Construction scene four-dimensional reconstruction method based on hierarchical representation and structure association graph

Abstract

The application discloses a construction scene four-dimensional reconstruction method based on hierarchical representation and structure association graph, relates to the field of scene reconstruction, solves the problem of insufficient accuracy of existing scene reconstruction, and comprises the following steps: S1, standard observation data is acquired; S2, consistency analysis is performed on image data to generate a final transient mask; an effective supervision area is generated, and a structure association graph is constructed; S3, camera poses under a current session are collected, predicted images are generated, image residuals are analyzed, evolution nodes are extracted, the structure association graph is associated, and evolution layer Gaussians are generated; S4, gradient descent optimization is performed through a standard splatting rendering process, and static layer Gaussians and evolution layer Gaussians are optimized; S5, evolution sequences of nodes are acquired, and four-dimensional structure evolution records are generated; and S6, consistency check is performed on the static layer Gaussians in combination with the four-dimensional structure evolution records; and the application can effectively improve the accuracy of scene reconstruction.

Description

Technical Field

[0001] This invention belongs to the field of scene reconstruction, specifically involving four-dimensional dynamic scene reconstruction technology, and more specifically, a four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagram. Background Technology

[0002] Existing deformable or four-dimensional Gaussian splash reconstruction methods generally assume that the scene has a unified gauge space. They use a continuous deformation field to represent the Gaussian primitives observed at each time step as deformation mappings of the gauge space primitives in the time dimension. This assumption implies that the scene content maintains topological continuity in time and that no components are added or disappeared.

[0003] Specifically, the existing technology has the following main shortcomings: 1. When topological changes such as the construction of new floors or the demolition of components occur in the construction scenario, a single continuous deformation field is difficult to accurately represent such discontinuous structural evolution. Geometric tearing, ghosting, or suspension artifacts are easily generated at the spatial boundary where the topological change occurs, and the reconstruction results deviate significantly from the actual construction process.

[0004] 2. Existing methods typically fail to distinguish between transient disturbances in the scene (such as on-site workers, temporarily parked vehicles, and mobile construction machinery) and the continuous structural evolution of the scene itself. Transient disturbances are included in the optimization along with the static background or evolving structure during multi-view supervision, polluting the Gaussian parameters corresponding to the static background layer and the structural evolution layer. This results in persistent ghosting and floating artifacts in the static areas in the rendering results, reducing the reconstruction accuracy.

[0005] 3. Existing methods generally provide reconstruction results in the form of independent characterization at each time step or in the form of continuous deformation fields. They lack explicit expression of the structural evolution relationship across time segments, making it difficult to directly support the structural level query needs common in construction progress management scenarios such as "when did a certain component appear" or "has a certain component been replaced or demolished". The usability in engineering management scenarios is limited.

[0006] To address this, a four-dimensional reconstruction method for construction scenarios based on hierarchical representation and structural association diagrams is proposed. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide a four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagrams, thereby improving the accuracy of scene reconstruction.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: a four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagrams, the specific working process of each step of which is as follows: Step S1: Acquire image data generated during the on-site acquisition session and generate standard observation data; Step S2: Based on standard observation data, collect image data and sparse point cloud data under the same session and perform consistency analysis; identify non-fixed categories in the image, combine the consistency analysis results to generate the final transient mask, provide feedback to the sparse point cloud, obtain the effective supervision region, and construct the structure association graph; Step S3: Generate an initial static layer Gaussian from standard observation data; acquire the camera pose in the current session, combine it with the initial static layer Gaussian to generate a predicted image, and analyze the image residuals; based on the image residuals, extract evolution nodes and associate them with the structure correlation graph; at the same time, extract the spatial range corresponding to the evolution nodes and generate an evolution layer Gaussian. Step S4: Statistically analyze the image data of the current session and historical sessions, and update the gradients of the static layer Gaussian and the evolved layer Gaussian; perform gradient descent optimization through the standard splash rendering process to optimize the static layer Gaussian and the evolved layer Gaussian. Step S5: Extract the evolution nodes, collect the Gaussian evolution layer and session timestamp corresponding to the evolution nodes to form the evolution sequence of the node, associate the evolution nodes with the corresponding evolution sequences to the structural association graph, and generate a four-dimensional structural evolution record. Step S6: Perform consistency verification by combining the static layer Gaussian with the four-dimensional structure evolution record, and correct the Gaussian based on the verification result to generate the four-dimensional reconstruction result.

[0009] Furthermore, the specific steps of step S1 are as follows: Step S11: Statistically analyze the image data generated in each on-site acquisition session, and obtain the camera pose and sparse point cloud for each session by performing independent motion reconstruction structure calculation; Step S12: Obtain the global static layer Gaussian rendering result of the construction site, extract the common feature parts of the image data generated by each site acquisition session and the global static layer Gaussian rendering result, and obtain the reference anchor point; Based on the reference anchor point, the camera pose of each session is mapped to the sparse point cloud in a unified coordinate system to obtain the camera pose and initial point cloud of each session unified to the global coordinate system, which constitutes the standard observation data.

[0010] Furthermore, the specific steps of step S2 are as follows: Step S21: Based on image data, collect multi-view images of the construction site; acquire sparse point cloud under the same session, reproject the 3D points on the sparse point cloud onto the multi-view images, extract the color information of the corresponding pixels of the 3D points at different views, and perform color consistency analysis; backproject the pixels on the multi-view images, and perform structural consistency analysis in combination with the sparse point cloud; based on color consistency analysis and structural consistency analysis, extract transient influence candidate points. Step S22: Input the multi-view images into the general object detection model for image recognition, extract transient objects at the construction site, perform intersection and verification between the transient objects and transient influence candidate points to obtain the final transient mask; based on the final transient mask, extract the effective supervision region in the sparse point cloud, identify the evolution structure of the effective supervision region, and construct a structure association graph.

[0011] Furthermore, the specific steps of step S21 are as follows: Step S211: Traverse the 3D point reprojection images and record the number of images as as; collect the color value of the corresponding pixel in each image reprojected by the 3D point and record it as scz. a ; For color value scz a Statistical analysis and mean calculation yield the color mean sjz; from the color value scz a The average color difference psc is obtained by calculating the average color difference of the three-dimensional point reprojection with the color mean sjz. The average color of the three-dimensional points corresponding to the reference anchor point is collected and a threshold is set to obtain the color threshold syz; If psc≤syz, the color of the image of the 3D point is considered to be consistent across all viewpoints. If psc > syz, it is determined that the colors of the three-dimensional point are inconsistent across different viewpoints, and the three-dimensional point is marked accordingly; Step S212: Obtain the 3D points of the sparse point cloud to obtain the original 3D points ysd. Map the original 3D points to the multi-view images, extract the corresponding pixels, and back-project the corresponding pixels onto the sparse point cloud to obtain the inverted 3D points fts. Collect the 3D spatial distance between the inverted 3D points fts and the original 3D points ysd. Statistically calculate the 3D spatial distance of the original 3D points on different images and extract the mean value to obtain the mean projection error twc. The average projection error of the three-dimensional points corresponding to the reference anchor point is collected and a threshold is set to obtain the structural threshold jyz; If twc≤jyz, the image structure of the three-dimensional point is determined to be consistent from all perspectives. If twc > jyz, it is determined that the image structure of the three-dimensional point is inconsistent from different viewpoints, and the three-dimensional point is marked. Based on the marked 3D points, 3D points that simultaneously exhibit color inconsistency and structural inconsistency are statistically analyzed to obtain candidate points for transient influence.

[0012] Furthermore, the specific steps of step S22 are as follows: Step S221: Perform image recognition on the multi-view images, extract transient objects at the construction site, record the corresponding pixels of the transient objects at the construction site, and obtain a transient candidate mask; map the transient influence candidate points to the multi-view images, perform intersection processing with the transient candidate masks, and obtain the final transient mask; Step S222: Based on the final transient mask, transient objects in the multi-view image are excluded, the effective supervision area of ​​the image is recorded, and the effective supervision area is fed back to the sparse point cloud; the evolution structure within the effective supervision area is collected; the evolution structure is statistically analyzed as nodes, and a structure association graph is constructed.

[0013] Furthermore, the specific steps of step S3 are as follows: Step S31: Extract different camera viewpoints based on the camera pose in the current session, use the initialized static layer Gaussian as the rendering data source, perform Gaussian splash rasterization rendering to obtain the predicted images under different viewpoints; according to the image data in the current session, collect the original images of the corresponding viewpoints; obtain the final transient mask of the corresponding image, fit the non-mask area of ​​the predicted image and the original image to obtain the image residual. Step S32: Collect the pixel luminance change of the image under illumination as the residual threshold; extract the corresponding image where the residual is greater than the residual threshold, record the change area of ​​the image, and determine the overlap between the change area and the corresponding area of ​​the node in the structure association graph. If there is an overlap between the node corresponding area and the change area, the change area is regarded as the continuation of the existing component and associated with the corresponding node; if there is no overlap between the node corresponding area and the change area, the change area is regarded as a newly added independent component, an evolution node is constructed, and stored in the structure association graph; at the same time, for each change area, collect the sparse point cloud of the current session and perform Gaussian initialization to obtain the Gaussian evolution layer corresponding to the change area.

[0014] Furthermore, the specific steps of step S4 are as follows: Step S41: Based on the image data of the current session and historical sessions, extract the effective supervised images, construct a time-series supervised image list, collect the static region gradient of each session, and obtain the gradient update accumulation of the static layer Gaussian; extract the evolution layer Gaussian of the current session, fit the evolution layer Gaussian to the construction scene, and obtain the gradient update of the evolution layer Gaussian. Step S42: Introduce an inter-layer competition suppression term, combining the gradient update accumulation of the static layer Gaussian and the gradient update of the evolved layer Gaussian, and apply a penalty to the static layer and the evolved layer Gaussian; perform gradient descent optimization through the standard splash rendering process, use a density-adaptive cloning and pruning strategy for the evolved layer Gaussian, and adopt a threshold-controlled density adjustment strategy for the static layer Gaussian to maintain the existing historical structure.

[0015] Furthermore, the specific steps of step S42 are as follows: Step S421: Collect the opacity of the static layer and the evolved layer Gaussian in the same spatial neighborhood and set an opacity threshold; if the opacity of both the static layer Gaussian and the evolved layer Gaussian is higher than the opacity threshold, introduce an inter-layer competition suppression term to the static layer Gaussian and the evolved layer Gaussian, apply a penalty to the static layer and the evolved layer Gaussian in the neighborhood, and separate the opacity of the static layer Gaussian and the evolved layer Gaussian. Step S422: Render the static Gaussian layer and the evolved Gaussian layer respectively. Combine the gradient update accumulation of the static Gaussian layer and the gradient update of the evolved Gaussian layer to obtain the photometric loss. Constrain the static Gaussian layer and the evolved Gaussian layer through independent regularization to obtain the regularization loss. Obtain the competition suppression loss from the inter-layer competition suppression term of the neighborhood Gaussian layer. Perform gradient descent optimization based on the various losses of the Gaussian layer.

[0016] Furthermore, the specific steps of step S5 are as follows: Step S51: Based on the structural association graph, traverse the evolution nodes, extract the nodes with continuous evolution, collect the corresponding regions of the nodes with continuous evolution, record the evolution layer Gaussian of the region and the current session timestamp, and construct the cross-dialogue evolution sequence of the nodes. Step S52: Extract the evolution nodes corresponding to the newly added independent components, collect and store the session timestamp and initial spatial range of the node, extract the spatial adjacency relationship between the node and the other evolution nodes on the structural association diagram, and establish edge relationships to connect with the other nodes. Step S53: Mark the nodes whose evolution layer Gaussian changes do not occur in two consecutive sessions, record the timestamp of the session corresponding to the disappearance of the change, and reduce the Gaussian opacity of the evolution layer corresponding to the node.

[0017] Furthermore, the specific steps of step S6 are as follows: Step S61: Using the static layer Gaussian as the common observation target across sessions, map the four-dimensional structure evolution record to the static layer Gaussian, record the evolution status of the cross-dialogue evolution of the evolution node on the static layer Gaussian, adjust the evolution node, perform global optimization on the joint reprojection error of the session pose of all sessions and the static layer Gaussian, and correct the inconsistency of the position of the reference anchor point between different sessions. Step S62: Based on the optimization results, propagate the corrected pose change to the Gaussian evolution layer corresponding to each session, so that the spatial position of the Gaussian evolution layer is consistent with the coordinate reference of the corrected static layer; record the overall pose correction magnitude, and according to the correction magnitude, add supplementary observations in the key reference anchor area for subsequent sessions. Generate four-dimensional reconstruction results by globally consistent corrected pose, static layer and all session Gaussian evolution layers.

[0018] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention explicitly decomposes the scene into a static background layer and an evolutionary structure layer and records the cross-session evolution relationship with a structural association diagram, replacing the implicit expression of topological changes by a single continuous deformation field, thereby reducing geometric tearing and afterimage artifacts caused when topological events such as the addition and removal of components occur.

[0019] 2. This invention uses a transient interference detection and layered masking mechanism to accurately identify transient objects at the construction site, mask these transient objects, ensure the data accuracy of the static background layer and the evolving structure layer, and reduce the contamination of Gaussian parameters by transient objects at the construction site.

[0020] 3. This invention explicitly records the addition, continuation, and disappearance events of components through a cross-temporal structural association diagram, supports the structural level query requirements in construction progress management scenarios, and improves the usability of reconstruction results in engineering management scenarios. Attached Figure Description

[0021] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0022] Figure 1 This is a schematic diagram of the method of the present invention; Figure 2 This is a schematic diagram of mask generation according to the present invention; Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0024] This application provides a method for four-dimensional reconstruction of construction scenes based on hierarchical representation and structural association graphs. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for four-dimensional reconstruction of construction scenes based on hierarchical representation and structural association graphs can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0025] Reference Figure 1 The diagram shown is a flowchart illustrating a four-dimensional reconstruction method for a construction scene based on hierarchical representation and structural association diagram, according to an embodiment of the present invention. In this embodiment, the four-dimensional reconstruction method for a construction scene based on hierarchical representation and structural association diagram includes: Step S1: Acquire image data generated by the on-site acquisition session, perform spatial calibration on the image data generated by each session, and generate standard observation data; Step S11: Statistically analyze the image data generated in each on-site acquisition session, and obtain the camera pose and sparse point cloud for each session by performing independent motion reconstruction structure calculation; It should be noted that a complete on-site image acquisition task is called a session. For example, if there is a weekly drone inspection, the inspection and shooting of each week is one session, and 20 weeks correspond to 20 sessions.

[0026] Step S12: Obtain the global static layer Gaussian rendering result of the construction site, extract the common feature parts of the image data generated by each site acquisition session and the global static layer Gaussian rendering result, and obtain the reference anchor point; It should be noted that the public features refer to areas around the construction site that have existed for a long time and have a stable appearance (such as site boundary fences, adjacent existing buildings, and fixed ground markers).

[0027] Based on the reference anchor point, the camera pose of each session is mapped to the sparse point cloud in a unified coordinate system to obtain the camera pose and initial point cloud of each session unified to the global coordinate system, which constitutes the standard observation data.

[0028] Please see Figure 2 Step S2: Based on standard observation data, image data and sparse point cloud under the same session are collected, and consistency analysis is performed on the image data; non-fixed categories in the image are identified, and the final transient mask is generated by combining the consistency analysis results; the final transient mask is used to feed back to the sparse point cloud to obtain the effective supervision region, and the effective supervision region is structurally labeled to construct a structural association graph. Step S21: Based on image data, collect multi-view images of the construction site; acquire sparse point cloud under the same session, reproject the 3D points on the sparse point cloud onto the multi-view images, extract the color information of the corresponding pixels of the 3D points at different views, and perform color consistency analysis; backproject the pixels on the multi-view images, and perform structural consistency analysis in combination with the sparse point cloud; based on color consistency analysis and structural consistency analysis, extract transient influence candidate points. Step S211: Traverse the 3D point reprojection images and record the number of images as as; collect the color value of the corresponding pixel in each image reprojected by the 3D point and record it as scz. a ; It should be noted that the color value is scz. a This represents the RGB value of the pixel corresponding to the a-th image, including the specific values ​​of different color channels.

[0029] For color value scz aStatistical analysis and mean calculation yield the color mean sjz; from the color value scz a The average color difference psc is obtained by calculating the average color difference of the three-dimensional point reprojection with the color mean sjz. ; It should be noted that: in the formula, scz a -sjz is the Euclidean distance between the color value and the color mean.

[0030] The average color of the three-dimensional points corresponding to the reference anchor point is collected and a threshold is set to obtain the color threshold syz; If psc≤syz, the color of the image of the 3D point is considered to be consistent across all viewpoints. If psc > syz, it is determined that the colors of the three-dimensional point are inconsistent across different viewpoints, and the three-dimensional point is marked accordingly; Step S212: Obtain the 3D points of the sparse point cloud to obtain the original 3D points ysd. Map the original 3D points to the multi-view images, extract the corresponding pixels, and back-project the corresponding pixels onto the sparse point cloud to obtain the inverted 3D points fts. Collect the 3D spatial distance between the inverted 3D points fts and the original 3D points ysd. Statistically calculate the 3D spatial distance of the original 3D points on different images and extract the mean value to obtain the mean projection error twc. The average projection error of the three-dimensional points corresponding to the reference anchor point is collected and a threshold is set to obtain the structural threshold jyz; If twc≤jyz, the image structure of the three-dimensional point is determined to be consistent from all perspectives. If twc > jyz, it is determined that the image structure of the three-dimensional point is inconsistent from different viewpoints, and the three-dimensional point is marked. Based on the marked 3D points, 3D points that simultaneously exhibit color inconsistency and structural inconsistency are statistically analyzed to obtain candidate points for transient influence.

[0031] It should be noted that by simultaneously constraining the selection of candidate points for transient influences through color consistency and structural consistency, the accuracy of the judgment of transient objects is ensured.

[0032] Step S22: Input the multi-view images into the general object detection model for image recognition, extract transient objects at the construction site, perform intersection and verification between the transient objects and transient influence candidate points to obtain the final transient mask; based on the final transient mask, extract the effective supervision region in the sparse point cloud, identify the evolution structure of the effective supervision region, and construct a structure association graph.

[0033] It should be noted that: general object detection models are deep learning vision models that can identify general real-world object categories, such as Faster R-CNN and Grounding DINO; transient objects at construction sites refer to objects that exist temporarily for a short period of time at the construction site, are movable, and do not belong to the main building.

[0034] Step S221: Perform image recognition on the multi-view images, extract transient objects at the construction site, record the corresponding pixels of the transient objects at the construction site, and obtain a transient candidate mask; map the transient influence candidate points to the multi-view images, perform intersection processing with the transient candidate masks, and obtain the final transient mask; It should be noted that by retaining the regions supported by both criteria as the final transient mask, the false detection rate of a single criterion is reduced.

[0035] Step S222: Based on the final transient mask, transient objects in the multi-view image are excluded, the effective supervision area of ​​the image is recorded, and the effective supervision area is fed back to the sparse point cloud; the evolution structure within the effective supervision area is collected; the evolution structure is statistically analyzed as nodes to construct a structure association graph; It should be noted that: Evolving structures are rigid building components that are gradually modified and modified by manual construction on the construction site and undergo temporal changes, such as floor slabs, walls, beams and columns, scaffolding, formwork, and tower crane bases.

[0036] Step S3: Generate an initial static layer Gaussian from standard observation data; acquire the camera pose in the current session, combine it with the initial static layer Gaussian to generate a predicted image, and analyze the image residuals; based on the image residuals, extract evolution nodes and associate them with the structure correlation graph; at the same time, extract the spatial range corresponding to the evolution nodes and generate an evolution layer Gaussian. It should be noted that: initializing the static layer Gaussian refers to the static layer Gaussian of the construction scene obtained based on different session states. For example, in the first acquisition session, the static layer Gaussian is generated based on the standard observation data corresponding to the first acquisition session. In non-first acquisition sessions, the static layer Gaussian of the previous session is extracted.

[0037] Step S31: Extract different camera viewpoints based on the camera pose in the current session, use the initialized static layer Gaussian as the rendering data source, perform Gaussian splash rasterization rendering to obtain the predicted images under different viewpoints; according to the image data in the current session, collect the original images of the corresponding viewpoints; obtain the final transient mask of the corresponding image, fit the non-mask area of ​​the predicted image and the original image to obtain the image residual. Step S32: Collect the pixel luminance change of the image under illumination as the residual threshold; extract the corresponding image where the residual is greater than the residual threshold, record the change area of ​​the image, and determine the overlap between the change area and the corresponding area of ​​the node in the structure association graph. If there is an overlap between the node corresponding area and the change area, the change area is regarded as the continuation of the existing component and associated with the corresponding node; if there is no overlap between the node corresponding area and the change area, the change area is regarded as a newly added independent component, an evolution node is constructed, and stored in the structure association graph; at the same time, for each change area, collect the sparse point cloud of the current session and perform Gaussian initialization to obtain the Gaussian evolution layer corresponding to the change area.

[0038] Step S4: Statistically analyze the image data of the current session and historical sessions, update the gradient of the static layer Gaussian, and update the gradient of the evolved layer Gaussian using the image data of the current session; perform gradient descent optimization through the standard splash rendering process to optimize the static layer Gaussian and the evolved layer Gaussian. Step S41: Based on the image data of the current session and historical sessions, extract the effective supervised images, construct a time-series supervised image list, collect the static region gradient of each session, and obtain the gradient update accumulation of the static layer Gaussian; extract the evolution layer Gaussian of the current session, fit the evolution layer Gaussian to the construction scene, and obtain the gradient update of the evolution layer Gaussian. Step S42: Introduce an inter-layer competition suppression term, combining the gradient update accumulation of the static layer Gaussian and the gradient update of the evolved layer Gaussian, and apply a penalty to the static layer and the evolved layer Gaussian; perform gradient descent optimization through the standard splash rendering process, use a density-adaptive cloning and pruning strategy for the evolved layer Gaussian, and adopt a threshold-controlled density adjustment strategy for the static layer Gaussian to maintain the existing historical structure.

[0039] Step S421: Collect the opacity of the static layer and the evolved layer Gaussian in the same spatial neighborhood and set an opacity threshold; if the opacity of both the static layer Gaussian and the evolved layer Gaussian is higher than the opacity threshold, introduce an inter-layer competition suppression term to the static layer Gaussian and the evolved layer Gaussian, apply a penalty to the static layer and the evolved layer Gaussian in the neighborhood, and separate the opacity of the static layer Gaussian and the evolved layer Gaussian. It should be noted that: the inter-layer competition suppression term prevents two Gaussian layers from being rendered at the same time with high transparency in the same area, avoiding the static background from being contaminated by the construction structure and the rendering of the evolving structure being disordered; the value of the inter-layer competition suppression term is equal to the sum of the products of the opacity of the static layer and the evolving layer in the neighborhood.

[0040] Step S422: Render the static Gaussian layer and the evolved Gaussian layer respectively. Combine the gradient update accumulation of the static Gaussian layer and the gradient update of the evolved Gaussian layer to obtain the photometric loss. Constrain the static Gaussian layer and the evolved Gaussian layer through independent regularization to obtain the regularization loss. Obtain the competition suppression loss from the inter-layer competition suppression term of the neighborhood Gaussian layer. Perform gradient descent optimization based on the various losses of the Gaussian layer.

[0041] Step S5: Based on the optimization process of the evolution layer Gaussian, the evolution nodes are extracted, the evolution layer Gaussian and session timestamps corresponding to the evolution nodes are collected to form the evolution sequence of the node, and the evolution nodes and corresponding evolution sequences are associated with the structural association graph to generate a four-dimensional structural evolution record. Step S51: Based on the structural association graph, traverse the evolution nodes, extract the nodes with continuous evolution, collect the corresponding regions of the nodes with continuous evolution, record the evolution layer Gaussian of the region and the current session timestamp, and construct the cross-dialogue evolution sequence of the nodes. Step S52: Extract the evolution nodes corresponding to the newly added independent components, collect and store the session timestamp and initial spatial range of the node, extract the spatial adjacency relationship between the node and the other evolution nodes on the structural association diagram, and establish edge relationships to connect with the other nodes. It should be noted that edge relationships include two types: node expansion or independent addition.

[0042] Step S53: Mark the nodes whose evolution layer Gaussian changes do not occur in two consecutive sessions, record the timestamp of the session corresponding to the disappearance of the change, and reduce the Gaussian opacity of the evolution layer corresponding to the node.

[0043] Step S6: Perform consistency verification by combining the static layer Gaussian with the four-dimensional structure evolution record, and correct the Gaussian based on the verification result to generate the four-dimensional reconstruction result; Step S61: Using the static layer Gaussian as the common observation target across sessions, map the four-dimensional structure evolution record to the static layer Gaussian, record the evolution status of the cross-dialogue evolution of the evolution node on the static layer Gaussian, adjust the evolution node, perform global optimization on the joint reprojection error of the session pose of all sessions and the static layer Gaussian, and correct the inconsistency of the position of the reference anchor point between different sessions. Step S62: Based on the optimization results, propagate the corrected pose change to the Gaussian evolution layer corresponding to each session, so that the spatial position of the Gaussian evolution layer is consistent with the coordinate reference of the corrected static layer; record the overall pose correction magnitude, and according to the correction magnitude, add supplementary observations in the key reference anchor area for subsequent sessions. Generate four-dimensional reconstruction results by globally consistent corrected pose, static layer and all session Gaussian evolution layers.

[0044] It should be noted that by adjusting the reference anchor point, the long-term consistency of the four-dimensional reconstruction results formed by long-term, multiple independent acquisition sessions on the overall coordinate reference is ensured, avoiding the accumulation of errors caused by incremental registration for each session.

[0045] Compared to the problems described in the background art, this invention explicitly decomposes the scene into a static background layer and an evolving structure layer, and records cross-session evolution relationships using a structural association graph. This replaces the implicit expression of topological changes by a single continuous deformation field, reducing geometric tearing and afterimage artifacts caused by topological events such as component addition and removal. Furthermore, this invention accurately identifies transient objects on the construction site through transient interference detection and a layered masking mechanism, masking these transient objects to ensure the data accuracy of the static background layer and the evolving structure layer, and reducing the contamination of Gaussian parameters by transient objects. Finally, this invention explicitly records component addition, continuation, and disappearance events through a cross-temporal structural association graph, supporting structural level query requirements in construction progress management scenarios and improving the usability of reconstruction results in engineering management scenarios. Therefore, the four-dimensional reconstruction method for construction scenes based on layered representation and structural association graphs provided by this invention can improve the accuracy of scene reconstruction.

[0046] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0047] Finally, it should be noted that deleting any one of the above embodiments does not affect the technical solutions of other embodiments. The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for four-dimensional reconstruction of construction scenes based on hierarchical representation and structural association diagrams, characterized in that, include: Step S1: Acquire image data generated during the on-site acquisition session and generate standard observation data; Step S2: Based on standard observation data, collect image data and sparse point cloud data under the same session and perform consistency analysis; identify non-fixed categories in the image, combine the consistency analysis results to generate transient masks, provide feedback to sparse point cloud data, obtain effective supervised regions, and construct structural association graphs; Step S3: Generate an initial static layer Gaussian from standard observation data; collect the camera pose in the current session, combine it with the initial static layer Gaussian to generate a predicted image, and analyze the image residuals; based on the image residuals, extract evolution nodes and associate them with the structure correlation graph; at the same time, extract the spatial range corresponding to the evolution nodes and generate an evolution layer Gaussian. Step S4: Statistically analyze the image data of the current session and historical sessions, and update the gradients of the static layer Gaussian and the evolved layer Gaussian; perform gradient descent optimization through the standard splash rendering process to optimize the static layer Gaussian and the evolved layer Gaussian. Step S5: Extract the evolution nodes, collect the Gaussian evolution layer and session timestamp corresponding to the evolution nodes to form the evolution sequence of the node, associate the evolution nodes with the corresponding evolution sequences to the structural association graph, and generate a four-dimensional structural evolution record. Step S6: Perform consistency verification by combining the static layer Gaussian with the four-dimensional structure evolution record, and correct the Gaussian based on the verification result to generate the four-dimensional reconstruction result.

2. The four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagram as described in claim 1, characterized in that, The specific steps of step S1 are as follows: Step S11: Statistically analyze the image data generated in each on-site acquisition session, and obtain the camera pose and sparse point cloud for each session by performing independent motion reconstruction structure calculation; Step S12: Obtain the global static layer Gaussian rendering result of the construction site, extract the common feature parts of the image data generated by each site acquisition session and the global static layer Gaussian rendering result, and obtain the reference anchor point; Based on the reference anchor point, the camera pose of each session is mapped to the sparse point cloud in a unified coordinate system to obtain the camera pose and initial point cloud of each session unified to the global coordinate system, which constitutes the standard observation data.

3. The four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagrams according to claim 1, characterized in that, The specific steps of step S2 are as follows: Step S21: Based on image data, collect multi-view images of the construction site; acquire sparse point cloud under the same session, reproject the 3D points on the sparse point cloud onto the multi-view images, extract the color information of the corresponding pixels of the 3D points at different views, and perform color consistency analysis; backproject the pixels on the multi-view images, and perform structural consistency analysis in combination with the sparse point cloud; based on color consistency analysis and structural consistency analysis, extract transient influence candidate points. Step S22: Input the multi-view images into the general target detection model for image recognition, extract transient objects at the construction site, and perform intersection and verification between the transient objects at the construction site and the transient influence candidate points to obtain the final transient mask; Based on the final transient mask, the effective supervision region in the sparse point cloud is extracted, the evolution structure of the effective supervision region is identified, and a structural association graph is constructed.

4. The four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagrams according to claim 3, characterized in that, The specific steps of step S21 are as follows: Step S211: Traverse the 3D point reprojection images and record the number of images as as; collect the color value of the corresponding pixel in each image reprojected by the 3D point and record it as scz. a ; For color value scz a Statistical analysis and mean calculation yield the color mean sjz; from the color value scz a The average color difference psc is obtained by calculating the average color difference of the three-dimensional point reprojection with the color mean sjz. The average color of the three-dimensional points corresponding to the reference anchor point is collected and a threshold is set to obtain the color threshold syz; If psc≤syz, the color of the image of the 3D point is considered to be consistent across all viewpoints. If psc > syz, it is determined that the colors of the three-dimensional point are inconsistent across different viewpoints, and the three-dimensional point is marked accordingly; Step S212: Obtain the 3D points of the sparse point cloud to obtain the original 3D points ysd. Map the original 3D points to the multi-view images, extract the corresponding pixels, and back-project the corresponding pixels onto the sparse point cloud to obtain the inverted 3D points fts. Collect the 3D spatial distance between the inverted 3D points fts and the original 3D points ysd. Statistically calculate the 3D spatial distance of the original 3D points on different images and extract the mean value to obtain the mean projection error twc. The average projection error of the three-dimensional points corresponding to the reference anchor point is collected and a threshold is set to obtain the structural threshold jyz; If twc≤jyz, the image structure of the three-dimensional point is determined to be consistent from all perspectives. If twc > jyz, it is determined that the image structure of the three-dimensional point is inconsistent from different viewpoints, and the three-dimensional point is marked. Based on the marked 3D points, 3D points that simultaneously exhibit color inconsistency and structural inconsistency are statistically analyzed to obtain candidate points for transient influence.

5. The four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagrams according to claim 3, characterized in that, The specific steps of step S22 are as follows: Step S221: Perform image recognition on the multi-view images, extract transient objects at the construction site, record the corresponding pixels of the transient objects at the construction site, and obtain a transient candidate mask; map the transient influence candidate points to the multi-view images, perform intersection processing with the transient candidate masks, and obtain the final transient mask; Step S222: Based on the final transient mask, transient objects in the multi-view image are excluded, the effective supervision area of ​​the image is recorded, and the effective supervision area is fed back to the sparse point cloud; the evolution structure within the effective supervision area is collected; the evolution structure is statistically analyzed as nodes, and a structure association graph is constructed.

6. The four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagrams according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S31: Extract different camera viewpoints based on the camera pose in the current session, use the initialized static layer Gaussian as the rendering data source, perform Gaussian splash rasterization rendering to obtain the predicted images under different viewpoints; according to the image data in the current session, collect the original images of the corresponding viewpoints; obtain the final transient mask of the corresponding image, fit the non-mask area of ​​the predicted image and the original image to obtain the image residual. Step S32: Collect the pixel luminance change of the image under illumination as the residual threshold; extract the corresponding image where the residual is greater than the residual threshold, record the change area of ​​the image, and determine the overlap between the change area and the corresponding area of ​​the node in the structural association diagram. If there is an overlap between the corresponding area of ​​the node and the change area, the change area is regarded as the continuation and evolution of the existing component and associated with the corresponding node. If no node's corresponding region overlaps with the changed region, the changed region is treated as a new independent component, an evolution node is constructed, and stored in the structural association graph; at the same time, for each changed region, the sparse point cloud of the current session is collected and Gaussian initialized to obtain the Gaussian evolution layer corresponding to the changed region.

7. The four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagram as described in claim 1, characterized in that, The specific steps of step S4 are as follows: Step S41: Based on the image data of the current session and historical sessions, extract the effective supervised images, construct a time-series supervised image list, collect the static region gradient of each session, and obtain the gradient update accumulation of the static layer Gaussian; extract the evolution layer Gaussian of the current session, fit the evolution layer Gaussian to the construction scene, and obtain the gradient update of the evolution layer Gaussian. Step S42: Introduce an inter-layer competition suppression term, combining the gradient update accumulation of the static layer Gaussian and the gradient update of the evolved layer Gaussian, and apply a penalty to the static layer and the evolved layer Gaussian; perform gradient descent optimization through the standard splash rendering process, use a density-adaptive cloning and pruning strategy for the evolved layer Gaussian, and adopt a threshold-controlled density adjustment strategy for the static layer Gaussian to maintain the existing historical structure.

8. The four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagram as described in claim 7, characterized in that, The specific steps of step S42 are as follows: Step S421: Collect the opacity of the static layer and the evolved layer Gaussian in the same spatial neighborhood and set an opacity threshold; if the opacity of both the static layer Gaussian and the evolved layer Gaussian is higher than the opacity threshold, introduce an inter-layer competition suppression term to the static layer Gaussian and the evolved layer Gaussian, apply a penalty to the static layer and the evolved layer Gaussian in the neighborhood, and separate the opacity of the static layer Gaussian and the evolved layer Gaussian. Step S422: Render the static Gaussian layer and the evolved Gaussian layer respectively. Combine the gradient update accumulation of the static Gaussian layer and the gradient update of the evolved Gaussian layer to obtain the photometric loss. Constrain the static Gaussian layer and the evolved Gaussian layer through independent regularization to obtain the regularization loss. Obtain the competition suppression loss from the inter-layer competition suppression term of the neighborhood Gaussian layer. Perform gradient descent optimization based on the various losses of the Gaussian layer.

9. The four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagram as described in claim 1, characterized in that, The specific steps of step S5 are as follows: Step S51: Based on the structural association graph, traverse the evolution nodes, extract the nodes with continuous evolution, collect the corresponding regions of the nodes with continuous evolution, record the evolution layer Gaussian of the region and the current session timestamp, and construct the cross-dialogue evolution sequence of the nodes. Step S52: Extract the evolution nodes corresponding to the newly added independent components, collect and store the session timestamp and initial spatial range of the node, extract the spatial adjacency relationship between the node and the other evolution nodes on the structural association diagram, and establish edge relationships to connect with the other nodes. Step S53: Mark the nodes whose evolution layer Gaussian changes do not occur in two consecutive sessions, record the timestamp of the session corresponding to the disappearance of the change, and reduce the Gaussian opacity of the evolution layer corresponding to the node.

10. The four-dimensional reconstruction method for construction scenes based on hierarchical representation and structural association diagrams according to claim 1, characterized in that, The specific steps of step S6 are as follows: Step S61: Using the static layer Gaussian as the common observation target across sessions, map the four-dimensional structure evolution record to the static layer Gaussian, record the evolution status of the cross-dialogue evolution of the evolution node on the static layer Gaussian, adjust the evolution node, perform global optimization on the joint reprojection error of the session pose of all sessions and the static layer Gaussian, and correct the inconsistency of the position of the reference anchor point between different sessions. Step S62: Based on the optimization results, propagate the corrected pose change to the Gaussian evolution layer corresponding to each session, so that the spatial position of the Gaussian evolution layer is consistent with the coordinate reference of the corrected static layer; record the overall pose correction magnitude, and according to the correction magnitude, add supplementary observations in the key reference anchor area for subsequent sessions. Generate four-dimensional reconstruction results by globally consistent corrected pose, static layer and all session Gaussian evolution layers.

Images