggering an exception handling process to re-request the multimodal large model to perform semantic recognition; when the three vectors form a left-handed coordinate system, inverting the lateral basis vectors to correct it to a right-handed coordinate system.

[0012] Preferably, step S4 further includes: when the determinant value of the rotation matrix is ​​approximately negative one, triggering the mirror component correction mechanism, inverting the column vectors in the rotation matrix corresponding to the side basis vectors, and eliminating the mirror components; the entire matrix operation of the rotation matrix uses double-precision floating-point numbers, and the rotation matrix is ​​orthogonalized before being applied to vertex transformation.

[0013] Preferably, step S5 further includes: after performing the rotation transformation, inputting the six-sided rendered image of the corrected model into the azimuth semantic recognition process again to verify that the normal vector of the front azimuth plane has been aligned to the front axis direction of the target world coordinate system and the normal vector of the upper azimuth plane has been aligned to the upper axis direction of the target world coordinate system, thus completing the quantitative verification of the alignment result; and recording the applied rotation matrix and target coordinate system configuration information in the metadata of the model file.

[0014] According to a second aspect of this application, an automatic alignment system for three-dimensional battlefield target models using the above-described automatic alignment method is provided, comprising: The preprocessing module is used to parse the input 3D model file of the battlefield target, calculate the axis to align the bounding box, normalize the geometric center of the model to the origin of the coordinate system by translation transformation, set up an orthogonal projection virtual camera based on the normal vector direction of the six azimuth planes of the bounding box, perform orthogonal projection rendering on the model from the six orthogonal directions, and generate six azimuth rendering images and a mapping table of the correspondence between the images and the azimuth planes of the bounding box. The azimuth semantic recognition module is used to construct a multimodal large model input by combining the six azimuth rendering images with structured text prompts, perform K independent semantic reasoning on the same set of inputs, perform azimuth majority voting on the K reasoning results, and adopt the recognition result that passes the vote with more than half the support rate; based on the recognition results and the corresponding mapping table, extract three semantic azimuth vectors, verify the orthogonality and right-hand coordinate system constraints, and construct a three-axis orthogonal semantic local coordinate system; The model coordinate axis correction and alignment module is used to construct the source coordinate system basis vector matrix in columns using the three unit basis vectors of the semantic local coordinate system, construct the target coordinate system basis vector matrix in columns using the corresponding standard axis of the target world coordinate system, solve the rotation matrix by multiplying the target basis vector matrix by the transpose of the source coordinate system basis vector matrix, perform legality verification on the rotation matrix, apply the rotation matrix to all vertex coordinates, vertex normals and face normals of the 3D model, and output the aligned and standardized 3D model.

[0015] According to a third aspect of this application, an electronic device is provided, comprising: a memory and a processor; the memory stores a computer program, and the processor executes the computer program to implement the above-described automatic alignment method for three-dimensional battlefield target models.

[0016] According to a fourth aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method for automatic alignment of three-dimensional battlefield target models.

[0017] The embodiments of the present invention have the following advantages: 1. This invention redefines the problem of normalizing the coordinate axes of a three-dimensional model as a semantic understanding problem, introduces the image semantic understanding capability of a multimodal large model, fundamentally solves the orientation failure problem caused by the lack of perception of the semantic orientation of the model by existing pure geometric methods, and can accurately handle complex asymmetric battlefield target models with inconsistent geometric distribution and semantic orientation.

[0018] 2. This invention completely eliminates the inherent symbolic ambiguity problem of PCA eigenvalue decomposition by directly predicting directed orientation vectors with clear semantic meanings, ensuring the uniqueness and determinism of the established semantic coordinate system in terms of direction, and distinguishing between front and back, top and bottom without manual intervention.

[0019] 3. This invention uses multi-view rendered images of the model as the input medium for semantic understanding. Multimodal large models naturally have the ability to focus on significant main structures in images and can automatically ignore the interference of auxiliary geometric noise, which significantly improves the robustness of the method on complex geometric structures and noisy models.

[0020] 4. This invention simultaneously predicts three key orientation vectors: front, top, and side, constructs a complete three-axis semantic local coordinate system, covers all three rotational degrees of freedom in three-dimensional space, achieves complete normalization of the coordinate axes, and completely solves the fundamental limitation of physical stability analysis methods that can only determine a single "up axis".

[0021] 5. This invention employs multiple independent reasoning and majority voting mechanisms to effectively suppress the randomness of large model reasoning. Combined with the direct construction method of basis vector matrix to efficiently solve the rotation matrix, it realizes the full automation of the coordinate axis orientation correction of the three-dimensional model, replacing the inefficient working mode that relies on manual correction, and has significant efficiency advantages. Attached Figure Description

[0022] Figure 1 This is a flowchart of the steps of an automatic alignment method for a three-dimensional battlefield target model according to the present invention; Figure 2 This is a functional block diagram of an automatic alignment system for three-dimensional battlefield target models according to the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device according to the present invention; Figure 4 This is a schematic diagram of the structure of a computer storage medium according to the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the above detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0025] Example 1: An Automatic Alignment Method for 3D Battlefield Target Models Based on the Semantic Meaning of Reconnaissance Image Azimuth

[0026] This method is deployed in the batch automated preprocessing scenario of a battlefield target 3D model asset library. It addresses the batch generation of battlefield target 3D models using AI 3D reconstruction technology (such as image-to-3D model generation). These models have random local coordinate system orientations and unknown three-degree-of-freedom rotational deviations from the predefined target world coordinate system, requiring fully automated standardization correction of coordinate axis orientations. This method employs three sequentially executed processing stages—preprocessing, azimuth plane semantic recognition, and coordinate axis correction and alignment—to achieve a fully automated transformation from an input model with arbitrary orientation to an output model whose orientation is completely consistent with the target world coordinate system. Figure 1 As shown, it includes the following steps: S1 parses the input 3D model file of the battlefield target, extracts the vertex coordinate set, calculates the axis aligned with the bounding box, and normalizes the geometric center of the model to the origin of the coordinates by translation transformation. Based on the normal vector directions of the six azimuth planes of the bounding box, an orthogonal projection virtual camera is set up, and orthogonal projection rendering is performed on the model from the six directions of positive X, negative X, positive Y, negative Y, positive Z, and negative Z respectively to generate six azimuth rendering images. In this step, an axis-aligned bounding box (AABB) is used as a standardized proxy structure for the geometric space of the 3D model. This maps the input 3D model, regardless of its source or coordinate system, to a standardized space with the model's geometric center as the origin. Using the six orthogonal azimuth planes of the bounding box as geometric references, multi-view orthogonal projection rendering transforms the 3D model's all-around appearance information into a sequence of two-dimensional images, providing standardized and complete visual input for the subsequent azimuth plane semantic recognition module. The specific steps are as follows: Step 11: 3D Model Reading and Parsing The system first reads and parses the input uncorrected 3D model file. Supported model formats include, but are not limited to, mainstream 3D model file formats such as OBJ, FBX, GLTF / GLB, PLY, and STL. During the parsing process, the system extracts the vertex coordinate array of the model. The model's memory representation structure is constructed using a patch index array, a normal vector array (if any), and texture coordinates and texture map data (if any), providing a data foundation for subsequent geometric processing.

[0027] For composite models containing multiple sub-meshes, the system uniformly includes the vertex coordinates of all sub-meshes in the subsequent bounding box calculation range to ensure that the bounding box can completely enclose the entire geometry of the model, avoiding problems such as the bounding box size being too small and the azimuth plane definition being incomplete due to ignoring some sub-meshes.

[0028] Step 12: Axis-Aligned Bounding Box (AABB) Calculation

[0029] The system traverses the set of all vertex coordinates of the model. Along the current coordinate system respectively The extreme values ​​of the vertex coordinate components along the three axes are calculated using the following formula:

[0030]

[0031]

[0032] This determines the coordinates of the eight corner points of the axis-aligned bounding box, and constructs the smallest axis-aligned cuboid bounding box that tightly encloses all the geometry of the model.

[0033] Step 13: Normalize the geometric center of the model

[0034] Calculate the geometric center coordinates of the axis-aligned bounding box :

[0035] With translation vector Perform a translation transformation on the coordinates of all vertices of the model:

[0036] After translation, the geometric center of the model is aligned with the world coordinate origin. Upon overlap, the bounding box is updated to a symmetrical cuboid centered at the origin. Thereafter, the outward normal vectors of the six facets of the bounding box can be directly used as directed direction vectors originating from the origin, without requiring additional translation corrections.

[0037] Step 14: Bounding box size normalization (optional)

[0038] To eliminate the impact of different model sizes on the model's proportion and visual clarity in the rendered view, the system can selectively perform size normalization on the model. The system also calculates the maximum side length of the bounding box along three axes. :

[0039] With scaling factor Perform a proportional scaling transformation on all vertex coordinates of the model to normalize the model to the unit bounding box range, ensuring that the visual appearance of the model is consistent in multi-view rendered images and improving the stability of semantic recognition of multimodal large models.

[0040] Step 15: Generation of multi-view orthogonal rendering images

[0041] After completing geometric center normalization (and optional size normalization), the system sets up six orthogonal projection virtual cameras based on the normal vector directions of the six azimuth planes of the bounding box, and renders the model from the following six orthogonal directions (as shown in Table 1), generating the corresponding azimuth rendering views: Table 1: Corresponding parameters for the six azimuth planes

[0042] Each camera uses orthographic projection, with the projection plane size set according to the side length of the bounding box in the corresponding direction to ensure the model is fully rendered without clipping. The model's texture maps, material colors, and basic lighting effects are preserved during rendering to provide the richest possible visual semantic information.

[0043] After rendering is complete, the system outputs six orientation rendering images with uniform resolution. The mapping table showing the correspondence between each image and the bounding box azimuth plane is then passed to the azimuth plane semantic recognition module.

[0044] Relevant technical measures include: 1. Using bounding box azimuth planes instead of direct point cloud analysis as the semantic recognition carrier. Compared to directly inputting the original point cloud or mesh data into a multimodal large model, this module uses bounding box azimuth plane-driven multi-view rendered images as the input carrier for semantic recognition, which has significant technical advantages: First, rendered images are the most mature input form for multimodal large models after large-scale training, and using images as input can maximize the semantic reasoning ability of multimodal large models; Second, there is a direct one-to-one correspondence between the normal vector of the bounding box azimuth plane and the candidate coordinate axis direction, and the semantic recognition result can be losslessly mapped to the direction vector without additional geometric calculations; Third, orthogonal rendered images eliminate the format dependency of the 3D model coordinate representation, enabling the system to be compatible with input models of any geometric representation.

[0045] 2. Orthographic projection replaces perspective projection to eliminate angular distortion.

[0046] This module explicitly uses orthographic projection in multi-view rendering, rather than the perspective projection mode more commonly used in 3D rendering. This improvement ensures that the visual proportions of each part of the model in the rendered image are proportional to its actual geometric size, eliminates the distortion effect of perspective shrinkage on the model's appearance, and enables large multimodal models to make semantic orientation judgments based on more objective and distortion-free visual information, thereby improving the accuracy and consistency of the recognition results.

[0047] 3. Decoupled design of geometric center normalization and size normalization.

[0048] This module designs geometric center normalization (translation to the origin) and size normalization (scaling to the unit bounding box) as two independent processing steps. Geometric center normalization is a mandatory step, while size normalization is optional. The significance of this decoupled design is that geometric center normalization is a necessary prerequisite for the correct calculation of the subsequent rotation matrix and is mandatory; while size normalization only affects the visual size of the model in the rendered image and has no impact on the calculation results of the direction vector. It can be flexibly selected for use based on the actual model size distribution, avoiding interference with downstream tasks that need to retain the true size information of the model in specific application scenarios.

[0049] S2, construct a multimodal large model input by combining the six directional rendering images with text prompts containing task description, directional semantic definition and structured output format requirements, perform K independent semantic reasoning on the same set of inputs, perform directional majority voting on the K reasoning results, and adopt the recognition result when the recognition results of the front, top and side directions pass the vote with more than half of the support rate respectively. In this step, the problem of normalizing the coordinate axis orientation of the 3D model is redefined from a traditional geometric calculation problem as a semantic understanding problem. Using the Multimodal Large Language Model (MLLM) as a semantic inference engine, the bounding box orientation planes carrying the three semantics of "front", "top" and "side" are identified by semantic parsing of the multi-view orthogonal rendered images generated by the preprocessing module. The corresponding directed normal vectors are extracted, and a three-axis orthogonal semantic local coordinate system that is highly consistent with the intrinsic semantics of the model is constructed.

[0050] Specifically, the steps include the following: Modal large model input construction: The azimuth surface semantic recognition module receives six azimuth rendering images from the preprocessing module. A mapping table showing the correspondence between the model and its bounding box orientation plane is used to construct prompt content for inputting multimodal large models.

[0051] The core task of input construction is to design effective prompting strategies (Prompt Engineering) to guide the multimodal large model to accurately understand the task objective and output structured recognition results. The input consists of the following two parts: Image input section: Six orientation rendering images are organized into an image sequence according to a uniform arrangement, and each image is assigned a number indicating its corresponding bounding box (e.g., "Image 1: "). "Directional View", "Image 2:" "Directional view", and so on, enables multimodal large models to clearly distinguish the spatial orientation of each input image during semantic reasoning.

[0052] Text prompts: Structured text prompts are designed to clearly describe the task objective, the generation method of the input image (six-view orthographic projection rendering), the semantic orientation categories to be recognized (front, top, side), and the expected output format (a structured correspondence between orientation plane numbers and semantic categories) to the multimodal large model. The design of the text prompts follows these principles: ① Clear and unambiguous task description to avoid ambiguity; ② Provision of sufficient contextual information to help the multimodal large model correctly understand the task scenario; ③ Requirement for a structured output format that is easy to parse programmatically, ensuring that the recognition results can be reliably extracted by subsequent processing flows.

[0053] Multimodal large-scale model semantic reasoning: The constructed image sequence and text prompts are input into the multimodal large-scale model, triggering it to perform comprehensive semantic reasoning on six directional rendered images. The reasoning process of the multimodal large-scale model can be functionally decomposed into the following three stages: Object category recognition stage: The multimodal large model first performs an overall analysis of the six input rendered images to identify the object category represented by the 3D model (such as drones, military vehicles, battlefield fortifications, etc.), activates prior knowledge related to the object category, and provides category constraints for subsequent orientation semantic judgment.

[0054] In the view-specific semantic analysis stage: the multimodal large model performs positional semantic analysis on each of the six rendered images, comprehensively considering the model's appearance features presented in each view (such as facial features, functional structure, user interface, symmetrical distribution, etc.) to determine the semantic positional category corresponding to each view. The candidate category set is as follows: .

[0055] Comprehensive Reasoning and Consistency Verification Phase: After completing the independent orientation semantic analysis of each view, the multimodal large model performs comprehensive reasoning and consistency verification on the analysis results of all six views to ensure that the recognition results meet the following constraints: ① The semantic category assignment results of the six views are not repeated pairwise (each semantic category corresponds to exactly one view); ② The bounding box faces corresponding to the views identified as "front" and "back" must be parallel to each other and have opposite normal vectors; ③ The view pairs identified as "top" and "bottom", "left" and "right" must also meet the above parallel constraints. If there are inconsistencies in the preliminary analysis results, the multimodal large model corrects them through internal reasoning iterations until a consistent recognition result that meets the above constraints is output.

[0056] S3. Based on the orientation plane recognition results output in step S2 and the bounding box orientation plane correspondence mapping table, extract the front semantic direction vector, the top semantic direction vector and the side semantic direction vector, verify that the dot product of each pair of the three vectors is approximately zero to confirm the orthogonality, verify the cross product relationship of the three vectors to confirm the right-hand coordinate system, and construct a three-axis orthogonal semantic local coordinate system with the geometric center of the model as the reference point. In this step, the output of the multimodal large model is structured text containing the correspondence between azimuth plane numbers and semantic categories. The system performs procedural parsing on this output text to extract the following key information: the view number corresponding to the front and the normal direction of its corresponding bounding box, the view number corresponding to the top and the normal direction of its corresponding bounding box, and the view number corresponding to the side and the normal direction of its corresponding bounding box.

[0057] During the parsing process, the system, based on the mapping table between the rendered image and the bounding box azimuth plane established by the preprocessing module, directly converts the view number output by the multimodal large model into the normal vector direction of the corresponding bounding box plane, completing the mapping transformation from semantic recognition results to geometric direction vectors. Specifically, the mapping rule for the azimuth plane normal vector direction is as follows: If the multimodal large model converts the image... If identified as "front", then the front semantic direction vector If the image If it is identified as "above", then the semantic direction vector above. And so on.

[0058] Since the six azimuth planes of the bounding box satisfy strict orthogonal constraints and consist of three pairs of mutually perpendicular parallel planes, theoretically, the outward normal vectors of the "front," "top," and "side" azimuth planes identified by the multimodal large model naturally satisfy the geometric constraint of mutual orthogonality. However, to ensure the numerical stability of subsequent coordinate axis alignment calculations, the system performs numerical analysis on the three extracted semantic azimuth vectors. Perform an orthogonality check by calculating the pairwise dot product of the three vectors:

[0059] If all the above dot product values ​​are close to zero (for axis-aligned direction vectors, the dot product result is exactly zero), then the verification passes; if the verification fails due to parsing errors, the system triggers the exception handling process (see the technical measures section for details), and re-requests the multimodal large model for semantic recognition or adopts the default correction strategy.

[0060] After successful verification, three mutually orthogonal semantic orientation vectors are used. , , Using the model's geometric center (i.e., the origin) as the reference point, a semantic local coordinate system strongly correlated with the model's intrinsic semantics is constructed. :

[0061] This semantic local coordinate system fully defines the three-axis orientation of the model at the semantic level, eliminating the uncertainty of all three rotational degrees of freedom in three-dimensional space. It is the core intermediate representation connecting the semantic recognition result and the coordinate axis alignment transformation. The semantic local coordinate system, along with the specific values ​​of the three basis vectors, is passed to the model coordinate axis correction and alignment module as the final output of this module.

[0062] The relevant technologies used in the above steps include: a. A multi-stage reasoning and voting mechanism to improve identification stability Because the reasoning process of a large language model has a certain degree of randomness, the result of a single reasoning iteration may be unstable. Therefore, the system performs multiple independent multimodal large model inferences on the same set of input images (e.g., K=3 or K=5 times), and performs majority voting on the multiple inference results, using the recognition result with the most votes as the final output. Voting is performed independently for the recognition results of the three directional planes. Only when a recognition result receives more than half of the consistent support across multiple inference iterations is it adopted as the final result; otherwise, an exception handling process is triggered.

[0063] b. Optimization measures for prompt words

[0064] To address the unique challenges of 3D model orientation semantic recognition, the system has specifically optimized the design of text prompts: ① The prompts explicitly state that the input image is an orthographic projection of the 3D model's six views, helping the multimodal large model establish a correct spatial understanding framework; ② The prompts include clear explanations of the semantic definitions of "front," "top," and "side" (e.g., "front refers to the side of an object facing the observer or user, such as the direction a person's face faces or the direction a car's front faces"), eliminating semantic ambiguity; ③ For special object categories where semantic orientation judgment is ambiguous (e.g., highly symmetrical spheres, cylinders, etc.), the prompts guide the multimodal large model to make judgments based on the object's functional characteristics or texture distribution, rather than simply relying on its geometric shape.

[0065] S4. Construct the source coordinate system basis vector matrix by column using the three unit basis vectors of the semantic local coordinate system, and construct the target coordinate system basis vector matrix by column using the corresponding standard axis of the target world coordinate system. Solve the rotation matrix by multiplying the target basis vector matrix by the transpose of the source coordinate system basis vector matrix, and perform orthogonality verification and determinant verification on the rotation matrix. In this step, the semantic local coordinate system output by the azimuth plane semantic recognition module is used. Using the source coordinate system and the predefined target world coordinate system. Using the target coordinate system, construct the optimal rigid body rotation transformation between the two coordinate systems by solving for the optimal rigid body rotation transformation. Rotation matrix This rotational transformation is then applied to all geometric data of the input 3D model, achieving precise alignment of the model's semantic coordinate axes with the target world coordinate system, and completing fully automatic correction of the 3D model's coordinate axis orientation. The specific steps are as follows: Step 1: Define the target world coordinate system The model coordinate axis correction and alignment module first explicitly defines the predefined target world coordinate system. The standard axial configuration. The axial definition of the target world coordinate system is a system-level global configuration parameter, which is set according to the coordinate system conventions of downstream application scenarios (as shown in Table 2).

[0066] Table 2: Commonly Used Standard Configurations of Target World Coordinate Systems

[0067] The axial configuration of the target world coordinate system is uniformly managed by the system parameter file, allowing users to switch flexibly according to specific application scenarios without modifying the core algorithm logic.

[0068] The following implementation process is illustrated using configuration A (Z-up, Y-front, X-right) as an example, where the standard axis of the target coordinate system is: The upward direction (up) corresponds to the target axis: ,Right now Axis; Front direction (front) corresponds to the target axis: ,Right now Axis; right direction (side) corresponds to the target axis: ,Right now axis.

[0069] Step 2: Standardization of semantic coordinate system basis vectors

[0070] Receive three semantic orientation vectors from the azimuth plane semantic recognition module. Then, the system first performs normalization to ensure that all three basis vectors are unit vectors:

[0071] Since the bounding box azimuth plane normal vectors constructed by the preprocessing module are all axis-aligned unit vectors (i.e., the components are only taken from the model), or The above normalization operation is an identity transformation under normal circumstances. This step is mainly used as a numerical stability guarantee measure.

[0072] Subsequently, the system verifies whether the three unit basis vectors constitute a right-handed coordinate system, that is, verifies the following cross product relationship:

[0073] If the verification result does not match the above formula (i.e., the three vectors form a left-handed coordinate system), the system will automatically adjust the settings. Invert: This ensures that the semantic local coordinate system is a standard right-handed coordinate system, consistent with the chirality convention of the target world coordinate system.

[0074] Step 3: Construction of the rotation matrix

[0075] The system uses the direct construction method of the basis vector matrix to solve for the rotation matrix. .

[0076] Construct the basis vector matrix of the original coordinate system

[0077] Arrange the three unit basis vectors of the semantic local coordinate system column-wise to construct... Source coordinate system basis vector matrix :

[0078] matrix The three columns correspond to the basis vectors in the "right," "front," and "top" directions in the semantic local coordinate system, respectively. Since the three basis vectors are mutually orthogonal and are all unit vectors, It is an orthogonal matrix that satisfies .

[0079] Construct the basis vector matrix of the target coordinate system : Adjust the three standard axis vectors of the target world coordinate system according to... Arrange the corresponding columns in order to construct Target coordinate system basis vector matrix :

[0080] Taking configuration A as an example, Exactly equal to the identity matrix The column order corresponds to "right". "、"front corresponds to "、"top corresponds to The axial mapping relationship of "".

[0081] Solving the rotation matrix Rotation matrix Transform the basis vectors of the semantic local coordinate system to the corresponding standard axes of the target world coordinate system, i.e., satisfy:

[0082] because If a matrix is ​​orthogonal, its inverse is equal to its transpose. ,therefore:

[0083] Taking configuration A as an example, Therefore:

[0084] Rotation matrix Satisfying orthogonality and determinant value This ensures the transformation is a pure rotation without mirror flipping, which is... The legal rotation matrix of a group, such as Figure 1 "rotation matrix" The steps are shown.

[0085] Step 4: Verify the validity of the rotation matrix

[0086] In the rotation matrix Before being applied to model transformation, the system performs a legality verification, which mainly includes the following checks: Orthogonality verification

[0087] in Let Frobenius norm be the matrix. Orthogonality error tolerance (usually set to) ).

[0088] Determinant verification

[0089] in For the determinant error tolerance (usually set to...) ).like If so, it means that the transformation contains a mirror component, and the rotation matrix needs to be corrected (see the technical measures section for details).

[0090] Verification of the rationality of rotation angles: The physical rationality is verified by calculating the rotation angles corresponding to the rotation matrix.

[0091] Rotation angle Should be in If the calculation result exceeds the specified range, exception handling will be triggered.

[0092] S5 applies the validated rotation matrix to all vertex coordinates, vertex normals, and face normals of the 3D model, completing a rigid body rotation transformation with the geometric center of the model as the reference point, and outputting a standardized 3D model whose coordinate axes are completely aligned with the target world coordinate system.

[0093] The specific method for this step is as follows: Vertex coordinate transformation: for the entire model Perform rotation transformation on the coordinates of each vertex:

[0094] in The column vector of vertex coordinates before transformation. This is the column vector of vertex coordinates after transformation.

[0095] Vertex normal vector transformation: If the model contains vertex normal vector data, perform the corresponding rotation transformation on all vertex normal vectors. The transformation matrix of the normal vector is the same as the transformation matrix of the vertex coordinates (because the rotation matrix is ​​an orthogonal matrix, and its inverse reverse matrix is ​​equal to itself):

[0096] in To transform the vertex normal vector, This is the transformed vertex normal vector.

[0097] Face normal vector transformation: If the model explicitly stores face normal vector data, a rotation transformation is also performed on it:

[0098] in This represents the total number of facets in the model. and The first and second parts are respectively the first and second parts before and after the transformation. The normal vector of each facet.

[0099] Other directional data transformations: If the model contains other directional geometric attribute data such as tangent vectors and bitangent vectors, perform corresponding rotation transformations on them to ensure that all geometric attribute data of the transformed model remain consistent in the new coordinate system.

[0100] Texture coordinates (UV coordinates) data are defined in the model's parametric domain (UV space), are independent of 3D spatial coordinates, do not require rotation transformations, and remain unchanged. Non-directional data such as face vertex index arrays and material properties also remain unchanged.

[0101] Post-processing and output of calibration results: Post-processing of calibration results includes bounding box update and coordinate system alignment verification.

[0102] After performing the rotation transformation, the system recalculates the axis-aligned bounding box based on the transformed vertex coordinates, updates the spatial boundary information of the model, and provides accurate space occupancy data for subsequent downstream tasks.

[0103] The system performs rapid verification of the correction results: six azimuth planes are extracted from the new bounding box of the transformed model, and their corresponding rendered views are input again into the azimuth plane semantic recognition module to perform semantic verification, confirming that the normal vector of the "frontal" azimuth plane has been aligned to the target. The axial direction, the normal vector of the "above" azimuth plane has been aligned to the target. Along the axis, the normal vector of the "side" azimuth plane has been aligned with the target. The axis direction is used to quantitatively verify the accuracy of the coordinate axis alignment.

[0104] If output is required, the results will be standardized into a model file. The coordinate-axis-corrected 3D model will be output in a standard file format, fully preserving the original model's geometric data, texture maps, material properties, and hierarchical structure. The applied rotation matrix will also be recorded in the model file's metadata. And target coordinate system configuration information, for subsequent traceability and reuse.

[0105] The relevant technical measures adopted in the above steps include: Mirror transformation detection and correction: When rare orientation plane confusion occurs in the semantic recognition results of a multimodal large model (such as recognizing the back face as the front face), it may cause the constructed semantic local coordinate system to be a left-handed coordinate system, which in turn affects the calculated transformation matrix. satisfy This includes mirror components. The system checks for this during the rotation matrix validity verification phase; once detected... The correction mechanism is immediately triggered: the column vectors corresponding to the "side" basis vectors in the rotation matrix are inverted to eliminate the mirror component, the transformation matrix is ​​corrected to a valid rotation matrix, the event is recorded and fed back to the azimuth semantic recognition module, and the re-recognition process is triggered to fundamentally correct the azimuth judgment result.

[0106] Numerical accuracy assurance for rotation transformation: To ensure the numerical accuracy of rotation transformation, the system takes the following measures: ① Double-precision floating-point numbers (float64) are used for matrix operations throughout the process to avoid the impact of accumulated errors from single-precision floating-point numbers on the results of large-scale vertex coordinate transformations; ② Before applying the rotation matrix to vertex coordinate transformations, orthogonalization correction (Gram-Schmidt orthogonalization or SVD correction) is performed on it to eliminate the small numerical errors introduced by floating-point operations and ensure that the rotation matrix strictly satisfies the orthogonality constraint.

[0107] Example 2: An Automatic Alignment System for 3D Battlefield Target Models Based on the Semantic Meaning of Reconnaissance Image Azimuth

[0108] This invention provides an automatic alignment system for 3D battlefield target models based on the semantics of reconnaissance image azimuth plane, such as... Figure 2 As shown, the system includes three functional units: a preprocessing module, an azimuth surface semantic recognition module, and a model coordinate axis correction and alignment module.

[0109] The preprocessing module performs geometric normalization and multi-view orthogonal rendering image generation operations on the 3D model of the battlefield target, corresponding to the processing content described in step S1 of Embodiment 1. The preprocessing module receives a 3D model file with arbitrary orientation as input and includes a 3D model parsing unit, an axis-aligned bounding box calculation unit, a geometric center normalization unit, an optional size normalization unit, and a multi-view orthogonal rendering unit, outputting six azimuth rendering images and a mapping table of the correspondence between the images and the bounding box azimuth planes.

[0110] The azimuth surface semantic recognition module is used to perform azimuth surface semantic reasoning and semantic local coordinate system construction operations based on a multimodal large model, corresponding to the processing content described in steps S2 and S3 of Embodiment 1. This module includes a multimodal large model input construction unit, a multiple reasoning execution unit, a majority voting statistics unit, a structured recognition result parsing unit, an orthogonality verification unit, and a semantic local coordinate system construction unit, outputting a three-axis orthogonal semantic local coordinate system and the specific values ​​of three semantic azimuth vectors.

[0111] The model coordinate axis correction and alignment module is used to perform rotation matrix construction, validity verification, and model geometric data rotation transformation operations, corresponding to the processing content described in steps S4 and S5 of Embodiment 1. This module includes a target coordinate system configuration reading unit, a basis vector normalization unit, a rotation matrix construction unit, a rotation matrix validity verification unit, a model geometric data rotation transformation unit, and a correction result post-processing and output unit, outputting a standardized 3D model file whose coordinate axes are completely aligned with the target world coordinate system.

[0112] The data interaction relationships between the modules are clear: the output of the preprocessing module (six azimuth rendering images and corresponding mapping tables) serves as the input of the azimuth surface semantic recognition module; the output of the azimuth surface semantic recognition module (semantic local coordinate system and three semantic azimuth vector values) serves as the input of the model coordinate axis correction and alignment module; the output of the model coordinate axis correction and alignment module (aligned standardized 3D model) is the final output of the entire system.

[0113] Example 3: Electronic Equipment

[0114] In some embodiments, the present invention provides an electronic device, such as... Figure 3As shown, the electronic device includes a processor, memory, and a communication interface, with each component interconnected via a bus. The memory includes read-only memory and random access memory, providing instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. When the processor executes the computer program stored in the memory, it implements all the steps of the aforementioned automatic alignment method for 3D battlefield target models based on the semantics of reconnaissance image azimuth planes. Specifically, this includes: parsing the 3D model file and calculating bounding boxes; calling the graphics rendering interface to complete multi-view orthogonal projection rendering; calling the multimodal large model through the API interface to perform semantic reasoning; performing voting statistics, vector extraction, and orthogonality verification; completing the rotation matrix construction and validity verification; and performing full geometric data rotation transformation of the model and outputting the results. The communication interface is used to realize information transmission between this electronic device and other devices, including receiving the input 3D model file to be processed and transmitting the corrected 3D model file to the downstream system.

[0115] Example 4: Computer-readable storage medium

[0116] like Figure 4 As shown, the present invention provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements all the steps of the above-described automatic alignment method for 3D battlefield target models based on the semantics of reconnaissance image azimuth. The computer-readable storage medium may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0117] Example 5: Specific Application Scenarios

[0118] 1. Model standardization processing in AI-driven 3D asset generation pipeline

[0119] With the rapid development of image-to-3D and text-to-3D generation models such as TripoSR, Zero123++, InstantMesh, and Stable Zero123, automated production pipelines for 3D assets in content industries such as games, film and television, and virtual reality are gradually being implemented. In a typical AI 3D content generation workflow, designers or content creators input text descriptions or reference images to trigger AI models to generate a large number of 3D assets in batches. These assets then need to be imported into 3D scene editors (such as Unity, Unreal Engine, Blender, etc.) for scene assembly and layout.

[0120] In the aforementioned production pipeline, this invention is deployed between the output end of the AI-generated model and the input end of the 3D scene editor, serving as an automated preprocessing node. It executes the SPAN processing pipeline on each batch of 3D assets generated by the AI ​​model, automatically identifying the semantic front, top, and side orientations of each model. It then uniformly corrects the coordinate axes of all generated models to a predefined target world coordinate system (such as Unreal Engine's Z-up, X-front coordinate convention), outputting a library of model assets with consistent coordinate axis orientation. AI generation pipelines typically generate multiple categories of 3D assets in the same batch, including buildings, vehicles, characters, furniture, and props, with significant differences in geometric shapes and semantic orientation patterns among these categories. This invention, through its multimodal large-scale model semantic understanding capabilities, eliminates the need to configure processing rules separately for different model categories. It achieves cross-category batch coordinate axis normalization within a unified processing framework, realizing truly universal batch processing.

[0121] 2. Standardization of Multi-Source 3D Model Integration in Industrial Digital Twin Platforms

[0122] In the fields of smart manufacturing and industrial digital twins, industrial enterprises typically need to integrate 3D models from multiple heterogeneous data sources into a unified digital twin platform to build complete digital factories or product digital twins. The data sources for these 3D models are extremely diverse, including: precise geometric models exported from CAD software (such as CATIA, SolidWorks, and AutoCAD); point cloud-converted models collected by 3D scanning equipment (such as laser scanners and structured light scanners); standard part 3D model libraries provided by suppliers; and 3D models generated from product photographs using AI reconstruction technology. Models from different sources use their own local coordinate systems with significantly different coordinate axis orientations. Direct integration into a unified platform can lead to chaotic model spatial poses, preventing them from correctly participating in advanced applications such as physical simulation, kinematic analysis, and human-computer interaction.

[0123] In the aforementioned industrial digital twin integration scenario, this invention serves as a multi-source model standardization processing node. It performs SPAN processing on industrial 3D models from different data sources and defined with different local coordinate systems in batches, uniformly correcting the coordinate axes of all models to the industrial standard coordinate system adopted by the digital twin platform (usually a right-handed coordinate system of Z-up, Y-front). This eliminates the problem of inconsistent coordinate axes among multi-source models and ensures that all models have the correct spatial pose under the unified platform coordinate system.

[0124] 3D models in industrial settings often possess highly complex geometric structures and contain numerous functional features (such as interface surfaces, mounting reference surfaces, and control panels). These functional orientation features are the core basis for determining the "correct posture" of industrial parts, but traditional PCA or physical stability methods cannot perceive them. This invention uses a multimodal large model to semantically understand the rendered view of industrial parts, identifying functional orientations that conform to engineers' cognitive habits, ensuring that the industrial model presents a standard posture in the digital twin platform consistent with the actual installation state.

[0125] The overall system operation process in conjunction with the scenario: Taking the "model standardization preprocessing operation process in the AI-driven 3D content generation pipeline" as an example: Table 3: Overall Configuration Parameters

[0126] Input: The external input for this application scenario is an uncorrected 3D model asset package output in batches by AI 3D content generation models (such as InstantMesh, TripoSR, etc.). The specific input specifications are as follows: Table 4: Input Specifications for Uncalibrated 3D Model Asset Packages

[0127] The steps and procedures are as follows: 1. Batch task initialization: First, scan the input directory and traverse all directories. Establish an ordered task queue for each GLB model file. Simultaneously initialize containing A parallel processing thread pool of worker threads, each worker thread independently processes one model task, and the threads share the API call interface for the multimodal large model (the concurrent call limit is set according to the API service quota, typically [value]). (Number of concurrent requests).

[0128] A batch output directory structure is pre-created, including model files with calibrated coordinate axes, JSON files containing transformation metadata for each model, and six-sided rendered view images of each model.

[0129] 2. Preprocessing module execution: Based on the batch number... Model task (Taking an AI-generated "red sports car" GLB model as an example) This article describes in detail the complete execution process of the preprocessing module.

[0130] First, the worker thread calls the model parsing library (using the Trimesh library, version ≥ 3.22.0) to read the input GLB file. The system then merges the vertex coordinates of the three sub-meshes into a unified global vertex coordinate array. This is used for subsequent bounding box calculations. After reading the model, the system uses a global vertex coordinate array. Perform axis-aligned bounding box calculations and calculate the direction of the outward normal vector. After the bounding box calculation is complete, the system further translates the geometric center of the bounding box until the center of the bounding box is precisely located at the origin. Then, based on the normalized model, the system configures six orthogonal projection virtual cameras and performs rendering (using the pyrender library or the blender offline rendering backend).

[0131] 3. Azimuth Semantic Module Execution: The system combines six rendered images with structured text prompts to construct the complete input for the multimodal large model. The multimodal large model (taking GPT-4V as an example, API call parameters: temperature=0.1, max_tokens=256) receives the image sequence and text prompts and performs semantic inference. Semantic inference is executed three times using different random seeds. Azimuth-specific voting is performed on the three inference results. If all votes pass unanimously, the final recognition result is adopted. Based on the image-azimuth correspondence mapping table, the recognition result is converted into the outer normal vector of the corresponding bounding box. For the extracted outer normal vectors, the pairwise dot product of the three semantic azimuth vectors is calculated to verify their orthogonality. If orthogonality is passed, the cross product relationship of the three normal vectors is verified to confirm whether a right-handed coordinate system is formed. Finally, a semantic local coordinate system is constructed.

[0132] 4. Model Coordinate Axis Correction and Alignment: The system first confirms the standard axis of the target coordinate system, based on the global configuration of this application scenario (Unreal Engine coordinate convention, Z-up, Y-front). After performing normalization verification on the three semantic direction vectors, the rotation matrix is ​​constructed. Source coordinate system basis vector matrix. The column vectors are arranged in the order of right, front, top, and the target coordinate system basis vector matrix. Arrange the column vectors in the order of right, front, top, and solve for them. After obtaining the rotation matrix R, a validity verification is performed. The model is then geometrically transformed according to the rotation matrix R, and a fast coordinate axis alignment verification is performed on the corrected model. The bounding box after transformation is then recalculated.

[0133] 5. Post-processing of calibration results: If downstream tasks require the model to maintain its original physical dimensions, the system performs an inverse scaling operation. The calibrated vertex coordinate array, normal vector array, and original material map, UV texture coordinates, and patch index array are repackaged into a GLB format file for output. (In batch processing...) All After each model task is completed, the system summarizes the processing results of each model and generates a batch processing quality inspection report.

[0134] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0135] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments of this application described herein can be implemented in sequences other than those illustrated or described herein.

[0136] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0137] For ease of description, spatial relative terms such as "above," "on top of," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "on top of" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways, such as rotated 90 degrees or in other orientations, and the spatial relative descriptions used herein will be interpreted accordingly.

[0138] In the detailed description above, reference has been made to the accompanying drawings, which form part of this document. In the drawings, similar symbols typically identify similar parts unless the context otherwise indicates otherwise. The illustrated embodiments described in the detailed specification, drawings, and claims are not intended to be limiting. Other embodiments may be used and other changes may be made without departing from the spirit or scope of the subject matter presented herein.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatic alignment of 3D battlefield target models, characterized in that, Includes the following steps: S1 parses the input 3D model file of the battlefield target, extracts the vertex coordinate set, calculates the axis aligned with the bounding box, and normalizes the geometric center of the model to the origin of the coordinates by translation transformation. Based on the normal vector directions of the six azimuth planes of the bounding box, an orthogonal projection virtual camera is set up, and orthogonal projection rendering is performed on the model from the six directions of positive X, negative X, positive Y, negative Y, positive Z, and negative Z respectively to generate six azimuth rendering images. S2, construct a multimodal large model input by combining the six directional rendering images with text prompts containing task description, directional semantic definition and structured output format requirements, perform K independent semantic reasoning on the same set of inputs, perform directional majority voting on the K reasoning results, and adopt the recognition result when the recognition results of the front, top and side directions pass the vote with more than half of the support rate respectively. The construction of the text prompt includes: stating in the text prompt that the input image is a six-view orthographic projection of a three-dimensional model, and adding a numbered identifier to each image corresponding to its bounding box surface; The text prompts include explicit definitions of the semantics of front, top, and side; the multimodal large model is required to output the correspondence between the azimuth plane number and the semantic category in a predefined JSON format; for symmetrical object categories where the semantic orientation judgment is ambiguous, the multimodal large model is guided to make judgments based on functional features or texture distribution. S3. Based on the orientation plane recognition result output in step S2 and the orientation plane correspondence mapping table of the bounding box, extract the front semantic direction vector, the top semantic direction vector and the side semantic direction vector, verify that the dot product of each pair of the three vectors is approximately zero to confirm the orthogonality, verify the cross product relationship of the three vectors to confirm the right-hand coordinate system, and construct a three-axis orthogonal semantic local coordinate system with the geometric center of the model as the reference point. S4. Construct the source coordinate system basis vector matrix by column using the three unit basis vectors of the semantic local coordinate system, and construct the target coordinate system basis vector matrix by column using the corresponding standard axis of the target world coordinate system. Solve the rotation matrix by multiplying the target basis vector matrix by the transpose of the source coordinate system basis vector matrix, and perform orthogonality verification and determinant verification on the rotation matrix. S5 applies the validated rotation matrix to all vertex coordinates, vertex normals, and face normals of the 3D model, completing a rigid body rotation transformation with the geometric center of the model as the reference point, and outputting a standardized 3D model whose coordinate axes are completely aligned with the target world coordinate system.

2. The automatic alignment method for three-dimensional battlefield target models according to claim 1, characterized in that, The orthographic projection rendering described in step S1 specifically includes: The projection plane size of each orthogonal projection virtual camera is set according to the side length of the corresponding direction of the bounding box to ensure that the model is fully presented in the rendered image without pruning; The rendering process preserves the model's texture maps, material colors, and basic lighting effects. After rendering, six orientation rendering images with uniform resolution and a mapping table showing the correspondence between the six orientation rendering images and the bounding box orientation plane are output.

3. The automatic alignment method for three-dimensional battlefield target models according to claim 1, characterized in that, Step S1 further includes: performing size normalization processing on the 3D model, calculating the maximum side length of the bounding box in the three-axis direction, performing a proportional scaling transformation on the coordinates of all vertices of the model using the reciprocal of the maximum side length as a scaling factor, and normalizing the model to the range of a unit bounding box.

4. The automatic alignment method for three-dimensional battlefield target models according to claim 1, characterized in that, Step S3 also includes: when the orthogonality verification of the three extracted semantic orientation vectors fails, an exception handling process is triggered to re-request the multimodal large model to perform semantic recognition; when the three vectors form a left-handed coordinate system, the lateral basis vectors are inverted and corrected to a right-handed coordinate system.

5. The automatic alignment method for three-dimensional battlefield target models according to claim 1, characterized in that, Step S4 also includes: when the determinant of the rotation matrix is ​​approximately negative one, triggering the mirror component correction mechanism, inverting the column vectors in the rotation matrix corresponding to the side basis vectors, and eliminating the mirror components; the entire matrix operation of the rotation matrix uses double-precision floating-point numbers, and the rotation matrix is ​​orthogonalized before being applied to vertex transformation.

6. The automatic alignment method for three-dimensional battlefield target models according to claim 1, characterized in that, Step S5 also includes: after performing the rotation transformation, inputting the six-sided rendered image of the corrected model into the azimuth semantic recognition process again to verify that the normal vector of the front azimuth plane has been aligned to the front axis direction of the target world coordinate system and the normal vector of the upper azimuth plane has been aligned to the upper axis direction of the target world coordinate system, thus completing the quantitative verification of the alignment result; and recording the applied rotation matrix and target coordinate system configuration information in the metadata of the model file.

7. A battlefield target 3D model automatic alignment system employing the battlefield target 3D model automatic alignment method according to any one of claims 1 to 6, characterized in that, include: The preprocessing module is used to parse the input 3D model file of the battlefield target, calculate the axis to align the bounding box, normalize the geometric center of the model to the origin of the coordinate system by translation transformation, set up an orthogonal projection virtual camera based on the normal vector direction of the six azimuth planes of the bounding box, perform orthogonal projection rendering on the model from the six orthogonal directions, and generate six azimuth rendering images and a mapping table of the correspondence between the images and the azimuth planes of the bounding box. The azimuth semantic recognition module is used to construct a multimodal large model input by combining the six azimuth rendering images with structured text prompts, perform K independent semantic reasoning on the same set of inputs, perform azimuth majority voting on the K reasoning results, and adopt the recognition result that passes the vote with more than half the support rate; based on the recognition results and the corresponding mapping table, extract three semantic azimuth vectors, verify the orthogonality and right-hand coordinate system constraints, and construct a three-axis orthogonal semantic local coordinate system; The model coordinate axis correction and alignment module is used to construct the source coordinate system basis vector matrix in columns using the three unit basis vectors of the semantic local coordinate system, construct the target coordinate system basis vector matrix in columns using the corresponding standard axis of the target world coordinate system, solve the rotation matrix by multiplying the target basis vector matrix by the transpose of the source coordinate system basis vector matrix, perform legality verification on the rotation matrix, apply the rotation matrix to all vertex coordinates, vertex normals and face normals of the 3D model, and output the aligned and standardized 3D model.

8. An electronic device, comprising: A memory and a processor; characterized in that the memory stores a computer program, and the processor executes the computer program to implement the automatic alignment method of a three-dimensional battlefield target model as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the automatic alignment method for three-dimensional battlefield target models as described in any one of claims 1 to 6.

Images