Bridge construction progress tracking method and system based on multi-source vision and digital twinning

Abstract

The application provides a bridge construction progress tracking method and system based on multi-source vision and digital twinning, relates to the field of bridge construction management, and comprises the following steps: constructing a multi-level inclined three-dimensional model of a construction scene as a digital base; using an EDA-DETR algorithm to analyze a construction monitoring video in real time to identify the type and accurate edge contour of the bridge construction structure; extracting the structural characteristics of the current construction stage through knowledge reasoning; based on the extracted structural characteristics, performing three-dimensional reconstruction to generate a local structure model, and registering and fusing the local structure model with the multi-level inclined three-dimensional model and comparing the local structure model with a bridge BIM design model to quantify the deviation between the actual construction progress and the planned progress; updating the construction state of each component in the multi-level inclined three-dimensional model according to the progress comparison result to complete data driving and state synchronization of the twinned scene. The application significantly improves the accurate perception and intelligent control level of the overall condition of the whole cycle of bridge construction.

Description

Technical Field

[0001] This invention relates to the field of bridge construction management technology, and more specifically, to a method and system for tracking bridge construction progress based on multi-source vision and digital twins. Background Technology

[0002] The core function of a bridge construction platform is to achieve digital, intelligent, and refined management of the entire bridge construction project process, such as tracking and monitoring the construction progress of the bridge structure. Failure to grasp the construction progress in a timely manner, resulting in deviations between the actual construction progress and the construction plan, or delays, will seriously affect the allocation of project resources and even the construction quality. Current bridge construction management technologies largely rely on manual on-site inspections and subjective experience-based judgment, which is not only inefficient and poses safety hazards but also makes it difficult to accurately quantify the progress of complex spatial structures. Therefore, how to overcome the inefficiency and subjective errors of traditional manual inspections and achieve high-precision, automated progress tracking and intelligent control throughout the entire bridge construction process has become a pressing technical problem to be solved in this field. Summary of the Invention

[0003] This invention aims to solve at least one of the technical problems existing in the prior art or related technologies, and discloses a bridge construction progress tracking method and system based on multi-source vision and digital twins. It not only realizes the automated closed loop from physical site perception to digital twin semantic understanding, but also fundamentally ensures the accuracy and timeliness of progress quantification through the dual verification of "multi-source vision recognition and three-dimensional reconstruction comparison", thus achieving the core purpose of intelligent tracking and control of the entire bridge construction cycle.

[0004] The first aspect of this invention discloses a bridge construction progress tracking method based on multi-source vision and digital twins, comprising: S1, multi-source data fusion and dynamic modeling of construction scenarios: acquiring geospatial data of the bridge construction site, integrating high-resolution remote sensing images, UAV oblique photography data, and construction monitoring videos, and constructing a multi-level oblique 3D model from macro-topography to micro-construction details; S2, intelligent identification of construction structures: using the EDA-DETR algorithm to perform real-time analysis of construction monitoring videos to identify the type and precise edge contour of the bridge construction structure; extracting structural features of the current construction stage through knowledge reasoning; wherein, the EDA-DETR algorithm is based on the RT-DETR algorithm, replacing the original backbone network with an M-based... The enhanced detection algorithm is based on the hierarchical architecture of ambaOut; S3, 3D reconstruction-driven quantitative comparison of construction progress: 3D reconstruction is performed based on the extracted structural features to generate a local structural model, and the local structural model is registered and fused with the multi-level inclined 3D model to form a local structural dynamic change model; the local structural dynamic change model is compared with the bridge BIM design model to quantify the deviation between the actual construction progress and the planned progress to obtain the construction progress result; S4, progress tracking closed loop and digital twin synchronous update: the quantified progress comparison result is fed back to the multi-level inclined 3D model to dynamically update the construction status attributes of the corresponding components to ensure that the multi-level inclined 3D model and the physical construction site achieve strict spatiotemporal synchronization.

[0005] In this technical solution, bridge construction strictly follows temporal and physical topology (e.g., steel reinforcement must be tied first, then formwork erected, and finally concrete poured; the formwork must be wrapped around the steel reinforcement). Therefore, the "knowledge reasoning" process of this invention is defined as: spatiotemporal topology matching based on construction prior rules: 1. Spatial reasoning: Calculate the positional relationships (containment, adjacency, support) between contours identified by EDA-DETR. 2. Temporal reasoning: Combine the state of this location in the previous frame / yesterday to eliminate impossible construction regressions. 3. Feature extraction: Infer what is currently being done (stage), and then extract the key parameters required for this stage (such as formwork height, steel reinforcement density).

[0006] According to the bridge construction progress tracking method based on multi-source vision and digital twin disclosed in this invention, preferably, in step S2, the hierarchical architecture of MambaOut specifically includes: A four-stage hierarchical design is adopted, with each stage constructing a feature pyramid by stacking Gated CNN modules; The meta-architecture computation process of the Gated CNN module includes: (1) (2) in, X'This represents the feature vector after normalizing the input vector X. Y This represents the final output feature vector of the entire Gated CNN module. Norm(·) represents normalization. TokenMixer(·) refers to the module that performs token mixing. W1, W2, and W3 are learnable parameters, where the MLP expansion rate is r. It is an activation function; The computation process of the token mixer module in Gated CNN and Mamba includes: (3) (4) Where Z is the input feature tensor, SSM refers to performing selective state space computation, Conv refers to the convolution operation, and TokenMixer... Mamba This refers to mixing Mamba tokens, using TokenMixer. GatedCNN This refers to token mixing during gated convolution.

[0007] In this technical solution, the present invention proposes for the first time to replace the backbone network with a hierarchical architecture based on MambaOut, and uses its unique gating mechanism to filter background noise such as scaffolding, thereby fundamentally overcoming the pain points of traditional algorithms in poor adaptability and blurred edges in harsh construction scenarios.

[0008] According to the bridge construction progress tracking method based on multi-source vision and digital twin disclosed in this invention, preferably, in step S2, the detection head of the EDA-DETR algorithm specifically includes: A new ultra-large target detection branch is added to the original three-scale detection head of the RT-DETR algorithm, forming a four-layer multi-scale detection system. Each detection head is constructed based on feature layers with different downsampling rates. The P3 / 8 feature layer has the highest spatial resolution and is suitable for detecting small components ranging from 8×8 to 16×16 pixels. After two downsampling operations, the P4 / 16 feature layer expands the channel dimension and enhances the semantic information of the features, and is responsible for detecting medium-sized components from 16×16 to 32×32 pixels; The P5 / 32 feature layer is downsampled three times to obtain the strongest semantic representation and the largest receptive field, which is specifically used for the recognition of large components with a resolution of 32×32 pixels or more. The newly added detection head is built based on a further downsampled feature layer, and the output feature map size is 1 / 64 of the original image, which is specifically designed to capture ultra-large-scale construction structures of 64×64 pixels or more.

[0009] In this technical solution, traditional target detection algorithms are prone to failure in bridge construction scenarios because bridge sites exhibit an extreme multi-scale "physical coexistence": the site contains both ultra-large-scale entities (such as integral piers, massive bridge abutments, and main span beams) at the hundreds of meters level, and densely distributed micro-scale components at the decimeter or even centimeter level (such as positioning fasteners, embedded parts, and individual steel cages); simultaneously, the crisscrossing scaffolding and dense safety nets on site constitute extremely strong high-frequency periodic texture interference, easily causing visual confusion in the attention mechanism of traditional algorithms. This invention, through a customized four-layer multi-scale detection head, utilizes the P3 / 8 feature layer with the highest spatial resolution and a newly added 1 / 64 downsampling ultra-large detection head for diversion processing, achieving simultaneous and accurate identification of "macro-giant structures" and "micro-tiny components" within the same field of view without any missed detections.

[0010] According to the bridge construction progress tracking method based on multi-source vision and digital twin disclosed in this invention, preferably, the step S2, which involves extracting structural features of the current construction stage through knowledge reasoning, specifically includes: Pre-construction of a priori semantic graph for bridge construction : (5) Among them, nodes Indicates the standard process status of construction components, edge Represents the spatiotemporal constraints between states; Extract the precise edge contours of each construction structure output by the EDA-DETR algorithm, calculate the spatial topological features between the contours, and generate a real-time scene topology map. : (6) Among them, nodes V S This indicates the process status in a real-time scenario. E S Represents the spatiotemporal constraint rules between real-time process states; Real-time scene topology map With prior semantic graph Subgraph isomorphic matching is performed, and Markov time-series state transition probability is calculated by combining the structural state attributes of the previous time step to infer the actual construction stage of the current bridge structure. Based on the inferred actual construction stage, the corresponding geometric parameters in the precise edge contour are extracted in a directional manner as the structural features of the current construction stage.

[0011] According to the bridge construction progress tracking method based on multi-source vision and digital twin disclosed in this invention, preferably, step S1 specifically includes: firstly, acquiring the geographic spatial range of the target bridge construction site, and based on this range, integrating remote sensing images and digital terrain models to construct a three-dimensional basic scene model that integrates macroscopic terrain, landform, and surrounding environmental information; based on the three-dimensional basic scene model, enabling a UAV to perform multi-view, multi-altitude oblique photography operations on the construction area along the optimal path, simultaneously acquiring high-resolution UAV images and their corresponding precise pose and positioning data; and by jointly processing the acquired image data and pose positioning data, using motion reconstruction structure and multi-view stereo vision technology, generating a multi-level oblique three-dimensional real-scene model with rich geometric details and realistic textures.

[0012] According to the bridge construction progress tracking method based on multi-source vision and digital twin disclosed in this invention, step S3 preferably includes: Based on the key feature information of all currently under construction structures extracted in step S2, each under construction structure is reconstructed in three dimensions to generate a corresponding local structural model. The generated local structural models are spatially registered and fused with the multi-level inclined 3D model to form a dynamic change model of the bridge's local structure that includes the latest state. Obtain the bridge BIM design model and extract the theoretical BIM component model under the current planned time node. Spatially overlay and compare the local structural dynamic change model with the theoretical BIM component model of the corresponding area. Perform a three-dimensional Boolean intersection operation on the registered local structural dynamic change model and the theoretical BIM component model to calculate the ratio of the actual reconstructed model volume to the total volume of the BIM planned model. Use this ratio as the quantitative percentage of the current actual construction progress. Based on the difference between this quantitative percentage and the planned construction progress, output the quantitative comparison result of the bridge structure construction progress.

[0013] According to the bridge construction progress tracking method based on multi-source vision and digital twin disclosed in this invention, preferably, step S4 specifically includes: The system receives the construction progress results generated in step S3 in real time and feeds them back to the multi-level inclined 3D model. The construction progress results include the completion status of each component, the progress deviation value, and potential risk warnings. By establishing unique component codes and attribute mapping rules, the construction status attributes, completion time attributes, and acceptance standard attributes of each component in the multi-level inclined 3D model are updated in real time, ensuring that the multi-level inclined 3D model and the physical construction site remain strictly synchronized.

[0014] The second aspect of the present invention discloses a bridge construction progress tracking system based on multi-source vision and digital twin, comprising: a memory for storing program instructions; and a processor for calling the program instructions stored in the memory to implement the bridge construction progress tracking method based on multi-source vision and digital twin as described in any of the above technical solutions.

[0015] The beneficial effects of the present invention include at least the following: 1. By constructing a multi-level 3D scene model, static terrain features are deeply integrated with the dynamic construction process, providing an accurate and regularly updated digital twin foundation for progress tracking; 2. An enhanced detection algorithm, EDA-DETR (Enhanced Detection Algorithm-DETR), for complex bridge construction structures is proposed. Addressing the challenges of varying target scales, complex backgrounds, and blurred edges in construction videos, this algorithm enhances the DETR framework, achieving high-precision, real-time identification and extraction of bridge construction structure types and edge contours. It overcomes the poor adaptability of traditional detection algorithms in construction scenarios, providing reliable feature input for automated progress analysis.

[0016] 3. By automatically registering, integrating, and comparing the dynamically generated local structural model with the BIM design model, the deviation between the actual progress and the planned progress is directly quantified, realizing automated closed-loop tracking of construction progress from image perception to three-dimensional semantic understanding, replacing the inefficient method that relies on manual interpretation.

[0017] 4. The configurable rapid delivery (deployment) method provided by this invention provides a complete "design-guide-verification-update" intelligent closed loop, ensuring precise control over the entire process from digital design to high-quality physical delivery. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a bridge construction progress tracking method based on multi-source vision and digital twin according to an embodiment of the present invention is shown.

[0019] Figure 2 A schematic diagram of the network structure of a target detection algorithm for a bridge construction progress tracking method based on multi-source vision and digital twins according to an embodiment of the present invention is shown.

[0020] Figure 3 A schematic diagram of the network structure of a Gated CNN module for an object detection algorithm according to an embodiment of the present invention is shown.

[0021] Figure 4 A schematic block diagram of a bridge construction progress tracking system based on multi-source vision and digital twin according to an embodiment of the present invention is shown. Detailed Implementation

[0022] To better understand the above-described objects, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention; however, the invention may be practiced in other ways different from those described herein, and therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] like Figure 1 As shown, an embodiment of the present invention discloses a method for tracking bridge construction progress based on multi-source vision and digital twin, comprising: S1, Multi-source data fusion and dynamic modeling of construction scenarios: First, by acquiring the geographic spatial range of the target bridge construction site, and integrating high-resolution satellite / aerial remote sensing imagery and a high-precision digital terrain model based on this range, a 3D basic scene model is constructed, fusing macroscopic terrain, landform, and surrounding environmental information. This provides a global spatial reference framework with geographic context for bridge construction. On this basis, intelligent planning of UAV mission routes is performed based on the aforementioned 3D basic scene model, enabling the UAV to perform multi-view, multi-altitude oblique photography operations along the optimal path over the construction area, simultaneously acquiring high-resolution UAV images and their corresponding precise pose and positioning data. By jointly processing the acquired multi-view images and pose-fixed positioning data, and utilizing motion reconstruction structures and multi-view stereo vision technology, a multi-level oblique 3D real-world model with rich geometric details and realistic textures is generated. This model not only represents the construction progress of the main bridge structure but also precisely reconstructs the spatial morphology and distribution characteristics of construction details such as formwork, supports, and temporary facilities. This constructs an integrated, measurable, and semantically meaningful 3D spatial benchmark for bridge construction, encompassing the terrain environment, the construction entity, and the macroscopic whole to the microscopic part. This provides a high-fidelity, dynamically updatable spatial model for subsequent visual recognition of construction status and progress comparison. It provides a high-fidelity, dynamically updatable three-dimensional spatial benchmark for visual recognition and progress comparison of subsequent construction status.

[0024] S2, Intelligent Recognition of Construction Structures: Construction monitoring videos are acquired regularly, and image sample training sets and image sample test sets for bridge construction structures are constructed based on the monitoring videos. The training set is then trained using the improved target detection algorithm EDA-DETR to achieve intelligent recognition of bridge construction structures.

[0025] Figure 2The specific network structure of the EDA-DETR algorithm proposed in this invention is shown: Based on RT-DETR, this invention replaces the original backbone network of RT-DETR with a hierarchical architecture based on MambaOut. This architecture adopts a four-stage hierarchical design, with each stage constructing a feature pyramid by stacking Gated CNN modules.

[0026] like Figure 3 As shown, the meta-architecture of the Gated CNN module can be viewed as a simplified integration of the token mixer and the multilayer perceptron (MLP) component in the MetaFormer framework. Given an input vector X, its meta-architecture can be expressed as Equation (1) and Equation (2).

[0027] (1) (2) in, X' This represents the feature vector after normalizing the input vector X. Y This represents the final output feature vector of the entire Gated CNN module. Norm(·) represents normalization. TokenMixer(·) refers to the module that performs token mixing. W1, W2, and W3 are learnable parameters, where the MLP expansion rate is r. It is the activation function, and the token mixer module structure of Gated CNN and Mamba is as shown in formulas (3) and (4).

[0028] (3) (4) Where Z is the input feature tensor, SSM refers to performing selective state space computation, Conv refers to the convolution operation, and TokenMixer... Mamba This refers to mixing Mamba tokens, using TokenMixer. GatedCNN This refers to token mixing during gated convolution.

[0029] like Figure 2As shown, the data flow linkage between the aforementioned token mixer module and the overall EDA-DETR architecture is as follows: The input construction monitoring video frame image first enters a hierarchical backbone network built based on MambaOut. In the four stages of this backbone network, the input feature tensor is sequentially processed by stacked Gated CNN modules for local feature extraction and token mixing operations. After layer-by-layer processing and feature transformation by the token mixer modules at each stage, the backbone network extracts and outputs a set of multi-scale feature maps. Subsequently, these high-order semantic feature maps generated by the token mixer are directly fed into the EDA-DETR Hybrid Encoder as input data streams for cross-scale feature fusion and intra-scale self-attention interaction. The fused feature sequence is further fed into the Transformer decoder for cross-attention operations with the object query tensor (ObjectQueries). Finally, the multi-scale detection head performs classification and regression operations, outputting the class probability and bounding box coordinates of the bridge construction structure, thereby realizing a closed-loop data flow from low-level feature extraction to end-to-end detection and inference.

[0030] Figure 2 The detection head of the EDA-DETR proposed in this invention is not a general, conventional configuration, but a four-layer multi-scale detection head specifically customized for the extreme scale spans of bridge construction. At the physical construction site, when a camera overlooks the entire bridge, a single span or main pier may occupy thousands of pixels in the image, while an embedded fastener may only occupy a few pixels. Therefore, the detection heads employ a tiered division of labor: the P3 / 8 feature layer, with the highest spatial resolution and low downsampling rate, is specifically responsible for capturing extremely small-scale components (such as rebar cage nodes and fasteners) ranging from 8×8 to 16×16 pixels, preventing them from "evaporating" in the deep network due to feature map sparsity; the subsequent P4 / 16 and P5 / 32 feature layers progressively process medium and large-scale components; most importantly, addressing the problem of traditional RT-DETR easily producing "fragmented false detections" (i.e., misidentifying a continuous beam as several independent objects) when dealing with "visual behemoths" such as entire spans and giant cable towers, this embodiment specifically adds an output size only equal to the original... Figure 1 The / 64 ultra-large target detection branch has an extreme receptive field, specifically designed to capture ultra-large-scale construction structures larger than 64×64 pixels, thus establishing a four-level detection mechanism at the algorithm's underlying layer that completely maps to the size distribution of bridge components in the physical world.

[0031] In this invention, the native detection heads of MambaOut (such as the fully connected layers / headers used for ImageNet classification) have been completely stripped (removed / cut off) during integration. MambaOut is used only as a backbone network in this invention. MambaOut (backbone network) is responsible for extracting multi-scale feature maps at the front end. Its output feature data stream is fused by the hybrid encoder of RT-DETR and finally input into the improved RT-DETR four-layer detection head for decoupled classification and bounding box regression calculation.

[0032] The EDA-DETR detection head proposed in this invention is a four-layer multi-scale detection head that enhances the original prediction mechanism of RT-DETR. At the detection head design level, considering the wide distribution of target scales at bridge construction sites, a new ultra-large target detection branch is added to the original three-scale detection head, forming a four-layer multi-scale detection system. Each detection head is constructed based on feature layers with different downsampling rates: the P3 / 8 feature layer has the highest spatial resolution and is suitable for detecting small components ranging from 8×8 to 16×16 pixels; the P4 / 16 feature layer, after two downsampling operations, expands the channel dimension and enhances the semantic information of the features, responsible for detecting medium-sized components ranging from 16×16 to 32×32 pixels; the P5 / 32 feature layer, after three downsampling operations, obtains the strongest semantic representation and the largest receptive field, specifically for recognizing large components larger than 32×32 pixels. The newly added detection head is constructed based on feature layers with further downsampling, and its output feature map size is 1 / 64 of the original image, specifically for capturing ultra-large-scale construction structures larger than 64×64 pixels.

[0033] After using EDA-DETR to output the type and precise edge contour of the bridge construction structure, this invention further extracts the structural features of the current construction stage through knowledge reasoning. The specific calculation process is as follows: Step A: Construct a priori construction semantic graph: Pre-construction of a priori semantic graph for bridge construction Among them, the node set Includes standard construction procedures for various bridge components at different stages of construction; edge set The constraints include two types of rules between states: spatial physical constraints (such as containment, support, and adjacency) and temporal topological constraints (such as pre-position and post-position), as shown in formula (5): (5) Step B: Real-time scene topology graph generation: The precise edge contours of each construction structure output by the EDA-DETR algorithm are extracted, and the spatial topological features between each contour are calculated. Specifically, by calculating the intersection region of the bounding rectangles of adjacent contours and the mask intersection-union ratio (IoU), it is determined whether there are spatial physical relationships such as enclosure, support, or adjacency between targets, and a real-time scene topology map is generated accordingly. As shown in formula (6): (6) Step C: Multimodal graph matching and timing state inference: The real-time scene topology map With the prior semantic graph Subgraph isomorphic matching is performed. To eliminate visual misjudgments caused by local occlusion or lighting artifacts on site, temporal constraints are introduced, and Markov temporal state transition probabilities are calculated based on the structural state attributes of the previous time step. If the current graph matching state belongs to a reverse process that does not conform to construction specifications, its state transition probability is reduced to zero, thereby accurately inferring the actual construction stage of the current bridge structure.

[0034] Step D: Targeted feature extraction based on inference results: Based on the inferred actual construction stage, irrelevant scaffolding or construction machinery interference in the background is filtered out, and the corresponding spatial geometric parameters (preferably the absolute elevation, verticality angle, and coordinates of the top surface fitting center point) in the precise edge contour are extracted in a directional manner as the key structural feature information of the current construction stage, and used as the data source for local three-dimensional reconstruction in the subsequent step S3.

[0035] S3, 3D reconstruction-driven quantitative comparison of construction progress: Based on the key feature information of all currently under construction structures extracted in step S2, each under construction structure is reconstructed in three dimensions to generate a corresponding local structural model. The generated local structural models are spatially registered and fused with the multi-level tilted 3D model in step S1 to form a dynamic change model of the local structure of the bridge that includes the latest state. Obtain the bridge's BIM design model and extract the theoretical BIM component models for the current planned timeframe. Spatially overlay and compare the dynamic change model of the local structure with the corresponding theoretical BIM component models to quantify the deviation between the actual construction progress and the planned progress.

[0036] The specific method for quantitative comparison is as follows: perform a three-dimensional Boolean intersection operation on the registered local structural dynamic change model and the theoretical BIM component model, calculate the ratio of the actual reconstructed model volume to the total volume of the BIM planned model, and use this ratio as the quantitative percentage of the current actual construction progress; based on the difference between this quantitative percentage and the planned construction progress, output the final quantitative comparison result of the bridge structure construction progress.

[0037] S4, progress tracking closed loop and synchronous updates with digital twin: The above steps are uniformly scheduled and coordinated by a central closed-loop management engine. The engine receives the quantitative comparison results of construction progress generated in step S3 in real time and automatically feeds these results—including the completion status of each component (e.g., "installed," "under construction," "pending acceptance"), progress deviation values, and potential risk warnings—back to the core digital twin database. Through the established unique component codes (IDs) and attribute mapping rules, the system updates the dynamic attributes of each component in the multi-level inclined 3D model in real time and accurately, including construction status, completion time, and acceptance criteria, ensuring strict synchronization between the 3D virtual model and the physical construction site. This forms a complete intelligent closed loop, ensuring accurate tracking throughout the entire process from multi-source visual perception to digital twin status synchronization.

[0038] like Figure 4 As shown, another embodiment of the present invention discloses a bridge construction progress tracking system 400 based on multi-source vision and digital twins, comprising: a memory 401 for storing program instructions; a processor 402 for calling the program instructions stored in the memory to implement the bridge construction progress tracking method based on multi-source vision and digital twins as described in the above embodiment; and a communication interface 403 for establishing a data communication link with external physical devices (including drones, construction monitoring cameras, and BIM center servers) to acquire geospatial data, multi-view image data, attitude positioning data, and construction monitoring video streams in real time. This system can be deployed on a cloud server or at an edge computing node on the construction site.

[0039] When the processor 402 calls the program instructions, it constructs and runs the following core functional modules in its internal logic to implement the method steps of any of the above embodiments: (1) Multi-source data modeling module: used to execute step S1, construct a multi-level tilted three-dimensional model that integrates macro-terrain and micro-construction details as a basic digital base; (2) Structural intelligent recognition module: used to execute step S2, call the EDA-DETR algorithm to intelligently recognize the video stream, and extract the structural features of the current construction stage through knowledge reasoning based on Markov state transition probability; (3) Progress quantification comparison module: used to execute step S3, perform three-dimensional reconstruction on the identified entities to generate a local structural model, and perform three-dimensional Boolean intersection operation with the theoretical BIM component model to output the progress quantification comparison result; (4) Twin synchronous update module: used to execute step S4, feed back the quantification comparison result to the digital twin database with GUID encoding as the index, and drive the dynamic overwriting and strict spatiotemporal synchronization of the state attributes of the multi-level tilted three-dimensional model.

[0040] In summary, this invention aims to address the problems of high reliance on manual inspection, low efficiency, and large subjective errors in traditional bridge construction progress tracking. It proposes a precise tracking strategy for complex bridge construction structures that integrates three aspects: EDA-DETR (algorithm layer), knowledge graph Markov chain (logic layer), and three-dimensional Boolean volume comparison (physical layer). Specifically: First, the geospatial extent of the construction site is acquired, and a 3D basic scene model is constructed by combining terrain data and high-resolution remote sensing imagery. Based on this model, UAV flight paths are planned, and imagery and positioning data are collected by the UAVs to generate a tilted 3D model with multi-level detailed features. Construction monitoring videos are acquired periodically, and the EDA-DETR (Enhanced Detection Algorithm-DETR) algorithm is proposed to intelligently identify bridge construction structures in the videos, extracting their type and edge information, and obtaining the key features of all currently under construction structures through knowledge reasoning. Based on this, each construction structure is 3D reconstructed to form a local structural model, which is then registered and fused with the multi-level tilted 3D model, and periodically updated to obtain a dynamic change model of the bridge's local structure. Finally, the bridge BIM design model is integrated, and by comparing the progress information of the dynamic change model and the design model, automated comparative analysis of the actual construction progress and planned construction progress is achieved, outputting the bridge structure construction progress tracking results. This invention not only realizes an automated closed loop from physical site perception to digital twin semantic understanding, but also fundamentally ensures the accuracy and timeliness of progress quantification through dual verification of "multi-source visual recognition and three-dimensional reconstruction comparison", thus achieving the core objective of intelligent tracking and control of the entire bridge construction cycle.

[0041] All or part of the steps in the various methods of the above embodiments can be implemented by a program controlling the relevant hardware. The program can be stored in a readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other readable medium that can be used to carry or store data.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for tracking bridge construction progress based on multi-source vision and digital twin, characterized in that, include: S1, Multi-source data fusion and dynamic modeling of construction scenarios: Obtain geospatial data of bridge construction sites, integrate high-resolution remote sensing images, UAV oblique photography data and construction monitoring videos, and construct a multi-level oblique 3D model from macro terrain to micro construction details. S2, Intelligent Recognition of Construction Structures: The EDA-DETR algorithm is used to analyze construction monitoring videos in real time to identify the type and precise edge contour of bridge construction structures; structural features of the current construction stage are extracted through knowledge reasoning; wherein, the EDA-DETR algorithm is an enhanced detection algorithm formed by replacing the original backbone network with a hierarchical architecture based on MambaOut on the basis of the RT-DETR algorithm. S3, 3D Reconstruction-Driven Quantitative Comparison of Construction Progress: Based on the extracted structural features, 3D reconstruction is performed to generate a local structural model, and the local structural model is registered and fused with the multi-level inclined 3D model to form a local structural dynamic change model; the local structural dynamic change model is compared with the bridge BIM design model to quantify the deviation between the actual construction progress and the planned progress, and to obtain the construction progress result. S4, Progress tracking closed loop and digital twin synchronous update: The quantified progress comparison results are fed back to the multi-level inclined 3D model to dynamically update the construction status attributes of the corresponding components, so as to ensure that the multi-level inclined 3D model and the physical construction site achieve strict spatiotemporal synchronization.

2. The bridge construction progress tracking method based on multi-source vision and digital twin as described in claim 1, characterized in that, In step S2, the hierarchical architecture of MambaOut specifically includes: A four-stage hierarchical design is adopted, with each stage constructing a feature pyramid by stacking Gated CNN modules; The meta-architecture computation process of the Gated CNN module includes: ； ； Where X' represents the feature vector after normalizing the input vector X, Y represents the final output feature vector of the entire Gated CNN module, Norm(·) represents normalization, TokenMixer(·) refers to the module that performs token mixing, W1, W2 and W3 are learnable parameters, and the MLP expansion rate is r; It is an activation function; The computation process of the token mixer module in Gated CNN and Mamba includes: ； ； Where Z is the input feature tensor, SSM indicates selective state space computation, Conv refers to the convolution operation, and TokenMixer... Mamba Indicates Mamba token mixing, TokenMixer GatedCNN This refers to token mixing during gated convolution.

3. The bridge construction progress tracking method based on multi-source vision and digital twin as described in claim 1, characterized in that, In step S2, the detection head of the EDA-DETR algorithm specifically includes: A new ultra-large target detection branch is added to the original three-scale detection head of the RT-DETR algorithm, forming a four-layer multi-scale detection system. Each detection head is constructed based on feature layers with different downsampling rates. The P3 / 8 feature layer has the highest spatial resolution and is suitable for detecting small components ranging from 8×8 to 16×16 pixels. After two downsampling operations, the P4 / 16 feature layer expands the channel dimension and enhances the semantic information of the features, and is responsible for detecting medium-sized components from 16×16 to 32×32 pixels; The P5 / 32 feature layer is downsampled three times to obtain the strongest semantic representation and the largest receptive field, which is specifically used for the recognition of large components with a resolution of 32×32 pixels or more. The newly added detection head is built based on a further downsampled feature layer, and the output feature map size is 1 / 64 of the original image, which is specifically designed to capture ultra-large-scale construction structures of 64×64 pixels or more.

4. The bridge construction progress tracking method based on multi-source vision and digital twin as described in claim 1, characterized in that, In step S2, the step of extracting the structural features of the current construction stage through knowledge reasoning specifically includes: Pre-construction of a priori semantic graph for bridge construction : ； Among them, nodes Indicates the standard process status of construction components, edge Represents the spatiotemporal constraints between states; Extract the precise edge contours of each construction structure output by the EDA-DETR algorithm, calculate the spatial topological features between each contour, and generate a real-time scene topology map. : ； Among them, node V S E represents the process status in a real-time scenario. S Represents the spatiotemporal constraint rules between real-time process states; Real-time scene topology map With prior semantic graph Subgraph isomorphic matching is performed, and Markov time-series state transition probability is calculated by combining the structural state attributes of the previous time step to infer the actual construction stage of the current bridge structure. Based on the inferred actual construction stage, the corresponding geometric parameters in the precise edge contour are extracted in a directional manner as the structural features of the current construction stage.

5. The bridge construction progress tracking method based on multi-source vision and digital twin as described in claim 1, characterized in that, Step S1 specifically includes: First, the geographic spatial range of the target bridge construction site is obtained, and based on this range, remote sensing images and digital terrain models are integrated to construct a three-dimensional basic scene model that integrates macro-topography, landform and surrounding environmental information. Based on the aforementioned three-dimensional basic scene model, the UAV performs multi-view, multi-altitude oblique photography operations on the construction area along the optimal path, simultaneously acquiring high-resolution UAV images and their corresponding precise pose and positioning data. By jointly processing the acquired image data and attitude positioning data, and utilizing motion reconstruction structure and multi-view stereo vision technology, a multi-level tilted 3D real-scene model with rich geometric details and realistic textures is generated.

6. The bridge construction progress tracking method based on multi-source vision and digital twin as described in claim 1, characterized in that, Step S3 specifically includes: Based on the key feature information of all currently under construction structures extracted in step S2, each under construction structure is reconstructed in three dimensions to generate a corresponding local structural model. The generated local structural models are spatially registered and fused with the multi-level inclined 3D model to form a dynamic change model of the bridge's local structure that includes the latest state. Obtain the bridge BIM design model and extract the theoretical BIM component model under the current planned time node. Spatially overlay and compare the local structural dynamic change model with the theoretical BIM component model of the corresponding area. Perform a three-dimensional Boolean intersection operation on the registered local structural dynamic change model and the theoretical BIM component model to calculate the ratio of the actual reconstructed model volume to the total volume of the BIM planned model. Use this ratio as the quantitative percentage of the current actual construction progress. Based on the difference between this quantitative percentage and the planned construction progress, output the bridge structure construction progress result.

7. The bridge construction progress tracking method based on multi-source vision and digital twin as described in claim 1, characterized in that, Step S4 specifically includes: The construction progress results generated in step S3 are received in real time and fed back to the multi-level inclined three-dimensional model. The construction progress results include the completion status of each component, the progress deviation value, and potential risk warnings. By establishing unique component codes and attribute mapping rules, the construction status attributes, completion time attributes, and acceptance standard attributes of each component in the multi-level inclined 3D model are updated in real time, ensuring that the multi-level inclined 3D model and the physical construction site remain strictly synchronized.

8. A bridge construction progress tracking system based on multi-source vision and digital twin, characterized in that, include: Memory, used to store program instructions; A processor is configured to invoke the program instructions stored in the memory to implement the bridge construction progress tracking method based on multi-source vision and digital twin as described in any one of claims 1 to 7.

Images