Exhibition stand assembly process visual guidance method based on three-dimensional modeling

Abstract

The invention discloses an exhibition stand assembly process visual guidance method based on three-dimensional modeling, and the method comprises the steps: collecting multi-source space-time perception data, and forming standardized multi-source space-time perception data; sequentially organizing the standardized multi-source space-time sensing data to form a component instance set; constructing an improved D-NeRF model to obtain a structure surface with continuous geometric representation, a component-level dynamic pose field and shielding complementation; constructing a structure state vector to form a structure state vector sequence; constructing a structure state evolution analysis module, and generating a short-time prediction state vector; and carrying out structure feasibility judgment on the next splicing operation to generate corresponding visual guidance information. According to the method, the improved D-NeRF model and the DM decomposition method are introduced, so that real-time fusion and accurate guidance of three-dimensional visual guidance and structural feasibility judgment in the exhibition stand assembling process are realized.

Description

Technical Field

[0001] This invention relates to the field of 3D modeling and intelligent assembly guidance technology, and in particular to a visualization guidance method for the booth assembly process based on 3D modeling. Background Technology

[0002] With the rapid development of the exhibition economy and commercial display activities, modular and detachable booths are widely used in exhibitions, commercial promotions, and temporary event setups. Traditional booth assembly processes typically rely on manual experience, two-dimensional construction drawings, or pre-made 3D demonstration animations for guidance. Construction workers need to install components step-by-step on-site, referring to the drawings or videos. Due to the complex structure of the booth, the large number of components, and the interdependence of assembly sequences, traditional guidance methods struggle to reflect the actual assembly status on-site in a timely manner. This can easily lead to problems such as unreasonable component assembly sequences, failure to detect spatial interference in advance, and insufficient structural stability, resulting in rework, low assembly efficiency, and even safety hazards.

[0003] Some existing technologies have begun to incorporate 3D modeling, augmented reality, or digital aids to visualize and guide the booth assembly process. However, these technologies often rely on pre-built ideal 3D models or fixed assembly animations, lacking the ability to dynamically perceive and continuously reconstruct the actual assembly process on-site. They cannot accurately reflect the continuous state changes of components during actual insertion, alignment, and fixing. Furthermore, existing visualization guidance technologies typically only provide static positional prompts, making it difficult to assess the structural feasibility of the next assembly operation based on structural state evolution. This results in a disconnect between 3D visualization guidance and the actual assembled structural state, failing to meet the high-precision, high-reliability assembly guidance requirements of complex booths.

[0004] Therefore, how to provide a visual guidance method for the booth assembly process based on 3D modeling is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a visualization guidance method for the booth assembly process based on 3D modeling. This invention utilizes multi-source spatiotemporal sensing data fusion, dynamic 3D field reconstruction, and structural state evolution analysis to continuously perceive and model the entire booth assembly process. The method introduces an improved D-NeRF model to achieve dynamic 3D reconstruction of the actual assembly process, obtaining a continuous geometric representation that changes over time, a component-level dynamic pose field, and occlusion-completed structural surfaces. Combined with structural state vector construction and DM-based structural state evolution analysis, the method characterizes and makes short-term predictions of the structural state changes during assembly. Based on this, the structural feasibility of the next assembly operation is determined, and corresponding visualization guidance information is generated. This method achieves dynamic alignment between 3D visualization guidance and the actual assembled structural state, enabling early identification of unreasonable assembly sequences and guidance of continuous geometric changes. It has the advantages of reliable guidance information, accurate assembly sequence judgment, and high overall assembly efficiency and safety.

[0006] A visualization guidance method for the booth assembly process based on 3D modeling according to an embodiment of the present invention includes: Collect multi-source spatiotemporal sensing data from the booth assembly site, preprocess the multi-source spatiotemporal sensing data, and form standardized multi-source spatiotemporal sensing data. Standardized multi-source spatiotemporal sensing data is organized sequentially according to time information data to form a continuous observation sequence. For each frame of observation data in the continuous observation sequence, the display component instance identification and tracking processing is performed to obtain the candidate pose information, visible boundary information and confidence information of each display component, forming a component instance set. By inputting the set of component instances and the continuous observation sequence into the improved D-NeRF model, the dynamic three-dimensional field during the booth assembly process is reconstructed, resulting in a continuous geometric representation that changes over time, a component-level dynamic pose field, and a structural surface with occlusion completion. Based on continuous geometric representation, component-level dynamic pose field and occlusion completion of structural surface, structural state vectors are constructed and a sequence of structural state vectors is formed. A structural state evolution analysis module based on DM decomposition is constructed. DM decomposition is performed on the structural state vector sequence to extract the dominant modes of structural state evolution and the corresponding mode evolution coefficients during the booth assembly process, and to generate short-time predicted state vectors of the structural state. Based on short-time predicted state vectors and combined with assembly constraint library, the structural feasibility of the next assembly operation is determined, and visual guidance information corresponding to feasible assembly operations is generated.

[0007] Optionally, the multi-source spatiotemporal sensing data specifically includes multi-view image data, three-dimensional point cloud data, and pose sensing data, and contains corresponding time information data.

[0008] Optionally, the preprocessing of multi-source spatiotemporal sensing data specifically includes time synchronization, spatial coordinate alignment, data format normalization, noise suppression, and anomaly removal.

[0009] Optionally, the set of forming component instances includes: Standardized multi-source spatiotemporal sensing data is framed and organized in chronological order, mapping multi-view image data, 3D point cloud data, and pose sensing data during the booth assembly process to the same time frame, forming a continuous observation data sequence arranged in chronological order. For each frame of observation data in the continuous observation data sequence, the multi-view image data is processed to identify the instance regions of different booth components. The instance regions corresponding to different booth components are identified and separated in the images from each view, and the two-dimensional visible region information of each booth component in the current observation frame is obtained. Based on the two-dimensional visible area information, instance-level association processing is performed on the three-dimensional point cloud data to divide the point cloud data corresponding to each booth component instance, thereby obtaining instance-level point cloud data corresponding to different booth component instances. Based on the instance-level point cloud data corresponding to each booth component instance, and combined with pose-aware data, candidate pose estimation processing is performed on each booth component instance to obtain the candidate spatial position and candidate spatial pose information of each booth component instance in the current observation frame. Instance-level association and tracking processing is performed on the booth component instances in adjacent observation frames. A consistent instance identifier is assigned to the same booth component in consecutive observation frames, and corresponding visible boundary information and confidence information are generated for each booth component instance, which are combined to form a component instance set.

[0010] Optionally, obtaining the time-varying continuous geometric representation, component-level dynamic pose field, and occlusion-completed structural surface includes: An improved D-NeRF model is constructed, which consists of a structure perception module, a phase time module, a dynamic harmonization module, an occlusion completion module, and a state output module. The structure perception module performs structure perception processing on the set of component instances, extracts candidate pose information, visible boundary information and confidence information of each booth component, and constructs a component state transition diagram by combining the temporal change relationship of each booth component in the continuous observation sequence, generating component structural condition features corresponding to each time frame in the continuous observation sequence. The stage time module generates time condition information based on time information data in the continuous observation sequence, and introduces a cyclic time gating structure to perform time-series modeling of structural changes between adjacent time frames. It associates time condition information with component structural condition features to obtain stage time conditions corresponding to each time frame. The dynamic harmonization module performs dynamic harmonization processing on the continuous observation sequence based on the stage time conditions and component state transition diagram. It introduces the insertion trajectory vector field to constrain and control the insertion direction and advancement distance of the booth components, generate dynamic three-dimensional deformation expression corresponding to each time frame, and form a dynamic three-dimensional field expression corresponding to the booth assembly process. The occlusion completion module performs occlusion completion processing, sets up an independent occlusion region detection channel, identifies the occluded spatial regions in the continuous observation sequence, and performs occlusion completion processing under the constraints of dynamic three-dimensional field expression, generating the completion structure surface corresponding to each time frame, and generating the volume density information and color information corresponding to the completion structure surface; The state output module performs volume rendering output based on volume density and color information, and combines dynamic 3D deformation expression, interpolation trajectory vector field and completed structural surface to output a continuous geometric representation that changes over time, a part-level dynamic pose field and occlusion-completed structural surface.

[0011] Optionally, forming the structural state vector sequence includes: For each time frame in the continuous observation sequence, the continuous geometric representation, component-level dynamic pose field and occlusion-completed structural surface of each time frame are organized according to the time information data to obtain the geometric and structural input set at the time frame level. Based on the component-level dynamic pose field in the geometric and structural input set at the time frame level, the pose state component set of each booth component instance in each time frame is extracted, and the pose state component sets of each booth component instance are combined to form a time frame-level pose state sub-vector. Based on the set of structural surfaces and component instances with occlusion completion in the geometric and structural input set at the time frame level, the connection state construction process is performed on the connection relationship between the booth components in each time frame to obtain the connection state component set, and then combined to form the time frame level connection state sub-vector. Based on the continuous geometric representation and interpolation-related structural surface relationships in the time-frame-level geometric and structural input set, interference and steady-state construction processing is performed on each time frame to obtain the interference state component set and the steady-state component set, which are then combined to form the time-frame-level structural safety sub-vector. For each time frame, the time frame-level pose state sub-vector, the time frame-level connection state sub-vector, and the time frame-level structural safety sub-vector are concatenated to form the structural state vector of the corresponding time frame. The structural state vectors of each time frame are then arranged in order according to the time information data to form a structural state vector sequence.

[0012] Optionally, the short-time predicted state vector for generating the structural state includes: A structural state evolution analysis module based on DM decomposition is constructed. The structural state evolution analysis module consists of a dynamic window sliding decomposition unit, a structural semantic label perception unit, a modal confidence adjustment unit, an abnormal modality compression unit, and a modal evolution trajectory prediction unit. The dynamic window sliding decomposition unit receives the structure state vector sequence, performs sliding segmentation on the structure state vector sequence according to the preset window length and sliding step size, and performs DM decomposition within each sliding window to generate a window-level modality set corresponding to each window. The structural semantic label perception unit receives the structural state vector sequence, constructs the structural semantic label perception path, generates structural semantic labels based on the pose state sub-vectors, connection state sub-vectors and structural safety sub-vectors in the structural state vector sequence, and associates the structural semantic labels with the window-level modality set to form a semantic modality; The modality confidence adjustment unit receives semantic modalities, generates modality confidence based on the consistency of semantic modalities in the time dimension, cross-window stability and reconstruction error index, and adjusts the modality evolution coefficients according to the modality confidence to obtain the adjusted modality evolution coefficients. The abnormal mode compression unit receives the modal evolution coefficients after regulation, identifies abnormal modes based on the abrupt change amplitude and outlier degree in the modal evolution coefficients, introduces the abnormal mode compression channel to perform compression processing on the modal coefficients corresponding to the abnormal modes, and obtains the dominant mode set; The modal evolution trajectory prediction unit receives the set of dominant modes, generates the structural state prediction results within the future prediction step based on the evolution trend of the dominant modes, and obtains the short-term predicted state vector of the structural state.

[0013] Optionally, the generation of visual guidance information corresponding to the assembly operation includes: Based on the structural state vector at the current moment, the short-term predicted state vector of the structural state, and the assembly constraint library, a candidate set of the next assembly operation to be evaluated is generated. For each candidate operation in the candidate set of the next assembly operation, the prediction time frame interval corresponding to the candidate operation is called in the short-time prediction state vector of the structural state, and the pose state sub-vector, connection state sub-vector and structural safety sub-vector are extracted to form the prediction structural state set corresponding to the candidate operation. For each candidate operation, a structural feasibility determination is performed on the predicted structural state set based on the assembly constraint library, and the feasibility determination result of the candidate operation is output. For candidate operations that are deemed feasible, visual guidance information corresponding to the candidate operations is generated based on continuous geometric representation, component-level dynamic pose field and occlusion completion of the structural surface. Visual guidance information is output to the assembly guidance terminal, and a feasibility determination record corresponding to the candidate operation is generated.

[0014] The beneficial effects of this invention are: This invention proposes a visualization guidance method for the booth assembly process based on 3D modeling. It comprehensively utilizes multi-source spatiotemporal sensing, 3D reconstruction, and structural state evolution analysis technologies to dynamically model and guide the entire booth assembly process in real time. Based on multi-view image data, 3D point cloud data, and pose perception data, this method acquires candidate poses, visible boundaries, and confidence information of booth components during assembly through instance identification and tracking. A continuous observation sequence is constructed, and an improved D-NeRF model is introduced to reconstruct the dynamic 3D field during assembly. This generates a continuous geometric representation that changes over time, a component-level dynamic pose field, and the structural surface after occlusion completion. This ensures that the 3D modeling results accurately reflect the on-site assembly state, rather than relying on preset or idealized animation models.

[0015] This invention constructs a structural state vector to uniformly express geometric, pose, connection relationships, and structural safety information. Based on DM decomposition, it establishes a structural state evolution analysis module to decompose and predict the structural state vector sequence, enabling analysis and short-term prediction of structural evolution trends during assembly. For operations deemed feasible, corresponding visual guidance information is generated. This invention effectively avoids problems such as unreasonable assembly sequence, structural interference, or insufficient stability, improving the accuracy and safety of assembly guidance. Furthermore, the synchronous output of continuous geometric and structural states makes the guidance process more intuitive, real-time, and reliable, suitable for efficient assembly and quality control of complex exhibition booth structures. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0017] Figure 1 This is a flowchart of a visualization guidance method for the booth assembly process based on 3D modeling, as proposed in this invention. Figure 2 This is a schematic diagram of the improved D-NeRF model for a visualization guidance method for the booth assembly process based on 3D modeling proposed in this invention. Figure 3 This is a structural diagram of the structural state evolution analysis module of a visualization guidance method for booth assembly process based on 3D modeling proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A visualization guidance method for the booth assembly process based on 3D modeling, comprising: Collect multi-source spatiotemporal sensing data from the booth assembly site, preprocess the multi-source spatiotemporal sensing data, and form standardized multi-source spatiotemporal sensing data. Standardized multi-source spatiotemporal sensing data is organized sequentially according to time information data to form a continuous observation sequence. For each frame of observation data in the continuous observation sequence, the display component instance identification and tracking processing is performed to obtain the candidate pose information, visible boundary information and confidence information of each display component, forming a component instance set. By inputting the set of component instances and the continuous observation sequence into the improved D-NeRF model, the dynamic three-dimensional field during the booth assembly process is reconstructed, resulting in a continuous geometric representation that changes over time, a component-level dynamic pose field, and a structural surface with occlusion completion. Based on continuous geometric representation, component-level dynamic pose field and occlusion completion of structural surface, structural state vectors are constructed and a sequence of structural state vectors is formed. A structural state evolution analysis module based on DM decomposition is constructed. DM decomposition is performed on the structural state vector sequence to extract the dominant modes of structural state evolution and the corresponding mode evolution coefficients during the booth assembly process, and to generate short-time predicted state vectors of the structural state. Based on short-time predicted state vectors and combined with assembly constraint library, the structural feasibility of the next assembly operation is determined, and visual guidance information corresponding to feasible assembly operations is generated.

[0020] In this embodiment, the multi-source spatiotemporal sensing data specifically includes multi-view image data, three-dimensional point cloud data, and pose sensing data, and contains corresponding time information data.

[0021] In this embodiment, the preprocessing of multi-source spatiotemporal sensing data specifically includes time synchronization, spatial coordinate alignment, data format normalization, noise suppression, and anomaly removal.

[0022] In this embodiment, the set of forming component instances includes: Standardized multi-source spatiotemporal sensing data is framed and organized in chronological order, mapping multi-view image data, 3D point cloud data, and pose sensing data during the booth assembly process to the same time frame, forming a continuous observation data sequence arranged in chronological order. For each frame of observation data in the continuous observation data sequence, the multi-view image data is processed to identify the instance regions of different booth components. The instance regions corresponding to different booth components are identified and separated in the images from each view, and the two-dimensional visible region information of each booth component in the current observation frame is obtained. Based on the two-dimensional visible area information, instance-level association processing is performed on the three-dimensional point cloud data. The point clouds corresponding to each booth component instance are divided into sub-instances, resulting in instance-level point cloud data corresponding to different booth component instances. Specifically, the division of the point clouds corresponding to each booth component instance involves: Based on the imaging parameter relationship between multi-view image data and 3D point cloud data, the information of the 2D visible area is mapped to the 3D point cloud coordinate space, and the correspondence between 2D image pixels and 3D point cloud points is established. According to the boundary range of each booth component instance in the 2D visible area, the mapped 3D point cloud points are filtered, and the 3D point cloud points whose projection positions fall into the corresponding 2D visible area are retained. Consistency constraints and fusion processing are performed on 3D point cloud points from different viewpoints, and point cloud points with inconsistent spatial positions and insufficient stability are eliminated. 3D point cloud points with continuous spatial positions and consistent distribution density are merged into the point cloud set corresponding to the same booth component instance. Based on the instance-level point cloud data corresponding to each booth component instance, and combined with pose-aware data, candidate pose estimation processing is performed on each booth component instance to obtain the candidate spatial position and candidate spatial pose information of each booth component instance in the current observation frame. Specifically, the candidate pose estimation processing for each booth component instance involves: Based on instance-level point cloud data, the geometric distribution features of each booth component instance in three-dimensional space are extracted. By analyzing the overall spatial distribution range, center position and main direction of the point cloud, the initial spatial position estimate of the booth component instance is generated. Combined with the pose information of the observation device recorded in the pose perception data, the spatial coordinates of the instance-level point cloud data are uniformly corrected, and the spatial orientation of the booth component instance is matched and adjusted to generate multiple candidate spatial poses that meet the current observation conditions. The spatial position estimation results and the spatial pose estimation results are combined to form the candidate spatial position and candidate spatial pose information of each booth component instance in the current observation frame. Instance-level association and tracking processing is performed on booth component instances in adjacent observation frames. A consistent instance identifier is assigned to the same booth component in consecutive observation frames, and corresponding visible boundary information and confidence information are generated for each booth component instance. These are combined to form a component instance set. Specifically, generating corresponding visible boundary information and confidence information for each booth component instance involves: In continuous observation frames, based on the projection results of each booth component instance in the multi-view image data, the visible area range of the booth component instance in the current observation frame is determined, and the visible area range is used as the visible boundary information of the booth component instance. By combining the positional continuity, attitude change stability, and spatial consistency of instance-level point cloud data of each booth component instance in adjacent observation frames, the stability of instance association results is evaluated, and confidence information reflecting the degree to which booth component instances are correctly identified and stably tracked in the current observation frame is generated.

[0023] In this embodiment, obtaining the time-varying continuous geometric representation, component-level dynamic pose field, and occlusion-completed structural surface includes: An improved D-NeRF model is constructed, which consists of a structure-aware module, a phase-time module, a dynamic harmonic reconciliation module, an occlusion completion module, and a state output module. Specifically, the construction of the improved D-NeRF model involves: The D-NeRF model incorporates candidate pose information, visible boundary information, and confidence information from the component instance set into its canonical space module to form a structure-aware module. The assembly stage semantics are integrated into the temporal representation of the D-NeRF model's temporal embedding module to form a stage-time module. While maintaining the ability to model time-continuous deformation, the deformation field module of the D-NeRF model introduces stage-time conditions and component state transition relationships to form a dynamic harmonization module. The NeRF network of the D-NeRF model is no longer used solely as a geometry and color fitting network but is upgraded and reconstructed into an occlusion completion module to structurally complete invisible areas caused by occlusion during assembly. Based on the volume rendering module of the D-NeRF model, a state output module is formed by structuring the dynamic 3D deformation representation, interpolation trajectory vector field, and completed structural surfaces, resulting in an improved D-NeRF model. The structure perception module performs structure perception processing on the set of component instances, extracting candidate pose information, visible boundary information, and confidence information for each booth component. It then constructs a component state transition diagram by combining the temporal changes of each booth component in the continuous observation sequence, generating component structural condition features corresponding to each time frame in the continuous observation sequence. Among these features: The construction of the component state transition diagram is specifically as follows: In the continuous observation sequence, the booth component instances are used as nodes in the graph, and the state changes of the same instance in adjacent observation frames are connected in chronological order to form a state transition relationship that reflects the evolution path of the component instances in the time dimension. Combining the spatial position change amplitude, attitude change continuity and visible boundary change of the component instances in adjacent time frames, the association relationship of instances that do not meet the continuity is removed to obtain the component state transition graph. The generated component structure condition features corresponding to each time frame in the continuous observation sequence are specifically as follows: For each time frame in the continuous observation sequence, the candidate pose information, visible boundary information and confidence information of each booth component instance in the time frame are aggregated, and combined with the state nodes corresponding to the time frame and the adjacent transition relationships in the component state transition diagram, the component structural condition features are formed. The stage time module generates time condition information based on time information data in a continuous observation sequence and introduces a cyclic time-series gating structure to perform time-series modeling of structural changes between adjacent time frames. It then associates the time condition information with component structural condition features to obtain the stage time conditions corresponding to each time frame, where: Cyclic temporal gating structures are temporal representation structures that express the state evolution relationship in a continuous time series. Their core feature is that by introducing state memory and updates in the time dimension, the structural state of the current time frame is not only related to the time information corresponding to the time frame, but also related to the structural change history in the previous time frame. Cyclic temporal gating structures express the continuous changes, stage transitions and stable intervals of structural states by establishing a recursive state transfer relationship between adjacent time frames. The specific steps for obtaining the stage time conditions corresponding to each time frame are as follows: Using time information data as the basic time index, the time frames in the continuous observation sequence are organized sequentially, and the component structure condition features corresponding to each time frame are input into the cyclic time-series gating structure. The cyclic time-series gating structure is used to perform temporal correlation on the structural change relationship between adjacent time frames. The fused time expression is associated with the component structure condition features of the corresponding time frame to form stage time conditions, and a one-to-one corresponding stage time condition is generated for each time frame in the continuous observation sequence. The dynamic harmonization module, based on stage time conditions and component state transition diagrams, performs dynamic harmonization processing on continuous observation sequences. It introduces an insertion trajectory vector field to constrain and control the insertion direction and advancement distance of the booth components, generating dynamic three-dimensional deformation representations corresponding to each time frame, and forming a dynamic three-dimensional field representation corresponding to the booth assembly process. Specifically: The insertion trajectory vector field refers to the geometric expression that describes the spatial motion trend of booth components during the assembly process. By establishing a continuous directional and displacement guidance relationship for the component insertion process in three-dimensional space, it depicts the overall motion direction followed by the component when transitioning from the initial spatial position to the target assembly position. The constraint and control of the insertion direction and advancement distance of the booth components are specifically as follows: Using the insertion trajectory vector field as a spatial motion reference condition, the booth components are guided to evolve along the spatial direction consistent with the structural connection relationship during dynamic three-dimensional deformation. The displacement change range of the components between adjacent time frames is limited. The structural evolution path corresponding to the component in the component state transition diagram is associated with the directional distribution in the insertion trajectory vector field, and the components gradually approach the target assembly position during the advancement process. The dynamic three-dimensional field representation corresponding to the booth assembly process is specifically as follows: Based on the stage time conditions, component state transition diagrams and insertion trajectory vector fields corresponding to each time frame, the spatial morphological changes of the booth components in the continuous observation sequence are uniformly expressed. The dynamic three-dimensional deformation generated in each time frame is continuously organized in the time dimension, and the continuity and directional consistency of spatial deformation are maintained between adjacent time frames to form a dynamic three-dimensional field expression. The occlusion completion module performs occlusion completion processing, sets up an independent occlusion region detection channel, identifies occluded spatial regions in a continuous observation sequence, and performs occlusion completion processing under the constraints of dynamic 3D field representation, generating a completed structure surface corresponding to each time frame, and generating volume density information and color information corresponding to the completed structure surface, wherein: The establishment of an independent occlusion area detection channel specifically refers to: The process of identifying occluded areas is distinguished from the process of dynamic 3D field modeling. A separate spatial detection path is established to describe the observation-missing areas. The occluded area detection channel takes multi-view image data, instance-level point cloud data and visible boundary information of components in the continuous observation sequence as input. It comprehensively analyzes the differences in spatial visibility under different views and identifies the spatial areas that cannot be directly observed in the current observation frame due to view limitations, component occlusion and structural overlap. The occlusion completion process performed under the constraints of the dynamic three-dimensional field representation specifically includes: Based on the occluded spatial region determined by the occluded region detection channel, the three-dimensional morphological changes of adjacent time frames in the dynamic three-dimensional field are used as a reference. Under the premise of maintaining the overall spatial continuity and structural consistency, the structural surface corresponding to the occluded region is completed. The volume density and color information corresponding to the generated and completed structure surface are specifically as follows: After the completed structure surface is generated, it is mapped to the volume representation space of the dynamic three-dimensional field. The volume density and color information distribution of the visible structure surfaces adjacent to the completed structure surface in the dynamic three-dimensional field are referenced to consistently describe the volume properties of the spatial location of the completed structure surface. By maintaining the consistency of the volume density change trend between the completed structure surface and the adjacent visible structure surfaces at continuous spatial locations, the color information corresponding to the completed structure surface is spatially extended and matched according to the color distribution characteristics of the adjacent visible structure surfaces to generate volume density and color information consistent with the overall expression of the dynamic three-dimensional field. The state output module performs volume rendering output based on volume density and color information, and combines dynamic 3D deformation representation, interpolated trajectory vector field, and completed structural surface to output a continuous geometric representation that changes over time, a part-level dynamic pose field, and occlusion-completed structural surface, wherein: The execution of the volume rendering output is specifically as follows: Volume density information and color information are used as volume attributes to describe the three-dimensional field. The continuous three-dimensional structure is expressed as a whole in the spatial coordinate system. By comprehensively processing the distribution relationship of volume density information and color information in space, a three-dimensional rendering result that reflects the real spatial structure morphology during the booth assembly process is formed. The output includes a continuous geometric representation that changes over time, a component-level dynamic pose field, and a structural surface with occlusion completion, specifically: The 3D rendering results generated at each time frame are sequentially associated according to the time information data to form a continuous geometric representation that reflects the continuous evolution of the booth assembly process. At the same time, the spatial evolution relationship of component instances in the dynamic 3D deformation expression and the insertion trajectory vector field is combined to extract and output the spatial position and spatial attitude changes of each booth component at different time frames, forming a component-level dynamic pose field. The structural surface after occlusion completion processing is output as an independent structure, and output together with the continuous geometric representation and the dynamic pose field.

[0024] In this embodiment, forming the structural state vector sequence includes: For each time frame in a continuous observation sequence, the continuous geometric representation, component-level dynamic pose field, and occlusion-completed structural surface of each time frame are organized according to the time information data to obtain a time frame-level geometric and structural input set. Specifically, the obtained time frame-level geometric and structural input set is as follows: Based on the time information data recorded in the continuous observation sequence, the three-dimensional results from different sources and in different forms of expression are time-aligned. The continuous geometric representations generated in the same time frame, the corresponding component-level dynamic pose fields, and the structural surfaces after occlusion completion are matched and merged. The geometric morphology information, component pose information, and structural surface information belonging to the same time frame are uniformly organized to form a geometric and structural input set of the overall structural state of the booth in the time frame. Based on the component-level dynamic pose field in the geometric and structural input set at the time frame level, the pose state component set of each booth component instance in each time frame is extracted, and the pose state component sets of each booth component instance are combined to form a time frame-level pose state sub-vector. Based on the set of structural surfaces and component instances with occlusion completion in the time-frame-level geometric and structural input set, connection state construction processing is performed on the connection relationships between booth components in each time frame to obtain a set of connection state components, which are then combined to form a time-frame-level connection state sub-vector. Specifically, the connection state construction processing on the connection relationships between booth components in each time frame involves: Based on the spatial distribution and mutual contact relationship of each booth component instance in the structural surface of the occlusion completion, it is determined whether there are structural contact, insertion and connection relationships between components. Combined with the candidate pose information and visible boundary information recorded in the component instance set, the relative position relationship and alignment status between components are analyzed. Component instance pairs that meet the requirements of spatial proximity and structural alignment are identified as being in the connected and waiting-to-be-connected state. The corresponding connection relationship is recorded in a structured way to form a connection state component set. The connection state component set is uniformly organized and encoded to generate a time frame-level connection state sub-vector. Based on the continuous geometric representations and interpolated structural surface relationships in the time-frame-level geometric and structural input sets, interference and steady-state construction processes are performed on each time frame to obtain the interference state component set and the steady-state component set, which are then combined to form the time-frame-level structural safety sub-vector, where: The structural surface relationships related to insertion refer to the spatial correspondence and geometric relationships formed between structural surfaces that are directly related to the insertion, fitting and assembly actions between booth components during the booth assembly process. The process of performing interference and steady-state construction on each time frame specifically includes: Based on the spatial distribution of each booth component instance in the continuous geometric representation, the system analyzes whether there is spatial overlap, mutual intrusion, or obstructed insertion paths between components in the current time frame. The state information reflecting the degree of spatial conflict between components is collected to form an interference state component set. At the same time, combined with the structural surface relationship related to insertion, the support relationship, contact stability, and overall structural balance of the booth components in the current time frame are evaluated. The state information reflecting whether the component structure is in a stable assembly state is collected to form a stable state component set. The interference state component set and the stable state component set are uniformly organized and encoded, and combined to form a time frame-level structural safety sub-vector.

[0025] For each time frame, the time frame-level pose state sub-vector, the time frame-level connection state sub-vector, and the time frame-level structural safety sub-vector are concatenated to form the structural state vector of the corresponding time frame. The structural state vectors of each time frame are then arranged in order according to the time information data to form a structural state vector sequence.

[0026] In this embodiment, the generation of the short-time predicted state vector of the structural state includes: A structural state evolution analysis module based on DM decomposition is constructed. This module comprises a dynamic window sliding decomposition unit, a structural semantic label perception unit, a modal confidence adjustment unit, an abnormal modality compression unit, and a modal evolution trajectory prediction unit. Specifically, the construction of this DM decomposition-based structural state evolution analysis module involves: The dynamic window sliding decomposition unit, the structural semantic label perception unit, the modal confidence adjustment unit, the abnormal modal compression unit, and the modal evolution trajectory prediction unit are connected in series to obtain the structural state evolution analysis module based on DM decomposition. The dynamic window sliding decomposition unit receives the structure state vector sequence, performs sliding segmentation on the structure state vector sequence according to the preset window length and sliding step size, and performs DM decomposition within each sliding window to generate a window-level modality set corresponding to each window, where: The preset window length is set to the minimum time length that can cover a complete local assembly action interval; The sliding step size is set to the minimum time step that can cover local assembly changes based on the continuity of the booth assembly action. The process of performing sliding segmentation on the structural state vector sequence according to the preset window length and sliding step size is as follows: Based on the temporal order of the structural state vector sequence, starting from the beginning of the structural state vector sequence, the window is shifted backward on the time axis according to a preset sliding step size, and a corresponding number of continuous structural state vectors are extracted to form an analysis window. The window extraction and shifting process is repeated, and the entire structural state vector sequence is organized into multiple sliding windows that partially overlap in time and are arranged sequentially. The specific steps for performing DM decomposition within each sliding window are as follows: For each subsequence of structural state vectors contained in a sliding window, the subsequence of structural state vectors is treated as an independent analysis object. The state change characteristics of the subsequence of structural state vectors within the time range of the sliding window are decomposed and processed. Modal expressions that characterize the evolution characteristics of structural state within the sliding window are extracted. Each sliding window corresponds to a set of window-level modes that reflect the evolution law of structural state within the sliding window, and the correspondence between windows in time order is maintained. The structural semantic label perception unit receives a sequence of structural state vectors, constructs a structural semantic label perception path, generates structural semantic labels based on the pose state sub-vectors, connection state sub-vectors, and structural safety sub-vectors in the structural state vector sequence, and associates the structural semantic labels with a window-level modality set to form a semantic modality, wherein: The construction of the semantic tag awareness path is specifically as follows: In the structural state vector sequence, according to the internal composition relationship of the structural state vector, a fixed correspondence is established between the pose state sub-vector, the connection state sub-vector, and the structural safety sub-vector. The three types of sub-vectors are organized in parallel within the same time frame, and the order consistency between time frames is maintained in the time dimension, forming a structural semantic mapping path that is connected by the structural state vector sequence. The generation of the structural semantic tags is specifically as follows: Based on the structural semantic tag perception path, the pose state sub-vector, connection state sub-vector and structural safety sub-vector are jointly identified in each time frame. The state information reflecting the spatial position relationship of the booth components, the connection status of the components and the structural safety status in the same time frame is combined and encoded to form structural semantic tags. The structural semantic tags corresponding to each time frame are arranged in order according to the time information data to obtain a structural semantic tag sequence that corresponds one-to-one with the structural state vector sequence. The modality confidence adjustment unit receives semantic modalities, generates modality confidence based on the consistency of semantic modalities in the time dimension, cross-window stability, and reconstruction error indicators, and adjusts the weights of the modality evolution coefficients according to the modality confidence to obtain the adjusted modality evolution coefficients, where: The generation of modality confidence based on the consistency of semantic modality in the time dimension, cross-window stability, and reconstruction error index is as follows: After the semantic modality is formed, the performance of the same semantic modality in adjacent time frames and adjacent sliding windows is aligned and compared along the time sequence corresponding to the structural state vector sequence. The modality confidence is formed by normalizing and integrating the changes of the semantic modality in consecutive time frames, the maintenance in different sliding windows, and the reconstruction deviation of the corresponding modality to the structural state vector sequence. The weighting of modal evolution coefficients based on modal confidence is specifically as follows: Modality confidence is used as the basis for weight adjustment, and a correspondence is established with the modality evolution coefficient of the corresponding semantic modality. Based on the relative magnitude of modality confidence, the modality evolution coefficient is adjusted proportionally. Semantic modalities with modality confidence higher than the confidence threshold maintain an evolution weight higher than the weight threshold, while the modality evolution coefficients corresponding to semantic modalities with modality confidence lower than the confidence threshold are suppressed accordingly, thus obtaining the set of modality evolution coefficients after adjustment. The anomalous mode compression unit receives the modal evolution coefficients after modulation, identifies anomalous modes based on the abrupt change amplitude and outlier degree in the modal evolution coefficients, introduces an anomalous mode compression channel to perform compression processing on the modal coefficients corresponding to the anomalous modes, and obtains the dominant mode set, where: The identification of anomalous modes based on the mutation amplitude and outlier degree in modal evolution coefficients specifically includes: In the modal evolution coefficient sequence after regulation, the changes in the evolution coefficients of each mode are compared and analyzed in chronological order. The change amplitude of the evolution coefficients of the same mode in adjacent time frames and adjacent sliding windows is used as a continuity reference. The evolution coefficients of the mode in the current time range are compared with the distribution of the evolution coefficients of the remaining modes as a whole. When the evolution coefficients of a mode deviate from its own historical change range in terms of time change amplitude, or significantly deviate from the overall distribution of the remaining modes in the same time window, the current mode is marked as an abnormal mode. The abnormal mode compression channel is an independent mode coefficient adjustment path set between the mode confidence control unit and the dominant mode output. It acts separately on the mode evolution coefficients identified as abnormal modes. Structurally, it is different from the main mode evolution path and is set in parallel with the coefficient transfer path of the normal mode. The modal evolution trajectory prediction unit receives the set of dominant modes, generates structural state prediction results within the future prediction step size based on the evolution trend of the dominant modes, and obtains the short-time predicted state vector of the structural state, where: The generation of the structural state prediction results within the future prediction step is specifically as follows: After obtaining the set of dominant modes, the changes in modal evolution coefficients of each dominant mode within the observed time range are organized chronologically. Evolutionary trend information reflecting the direction and magnitude of structural state changes is extracted. While maintaining the temporal continuity of the modes, the evolutionary trend is extended to the future prediction step size to obtain the predicted evolution state of each dominant mode in future time frames. The predicted evolution states of each dominant mode within the future prediction time range are then comprehensively organized to form the structural state prediction result corresponding to the future prediction step size, where: The future prediction step size range refers to the number of short-term prediction time frames that cover the next assembly action; The short-time predicted state vector of the structural state is obtained as follows: Based on the structural state prediction results formed within the future prediction step, the predicted pose-related prediction states, connection relationship prediction states, and structural safety-related prediction states are organized and spliced ​​according to the composition method of the structural state vector to form a prediction state expression consistent with the structural state vector format. The predicted state is then arranged in the order of future time frames to obtain the short-term prediction state vector of the structural state belonging to the current assembly action.

[0027] In this embodiment, the generation of visual guidance information corresponding to the assembly operation includes: Based on the current structural state vector, the short-time predicted structural state vector, and the assembly constraint library, a candidate set of the next assembly operation to be evaluated is generated, where: The assembly constraint library refers to the set of constraint information describing the assembly relationships between various components of the booth. It includes the connection methods of the booth components, assembly dependencies, allowed assembly sequences, and structural constraint information related to component insertion, fixing, and support. It is stored in the form of structured data and maintains a corresponding relationship with the component identifiers in the set of booth component instances. The generation of the candidate set of the next assembly operation to be evaluated specifically includes: Based on the structural state vector at the current moment, and combined with the future structural change trend reflected in the short-term predicted state vector of the structural state, the assembly operation items that satisfy the assembly dependency relationship in the current structural state and have not been executed are extracted from the assembly constraint library. The assembly operation items are then organized into multiple next assembly operation options according to the component instance identifier, forming a candidate set for the next assembly operation. For each candidate operation in the candidate set of the next assembly operation, the prediction time frame interval corresponding to the candidate operation is called in the short-time prediction state vector of the structural state, and the pose state sub-vector, connection state sub-vector and structural safety sub-vector are extracted to form the prediction structural state set corresponding to the candidate operation. For each candidate operation, a structural feasibility determination process is performed on the predicted structural state set based on the assembly constraint library, and the feasibility determination result of the candidate operation is output. Specifically, the structural feasibility determination process based on the assembly constraint library on the predicted structural state set is as follows: For each candidate operation, the information corresponding to the candidate operation in the assembly constraint library is called. The pose state sub-vector, connection state sub-vector, and structural safety sub-vector in the predicted structural state set corresponding to the candidate operation are compared with the assembly constraints item by item. Within the prediction time frame covered by the predicted structural state set, it is determined whether the structural state corresponding to the candidate operation meets the assembly relationship requirements in the assembly constraint library. If the corresponding assembly relationship requirements are met, the current candidate operation is output as feasible; otherwise, an infeasible decision is output. For candidate operations deemed feasible, visual guidance information corresponding to the candidate operations is generated based on continuous geometric representation, component-level dynamic pose field, and occlusion-completed structural surface. The generation of this visual guidance information specifically involves: After a candidate operation is deemed feasible, the continuous geometric representation, component-level dynamic pose field, and occlusion-completed structural surface related to the candidate operation are invoked. The booth component involved in the candidate operation is taken as the target object. The spatial correspondence between the current structural state and the target assembly state of the target component is determined in the continuous geometric representation. Based on the component-level dynamic pose field, the pose change direction and magnitude of the target component during the execution of the candidate operation are extracted. Combined with the occlusion-completed structural surface, the area of ​​the target component that may be occluded during the execution of the candidate operation is fully structurally expressed to form visual guidance information. Visual guidance information is output to the assembly guidance terminal, and a feasibility determination record corresponding to the candidate operation is generated.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a modular commercial booth assembly scenario in a large city convention center. The booth, used for a three-day industry exhibition, occupies approximately 120 square meters. Its structure consists of a base frame, column components, beam components, display panels, and various plug-in connectors, totaling over 160 parts. The assembly process must be completed within 8 hours. Traditional assembly methods rely primarily on two-dimensional construction drawings and manual experience. In previous similar projects, on-site construction personnel commonly encountered difficulties in determining the assembly sequence, improper component insertion, and rework during the process. Particularly during the column and beam insertion stage, incorrect sequence or spatial interference often led to increased disassembly and reassembly times, compromising overall efficiency and structural safety.

[0029] The method of this invention is implemented and applied in this scenario. Multiple fixed cameras and mobile terminals are deployed at the assembly site to collect multi-view image data, 3D point cloud data, and terminal pose data in real time, and the data is processed synchronously. During the assembly process, the booth components are continuously identified and tracked. Based on an improved D-NeRF model, dynamic 3D reconstruction of the assembly process is performed, generating continuous geometric representations and component-level dynamic pose fields. By constructing structural state vectors and introducing structural state evolution analysis based on DM decomposition, the current and short-term future structural states are predicted. Combined with an assembly constraint library, the structural feasibility of the next assembly operation is determined, and only assembly operations deemed feasible and corresponding visual guidance information are pushed to the construction personnel.

[0030] In practical applications, the embodiments compared the effects of traditional assembly methods with the method of the present invention. The results show that, using the method of the present invention, the overall assembly time of the booth was reduced from an average of 7.6 hours to 6.1 hours, improving assembly efficiency by approximately 19.7%. The number of rework steps in key connection steps was reduced from an average of 12 times to 3 times, and structural interference problems were reduced by more than 75%. In the assembly data recorded on-site, the method of the present invention identified and prevented 5 potential unreasonable assembly sequences in advance, avoiding the risk of structural instability. The structural inspection results after assembly showed that the alignment error of each key connection part was controlled within 2 mm, significantly better than the results of previous manual assembly, verifying the significant beneficial effects of the present invention in improving assembly efficiency, structural safety, and guidance accuracy.

[0031] Table 1. Comparison of the overall effects of the booth assembly process using traditional methods and the method of this invention.

[0032] As can be seen from the data in Table 1, there are significant differences between the traditional assembly method and the method of this invention in terms of booth assembly efficiency, structural accuracy, and safety. Regarding assembly efficiency, the traditional method takes 7.6 hours to complete the overall booth assembly, while the method of this invention reduces this to 6.1 hours, a reduction of approximately 1.5 hours. This demonstrates that the visualization guidance based on 3D modeling effectively reduces unnecessary operations and waiting time. The average installation time per component decreased from 2.85 minutes to 2.12 minutes, and the time for critical structural stages was reduced from 3.4 hours to 2.5 hours, indicating that the method of this invention has a particularly significant guiding effect on the assembly rhythm during complex assembly stages.

[0033] Regarding rework, the traditional assembly process involves a total of 12 rework cycles, with 8 of them related to insertion and alignment. However, under the guidance of the method of this invention, the total number of rework cycles is reduced to 3, with only 2 related to insertion, resulting in a rework rate reduction of approximately 75%. This demonstrates that the present invention can significantly reduce rework problems caused by misjudgment of the assembly sequence or spatial interference, thereby improving the stability of the assembly process from the source.

[0034] Regarding structural accuracy, the average alignment error of key connection parts was reduced from 4.8 mm in the traditional method to 1.9 mm, and the maximum alignment error was reduced from 7.2 mm to 2.6 mm, demonstrating the significant advantage of this invention in controlling the alignment accuracy of the joints. This invention's method can significantly improve structural accuracy and safety while ensuring assembly efficiency, achieving efficient, reliable, and predictable guidance for the assembly process of complex exhibition booths.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A visualization guidance method for the booth assembly process based on 3D modeling, characterized in that, include: Collect multi-source spatiotemporal sensing data from the booth assembly site, preprocess the multi-source spatiotemporal sensing data, and form standardized multi-source spatiotemporal sensing data. Standardized multi-source spatiotemporal sensing data is organized sequentially according to time information data to form a continuous observation sequence. For each frame of observation data in the continuous observation sequence, the display component instance identification and tracking processing is performed to obtain the candidate pose information, visible boundary information and confidence information of each display component, forming a component instance set. By inputting the set of component instances and the continuous observation sequence into the improved D-NeRF model, the dynamic three-dimensional field during the booth assembly process is reconstructed, resulting in a continuous geometric representation that changes over time, a component-level dynamic pose field, and a structural surface with occlusion completion. Based on continuous geometric representation, component-level dynamic pose field and occlusion completion of structural surface, structural state vectors are constructed and a sequence of structural state vectors is formed; A structural state evolution analysis module based on DM decomposition is constructed. DM decomposition is performed on the structural state vector sequence to extract the dominant modes of structural state evolution and the corresponding mode evolution coefficients during the booth assembly process, and to generate short-time predicted state vectors of the structural state. Based on short-time predicted state vectors and combined with assembly constraint library, the structural feasibility of the next assembly operation is determined, and visual guidance information corresponding to feasible assembly operations is generated.

2. The method for visual guidance of booth assembly process based on 3D modeling according to claim 1, characterized in that, The multi-source spatiotemporal sensing data specifically includes multi-view image data, 3D point cloud data, and pose sensing data, and contains corresponding time information data.

3. The method for visual guidance of booth assembly process based on 3D modeling according to claim 1, characterized in that, The preprocessing of multi-source spatiotemporal sensing data specifically includes time synchronization, spatial coordinate alignment, data format normalization, noise suppression, and anomaly removal.

4. The method for visual guidance of booth assembly process based on 3D modeling according to claim 1, characterized in that, The set of forming component instances includes: Standardized multi-source spatiotemporal sensing data is framed and organized in chronological order, mapping multi-view image data, 3D point cloud data, and pose sensing data during the booth assembly process to the same time frame, forming a continuous observation data sequence arranged in chronological order. For each frame of observation data in the continuous observation data sequence, the multi-view image data is processed to identify the instance regions of different booth components. The instance regions corresponding to different booth components are identified and separated in the images from each view, and the two-dimensional visible region information of each booth component in the current observation frame is obtained. Based on the two-dimensional visible area information, instance-level association processing is performed on the three-dimensional point cloud data to divide the point cloud data corresponding to each booth component instance, thereby obtaining instance-level point cloud data corresponding to different booth component instances. Based on the instance-level point cloud data corresponding to each booth component instance, and combined with pose-aware data, candidate pose estimation processing is performed on each booth component instance to obtain the candidate spatial position and candidate spatial pose information of each booth component instance in the current observation frame. Instance-level association and tracking processing is performed on the booth component instances in adjacent observation frames. A consistent instance identifier is assigned to the same booth component in consecutive observation frames, and corresponding visible boundary information and confidence information are generated for each booth component instance, which are combined to form a component instance set.

5. The method for visual guidance of booth assembly process based on 3D modeling according to claim 1, characterized in that, The obtained continuous geometric representation, component-level dynamic pose field, and occlusion-completed structural surface that varies with time include: An improved D-NeRF model is constructed, which consists of a structure perception module, a phase time module, a dynamic harmonization module, an occlusion completion module, and a state output module. The structure perception module performs structure perception processing on the set of component instances, extracts candidate pose information, visible boundary information and confidence information of each booth component, and constructs a component state transition diagram by combining the temporal change relationship of each booth component in the continuous observation sequence, generating component structural condition features corresponding to each time frame in the continuous observation sequence. The stage time module generates time condition information based on time information data in the continuous observation sequence, and introduces a cyclic time gating structure to perform time-series modeling of structural changes between adjacent time frames. It associates time condition information with component structural condition features to obtain stage time conditions corresponding to each time frame. The dynamic harmonization module performs dynamic harmonization processing on the continuous observation sequence based on the stage time conditions and component state transition diagram. It introduces the insertion trajectory vector field to constrain and control the insertion direction and advancement distance of the booth components, generate dynamic three-dimensional deformation expression corresponding to each time frame, and form a dynamic three-dimensional field expression corresponding to the booth assembly process. The occlusion completion module performs occlusion completion processing, sets up an independent occlusion region detection channel, identifies the occluded spatial regions in the continuous observation sequence, and performs occlusion completion processing under the constraints of dynamic three-dimensional field expression, generating the completion structure surface corresponding to each time frame, and generating the volume density information and color information corresponding to the completion structure surface; The state output module performs volume rendering output based on volume density and color information, and combines dynamic 3D deformation expression, interpolation trajectory vector field and completed structural surface to output a continuous geometric representation that changes over time, a part-level dynamic pose field and occlusion-completed structural surface.

6. The method for visual guidance of booth assembly process based on 3D modeling according to claim 1, characterized in that, The formation of the structural state vector sequence includes: For each time frame in the continuous observation sequence, the continuous geometric representation, component-level dynamic pose field and occlusion-completed structural surface of each time frame are organized according to the time information data to obtain the geometric and structural input set at the time frame level. Based on the component-level dynamic pose field in the geometric and structural input set at the time frame level, the pose state component set of each booth component instance in each time frame is extracted, and the pose state component sets of each booth component instance are combined to form a time frame-level pose state sub-vector. Based on the set of structural surfaces and component instances with occlusion completion in the geometric and structural input set at the time frame level, the connection state construction process is performed on the connection relationship between the booth components in each time frame to obtain the connection state component set, and then combined to form the time frame level connection state sub-vector. Based on the continuous geometric representation and interpolation-related structural surface relationships in the time-frame-level geometric and structural input set, interference and steady-state construction processing is performed on each time frame to obtain the interference state component set and the steady-state component set, which are then combined to form the time-frame-level structural safety sub-vector. For each time frame, the time frame-level pose state sub-vector, the time frame-level connection state sub-vector, and the time frame-level structural safety sub-vector are concatenated to form the structural state vector of the corresponding time frame. The structural state vectors of each time frame are then arranged in order according to the time information data to form a structural state vector sequence.

7. The method for visual guidance of booth assembly process based on 3D modeling according to claim 1, characterized in that, The short-time predicted state vector for generating the structural state includes: A structural state evolution analysis module based on DM decomposition is constructed. The structural state evolution analysis module consists of a dynamic window sliding decomposition unit, a structural semantic label perception unit, a modal confidence adjustment unit, an abnormal modality compression unit, and a modal evolution trajectory prediction unit. The dynamic window sliding decomposition unit receives the structure state vector sequence, performs sliding segmentation on the structure state vector sequence according to the preset window length and sliding step size, and performs DM decomposition within each sliding window to generate a window-level modality set corresponding to each window. The structural semantic label perception unit receives the structural state vector sequence, constructs the structural semantic label perception path, generates structural semantic labels based on the pose state sub-vectors, connection state sub-vectors and structural safety sub-vectors in the structural state vector sequence, and associates the structural semantic labels with the window-level modality set to form a semantic modality; The modality confidence adjustment unit receives semantic modalities, generates modality confidence based on the consistency of semantic modalities in the time dimension, cross-window stability and reconstruction error index, and adjusts the modality evolution coefficients according to the modality confidence to obtain the adjusted modality evolution coefficients. The abnormal mode compression unit receives the modal evolution coefficients after regulation, identifies abnormal modes based on the abrupt change amplitude and outlier degree in the modal evolution coefficients, introduces the abnormal mode compression channel to perform compression processing on the modal coefficients corresponding to the abnormal modes, and obtains the dominant mode set; The modal evolution trajectory prediction unit receives the set of dominant modes, generates the structural state prediction results within the future prediction step based on the evolution trend of the dominant modes, and obtains the short-term predicted state vector of the structural state.

8. The method for visual guidance of booth assembly process based on 3D modeling according to claim 1, characterized in that, The generated and assembled visual guidance information includes: Based on the structural state vector at the current moment, the short-term predicted state vector of the structural state, and the assembly constraint library, a candidate set of the next assembly operation to be evaluated is generated. For each candidate operation in the candidate set of the next assembly operation, the prediction time frame interval corresponding to the candidate operation is called in the short-time prediction state vector of the structural state, and the pose state sub-vector, connection state sub-vector and structural safety sub-vector are extracted to form the prediction structural state set corresponding to the candidate operation. For each candidate operation, a structural feasibility determination is performed on the predicted structural state set based on the assembly constraint library, and the feasibility determination result of the candidate operation is output. For candidate operations that are deemed feasible, visual guidance information corresponding to the candidate operations is generated based on continuous geometric representation, component-level dynamic pose field and occlusion completion of the structural surface. Visual guidance information is output to the assembly guidance terminal, and a feasibility determination record corresponding to the candidate operation is generated.

Images