A three-dimensional model animation video generation method

Abstract

This invention relates to the field of 3D animation technology and provides a method for generating 3D model animation videos. The method includes: acquiring basic model data and video generation requirement data of a target product to form a set of model data to be processed; performing structural analysis on the set of model data to obtain a model structure constraint diagram, and performing semantic decomposition on the video generation requirement data to obtain a set of motion intent units; binding the set of motion intent units to the model structure constraint diagram to obtain a structured animation task chain; using the structural relationships and spatial constraints in the model structure constraint diagram as constraints, performing motion trajectory processing on the structured animation task chain to obtain a constrained motion sequence; generating a sequence of motion-related shots and a sequence of synchronized annotations, and performing joint editing processing to obtain a 3D model animation video. This invention can improve the structural accuracy, presentation coherence, and narration synchronization of 3D model animation generation.

Description

Technical Field

[0001] This invention relates to the field of 3D animation technology, specifically a method for generating 3D model animation videos. Background Technology

[0002] In scenarios such as industrial product demonstrations, assembly training, and after-sales maintenance instructions, existing 3D models typically include not only the geometric shapes of components but also component hierarchies, connection relationships, assembly relationships, functional descriptions, and target explanation scripts. How to automatically generate 3D model animation videos that conform to the actual structural constraints of the product and can clearly demonstrate disassembly and assembly relationships and working processes based on this structural and explanatory information is a technical problem that existing technologies need to solve.

[0003] Current technologies typically involve designers importing 3D models into animation software, manually setting keyframes for components, motion paths, camera angles, materials, lighting, and text annotations, and then outputting animated videos through a rendering engine. Alternatively, template-based animation generation methods can be used, where manually selected components perform fixed actions in a specified order based on preset animation templates such as rotation, explosion, assembly, and path movement, and are combined with fixed camera angles or simple automatic camera movements to generate a demonstration video. Some solutions also use text-generated videos or motion-generated models to generate animation clips based on natural language prompts.

[0004] However, in the process of generating animations for industrial equipment, mechanical products, and complex assemblies, existing template-based or artificial intelligence generation methods mostly rely on the user-selected animation type, manually specified sequence, or text prompts as the basis for action generation. They fail to take the component hierarchy, connection relationships, kinematic pair relationships, assembly sequence, and functional dependencies in the 3D model as constraints for action generation. This leads to problems such as clipping, incorrect assembly / disassembly sequence, non-functional movement, and unclear action meaning in the generated component movements. As a result, although the animation video can present dynamic effects, it cannot accurately express the real structural relationships and working mechanisms of the product. Summary of the Invention

[0005] When generating product animation videos based on 3D models, existing technologies usually rely on manual setting of actions, shots, and annotations, or the application of fixed animation templates. It is difficult to automatically determine the sequence of component actions, movement trajectories, and shot display methods according to the structural relationships of the model and the needs of the explanation. Even if actions, shots, and annotations are generated separately, there is a lack of coordination constraints among the three in terms of timeline, screen space, and narration rhythm, which can easily lead to problems such as unreasonable structure, unclear display of key components, and asynchronous explanation and annotation in the animation process.

[0006] To at least partially solve the above problems, this invention proposes a method for generating 3D model animation videos, comprising: Obtain the basic model data and video generation requirement data of the target product, and form a set of model data to be processed based on the basic model data and video generation requirement data. Structural parsing is performed on the dataset of models to be processed to obtain a model structure constraint graph; semantic decomposition is performed on the video generation requirement data to obtain a set of action intent units. The set of action intent units is bound to the model structure constraint graph to obtain a structured animation task chain; The structured animation task chain is processed to obtain a constrained action sequence; Based on the constrained action sequence, generate an action-related shot sequence and a synchronous annotation sequence. Perform joint arrangement processing on the constrained action sequence, action-related shot sequence and synchronous annotation sequence to obtain a 3D model animation video.

[0007] As a preferred implementation, structural parsing is performed on the dataset of models to be processed to obtain a model structure constraint graph, including: The model's basic data in the dataset to be processed is subjected to component layering, and each component is divided into different levels according to the hierarchical relationships in the component layer data. Connection relationship identification is performed on the components, and component pairs with connection relationships and their connection types are determined based on the spatial adjacency and assembly relationship descriptions between components. Spatial bounding box calculation is performed on the components, and their geometric bounding boxes and center point coordinates are determined based on the mesh vertex coordinates of the components. Movable direction identification is performed on the components, and the allowed movement modes, assembly dependency directions, and prohibited crossing areas of the components are determined based on the connection types and assembly order. The processing results of component layering, connection relationship identification, spatial bounding box calculation, and movable direction identification are encapsulated in a graph structure to obtain the model structure constraint graph.

[0008] In a preferred embodiment, semantic decomposition is performed on the video generation requirement data to obtain a set of action intent units, including: The narration script in the video generation requirement data is segmented into sentences according to a preset sentence identifier. Keyword recognition is performed on each sentence, matching the words in the sentence with a preset keyword lexicon to obtain keyword recognition results including action words, component nouns, direction words, and amplitude words. Action word recognition is performed on the keyword recognition results, mapping the identified action words to predefined action types. Target component recognition is performed on the keyword recognition results, matching the identified component nouns with the component names in the model's basic data to obtain candidate target component information corresponding to that sentence. For each sentence, an action intent unit is constructed. The action word recognition and candidate target component information, as well as the action direction and action amplitude jointly determined by the model structure constraint diagram and action type, are encapsulated into the action intent unit and arranged in order according to the sentence number to obtain the action intent unit set.

[0009] As a preferred implementation, the set of motion intent units is bound to a model structure constraint graph to obtain a structured animation task chain, including: For each action intent unit, the candidate target component information is first directly matched with the component name attributes of each node in the model structure constraint diagram. If direct matching fails, a set of candidate components is determined based on component function words, bill of materials level, adjacent components, and geometric position. The matching confidence score is calculated for each candidate component in the set, and the target component node is determined from the set based on the matching confidence score. When the matching confidence score is lower than a preset confidence threshold, the corresponding action intent unit is marked as pending review. Each successfully bound action intent unit is associated and encapsulated with the determined target component node and the corresponding structural relationship edge, and arranged in order according to the action sequence to obtain the structured animation task chain.

[0010] In a preferred embodiment, motion trajectory processing is performed on the structured animation task chain to obtain a constrained motion sequence, including: Using the structural relationships and spatial constraints in the model structure constraint diagram as constraints, trajectory generation is performed on each animation subtask in the structured animation task chain. An initial motion trajectory is generated based on the action type, target component, action direction, and action amplitude in the animation subtask. Collision detection is performed on the initial motion trajectory, and discrete sampling is performed along the initial motion trajectory to detect whether geometric interference occurs between the target component and non-target components. The interpenetration rate is determined based on the proportion of sampling points with geometric interference to the total sampling points. When the interpenetration rate is not lower than a preset trajectory qualification threshold, trajectory correction is performed on the initial motion trajectory by adjusting the trajectory direction, adding intermediate avoidance points, or shortening the single movement step length, and the collision detection is performed iteratively until the interpenetration rate is lower than the trajectory qualification threshold. The keyframe trajectories after trajectory correction are arranged in order according to the execution sequence of the animation subtasks to obtain the constrained action sequence.

[0011] As a preferred embodiment, generating an action-related shot sequence based on the constrained action sequence includes: Each constraint action subsequence in the constraint action sequence is assigned a corresponding shot segment. Candidate shot parameters are generated based on the geometric bounding box size, action type, and motion direction of the target component in the constraint action subsequence. The candidate shot parameters are then filtered based on the exposed surface ratio of the target component under the candidate shot parameters to determine the shot position, shot orientation, focal length parameter, and shot motion mode of the shot segment. When the target components of two adjacent constraint action subsequences are different, a shot transition frame sequence is generated between the two shot segments. All shot segments and shot transition frame sequences are arranged in chronological order to obtain the action-related shot sequence. For each constraint action subsequence in the constraint action sequence, a corresponding annotation frame is generated. Text annotations, component highlighting annotations, and motion trajectory annotations are generated based on the explanatory text fragments, target component geometric information, and motion paths corresponding to the constraint action subsequence. The annotation display position is determined based on the screen projection area of ​​the target component in the corresponding shot segment, and the annotation display position is adjusted when it overlaps with the screen projection area of ​​the target component. Using the action time period of the constraint action subsequence as the time axis alignment reference, all annotation frames are arranged in chronological order to obtain the synchronous annotation sequence.

[0012] As a preferred embodiment, joint arrangement processing is performed on the constrained action sequence, action-related shot sequence, and synchronized annotation sequence to obtain a 3D model animation video, including: Timeline alignment is performed on the constrained action sequence, action-related shot sequence, and synchronized annotation sequence, using the target duration in the video generation requirement data as the total timeline. The constrained action sequence, action-related shot sequence, and synchronized annotation sequence are aligned to the same timeline according to their respective action time periods. The timeline-aligned constrained action sequence is used as the base animation layer, the action-related shot sequence as the shot view layer, and the synchronized annotation sequence as the annotation overlay layer. The layers are overlaid in the order of the base animation layer, the shot view layer, and the annotation overlay layer. The overlaid layer data is rendered frame by frame according to the output resolution parameters in the video generation requirement data, and the rendered image frames are encapsulated and video encoded in timeline order to obtain the 3D model animation video.

[0013] As a preferred embodiment, it further includes: Cross-sequence consistency checks are performed on the constrained action sequence, action-related shot sequence, and synchronized annotation sequence. Deviations are calculated from the dimensions of temporal consistency, spatial consistency, and semantic presentation consistency to obtain cross-sequence consistency evaluation results. The deviations in the cross-sequence consistency evaluation results are compared with the corresponding preset consistency thresholds. When any deviation exceeds the corresponding threshold, the constrained action sequence, action-related shot sequence, and synchronous annotation sequence are adjusted to ensure that the action execution state, shot display range, and annotation presentation timing are coordinated and matched within a unified time axis and a unified screen space. When the time axis consistency deviation exceeds the corresponding threshold, the timestamps of the key action nodes in the constrained action sequence are used as the reference time axis to offset and correct the corresponding timestamps in the action-related shot sequence and synchronous annotation sequence. When the spatial consistency deviation exceeds the corresponding threshold, the lens position and lens orientation of the corresponding lens segment in the action-related shot sequence are recalculated so that the target component is located in the center area of ​​the lens cone, and the label display position of the corresponding label frame in the synchronous label sequence is adjusted so that the label display position avoids the projection area of ​​the target component. When the semantic presentation consistency deviation exceeds the corresponding threshold, the explanatory text fragment of the corresponding annotation frame in the synchronous annotation sequence is replaced or supplemented to ensure that it is semantically consistent with the action type and target component name at that moment. The cross-sequence consistency check is re-executed on the constrained action sequence, action-related shot sequence, and synchronized annotation sequence after the collaborative alignment adjustment. If there are still deviations exceeding the corresponding threshold, the collaborative alignment adjustment is executed cyclically until all deviations are lower than the corresponding threshold or the number of iterations reaches the preset maximum number of iterations.

[0014] Compared with the prior art, the present invention has the following advantages: 1. By acquiring the basic model data of the target product and the video generation requirement data, a set of model data to be processed is formed. Then, the structure of the set of model data to be processed is further analyzed to obtain the model structure constraint diagram. This allows the component hierarchy, structural relationship and spatial constraints in the 3D model to be used as the basis for animation generation, avoiding reliance on manual experience or fixed templates to generate actions, and improving the structural accuracy of 3D model animation generation.

[0015] 2. By performing semantic decomposition on the video generation requirement data, a set of action intent units is obtained, and the set of action intent units is bound to the model structure constraint diagram to obtain a structured animation task chain. This enables the action requirements in the narration script to establish a correspondence with specific model parts, structural relationships and action objects, solving the problem that natural language narration requirements are difficult to directly drive the generation of 3D model animation.

[0016] 3. Based on the structural relationships and spatial constraints in the model structure constraint diagram, the structured animation task chain is processed to obtain a constrained action sequence, so that the motion trajectory of the parts can meet the requirements of assembly relationship, motion direction and spatial avoidance, and reduce the situation of parts intersecting, incorrect motion direction or unreasonable action sequence in the animation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0018] Figure 1 This is an exemplary flowchart of the three-dimensional model animation video generation method provided in the embodiments of the present invention; Figure 2 This is a comparison diagram of the effects of the three-dimensional model animation video generation method provided in the embodiments of the present invention and the prior art, wherein gray bars represent the prior art and black bars represent the present invention. Detailed Implementation

[0019] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0020] Example 1: To achieve the above objectives, please refer to Figures 1 to 2 This invention provides a method for generating 3D model animation videos, the method comprising the following steps: Constructing animation source data: Obtain the basic model data and video generation requirement data of the target product to form a set of model data to be processed; Model Structure and Action Intent Parsing: Perform structure parsing on the model data set to be processed to obtain the model structure constraint graph; perform semantic decomposition on the video generation requirement data to obtain the action intent unit set. Animation task structure binding: The set of action intention units is bound to the model structure constraint diagram to obtain a structured animation task chain; Constrained motion trajectory generation: Perform motion trajectory processing on the structured animation task chain to obtain a constrained motion sequence; Video arrangement and generation: Based on the constrained action sequence, generate an action-related shot sequence and a synchronous annotation sequence. Perform joint arrangement processing on the constrained action sequence, action-related shot sequence and synchronous annotation sequence to obtain a 3D model animation video.

[0021] In this embodiment, the specific content of constructing the animation source data includes: The model's basic data is read from the digital asset repository of the target product. The model's basic data consists of two parts: 3D model data and model-related information. When reading the 3D model data, the format type of the input file is identified. Specifically, the file extension field of the input file is extracted, and the extracted file extension field is matched with a preset list of supported format extensions. The preset list of supported format extensions includes OBJ, FBX, GLTF, STEP, and general 3D format file extensions obtained by format conversion processing after being exported from computer-aided design software. If the match is successful, the format type of the input file is determined. After identifying the file format, the corresponding format parser is called to parse the file, extracting four types of structured data sub-items: mesh data, component hierarchy data, material data, and coordinate data. Mesh data refers to a set of polygonal meshes composed of vertex coordinates, edges, and faces, used to accurately describe the geometric outline of each component of the target product. Component hierarchy data refers to component subordination data organized in a tree structure, used to describe the parent-child hierarchical relationships and assembly hierarchy among the components of the target product. Material data refers to surface attribute description data bound to the geometry of each component, including color parameters, reflectivity parameters, roughness parameters, and texture map indexes, used to restore the visual appearance of the target product in subsequent rendering processes. Coordinate data refers to the position coordinates, rotation angle, and scaling parameters of each component in the three-dimensional world coordinate system, used to determine the absolute and relative positions of each component in three-dimensional space. After extracting the 3D model data, the model's auxiliary information is further read. This auxiliary information refers to non-geometric textual description data that accompanies the 3D model of the target product. The auxiliary information is read item by item according to its source type. Specifically, the bill of materials table is read to obtain the correspondence between component numbers and component names; the component name list is read to establish the mapping between component identifiers and natural language names; the assembly relationship description between components is read to obtain the installation sequence and connection method of each component; and the maintenance manual, product manual, or existing process description text is read to obtain auxiliary semantic information such as operation steps and precautions. After the above auxiliary information is read, the auxiliary information items from each source are associated and bound with the corresponding component geometric data in the 3D model data, using the component number as the index key, to form complete basic model data.

[0022] The system receives video generation requirement data submitted by users through an interactive interface or structured input interface. This video generation requirement data characterizes the user's desired video content goals and display requirements, and consists of six sub-items: target video purpose, narration script, target duration, output resolution, camera style, and key display components. Specifically, the target video purpose refers to the user's description of the usage scenario of the generated video; the narration script is the video content description text provided by the user in natural language; the target duration is the total length of the final output video specified by the user; the output resolution is the pixel size specification of the final video frame specified by the user; the camera style is the user's descriptive instructions on the video camera movement and perspective preferences; and the key display components are one or more target components explicitly specified by the user that need to be highlighted in the video. After receiving the component names in natural language, the system performs string matching with the component name list in the model's attached information to establish an association between the key display component names and the corresponding component geometric data in the 3D model data. The basic model data and video generation requirement data obtained in the above steps are aligned and formatted according to the predefined data structure specifications. Specifically, the extracted sub-items of the 3D model data, the associated items of the model's auxiliary information, the purpose of the target video, the narration script, the target duration, the output resolution, the shot style, and the key display components are sequentially filled into a unified structured data container. The data sub-items are cross-linked and labeled with the component number as the index key to form a set of model data to be processed.

[0023] In this embodiment, the specific content of the model structure and action intent parsing includes: A subset of basic model data is extracted from the dataset of models to be processed. Four processing operations are performed on the subset of basic model data in sequence: component layering, connection relationship identification, spatial bounding box calculation, and motion direction identification. This transforms the static 3D model into structural constraint data that can participate in animation generation. Finally, a model structural constraint diagram is output. The model structural constraint diagram is used to represent the structural relationships, spatial relationships, and motion restriction relationships between the components in the 3D model. It is composed of a set of nodes and a set of edges. When performing component hierarchical operations, the component hierarchy data in the model's basic data is retrieved. Based on the existing tree-like hierarchical relationships in the component hierarchy data, each component is sequentially divided into three hierarchical categories: root-level components, intermediate-level components, and leaf-level components. Root-level components are those without parent nodes in the tree structure; leaf-level components are those without child nodes; and intermediate-level components are those between root-level and leaf-level components. When a component lacks a hierarchical identifier in the component hierarchy data, the bill of materials table in the model's supplementary information is retrieved, and the component's hierarchy is completed according to the hierarchy number in the bill of materials table. When performing the connection relationship identification operation, the relative spatial relationships between each component after layering are calculated pairwise. Specifically, for any two components, the boundary surface coordinates of their bounding boxes are extracted, and the distance between the closest points of the two bounding boxes in the three-dimensional world coordinate system is calculated. The formula for calculating the closest point distance D is as follows: , in This represents the axial distance between the geometric bounding boxes of two components along the X-axis. This represents the axial distance between the geometric bounding boxes of two components along the Y-axis. This represents the axial distance between the geometric bounding boxes of two components along the Z-axis; when , , When all three values ​​are 0, it means that the geometric bounding boxes of the two parts have intersected or fit together in three-dimensional space, and the nearest point distance D is 0. When the distance between the nearest points is less than a preset connection threshold, it is determined that there is a connection relationship between the two components; further, according to the assembly relationship description between components in the model's supplementary information, the connection type of the component pairs that are determined to have a connection relationship is marked. The connection type includes bolted connection, snap-fit ​​connection, welded connection, hinge connection, sliding fit connection and plug-in connection. In one embodiment, the connection threshold is 0.1mm. When performing the spatial bounding box calculation operation, the mesh data of each component is traversed, and the coordinates of all vertices of the component are extracted. The maximum and minimum values ​​of the vertex coordinates are taken in the X-axis, Y-axis, and Z-axis directions of the three-dimensional world coordinate system. The six boundary planes formed by the obtained maximum and minimum values ​​in the three axes together form an axis-aligned bounding box, which is used as the geometric bounding box of the component. Further, the arithmetic mean of the coordinates of the eight vertices of the geometric bounding box is taken to obtain the coordinates of the center point of the component. When performing the movable direction recognition operation, based on the connection type marked in the connection relationship recognition operation, the allowed movement mode of each component in the assembly relationship is determined. The allowed movement mode includes one or more combinations of six basic movement modes: translation along the X-axis, translation along the Y-axis, translation along the Z-axis, rotation around the X-axis, rotation around the Y-axis, and rotation around the Z-axis. At the same time, based on the assembly sequence recorded in the assembly relationship description between components, the assembly dependency direction of the component during the disassembly and assembly process is determined. The assembly dependency direction is used to characterize the entry and exit direction that the component must follow relative to its adjacent components during disassembly and assembly. Further, interference judgment is performed between the allowed movement path of the component and the geometric bounding boxes of other adjacent components, and the spatial area where geometric interference occurs is marked as a prohibited crossing area. The geometric interference includes two judgment methods: geometric bounding box overlap and mesh interpenetration. First, the geometric bounding box overlap judgment is performed. When there is an overlapping area, the mesh interpenetration judgment is further performed. When there is triangular facet intersection, it is determined that geometric interference has occurred. After completing the above four processing operations, the model structure constraint graph is constructed according to the graph structure definition method of nodes and edges. Specifically, each component is abstracted as a node, and the following node attributes are configured for this node: component name, component number, geometric bounding box, center point coordinates, material category, exposed surface ratio, and visibility priority. Two components with a connection relationship are abstracted as an edge, and the following edge attributes are configured for this edge: connection type, adjacent distance, assembly dependency direction, allowed movement mode, and prohibited crossing area. Among them, the component name and component number are obtained by looking up the component name list and material list table in the model's attached information. The geometric bounding box and center point coordinates are obtained by the spatial bounding box calculation operation. The material category is mapped from the material identifier field in the material data in the model's basic data. The connection type, assembly dependency direction, and allowed movement mode are obtained by the connection relationship identification operation and the movable direction identification operation. The adjacent distance is taken as the nearest point distance calculated in the corresponding connection relationship identification operation. The prohibited crossing area is obtained from the interference judgment result in the movable direction identification operation. For the exposed surface ratio in the node attributes, the set of triangular facets on the surface of the component is traversed piece by piece. For each triangular facet, a ray is emitted from the geometric center of the triangular facet towards its normal direction. It is determined whether the ray intersects with the geometric bounding box of the adjacent component. If no intersection occurs, the triangular facet is determined to be an unoccluded facet. If an intersection occurs, the triangular facet is determined to be an occluded facet. The sum of the areas of all unoccluded triangular facets of the component is calculated. The sum of the areas is divided by the total area of ​​all triangular facets of the component. The resulting ratio is the exposed surface ratio of the component. The visibility priority of the node attributes is determined based on the number of times the component appears in the explanation script, its hierarchical position in the bill of materials table, and the user-specified level of importance, specifically calculated using the following formula: , Where i represents the component number. This represents the visibility priority value of the i-th component. The normalized frequency of the component name or its synonym of the i-th component in the explanation script is obtained by counting the total number of times the component name and its synonym appear in the explanation script and dividing the total number of appearances by the total number of appearances of all component names and their synonyms in the explanation script. The normalized result is between 0 and 1. This indicates the normalized importance of the i-th component in the bill of materials table or component hierarchy data. The original importance value is the reciprocal of the hierarchy depth of the component in the hierarchy tree. The original importance values ​​of all components are normalized to the maximum value to obtain a normalized result with a value between 0 and 1. This indicates the user-specified level of importance; when the highlighted display components in the video generation requirement data include this component, The value is 1, when the key display components in the video generation requirement data do not include this component. The value is 0.5; wherein, the source of the synonym is the synonym field corresponding to the list of component names in the model's supplementary information or a preset synonym dictionary; the larger the visible priority value, the higher the shot weight assigned to the component in the subsequent action-related shot sequence generation stage; All nodes and edges obtained above are encapsulated according to the graph structure storage format to form the model structure constraint graph; The script is segmented using Chinese periods, semicolons, question marks, exclamation marks, and line breaks as segmentation identifiers to obtain a sequence of sentence texts, which are then assigned sequence numbers according to their order of appearance. When the length of a single sentence exceeds the maximum sentence length threshold, it is segmented again using Chinese commas as secondary segmentation identifiers. In one embodiment, the maximum sentence length threshold is 80 characters. Further, each sentence is segmented into words, and the resulting word sequence is matched against a preset keyword lexicon. The keyword lexicon includes four sub-lexicons: action word lexicon, component noun lexicon, direction word lexicon, and amplitude word lexicon. The matched keywords are categorized according to their respective sub-lexicons. The entries in the component noun lexicon are derived from the list of component names and their synonym fields in the model's supplementary information. The extracted action words are matched with a preset action type mapping table to map natural language action words to predefined action types. The predefined action types include nine basic action types: rotation, translation, disassembly, assembly, slicing, transparency, explosion, local magnification, and state switching. The extracted component names are matched with the component name list and synonym field in the model's basic data to obtain a set of candidate component names. When multiple candidate component names exist, all candidate component names are written into the candidate target component field of the corresponding action intent unit. When no candidate component names exist, the candidate target component field of the sentence is marked as pending completion, and the sentence is added to the queue for manual review. For each sentence, a motion intent unit is constructed. The motion intent unit includes seven fields: candidate target component, motion type, motion direction, motion amplitude, motion start condition, motion end condition, and explanatory text fragment. The candidate target component field is filled with the set of candidate component names obtained by the target component recognition operation. The motion type field is filled with the motion type determined by the action word recognition operation. The explanatory text fragment field is filled with the original text content of the sentence. The motion start condition field and the motion end condition field are filled with the start timestamp and end timestamp, respectively, obtained by allocating the target duration in the video generation requirement data according to the proportion of character length of each sentence. When the video generation requirement data does not contain a target duration, a default target duration is generated based on the number of sentences and the number of motion intent units in the explanatory script, and a preset motion duration is assigned to each motion intent unit. In one embodiment, the preset motion duration of each motion intent unit is 3 seconds. Specifically, the action direction field in the action intent unit first retrieves the node corresponding to the candidate target component and the set of edges connected to that node in the model structure constraint graph, and searches the set of edges to see if there is a value for the assembly dependency direction attribute. If there is a value for the assembly dependency direction attribute, the assembly dependency direction is used as the action direction value. If there is no value for the assembly dependency direction attribute in the set of edges or the target component has no connecting edges in the model structure constraint graph, the coordinates of the center point of the target component and the coordinates of the center point of the model are retrieved. The coordinates of the center point of the model are obtained by taking the arithmetic mean of the coordinates of the center points of all nodes in the model structure constraint graph, and the unit vector obtained by normalizing the direction vector from the center point of the model to the center point of the target component is used as the action direction value. For the motion amplitude field in the motion intent unit, the geometric bounding box of the target component is retrieved, and the diagonal length L of the geometric bounding box is calculated using the following formula: , in , These represent the maximum and minimum coordinate values ​​of the bounding box along the X-axis, respectively; , These represent the maximum and minimum coordinate values ​​of the bounding box along the Y-axis, respectively; , These represent the maximum and minimum coordinate values ​​of the bounding box along the Z-axis, respectively; Furthermore, the amplitude of the action is differentiated according to the action type: when the action type is translation, disassembly, or explosive unfolding, the displacement distance of the action amplitude is taken as 1.2 times the diagonal length L; when the action type is rotation, the rotation axis is determined according to the allowed movement mode of the corresponding side in the model structure constraint diagram, and the rotation angle is determined according to the modifier of the action word in the clause; when the action type is sectioning, the position of the sectioning plane is taken as the center point of the geometric bounding box of the target component, and the normal of the sectioning plane is taken as the unit vector of the action direction; when the action type is transparency, the material transparency parameter is gradually changed from 1.0 to a preset transparency target value. In one embodiment, the transparency target value is 0.2; when the action type is local magnification, the magnification factor is taken as 2 times; when the action type is state switching, the state parameters before and after the switching are determined according to the state description recorded in the process description text in the model's supplementary information. All the action intent units constructed above are arranged in order according to the clause number to which they belong, forming the action intent unit set. In this embodiment, the specific content of the animation task structure binding includes: The candidate target component field in the action intent unit is matched with the component name, component number, synonym field, and functional description of each node in the model structure constraint diagram. When a match is successful, the corresponding node is determined as the target component node. When a match fails or multiple candidate nodes exist, the set of candidate component nodes is determined sequentially according to component function words, bill of materials level, adjacent components, and geometric position. The matching confidence S is calculated for each candidate node in the set of candidate component nodes, and the matching confidence S is calculated using the following formula: , in, The name similarity is calculated by normalizing the edit distance between the target component name and the candidate node component name in the action intent unit. The functional similarity is calculated by using word vector cosine similarity between the functional description in the action intent unit and the functional word text associated with the candidate node. This indicates structural adjacency consistency. The value is 1 when the candidate node and the target component node of the previous action intention unit are directly connected by an edge in the model structure constraint graph, and 0 otherwise. To indicate context consistency, the value is 1 when the action type of the candidate node is executable and the action type of the action intent unit matches the allowed movement mode of the corresponding edge in the model structure constraint graph; otherwise, the value is 0. , , , These represent the weighting coefficients for the four similarity factors mentioned above. In one implementation, Take 0.4, Take 0.3, Take 0.2, Take 0.1; The candidate node with the highest matching confidence value in the candidate component node set is determined as the target component node of the action intent unit. When multiple candidate nodes have the same matching confidence value, the candidate node with the highest visual priority is further selected as the target component node. When the highest matching confidence value is lower than a preset confidence threshold, the action intent unit is added to the queue for manual review. In one embodiment, the default value of the confidence threshold is 0.70. For each successfully bound motion intent unit, the motion intent unit is encapsulated with the determined target component node, the corresponding structural relationship edge, the action sequence number, and the explanatory text fragment to form an animation subtask. All animation subtasks are arranged in order according to the sentence sequence number of each motion intent unit in the motion intent unit set to form the structured animation task chain. The structured animation task chain is passed as a whole to the subsequent motion trajectory processing step as input data for generating the constrained motion sequence.

[0024] In this embodiment, the specific content of the constraint motion trajectory generation includes: Using the structural relationship edges and spatial constraint attributes in the model structure constraint diagram as constraints, each animation subtask in the structured animation task chain is sequentially processed with three operations: trajectory generation, collision detection, and trajectory correction. Specifically, the trajectory generation uses the coordinates of the center point of the target component as the trajectory starting point, the unit vector of the action direction as the trajectory direction vector, and the displacement distance or rotation angle corresponding to the action amplitude as the trajectory termination condition. An initial motion trajectory sequence composed of several keyframe poses is generated according to a preset keyframe sampling interval. When the action type is translation, separation, or explosion unfolding, the initial motion trajectory sequence is an equidistant straight line trajectory of the center point of the target component along the action direction. When the action type is rotation, the initial motion trajectory sequence is an equiangular rotation trajectory of the target component around the rotation axis specified by the corresponding edge in the model structure constraint diagram. When the action type is cutting, transparency, local magnification, or state switching, the initial motion trajectory sequence is a gradient interpolation sequence of corresponding parameters. In one embodiment, the keyframe sampling interval is 40ms. The collision detection involves discretely sampling the pose of the target component along the initial motion trajectory sequence to obtain a set of sampling points. For each sampling point, the geometric bounding box of the target component at that sampling point pose is compared with the geometric bounding boxes of other non-target component nodes in the model structure constraint diagram. When any two geometric bounding boxes overlap in three-dimensional space, a mesh interpenetration determination is further performed. This mesh interpenetration determination is obtained by detecting the intersection of triangular faces in the target component mesh data with triangular faces in the adjacent non-target component mesh data. When triangular faces intersect, the sampling point is determined to be an interpenetration sampling point. The proportion of interpenetration sampling points to the total number of sampling points is counted and recorded as the interpenetration rate. The trajectory correction compares the interlacing rate with a preset trajectory qualification threshold. When the interlacing rate is not lower than the trajectory qualification threshold, a trajectory correction sub-operation is executed. The trajectory correction sub-operation includes one or more of the following three correction methods: The first correction method is to adjust the trajectory direction, specifically, to deflect the motion direction unit vector along a direction perpendicular to the normal of the prohibited crossing area, with the deflection angle ranging from 5 degrees to 30 degrees; The second correction method is to add intermediate avoidance points, specifically, to insert several non-interlacing transition poses as intermediate keyframes before and after the interlacing sampling point, so that the target... The component first moves along the avoidance direction to the middle avoidance point pose and then continues to move towards the termination pose; the third correction method is to shorten the single movement step size, specifically, reduce the keyframe sampling interval to half of the original value and regenerate the trajectory; the above three correction methods are executed in sequence and iteratively until the interlacing rate is lower than the trajectory qualification threshold and the iteration stops. In one embodiment, the default value of the trajectory qualification threshold is 2%; the maximum number of iterations of the trajectory correction sub-operation is 10 times. If the interlacing rate requirement is still not met after exceeding the maximum number of iterations, the animation sub-task is added to the queue for manual review; After completing the above three processing operations, for each animation subtask, the keyframe trajectory obtained after trajectory correction is encapsulated with the target component, action type, action time period, motion direction, and keyframe state to form a constrained action subsequence. The action time period is determined by the action start condition and action end condition in the animation subtask, and the keyframe state includes the target component pose parameters, material transparency parameters, and scaling parameters corresponding to each keyframe. All constrained action subsequences are arranged in order according to the execution order of each animation subtask in the structured animation task chain to form the constrained action sequence.

[0025] In this embodiment, the specific content generated by video arrangement includes: Each constraint action subsequence in the constraint action sequence is assigned a corresponding shot segment. Specifically, based on the target component's geometric bounding box size, action type, and direction of movement in the constraint action subsequence, the shot segment's shot type, shot position, shot orientation, focal length parameter, and shot movement mode are determined. The shot type includes surround shot, zoom-in shot, zoom-out shot, top-down fixed shot, and follow shot. The shot position is calculated using the following formula: , in, Represents the lens position coordinate vector. This represents the coordinate vector of the center point of the target component. Indicates the diagonal length of the geometric bounding box of the target component; The direction vector is defined as the unit direction vector from the center point of the target component to the camera position. It is determined by the camera style parameters in the video generation requirement data and the exposed surface ratio of the target component, so that the camera orientation falls in the direction with the largest exposed surface ratio to ensure the visual visibility of the target component; k represents the camera distance magnification, and in one embodiment, k is 1.5; the camera orientation is the direction vector from the camera position to the center point of the target component; when the target components of two adjacent constraint action sub-sequences are different, a camera transition frame sequence is inserted between the two camera segments. The camera transition frame sequence is generated by linear interpolation, and in one embodiment, the camera transition duration is 0.5 seconds; all camera segments and camera transition frame sequences are arranged in chronological order to form the action-related camera sequence; For each constraint action subsequence in the constraint action sequence, a corresponding annotation frame is generated. The annotation frame includes three types of annotation sub-items: action description annotation, component highlight annotation, and motion trajectory annotation. The text content of the action description annotation is filled in by the explanatory text fragment in the corresponding animation sub-task and is displayed synchronously at a preset position on the video screen according to the action time period. The component highlight annotation is achieved by superimposing a preset highlight color on the material color parameter of the target component. In one embodiment, the transparency of the highlight color is set to 0.3. The motion trajectory annotation is achieved by drawing the motion path of the center point of the target component in the video screen with continuous line segments. Each annotation sub-item is aligned with the action time period of the constraint action subsequence as the time axis alignment reference, and all annotation frames are arranged in chronological order to form the synchronous annotation sequence.

[0026] The joint arrangement involves three sub-steps: timeline alignment, layer overlay, and rendering output. In the timeline alignment sub-step, the target duration in the video generation requirement data is used as the total timeline, and the constrained action sequence, action-related shot sequence, and synchronized annotation sequence are aligned to the same timeline according to their respective action time periods or shot time intervals. In the layer overlay sub-step, the constrained action sequence is used as the base animation layer, the action-related shot sequence as the shot view layer, and the synchronized annotation sequence as the annotation overlay layer, and the layers are overlaid in the order of base animation layer, shot view layer, and annotation overlay layer. In the rendering output sub-step, the overlaid layer data is rendered frame by frame according to the output resolution parameters in the video generation requirement data. During each frame rendering process, the material data in the model base data is called for surface shading, and the rendered image frames are encapsulated using video encoding in timeline order to obtain the 3D model animation video. In one embodiment, the video encoding frame rate is one of 24 frames per second, 30 frames per second, or 60 frames per second.

[0027] like Figure 2The image shows a comparison of the effects of a method for generating 3D model animation videos. The horizontal axis lists key performance indicators, and the vertical axis represents the exemplified performance score, ranging from 0 to 100%. Higher values ​​indicate better performance. The aim is to intuitively demonstrate the expected improvement of the invention in key capabilities compared to typical prior art.

[0028] Example 2: In Embodiment 1 of the present invention, by parsing the model structure and action intent, binding the animation task structure, generating the action trajectory under structural constraints, and arranging the shot annotations synchronously, the three-dimensional model, auxiliary information, and video generation requirements of the target product can be transformed into a three-dimensional model animation video, so that the animation action has a clear target component, structural source, and motion trajectory.

[0029] However, in the actual generation process, although the constraint action sequence, action-related shot sequence and synchronous annotation sequence meet their respective generation conditions, there may still be a problem of insufficient cross-sequence consistency among the three. This results in the action execution state, shot display range and annotation presentation timing not being coordinated and matched within a unified time axis and unified screen space, which in turn affects the viewer's continuous understanding of the product structure relationship, component movement process and explanation content.

[0030] In this embodiment, to address the deficiency of insufficient cross-sequence consistency among the constraint action sequence, action-related shot sequence, and synchronous annotation sequence in Embodiment 1, three operations are performed sequentially: cross-sequence consistency verification, collaborative alignment adjustment, and joint rendering processing, so that the action execution state, shot display range, and annotation presentation timing can be coordinated and matched within a unified timeline and a unified screen space. The action execution state refers to the motion stage and target component pose of each constrained action sub-sequence at that moment; the camera display range refers to the three-dimensional spatial area covered by the camera's view frustum at that moment; the annotation presentation timing refers to the start timestamp and duration of each annotation frame; The cross-sequence consistency check calculates the deviations of the constrained action sequence, action-related shot sequence, and synchronized annotation sequence from three dimensions: temporal consistency, spatial consistency, and semantic presentation consistency, to obtain the cross-sequence consistency evaluation result; temporal consistency deviation. The calculation formula is: , in, This represents the time axis consistency deviation value, M represents the total number of critical action nodes in the constraint action sequence, and j represents the sequence number of the critical action node. This represents the timestamp of the j-th critical action node in the constrained action sequence. This represents the start timestamp of the shot segment corresponding to the j-th key action node in the action-related shot sequence. This represents the start timestamp of the annotation frame corresponding to the j-th key action node in the synchronous annotation sequence; Image spatial consistency deviation The calculation formula is: , in, This indicates the value of the image space consistency deviation. This represents the visible area ratio of the target component corresponding to the j-th critical action node within the lens's view frustum at that moment. This visible area ratio is obtained by projecting the target component's geometric bounding box onto the lens's view frustum plane and calculating the proportion of the projected area within the view frustum to the total projected area of ​​the target component. Furthermore, this visible area ratio is compared with the annotation display position of the corresponding annotation frame in the synchronous annotation sequence for overlap determination. When the annotation display position does not overlap with the target component's projected area, [the following is omitted as it is not part of the original text]. The value is multiplied by 0.5 for attenuation; Semantic presentation consistency deviation The calculation formula is: , in, This indicates the value of the semantic consistency deviation. The semantic matching degree between the explanatory text segment corresponding to the j-th key action node and the action type and target component name at that moment is calculated using word vector cosine similarity. When performing the collaborative alignment adjustment operation, the calculated , , The values ​​are compared with preset timeline consistency thresholds, spatial consistency thresholds, and semantic presentation consistency thresholds. When any deviation exceeds the corresponding threshold, adjustments are made according to the following rules: When the timeline consistency threshold is exceeded, the timestamps of key action nodes in the constrained action sequence are used as the reference timeline. The starting timestamps of the corresponding shot segments in the action-related shot sequence and the starting timestamps of the corresponding labeled frames in the synchronized annotation sequence are offset and corrected to align them with the reference timeline. The time offset is measured as follows: and , The difference; when When the spatial consistency threshold is exceeded, the camera position and orientation of the corresponding shot segment in the action-related shot sequence are recalculated to ensure the target component is located in the center region of the camera's view frustum. Simultaneously, the display position of the annotations in the corresponding annotation frames of the synchronized annotation sequence is adjusted to the adjacent blank area of ​​the target component's projection region. When the semantic presentation consistency threshold is exceeded, the explanatory text fragments of the corresponding annotation frames in the synchronous annotation sequence are replaced or supplemented to ensure semantic consistency with the action type and target component name at that moment. In one implementation, the time axis consistency threshold is 0.1 seconds, the screen space consistency threshold is 0.2 seconds, and the semantic presentation consistency threshold is 0.2 seconds. The maximum number of iterations for the above collaborative alignment adjustment operation is 5 times. If the threshold requirement is not met even after exceeding the maximum number of iterations, the corresponding key action node is added to the queue for manual review. When performing joint rendering processing, the constrained action sequence, action-related shot sequence, and synchronized annotation sequence, which have been adjusted through collaborative alignment, are overlaid and rendered in frames according to a unified time axis and a unified screen space coordinate system to obtain the 3D model animation video.

[0031] The embodiments of the present invention described above are subject to modification and change of method by those skilled in the art without departing from the embodiments and broader aspects of the present invention. The appended claims are intended to include all such modifications and changes of method that do not depart from the present invention.

Claims

1. A method for generating 3D model animation videos, characterized in that, include: Obtain the basic model data and video generation requirement data of the target product, and form a set of model data to be processed based on the basic model data and video generation requirement data. Structural parsing is performed on the dataset of models to be processed to obtain a model structure constraint graph; semantic decomposition is performed on the video generation requirement data to obtain a set of action intent units. The set of action intent units is bound to the model structure constraint graph to obtain a structured animation task chain; The structured animation task chain is processed to obtain a constrained action sequence; Based on the constrained action sequence, generate an action-related shot sequence and a synchronous annotation sequence. Perform joint arrangement processing on the constrained action sequence, action-related shot sequence and synchronous annotation sequence to obtain a 3D model animation video.

2. The method for generating 3D model animation video according to claim 1, characterized in that, Perform structure parsing on the dataset of models to be processed to obtain a model structure constraint graph, including: The model's basic data in the dataset to be processed is subjected to component layering, and each component is divided into different levels according to the hierarchical relationships in the component layer data. Connection relationship identification is performed on the components, and component pairs with connection relationships and their connection types are determined based on the spatial adjacency and assembly relationship descriptions between components. Spatial bounding box calculation is performed on the components, and their geometric bounding boxes and center point coordinates are determined based on the mesh vertex coordinates of the components. Movable direction identification is performed on the components, and the allowed movement modes, assembly dependency directions, and prohibited crossing areas of the components are determined based on the connection types and assembly order. The processing results of component layering, connection relationship identification, spatial bounding box calculation, and movable direction identification are encapsulated in a graph structure to obtain the model structure constraint graph.

3. The method for generating 3D model animation video according to claim 1, characterized in that, Semantic decomposition is performed on the video generation requirement data to obtain a set of action intent units, including: The narration script in the video generation requirement data is segmented into sentences according to a preset sentence identifier. Keyword recognition is performed on each sentence, matching the words in the sentence with a preset keyword lexicon to obtain keyword recognition results including action words, component nouns, direction words, and amplitude words. Action word recognition is performed on the keyword recognition results, mapping the identified action words to predefined action types. Candidate target component recognition is performed on the keyword recognition results, matching the identified component nouns with the component names in the model's basic data to obtain the candidate target component information corresponding to that sentence. For each sentence, an action intent unit is constructed. The action word recognition and candidate target component information, as well as the action direction and action amplitude jointly determined by the model structure constraint diagram and action type, are encapsulated into the action intent unit and arranged in order according to the sentence number to obtain the action intent unit set.

4. The method for generating 3D model animation video according to claim 1, characterized in that, Binding the set of motion intent units to the model structure constraint graph yields a structured animation task chain, including: For each action intent unit, the candidate target component information is first directly matched with the component name attributes of each node in the model structure constraint diagram. If direct matching fails, a set of candidate components is determined based on component function words, bill of materials level, adjacent components, and geometric position. The matching confidence score is calculated for each candidate component in the set, and the target component node is determined from the set based on the matching confidence score. When the matching confidence score is lower than a preset confidence threshold, the corresponding action intent unit is marked as pending review. Each successfully bound action intent unit is associated and encapsulated with the determined target component node and the corresponding structural relationship edge, and arranged in order according to the action sequence to obtain the structured animation task chain.

5. The method for generating 3D model animation video according to claim 1, characterized in that, The structured animation task chain is processed to obtain a constrained action sequence, including: Using the structural relationships and spatial constraints in the model structure constraint diagram as constraints, trajectory generation is performed on each animation subtask in the structured animation task chain. An initial motion trajectory is generated based on the action type, target component, action direction, and action amplitude in the animation subtask. Collision detection is performed on the initial motion trajectory, and discrete sampling is performed along the initial motion trajectory to detect whether geometric interference occurs between the target component and non-target components. The interpenetration rate is determined based on the proportion of sampling points with geometric interference to the total sampling points. When the interpenetration rate is not lower than a preset trajectory qualification threshold, trajectory correction is performed on the initial motion trajectory by adjusting the trajectory direction, adding intermediate avoidance points, or shortening the single movement step length, and the collision detection is performed iteratively until the interpenetration rate is lower than the trajectory qualification threshold. The keyframe trajectories after trajectory correction are arranged in order according to the execution sequence of the animation subtasks to obtain the constrained action sequence.

6. The method for generating 3D model animation video according to claim 1, characterized in that, Generate an action-related shot sequence based on the constrained action sequence, including: Each constraint action subsequence in the constraint action sequence is assigned a corresponding shot segment. Candidate shot parameters are generated based on the bounding box size, action type, and motion direction of the target component in the constraint action subsequence. The candidate shot parameters are then filtered based on the exposed surface ratio of the target component under the candidate shot parameters to determine the shot position, shot orientation, focal length parameter, and shot motion mode of the shot segment. When the target components of two adjacent constraint action subsequences are different, a shot transition frame sequence is generated between the two shot segments. All shot segments and shot transition frame sequences are arranged in chronological order to obtain the action-related shot sequence.

7. The method for generating 3D model animation video according to claim 6, characterized in that, Also includes: For each constraint action subsequence in the constraint action sequence, a corresponding annotation frame is generated. Text annotations, component highlighting annotations, and motion trajectory annotations are generated based on the explanatory text fragments, target component geometric information, and motion paths corresponding to the constraint action subsequence. The annotation display position is determined based on the screen projection area of ​​the target component in the corresponding shot segment, and the annotation display position is adjusted when it overlaps with the screen projection area of ​​the target component. Using the action time period of the constraint action subsequence as the time axis alignment reference, all annotation frames are arranged in chronological order to obtain the synchronous annotation sequence.

8. The method for generating 3D model animation video according to claim 1, characterized in that, The constrained action sequence, action-related shot sequence, and synchronized annotation sequence are jointly arranged to obtain a 3D model animation video, including: Timeline alignment is performed on the constrained action sequence, action-related shot sequence, and synchronized annotation sequence, using the target duration in the video generation requirement data as the total timeline. The constrained action sequence, action-related shot sequence, and synchronized annotation sequence are aligned to the same timeline according to their respective action time periods. The timeline-aligned constrained action sequence is used as the base animation layer, the action-related shot sequence as the shot view layer, and the synchronized annotation sequence as the annotation overlay layer. The layers are overlaid in the order of the base animation layer, the shot view layer, and the annotation overlay layer. The overlaid layer data is rendered frame by frame according to the output resolution parameters in the video generation requirement data, and the rendered image frames are encapsulated and video encoded in timeline order to obtain the 3D model animation video.

9. The method for generating three-dimensional model animation videos according to claim 1, characterized in that, Also includes: Cross-sequence consistency checks are performed on the constrained action sequence, action-related shot sequence, and synchronized annotation sequence. Deviations are calculated from the dimensions of temporal consistency, spatial consistency, and semantic presentation consistency to obtain cross-sequence consistency evaluation results. The deviations in the cross-sequence consistency evaluation results are compared with the corresponding preset consistency thresholds. When any deviation exceeds the corresponding threshold, the constrained action sequence, action-related shot sequence, and synchronous annotation sequence are adjusted to coordinate the action execution state, shot display range, and annotation presentation timing within a unified time axis and a unified screen space.

10. The method for generating three-dimensional model animation videos according to claim 9, characterized in that, Perform collaborative alignment adjustments on the constrained action sequence, action-related shot sequence, and synchronized annotation sequence, including: When the time axis consistency deviation exceeds the corresponding threshold, the timestamps of key action nodes in the constrained action sequence are used as the reference time axis to offset and correct the corresponding timestamps in the action-related shot sequence and the synchronous annotation sequence. When the spatial consistency deviation exceeds the corresponding threshold, the lens position and lens orientation of the corresponding lens segment in the action-related shot sequence are recalculated so that the target component is located in the center area of ​​the lens cone, and the label display position of the corresponding label frame in the synchronous label sequence is adjusted so that the label display position avoids the projection area of ​​the target component. When the semantic presentation consistency deviation exceeds the corresponding threshold, the explanatory text fragment of the corresponding annotation frame in the synchronous annotation sequence is replaced or supplemented to ensure that it is semantically consistent with the action type and target component name at that moment. The cross-sequence consistency check is re-executed on the constrained action sequence, action-related shot sequence, and synchronized annotation sequence after the collaborative alignment adjustment. If there are still deviations exceeding the corresponding threshold, the collaborative alignment adjustment is executed cyclically until all deviations are lower than the corresponding threshold or the number of iterations reaches the preset maximum number of iterations.

Images