Three-dimensional fence crossing recognition system and method based on building information model

Abstract

The application discloses a three-dimensional fence out-of-bound recognition system and method based on a building information model, belongs to the technical field of out-of-bound behavior recognition, and solves the problem that the fence boundary lacks three-dimensional space information and cannot adapt to a complex three-dimensional construction environment in the prior art method.The method comprises the following steps: recognizing and identifying a building information model based on an edge computing node, generating a three-dimensional dynamic fence in combination with a target component set, performing edge-side preprocessing on real-time monitoring data, performing 3D target detection on a behavior fusion feature set based on a 3D target tracking model, and outputting a three-dimensional fence out-of-bound recognition result.The application realizes deep fusion of the fence boundary and a BIM component and dynamic adaptation with the construction progress by importing the building information model into the edge computing node for recognition and identification, realizes a leap from two-dimensional plane out-of-bound detection to three-dimensional space occupation determination based on the 3D target tracking model, and realizes accurate detection and space occupation determination of a moving target in a complex three-dimensional space.

Description

Technical Field

[0001] This invention belongs to the field of boundary crossing behavior recognition technology, specifically relating to a three-dimensional fence boundary crossing recognition system and method based on building information modeling. Background Technology

[0002] Construction safety management is a core aspect of engineering construction. Traditional construction site safety fencing primarily employs physical isolation methods, such as setting up warning tapes, barriers, and safety nets, supplemented by manual patrols for safety control. With the development of information technology, electronic fence technology has gradually been introduced into the field of construction site safety management, achieving automated monitoring of fenced areas through video surveillance, sensors, and other technological means.

[0003] Traditional electronic fences are usually defined based on two-dimensional planar diagrams, which cannot accurately reflect the complex three-dimensional spatial relationships of the construction site, easily leading to false alarms or missed alarms. Moreover, existing monitoring systems are often independent of the BIM management system, failing to effectively utilize the rich component information and spatial topology relationships contained in the BIM model. This results in isolated early warning information that cannot be quickly linked to specific building components or construction tasks, making it difficult for managers to make rapid emergency decisions and allocate resources.

[0004] Chinese patent CN117409362A discloses a method and system for identifying personnel crossing boundaries within a demarcated fenced area. The method includes: acquiring surveillance video data within the demarcated fenced area; labeling normal personnel behavior and personnel crossing boundaries on the surveillance video data; segmenting the surveillance video data into multiple segments according to the labeling information and assigning a label to each segment to form label data; acquiring personnel data from the surveillance video data and combining it with the label data to form a training dataset; setting a personnel crossing boundary recognition model; inputting the training dataset into the personnel crossing boundary recognition model for training; and identifying personnel crossing boundaries using the trained personnel crossing boundary recognition model. However, existing methods lack three-dimensional spatial information about the fence boundary, making them unsuitable for complex three-dimensional construction environments. This leads to a significant decrease in the accuracy of boundary crossing recognition in complex three-dimensional environments. Furthermore, the boundary crossing judgment lacks a verification mechanism related to BIM semantics, the warning information has a low degree of structure, and it is not associated with the component attributes in the building information model, making it impossible to output the specific component information involved in the boundary crossing. To address these problems, we propose a three-dimensional fence boundary crossing recognition system and method based on building information model. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a three-dimensional fence boundary recognition system and method based on building information modeling, which solves the problem that existing methods lack three-dimensional spatial information on fence boundaries and cannot adapt to complex three-dimensional construction environments.

[0006] This invention is implemented as follows: a 3D fence boundary crossing identification method based on Building Information Modeling (BIM), wherein the 3D fence boundary crossing identification method based on BIM includes: The building information model of the target construction site is obtained and imported into the edge computing node. The building information model is parsed and identified based on the edge computing node to obtain the target component set. The target component set is then combined to generate a three-dimensional dynamic fence. Based on the real-time acquisition of real-time monitoring data within the target construction site by distributed sensor network nodes, edge-side preprocessing of the real-time monitoring data is performed, and the preprocessed behavior fusion feature set is output. Load the behavior fusion feature set, pre-build a deep learning-based 3D target tracking model, perform 3D target detection on the behavior fusion feature set based on the 3D target tracking model, and output the 3D fence crossing recognition result.

[0007] Preferably, the method for parsing and identifying building information models based on edge computing nodes includes: Load the target construction site building information model, identify and analyze the target components in the target construction site building information model, read the type attributes and geometric coordinate information of the target components, combine the type attributes and geometric coordinate information of the target components, and perform preliminary screening of the target components based on the preset dangerous area screening rules to generate a preliminary screening set of components; The target components in the building information model of the target construction site are traversed twice to identify the number of adjacent components associated with the target components, the spatial influence range, the process criticality, and the hidden risk attributes. The number of adjacent components associated with the target components, the spatial influence range, the process criticality, and the hidden risk attributes are normalized to obtain the degree of association of adjacent components, the degree of spatial influence, the degree of process criticality, and the degree of hidden risk. The order relation analysis method is used to assign weights to obtain the component screening weights of the target components. The formula for calculating the component screening weight is as follows: , , , in, Indicates the component selection weight. These represent the degree of association and adjacency of components, the degree of spatial influence, the degree of process criticality, and the degree of hidden risk, respectively. These are the importance weights for associated adjacent components, spatial influence, process criticality, and hidden risks, respectively. This indicates the number of adjacent components associated with the current target component and the maximum number of adjacent components associated with it in the building information model of the target construction site. These represent the volume of spatial influence of the current target component and the maximum volume of spatial influence of the component in the model, respectively. These represent the process criticality score of the current target component and the maximum process criticality score of the component in the model, respectively. These represent the current target component's hidden risk score and the maximum hidden risk score of the component in the model, respectively. Determine whether the component screening weight of the target component exceeds the preset screening threshold; If the component screening weight of the target component exceeds the preset screening threshold, the target component will be merged into the initial screening set of components to obtain the target component set.

[0008] Preferably, the method for generating a three-dimensional dynamic fence by combining a set of target components includes: Load the target component set, and generate the target component spatial composition matrix based on the geometric coordinate information and type attributes of the target components in the target component set; The three-dimensional initial boundary of the target component is generated based on the spatial composition matrix of the target component, and the spatial influence degree, process criticality degree, and hidden risk degree of the target component are identified. Based on the normalized spatial influence degree, process criticality degree, and hidden risk degree of the target component, the three-dimensional initial boundary of the target component is subjected to spatial expansion redundancy processing to generate the three-dimensional redundant boundary of the target component. Obtain the three-dimensional redundant boundary of the target component, perform a Boolean union operation on the three-dimensional edge region of the target component based on the edge computing node, eliminate the overlapping boundary of the three-dimensional redundant boundary of the target component, and obtain the three-dimensional redundant boundary of the target component after the edge overlap region is eliminated. After eliminating the overlapping areas of the loading edge, the target component's three-dimensional redundant boundary is vertically stretched and height-blended to obtain the merged three-dimensional fence of the component. Obtain the merged 3D fence of the components, establish a component attribute mapping table for the 3D fence of the components, and establish the spatial topological relationship between the target components. Based on the component attribute mapping table and the spatial topological relationship, construct a dynamic fence constraint set, embed the dynamic fence constraint set into the 3D fence of the components, and obtain the 3D dynamic fence.

[0009] Preferably, the method for edge-side preprocessing of real-time monitoring data includes: It receives asynchronous real-time monitoring data from distributed sensor network nodes and timestamps the real-time monitoring data based on the local clock of the edge computing node. The real-time monitoring data after timestamp alignment is obtained, and the real-time monitoring data is transformed into the component coordinate system of the dynamic fence constraint set. With the sensor network node constraints as a priori, the mapping relationship between the real-time monitoring data and the target components in the three-dimensional dynamic fence is established to obtain the global monitoring data point set. Noise filtering and load reduction processing are performed on the global monitoring data point set to obtain the behavioral feature dataset; Obtain the behavioral feature dataset, extract behavioral features from the dataset, and output the preprocessed behavioral fusion feature set.

[0010] Preferably, the method for noise filtering and load reduction processing of the global monitoring data point set includes: Obtain the global monitoring data point set, perform spatial distribution density statistics on the global monitoring data point set, identify and remove global monitoring data points whose number of neighboring points is lower than the preset neighboring point threshold, and output the global monitoring data point set after outlier removal; The trajectory continuity of the moving target in the global monitoring data point set is verified, and global monitoring data points with inconsistent trajectory continuity are removed to obtain a global monitoring data point set with continuous trajectory. Based on the mapping relationship between real-time monitoring data and target components in the 3D dynamic fence, global monitoring data points that exceed the 3D redundant boundary of the target component are removed from the global monitoring data point set, resulting in a global monitoring data point set after component spatial constraint filtering. The sensor confidence level is determined by the historical calibration accuracy and transmission stability of the sensor network nodes associated with the global monitoring data points. Based on the sensor confidence level and the component screening weight of the target components associated with the global monitoring data points, the filtering confidence weight of the global monitoring data points is calculated, and it is determined whether the filtering confidence weight of the global monitoring data points exceeds the preset filtering confidence threshold. The formula for calculating the filter confidence weight is as follows: In the formula, Indicates the filter confidence weights. These represent the component selection weights and sensor confidence levels, respectively. The confidence adjustment coefficient is... These represent the node's historical calibration normalization accuracy, normalized transmission stability, and normalized environmental adaptability, respectively. These represent the weighting coefficients for the node's historical calibration normalization accuracy, normalized transmission stability, and normalized environmental adaptability, respectively. If the filtering confidence weight of the global monitoring data points exceeds the preset filtering confidence threshold, the global monitoring data points are subjected to mean filtering, and the global monitoring data points after mean filtering are output. If the filter confidence weight of the global monitoring data points does not exceed the preset filter confidence threshold, the global monitoring data points are subjected to extended Kalman filtering, and the global monitoring data points after extended Kalman filtering are output. The filtered global monitoring data point set is obtained, and spatial voxel downsampling is performed on the global monitoring data point set. The uniform voxel point group where the global monitoring data point is located is replaced with the voxel center point to obtain the behavior feature dataset after spatial voxel downsampling.

[0011] Preferably, the method for extracting behavioral features from the behavioral feature dataset includes: Obtain the behavioral feature dataset, extract spatial geometric features from the behavioral feature data in the dataset, construct a geometric feature covariance matrix based on the spatial geometric features of the behavioral feature data, and perform feature decomposition on the geometric feature covariance matrix. The minimum eigenvalue of the geometric feature covariance matrix is ​​used as the behavioral feature normal vector. The mapping relationship of the target components in the 3D dynamic fence associated with the behavioral feature normal vector is used as a priori to perform height-level labeling on the behavioral feature normal vector and output the height-level behavioral feature normal vector. Based on the optical flow estimation network, temporal motion features are extracted from the behavioral feature data in the behavioral feature dataset, and the temporal motion features of the moving target are output. Based on the mapping relationship index of target components in the 3D dynamic fence associated with behavioral feature normal vectors, the component attribute mapping table is queried to obtain the BIM semantic tags associated with behavioral feature normal vectors, the BIM semantic tags within the field of view of the target component are counted, and scene-level semantic context features are generated. After loading the hierarchical behavioral feature normal vectors, temporal motion features, and semantic context features, the behavioral feature normal vectors, temporal motion features, and semantic context features are standardized. Based on the cross-modal attention mechanism, the correlation features between behavioral feature data are calculated. The correlation features are then residually connected with the behavioral feature normal vectors, temporal motion features, and semantic context features to obtain the fused behavioral fusion features. Based on the behavioral fusion features, a behavioral fusion feature set is generated.

[0012] Preferably, the 3D target tracking model uses a deep neural network model as the initial model and further includes an input layer and an output layer. A feature decoder is introduced between the input layer and the deep neural network model, and a 3D target detection head and a target association tracking block are introduced between the deep neural network model and the output layer. The feature decoder includes a multilayer perceptron and a fully connected layer. The multilayer perceptron is used to perform nonlinear transformation on the behavior fusion features and extract high-dimensional hidden layer features of the behavior fusion features. The fully connected layer is used to obtain high-dimensional hidden layer features, perform nonlinear transformation and dimensionality reduction on the high-dimensional hidden layer features, and output the dimensionality-reduced 3D feature tensor. The deep neural network model includes a voxelization layer. The system comprises a 3D sparse convolutional layer and a 3D pooling layer. The voxelization layer is used to obtain a 3D feature tensor, which is then divided into a regular 3D voxel grid. The 3D sparse convolutional layer performs sparse convolution on the regular 3D voxel grid to extract local spatial features. The 3D pooling layer is used to downsample the local spatial features to generate aggregated global features. The 3D target detection head includes an anchor frame generation layer, a target detection layer, a target classification layer, and a confidence prediction layer. The anchor frame generation layer is used to divide the 3D search space based on the spatial distribution density of the aggregated global features in the component coordinate system of the dynamic fence constraint set, using voxelization, and then clustering based on DBSCAN. The analysis method generates 3D candidate boxes for moving targets. The target classification layer is used to classify the existence of targets by aggregating global features of the 3D candidate boxes. 3D bounding box regression analysis is performed on valid targets in the 3D candidate boxes, outputting the 3D center coordinates, target size, target category, and motion descriptor of the moving target analysis results. Based on the confidence prediction layer, the overall confidence of the moving target analysis results is comprehensively predicted, and moving target analysis results exceeding a preset confidence threshold are filtered out. The target association tracking block includes a constraint index layer, a geometric fence judgment layer, and an attribute verification layer. The constraint index layer is used to obtain the moving target analysis results and index the association based on the moving target analysis results. The three-dimensional dynamic fence is constructed and its coordinates are aligned with the coordinate system of the component in the dynamic fence constraint set in the three-dimensional dynamic fence to determine the associated target components in the three-dimensional dynamic fence. The geometric fence judgment layer is used to project the moving target analysis results onto the three-dimensional dynamic fence of the associated target components, perform spatial occupation judgment on the moving target analysis results, and generate a fence occupation voxel set. The attribute verification layer is used to traverse the dynamic fence constraint set of the three-dimensional dynamic fence, generate whether the boundary is crossed, the boundary type identification result, and the associated target component attributes, and output the three-dimensional fence boundary identification result containing the whether the boundary is crossed, the boundary type identification result, and the associated target component attributes.

[0013] Preferably, the method for 3D target detection based on a behavior fusion feature set using a 3D target tracking model includes: The behavior fusion features are subjected to nonlinear transformation to extract high-dimensional hidden layer features. The high-dimensional hidden layer features are then subjected to nonlinear transformation and dimensionality reduction to output the dimensionality-reduced three-dimensional feature tensor. Obtain a three-dimensional feature tensor, divide the three-dimensional feature tensor into a regular three-dimensional voxel grid, and perform sparse convolution processing on the regular three-dimensional voxel grid through a three-dimensional sparse convolution layer to extract local spatial features. Then, downsample the local spatial features to generate aggregated global features. The aggregated global features are represented as follows: These represent aggregated global features and three-dimensional feature tensors, respectively. , , These are respectively 3D pooling processing, 3D sparse convolution processing, and voxelization processing; Based on the spatial distribution density of aggregated global features in the component coordinate system of the dynamic fence constraint set, the three-dimensional search space is divided by voxelization, and three-dimensional candidate boxes of moving targets are generated based on the DBSCAN clustering analysis method. The aggregated global features of the three-dimensional candidate boxes are used to classify the existence of targets, and three-dimensional bounding box regression analysis is performed on the effective targets of the three-dimensional candidate boxes. The three-dimensional center coordinates, target size, target category and motion descriptor of the moving target analysis results are output. Based on the overall confidence level of the predicted moving target analysis results, moving target analysis results exceeding a preset confidence threshold are filtered out. The formula for calculating the overall confidence level is as follows: in, These represent the overall confidence level, the existence classification confidence level, the location confidence level, the tracking consistency confidence level, and the motion rationality confidence level, respectively. These are the location confidence weight index, the tracking consistency confidence weight index, and the motion rationality confidence weight index, respectively. Represents the component probability distribution. These are the classification weight matrix, the target size coverage feature vector, and the component bias vector, respectively. Obtain the moving target analysis results, index the associated 3D dynamic fence based on the moving target analysis results, and align the coordinates with the component coordinate system of the dynamic fence constraint set in the 3D dynamic fence to determine the associated target components in the 3D dynamic fence. The moving target analysis results are projected onto the three-dimensional dynamic fence of the associated target components, and the spatial occupancy of the moving target analysis results is determined to generate a set of fence occupancy voxels. Traverse the dynamic fence constraint set of the 3D dynamic fence, generate boundary violation verification results, boundary violation type identification results, and associated target component attributes, and output the 3D fence boundary violation identification results containing the boundary violation verification results, boundary violation type identification results, and associated target component attributes.

[0014] On the other hand, the present invention also provides a three-dimensional fence boundary crossing recognition system based on building information modeling, the three-dimensional fence boundary crossing recognition system based on building information modeling comprising: The dynamic fence generation module is used to acquire the building information model of the target construction site, import the building information model into the edge computing node, parse and identify the building information model based on the edge computing node to obtain the target component set, and generate a three-dimensional dynamic fence by combining the target component set. The edge data processing module collects real-time monitoring data from the target construction site based on distributed sensor network nodes, performs edge-side preprocessing on the real-time monitoring data, and outputs a preprocessed behavior fusion feature set. The boundary crossing detection module is used to load the behavior fusion feature set, pre-build a deep learning-based 3D target tracking model, perform 3D target detection on the behavior fusion feature set based on the 3D target tracking model, and output the 3D fence boundary crossing detection result.

[0015] Preferably, the edge data processing module includes: The timestamp alignment unit is used to receive asynchronous real-time monitoring data from distributed sensor network nodes and align the real-time monitoring data with the local clock of the edge computing node as a reference. The coordinate transformation unit is used to acquire real-time monitoring data after timestamp alignment, transform the real-time monitoring data to the component coordinate system of the dynamic fence constraint set, and establish the mapping relationship between the real-time monitoring data and the target components in the three-dimensional dynamic fence with the sensor network node constraints as a priori, so as to obtain the global monitoring data point set. The noise filtering unit is used to filter out noise and reduce load on the global monitoring data point set to obtain the behavioral feature dataset. The behavior feature recognition unit is used to acquire a behavior feature dataset, extract behavior features from the behavior feature dataset, and output a preprocessed behavior fusion feature set.

[0016] Compared with the prior art, the embodiments of this application have the following main advantages: In this embodiment of the invention, a three-dimensional dynamic fence is constructed by importing the Building Information Model (BIM) into edge computing nodes for parsing and identification. This is achieved by combining component type attributes, geometric coordinate information, and multi-dimensional indicators such as the number of adjacent components, spatial influence range, process criticality, and hidden risk attributes for order relationship analysis and weighted filtering. This constructs a fence containing height layering constraints, horizontal projection constraints, component association constraints, and sensor network node constraints. This achieves deep integration between the fence boundary and BIM components and dynamic adaptation with construction progress, solving the technical problems of existing two-dimensional fences lacking spatial depth information, being unable to adapt to complex three-dimensional construction environments, and having statically fixed fence boundaries that are disconnected from the actual construction status. Furthermore, it is based on deep learning. The 3D target tracking model extracts and aggregates global features through a 3D sparse convolutional network to generate 3D candidate boxes. It then combines comprehensive confidence prediction to filter valid targets. Finally, it integrates and verifies the target with a 3D dynamic fence through a constraint index layer, a geometric fence judgment layer, and an attribute verification layer. The output includes a structured boundary violation verification result, a boundary violation type identification result, and related target component attributes. This not only achieves a leap from 2D plane boundary violation detection to 3D spatial occupation determination, but also enables accurate detection and spatial occupation determination of moving targets in complex 3D space. As a result, it significantly improves the intelligence level and risk prevention and control capabilities of construction site safety management, ensuring the accuracy of construction site safety management.

[0017] In this embodiment of the invention, when analyzing and identifying building information models based on edge computing nodes, the target components are first rapidly screened according to preset hazardous area screening rules, effectively narrowing the data range for subsequent detailed analysis and reducing computational load. Then, through a second traversal, multi-dimensional semantic information such as the number of adjacent components, spatial influence range, process criticality, and hidden risk attributes of the components are deeply mined, and weights are assigned using the order relation analysis method to obtain component screening weights that comprehensively reflect the safety risk level of the components, realizing a leap from single geometric boundary identification to multi-attribute fusion intelligent screening. Finally, a threshold judgment mechanism is used to accurately lock the set of target components with high safety risks. This not only makes full use of the low latency and high real-time performance advantages of edge computing nodes, avoiding network transmission bottlenecks and privacy leakage risks caused by centralized cloud processing, but also significantly improves the targeting and accuracy of 3D dynamic fence construction through component-level fine screening, so that the fence boundary can closely fit the actual distribution of hazardous sources on the construction site.

[0018] In this embodiment of the invention, the three-dimensional dynamic fence overcomes the rigidity of traditional electronic fence boundaries by constructing a spatial composition matrix of the target components and introducing spatial expansion redundancy processing based on spatial influence, process criticality, and hidden risk. This achieves adaptive wrapping of the geometric contours of complex building components and intelligent reservation of dangerous buffer zones. Furthermore, by performing Boolean union operations on the three-dimensional edge regions through edge computing nodes, overlapping redundancy at the boundaries of dangerous areas of multiple components is effectively eliminated. Combined with vertical stretching and height fusion to generate a merged three-dimensional fence for the components, the fragmentation of spatial computing and resource overhead are significantly reduced. Finally, by solidifying the component attribute mapping table and spatial topology relationship into a dynamic fence constraint set containing height layering, horizontal projection, component association, and sensor node constraints and embedding it into geometric entities, the fence not only has accurate three-dimensional spatial definition capabilities but also carries rich BIM semantics and engineering logic, thereby enabling dynamic adjustment and closed-loop verification according to construction progress and site conditions.

[0019] In this embodiment of the invention, noise filtering and load reduction processing of the global monitoring data point set achieves multi-dimensional and multi-level refined data cleaning. Furthermore, a filtering confidence weight calculation mechanism that integrates sensor confidence and component screening weights is introduced. Based on data quality, a hierarchical processing strategy of mean filtering or extended Kalman filtering is dynamically selected, ensuring that high-confidence data retains more original details and low-confidence data receives stronger noise suppression. This effectively balances the contradictory relationship between data fidelity and denoising intensity. Simultaneously, spatial voxel downsampling processing replaces dense point clusters with voxel center points, significantly reducing the computational load of subsequent feature extraction and 3D target tracking. Finally, noise filtering and load reduction processing of the global monitoring data point set fully utilizes the real-time processing capabilities of edge computing nodes, shifting the burden of centralized cloud preprocessing forward. This not only significantly improves the timeliness of data processing and system response speed but also ensures accurate matching of data quality and component risk levels through differentiated filtering strategies based on BIM semantic association.

[0020] In this embodiment of the invention, the 3D target tracking model adopts a modular architecture design. Using a deep neural network model as the initial model, it performs nonlinear transformation and dimensionality reduction on the behavior fusion features through a multilayer perceptron and fully connected layers in the feature decoder, outputting a 3D feature tensor adapted to 3D spatial processing. This effectively solves the dimensionality adaptation problem between high-dimensional behavior features and the 3D detection network. Furthermore, through the cascaded operation of voxelization layers, 3D sparse convolutional layers, and 3D pooling layers, sparse convolution processing is performed on a regular 3D voxel grid, and downsampling is used to generate aggregated global features. This fully utilizes the computational optimization advantages of sparse convolution for non-empty voxels, significantly reducing computational redundancy and memory consumption in 3D point cloud processing, and improving edge computing performance. The system leverages the real-time processing capabilities of the nodes. The 3D target detection head, based on DBSCAN clustering analysis, generates 3D candidate boxes in the component coordinate system of the dynamic fence constraint set. It then filters high-confidence moving target analysis results through target classification, bounding box regression, and comprehensive confidence prediction, achieving accurate target positioning and reliable detection from feature space to 3D physical space. The target association tracking block, through a layered verification mechanism of constraint index layer, geometric fence judgment layer, and attribute verification layer, aligns the detection results with the 3D dynamic fence in coordinates, determines spatial occupancy, and traverses the dynamic fence constraint set. The final output includes structured boundary violation verification results, boundary violation type identification results, and associated target component attributes. The 3D target tracking model not only achieves deep fusion verification of 3D target detection and 3D dynamic fences but also, through a detection and verification mechanism in the BIM component coordinate system, ensures that boundary violation judgment results possess clear spatial attribution and semantic association. This effectively solves the technical defects of traditional 2D target tracking models, such as lack of 3D spatial perception, inability to semantically couple with the building information model, and low structured boundary violation warning information. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the implementation process of the 3D fence boundary crossing recognition method based on building information modeling provided by the present invention.

[0022] Figure 2 A schematic diagram illustrating the implementation process of a method for parsing and recognizing building information models based on edge computing nodes is shown.

[0023] Figure 3 A schematic diagram illustrating the process of generating a 3D dynamic fence by combining a set of target components is shown.

[0024] Figure 4 A schematic diagram illustrating the implementation process of edge-side preprocessing for real-time monitoring data is shown.

[0025] Figure 5 A schematic diagram of the process for noise filtering and load reduction of a global monitoring data point set is shown.

[0026] Figure 6 The diagram illustrates the implementation process of a method for extracting behavioral features from a behavioral feature dataset.

[0027] Figure 7 A schematic diagram of the implementation process of a method for 3D target detection based on behavior fusion feature set using a 3D target tracking model is shown.

[0028] Figure 8 A schematic diagram of a 3D fence crossing recognition system based on building information modeling is shown. Detailed Implementation

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0030] Existing methods for fence boundaries lack three-dimensional spatial information, making them unsuitable for complex three-dimensional construction environments. This leads to a significant decrease in boundary violation recognition accuracy in complex three-dimensional environments. Furthermore, boundary violation judgment lacks a verification mechanism that correlates with BIM semantics, resulting in low-structured warning information that is not linked to component attributes in the Building Information Model (BIM) and cannot output specific component information related to the boundary violation. To address these issues, we propose a three-dimensional fence boundary violation recognition system and method based on the BIM model. In this embodiment, the BIM model is imported into edge computing nodes for parsing and recognition. Combining component type attributes, geometric coordinate information, and multi-dimensional indicators such as the number of adjacent components, spatial influence range, process criticality, and hidden risk attributes, order relationship analysis and weighted filtering are performed to construct a three-dimensional dynamic fence containing height layer constraints, horizontal projection constraints, component association constraints, and sensor network node constraints. This achieves deep integration of the fence boundary with BIM components and dynamic adaptation with construction progress, solving the technical problems of existing two-dimensional fences lacking spatial depth information, being unsuitable for complex three-dimensional construction environments, and having statically fixed fence boundaries that are disconnected from the actual construction status. Moreover, it is based on deep learning. The 3D target tracking model extracts and aggregates global features through a 3D sparse convolutional network to generate 3D candidate boxes. It then combines comprehensive confidence prediction to filter valid targets. Finally, it integrates and verifies the target with a 3D dynamic fence through a constraint index layer, a geometric fence judgment layer, and an attribute verification layer. The output includes a structured boundary violation verification result, a boundary violation type identification result, and related target component attributes. This not only achieves a leap from 2D plane boundary violation detection to 3D spatial occupation determination, but also enables accurate detection and spatial occupation determination of moving targets in complex 3D space. As a result, it significantly improves the intelligence level and risk prevention and control capabilities of construction site safety management, ensuring the accuracy of construction site safety management.

[0031] This invention provides a three-dimensional fence boundary crossing recognition method based on Building Information Modeling (BIM). Figure 1 A schematic diagram illustrating the implementation process of a 3D fence boundary crossing detection method based on Building Information Modeling (BIM) is shown. The 3D fence boundary crossing detection method based on BIM specifically includes: S10: Obtain the building information model of the target construction site, import the building information model into the edge computing node, parse and identify the building information model based on the edge computing node to obtain the target component set, and generate a three-dimensional dynamic fence by combining the target component set. S20: Based on distributed sensor network nodes, real-time monitoring data is collected in the target construction site. The real-time monitoring data is preprocessed at the edge and outputs the preprocessed behavior fusion feature set. S30: Load the behavior fusion feature set, pre-build a deep learning-based 3D target tracking model, perform 3D target detection on the behavior fusion feature set based on the 3D target tracking model, and output the 3D fence crossing recognition result.

[0032] In this embodiment of the invention, the target construction site can be a deep foundation pit operation area for building construction projects, an exterior curtain wall installation area for high-rise buildings, a steel structure hoisting area for large industrial plants, a confined space operation section for urban integrated pipe corridors, or a high-risk area for heavy equipment installation in nuclear power / chemical projects. These scenarios generally have characteristics such as three-dimensional cross-operations, numerous openings near edges, dense large machinery, and dynamic changes in dangerous areas, which urgently require high-precision three-dimensional spatial safety management and control. The distributed sensor network nodes are mainly composed of UWB ultra-wideband positioning base stations, millimeter-wave radar, depth vision cameras, smart wearable devices equipped with IMUs, laser LiDAR scanners, and environmental sensing terminals. They are connected to edge computing nodes through wired or wireless means to form a multi-source heterogeneous perception system covering the entire area, which is used to collect real-time three-dimensional position, motion trajectory, and attitude data of personnel and machinery.

[0033] In this embodiment of the invention, a three-dimensional dynamic fence is constructed by importing the Building Information Model (BIM) into edge computing nodes for parsing and identification. This is achieved by combining component type attributes, geometric coordinate information, and multi-dimensional indicators such as the number of adjacent components, spatial influence range, process criticality, and hidden risk attributes for order relationship analysis and weighted filtering. This constructs a fence containing height layering constraints, horizontal projection constraints, component association constraints, and sensor network node constraints. This achieves deep integration between the fence boundary and BIM components and dynamic adaptation with construction progress, solving the technical problems of existing two-dimensional fences lacking spatial depth information, being unable to adapt to complex three-dimensional construction environments, and having statically fixed fence boundaries that are disconnected from the actual construction status. Furthermore, it is based on deep learning. The 3D target tracking model extracts and aggregates global features through a 3D sparse convolutional network to generate 3D candidate boxes. It then combines comprehensive confidence prediction to filter valid targets. Finally, it integrates and verifies the target with a 3D dynamic fence through a constraint index layer, a geometric fence judgment layer, and an attribute verification layer. The output includes a structured boundary violation verification result, a boundary violation type identification result, and related target component attributes. This not only achieves a leap from 2D plane boundary violation detection to 3D spatial occupation determination, but also enables accurate detection and spatial occupation determination of moving targets in complex 3D space. As a result, it significantly improves the intelligence level and risk prevention and control capabilities of construction site safety management, ensuring the accuracy of construction site safety management.

[0034] This invention provides a method for parsing and recognizing building information models based on edge computing nodes. Figure 2 This diagram illustrates the implementation flow of a method for parsing and recognizing Building Information Models (BIM) based on edge computing nodes. The method specifically includes: S101, Load the target construction site building information model (BIM), identify and analyze the target components in the target construction site building information model, read the type attributes and geometric coordinate information of the target components, combine the type attributes and geometric coordinate information of the target components, and perform preliminary screening of the target components based on the preset dangerous area screening rules to generate a preliminary screening set of components. Here, the target components refer to the building element objects that are directly or indirectly related to the potential dangerous areas and are selected from the target construction site building information model (BIM). Target components include, but are not limited to, beams, slabs, columns, equipment foundations, openings, foundation pit support structures, etc. The target construction site building information model (BIM) refers to a digital three-dimensional model created for a specific engineering project that is under construction or about to be constructed, containing geometric and non-geometric information of the entire life cycle of the project. S102 involves a second traversal of the target component in the building information model of the target construction site. This process identifies the number of adjacent components associated with the target component, its spatial influence range, process criticality, and hidden risk attributes. The system then normalizes these attributes to obtain the degree of association between adjacent components, spatial influence range, process criticality, and hidden risk. An ordinal relation analysis method is used to assign weights to these weights, resulting in a component screening weight. This component screening weight transforms qualitative attributes that are difficult to compare directly into quantifiable values. The system determines whether to include a component in the final monitoring scope based on whether its screening weight exceeds a preset threshold, thus achieving intelligent, data-driven risk prediction. The formula for calculating the component screening weight is as follows: , , , in, Indicates the component selection weight. These represent the degree of association and adjacency of components, the degree of spatial influence, the degree of process criticality, and the degree of hidden risk, respectively. These are the importance weights for associated adjacent components, spatial influence, process criticality, and hidden risks, respectively. This indicates the number of adjacent components associated with the current target component and the maximum number of adjacent components associated with it in the building information model of the target construction site. These represent the volume of spatial influence of the current target component and the maximum volume of spatial influence of the component in the model, respectively. These represent the process criticality score of the current target component and the maximum process criticality score of the component in the model, respectively. These represent the current target component's hidden risk score and the maximum hidden risk score of the component in the model, respectively. S103, determine whether the component screening weight of the target component exceeds the preset screening threshold; S104. If the component screening weight of the target component exceeds the preset screening threshold, the target component is merged into the component initial screening set to obtain the target component set.

[0035] In this embodiment of the invention, when analyzing and identifying building information models based on edge computing nodes, the target components are first rapidly screened according to preset hazardous area screening rules, effectively narrowing the data range for subsequent detailed analysis and reducing computational load. Then, through a second traversal, multi-dimensional semantic information such as the number of adjacent components, spatial influence range, process criticality, and hidden risk attributes of the components are deeply mined, and weights are assigned using the order relation analysis method to obtain component screening weights that comprehensively reflect the safety risk level of the components, realizing a leap from single geometric boundary identification to multi-attribute fusion intelligent screening. Finally, a threshold judgment mechanism is used to accurately lock the set of target components with high safety risks. This not only makes full use of the low latency and high real-time performance advantages of edge computing nodes, avoiding network transmission bottlenecks and privacy leakage risks caused by centralized cloud processing, but also significantly improves the targeting and accuracy of 3D dynamic fence construction through component-level fine screening, so that the fence boundary can closely fit the actual distribution of hazardous sources on the construction site.

[0036] This invention provides a method for generating a three-dimensional dynamic fence by combining a set of target components. Figure 3 This diagram illustrates the implementation flow of a method for generating a 3D dynamic fence by combining a set of target components. The method specifically includes: S201, Load the target component set, and generate the target component spatial composition matrix based on the geometric coordinate information and type attributes of the target components in the target component set; S202, generate the three-dimensional initial boundary of the target component based on the spatial composition matrix of the target component, and identify the spatial influence degree, process criticality degree, and hidden risk degree of the target component. Based on the normalized spatial influence degree, process criticality degree, and hidden risk degree of the target component, perform spatial expansion redundancy processing on the three-dimensional initial boundary of the target component to generate the three-dimensional redundant boundary of the target component. S203, obtain the three-dimensional redundant boundary of the target component, perform a Boolean union operation on the three-dimensional redundant boundary of the target component based on the edge computing node, eliminate the overlapping boundary of the three-dimensional redundant boundary of the target component, and obtain the three-dimensional redundant boundary of the target component after the edge overlapping region is eliminated. S204, after the target component's three-dimensional redundant boundary is eliminated by loading edge overlap area, the target component's three-dimensional redundant boundary is stretched vertically and height-blended to obtain the merged component's three-dimensional fence. S205, obtain the merged component 3D fence, establish a component attribute mapping table for the component 3D fence, and establish the spatial topological relationship between the target components. Based on the component attribute mapping table and the spatial topological relationship, construct a dynamic fence constraint set (wherein, the dynamic fence constraint set includes height layer constraint, horizontal projection constraint, component association constraint, and sensor network node constraint). Embed the dynamic fence constraint set into the component 3D fence to obtain the 3D dynamic fence. The dynamic fence constraint set refers to a comprehensive rule system (or database) built by integrating the component attribute mapping table and the spatial topological relationship between target components after generating the merged 3D component fence. It endows the geometric fence with engineering semantics and logical judgment capabilities and is the core rule base for realizing intelligent boundary crossing identification. Among them, the height layer constraint is a Z-axis interval rule defined by the floor elevation of the BIM model and the vertical geometric information of the components. It is used to accurately distinguish dangerous areas of different construction floors or height levels and avoid misjudgment caused by three-dimensional cross-operations. The horizontal projection constraint is a two-dimensional closed boundary generated based on the geometric contour of the target component on the XY plane. It serves as the basis for judging whether the target has invaded the dangerous area on the horizontal plane. The component association constraint is a topological logic established based on the adjacency, load-bearing or process dependency relationship between components in the BIM model. It is used to trace related structural entities or determine the responsibility of composite dangerous areas when boundary crossing occurs. The sensor network node constraint is a confidence rule formed by combining the calibration accuracy, transmission stability and environmental adaptability of distributed sensors. It is used to perform weighted screening and coordinate registration of the reliability of monitoring data in the data preprocessing and fusion verification stages.

[0037] In this embodiment of the invention, the three-dimensional dynamic fence overcomes the rigidity of traditional electronic fence boundaries by constructing a spatial composition matrix of the target components and introducing spatial expansion redundancy processing based on spatial influence, process criticality, and hidden risk. This achieves adaptive wrapping of the geometric contours of complex building components and intelligent reservation of dangerous buffer zones. Furthermore, by performing Boolean union operations on the three-dimensional edge regions through edge computing nodes, overlapping redundancy at the boundaries of dangerous areas of multiple components is effectively eliminated. Combined with vertical stretching and height fusion to generate a merged three-dimensional fence for the components, the fragmentation of spatial computing and resource overhead are significantly reduced. Finally, by solidifying the component attribute mapping table and spatial topology relationship into a dynamic fence constraint set containing height layering, horizontal projection, component association, and sensor node constraints and embedding it into geometric entities, the fence not only has accurate three-dimensional spatial definition capabilities but also carries rich BIM semantics and engineering logic, thereby enabling dynamic adjustment and closed-loop verification according to construction progress and site conditions.

[0038] This invention provides a method for edge-side preprocessing of real-time monitoring data. Figure 4This diagram illustrates a method for edge-side preprocessing of real-time monitoring data. The method specifically includes: S301 receives asynchronous real-time monitoring data from distributed sensor network nodes and timestamps the real-time monitoring data based on the local clock of the edge computing node. The real-time monitoring data is collected through UWB positioning base stations, millimeter-wave radar, depth vision cameras and smart wearable devices equipped with IMU. The data types of the real-time monitoring data include, but are not limited to, spatiotemporal location data, point cloud set feature data, kinematic inertial data, visual semantics and optical flow data. S302, acquire the real-time monitoring data after timestamp alignment, convert the real-time monitoring data to the component coordinate system of the dynamic fence constraint set, and establish the mapping relationship between the real-time monitoring data and the target components in the three-dimensional dynamic fence with the sensor network node constraints as a priori, so as to obtain the global monitoring data point set; S303 performs noise filtering and load reduction processing on the global monitoring data point set to obtain a behavioral feature dataset; S304, Obtain the behavioral feature dataset, extract behavioral features from the behavioral feature dataset, and output the preprocessed behavioral fusion feature set.

[0039] In this embodiment of the invention, when performing edge-side preprocessing on real-time monitoring data, the asynchronous data of distributed sensor network nodes are first timestamped based on the local clock of the edge computing node, effectively solving the problem of temporal disorder caused by differences in acquisition frequency and latency of multi-source heterogeneous data. Then, by converting the data to the component coordinate system of the dynamic fence constraint set and establishing a mapping relationship with the target component, accurate spatial registration of physical space monitoring data and BIM digital model is achieved, ensuring the engineering semantic relevance of subsequent analysis. Subsequently, by performing noise filtering and load reduction processing on the global monitoring data point set, outliers and redundant information are eliminated, significantly reducing data transmission bandwidth and cloud computing pressure. Finally, through in-depth behavioral feature extraction and fusion, the original, messy sensor data is transformed into a high signal-to-noise ratio behavioral fusion feature set, thereby completing a high-quality transformation from raw signals to decisionable features at the edge.

[0040] This invention provides a method for noise filtering and load reduction processing of a global monitoring data point set. Figure 5 This diagram illustrates the implementation flow of a method for noise filtering and load reduction of a global monitoring data point set. The method specifically includes: S3031, acquire the global monitoring data point set, perform spatial distribution density statistics on the global monitoring data point set, identify and remove global monitoring data points whose number of neighboring points is lower than the preset neighboring point threshold, and output the global monitoring data point set after outlier removal. By counting the number of points in the spatial neighborhood of each monitoring point and removing isolated points below the preset threshold, random noise and false targets caused by sensor multipath effect, electromagnetic interference or false detection are effectively filtered out. S3032, Perform trajectory continuity verification on the moving target in the global monitoring data point set, remove global monitoring data points with inconsistent trajectory continuity, and obtain a global monitoring data point set with continuous trajectory. By performing continuity logic judgment on the trajectory of the moving target, data points that do not conform to the laws of physical motion are removed. S3033, based on the mapping relationship between real-time monitoring data and target components in the 3D dynamic fence, removes global monitoring data points that exceed the 3D redundant boundary of the target component from the global monitoring data point set, and obtains the global monitoring data point set after component spatial constraint filtering. Among them, the engineering semantic constraints of the BIM model are introduced, which can effectively shield useless data in non-interested areas, greatly reduce the processing range of subsequent algorithms, and realize the focusing of the region of interest (ROI) based on the engineering scene. S3034 determines the sensor confidence level based on the historical calibration accuracy and transmission stability of sensor network nodes associated with global monitoring data points. Based on the sensor confidence level and the component selection weights of the target components associated with the global monitoring data points, a filtering confidence weight is calculated for the transmitted global monitoring data points. This filtering confidence weight combines the historical calibration accuracy and transmission stability of the sensor nodes with the selection weights of the associated target components to calculate a comprehensive filtering confidence weight. This overcomes the dependence of traditional filtering on a single data source, enabling quantitative evaluation and hierarchical processing of data quality, and providing a scientific basis for the selection of subsequent differentiated filtering strategies. S3035, determine whether the filtering confidence weight of the global monitoring data points exceeds the preset filtering confidence threshold; The formula for calculating the filter confidence weight is as follows: In the formula, Indicates the filter confidence weights. These represent the component selection weights and sensor confidence levels, respectively. The confidence adjustment coefficient is... These represent the node's historical calibration normalization accuracy, normalized transmission stability, and normalized environmental adaptability, respectively. These represent the weighting coefficients for the node's historical calibration normalization accuracy, normalized transmission stability, and normalized environmental adaptability, respectively. S3036, If the filtering confidence weight of the global monitoring data points exceeds the preset filtering confidence threshold, the global monitoring data points are subjected to mean filtering and the mean-filtered global monitoring data points are output. S3037, if the filter confidence weight of the global monitoring data points does not exceed the preset filter confidence threshold, extended Kalman filtering is applied to the global monitoring data points, and the global monitoring data points after extended Kalman filtering are output. Based on the comparison result between the filter confidence weight and the preset threshold, mean filtering or extended Kalman filtering is intelligently selected: for high-confidence data, the computationally simple mean filtering is used to reduce the load; for low-confidence, high-uncertainty data, EKF is used for optimal estimation and noise suppression. This adaptive filtering mechanism optimizes the allocation of computing resources of edge computing nodes to the maximum extent while ensuring positioning accuracy. S3038: Obtain the filtered global monitoring data point set, and perform spatial voxel downsampling on the global monitoring data point set. Replace the uniform voxel clusters containing the global monitoring data points with voxel center points to obtain the behavior feature dataset after spatial voxel downsampling. Specifically, by using voxelized mesh processing and replacing dense point clusters with voxel center points, exponential compression of data volume is achieved while preserving the target's geometric contour and spatial distribution features. This not only significantly reduces the computational complexity of subsequent behavior feature extraction and 3D target detection, but also enables the system to transmit more effective information within limited bandwidth.

[0041] In this embodiment of the invention, noise filtering and load reduction processing of the global monitoring data point set achieves multi-dimensional and multi-level refined data cleaning. Furthermore, a filtering confidence weight calculation mechanism that integrates sensor confidence and component screening weights is introduced. Based on data quality, a hierarchical processing strategy of mean filtering or extended Kalman filtering is dynamically selected, ensuring that high-confidence data retains more original details and low-confidence data receives stronger noise suppression. This effectively balances the contradictory relationship between data fidelity and denoising intensity. Simultaneously, spatial voxel downsampling processing replaces dense point clusters with voxel center points, significantly reducing the computational load of subsequent feature extraction and 3D target tracking. Finally, noise filtering and load reduction processing of the global monitoring data point set fully utilizes the real-time processing capabilities of edge computing nodes, shifting the burden of centralized cloud preprocessing forward. This not only significantly improves the timeliness of data processing and system response speed but also ensures accurate matching of data quality and component risk levels through differentiated filtering strategies based on BIM semantic association.

[0042] This invention provides a method for extracting behavioral features from a behavioral feature dataset. Figure 6This diagram illustrates the implementation flow of a method for extracting behavioral features from a behavioral feature dataset. The method specifically includes: S3041, Obtain the behavioral feature dataset, extract spatial geometric features from the behavioral feature data in the behavioral feature dataset, construct a geometric feature covariance matrix based on the spatial geometric features of the behavioral feature data, and perform feature decomposition on the geometric feature covariance matrix. S3042 uses the minimum eigenvalue of the geometric feature covariance matrix as the behavioral feature normal vector, and takes the mapping relationship of the target components in the 3D dynamic fence associated with the behavioral feature normal vector as a priori to perform height-layer labeling on the behavioral feature normal vector, and outputs the height-layered behavioral feature normal vector. Among them, the behavioral feature normal vector corresponding to the minimum eigenvalue accurately describes the spatial surface orientation and local geometric structure of the monitoring data. At the same time, the height-layer labeling is performed with the BIM component mapping relationship as a priori, so that the geometric features are tightly coupled with the height-layer constraints of the 3D dynamic fence, realizing the leap from two-dimensional pixel description to three-dimensional spatial semantic description. S3043, based on the optical flow estimation network, extracts temporal motion features from the behavioral feature dataset and outputs the temporal motion features of the moving target; S3044, based on the mapping relationship index query of target components in the 3D dynamic fence associated with behavioral feature normal vectors, the component attribute mapping table is queried to obtain the BIM semantic tags associated with behavioral feature normal vectors, the BIM semantic tags within the field of view of the target components are counted, and scene-level semantic context features are generated. S3045 loads the hierarchical behavioral feature normal vectors, temporal motion features, and semantic context features. It standardizes the behavioral feature normal vectors, temporal motion features, and semantic context features, and calculates the correlation features between behavioral feature data based on the cross-modal attention mechanism. It performs residual connection between the correlation features and the behavioral feature normal vectors, temporal motion features, and semantic context features to obtain the fused behavioral fusion features. It generates a behavioral fusion feature set based on the behavioral fusion features.

[0043] In this embodiment of the invention, when extracting behavioral features from the behavioral feature dataset, not only is the low latency advantage of edge computing nodes fully utilized to achieve real-time processing of feature extraction, but also a multimodal feature fusion strategy guided by BIM semantic priors is used to enable behavioral features to have clear three-dimensional spatial attribution and component semantic association. This significantly improves the understanding depth and discrimination accuracy of the subsequent 3D target tracking model for personnel behavior in the construction scene, and provides a high-dimensional, strong semantic, and spatiotemporally consistent feature expression foundation for three-dimensional fence boundary crossing recognition. This effectively solves the technical defects of traditional feature extraction methods that lack spatial depth perception and architectural semantic fusion.

[0044] In this embodiment of the invention, the 3D target tracking model uses a deep neural network model as the initial model and further includes an input layer and an output layer. A feature decoder is introduced between the input layer and the deep neural network model, and a 3D target detection head and a target association tracking block are introduced between the deep neural network model and the output layer. The feature decoder includes a multilayer perceptron and a fully connected layer. The multilayer perceptron is used to perform nonlinear transformation on the behavior fusion features and extract high-dimensional hidden layer features of the behavior fusion features. The fully connected layer is used to obtain high-dimensional hidden layer features, perform nonlinear transformation and dimensionality reduction on the high-dimensional hidden layer features, and output the dimensionality-reduced 3D feature tensor. The deep neural network model includes a volumetric model. The system comprises a voxelization layer, a 3D sparse convolutional layer, and a 3D pooling layer. The voxelization layer is used to obtain a 3D feature tensor, which is then divided into a regular 3D voxel grid. The 3D sparse convolutional layer performs sparse convolution on this grid to extract local spatial features. The 3D pooling layer downsamples these local features to generate aggregated global features. The 3D target detection head includes an anchor frame generation layer, a target detection layer, a target classification layer, and a confidence prediction layer. The anchor frame generation layer, based on the spatial distribution density of the aggregated global features in the component coordinate system of the dynamic fence constraint set, uses voxelization to divide the 3D search space and, based on DBSCAN... Clustering analysis generates 3D candidate boxes for moving targets. The target classification layer classifies the target presence based on the aggregated global features of the 3D candidate boxes. 3D bounding box regression analysis is performed on valid targets within the 3D candidate boxes, outputting the 3D center coordinates, target size, target category, and motion descriptor of the moving target analysis results. A confidence prediction layer comprehensively predicts the overall confidence of the moving target analysis results, filtering out moving target analysis results exceeding a preset confidence threshold. The target association tracking block includes a constraint index layer, a geometric fence judgment layer, and an attribute verification layer. The constraint index layer is used to obtain the moving target analysis results and index the relevant parameters based on the moving target analysis results. The system constructs a 3D dynamic fence and aligns it with the coordinate system of the component in the dynamic fence constraint set to determine the associated target components in the 3D dynamic fence. The geometric fence judgment layer projects the moving target analysis results onto the 3D dynamic fence of the associated target components, performs spatial occupancy determination on the moving target analysis results, and generates a fence occupancy voxel set. The attribute verification layer traverses the dynamic fence constraint set of the 3D dynamic fence, generates boundary violation verification results, boundary violation type identification results, and associated target component attributes, and outputs a 3D fence boundary violation identification result containing boundary violation verification results, boundary violation type identification results, and associated target component attributes.

[0045] This invention provides a method for 3D target detection based on a behavior fusion feature set using a 3D target tracking model. Figure 7This diagram illustrates the implementation flow of a method for 3D object detection based on a 3D object tracking model and a behavior fusion feature set. The method specifically includes: S401 performs a nonlinear transformation on the behavior fusion features, extracts the high-dimensional hidden layer features of the behavior fusion features, obtains the high-dimensional hidden layer features, performs a nonlinear transformation and dimensionality reduction on the high-dimensional hidden layer features, and outputs the dimensionality-reduced three-dimensional feature tensor. S402, obtain the three-dimensional feature tensor, divide the three-dimensional feature tensor into a regular three-dimensional voxel grid, and perform sparse convolution processing on the regular three-dimensional voxel grid through a three-dimensional sparse convolution layer to extract local spatial features, downsample the local spatial features, and generate aggregated global features. The aggregated global features are represented as follows: These represent aggregated global features and three-dimensional feature tensors, respectively. , , These are respectively 3D pooling processing, 3D sparse convolution processing, and voxelization processing; S403, based on the spatial distribution density of aggregated global features in the component coordinate system of the dynamic fence constraint set, uses voxelization to divide the three-dimensional search space, and generates three-dimensional candidate boxes of moving targets based on DBSCAN clustering analysis. It classifies the existence of targets by aggregated global features of the three-dimensional candidate boxes, performs three-dimensional bounding box regression analysis on the effective targets of the three-dimensional candidate boxes, and outputs the three-dimensional center coordinates, target size, target category and motion descriptor of the moving target analysis results. S404, comprehensively predict the confidence level of moving target analysis results and filter moving target analysis results that exceed the preset confidence threshold; The formula for calculating the overall confidence level is as follows: in, These represent the overall confidence level, the existence classification confidence level, the location confidence level, the tracking consistency confidence level, and the motion rationality confidence level, respectively. These are the positional confidence weight index, the tracking consistency confidence weight index, and the motion rationality confidence weight index, respectively. Represents the component probability distribution. These are the classification weight matrix, the target size coverage feature vector, and the component bias vector, respectively. S405, Obtain the moving target analysis results, index the associated 3D dynamic fence based on the moving target analysis results, and align the coordinates with the component coordinate system of the dynamic fence constraint set in the 3D dynamic fence to determine the associated target component in the 3D dynamic fence. S406, Project the moving target analysis results onto the three-dimensional dynamic fence of the associated target components, determine the spatial occupation of the moving target analysis results, and generate a set of fence occupation voxels; S407: Traverse the dynamic fence constraint set of the 3D dynamic fence, generate boundary violation verification results, boundary violation type identification results, and associated target component attributes, and output the 3D fence boundary violation identification results containing the boundary violation verification results, boundary violation type identification results, and associated target component attributes.

[0046] In this embodiment of the invention, the 3D target tracking model adopts a modular architecture design. Using a deep neural network model as the initial model, it performs nonlinear transformation and dimensionality reduction on the behavior fusion features through a multilayer perceptron and fully connected layers in the feature decoder, outputting a 3D feature tensor adapted to 3D spatial processing. This effectively solves the dimensionality adaptation problem between high-dimensional behavior features and the 3D detection network. Furthermore, through the cascaded operation of voxelization layers, 3D sparse convolutional layers, and 3D pooling layers, sparse convolution processing is performed on a regular 3D voxel grid, and downsampling is used to generate aggregated global features. This fully utilizes the computational optimization advantages of sparse convolution for non-empty voxels, significantly reducing computational redundancy and memory consumption in 3D point cloud processing, and improving edge computing performance. The system leverages the real-time processing capabilities of the nodes. The 3D target detection head, based on DBSCAN clustering analysis, generates 3D candidate boxes in the component coordinate system of the dynamic fence constraint set. It then filters high-confidence moving target analysis results through target classification, bounding box regression, and comprehensive confidence prediction, achieving accurate target positioning and reliable detection from feature space to 3D physical space. The target association tracking block, through a layered verification mechanism of constraint index layer, geometric fence judgment layer, and attribute verification layer, aligns the detection results with the 3D dynamic fence in coordinates, determines spatial occupancy, and traverses the dynamic fence constraint set. The final output includes structured boundary violation verification results, boundary violation type identification results, and associated target component attributes. The 3D target tracking model not only achieves deep fusion verification of 3D target detection and 3D dynamic fences but also, through a detection and verification mechanism in the BIM component coordinate system, ensures that boundary violation judgment results possess clear spatial attribution and semantic association. This effectively solves the technical defects of traditional 2D target tracking models, such as lack of 3D spatial perception, inability to semantically couple with the building information model, and low structured boundary violation warning information.

[0047] On the other hand, embodiments of the present invention also provide a three-dimensional fence boundary crossing recognition system based on building information modeling. Figure 8A schematic diagram of a 3D fence boundary crossing detection system based on Building Information Modeling (BIM) is shown. The BIM-based 3D fence boundary crossing detection system specifically includes: The dynamic fence generation module 100 is used to acquire the building information model of the target construction site, import the building information model into the edge computing node, parse and identify the building information model based on the edge computing node to obtain the target component set, and generate a three-dimensional dynamic fence by combining the target component set. The edge data processing module 200 collects real-time monitoring data from the target construction site based on distributed sensor network nodes, performs edge-side preprocessing on the real-time monitoring data, and outputs the preprocessed behavior fusion feature set. The edge data processing module 200 includes: The timestamp alignment unit 210 is used to receive asynchronous real-time monitoring data from distributed sensor network nodes and perform timestamp alignment on the real-time monitoring data based on the local clock of the edge computing node. The coordinate transformation unit 220 is used to acquire real-time monitoring data after timestamp alignment, transform the real-time monitoring data to the component coordinate system of the dynamic fence constraint set, and establish the mapping relationship between the real-time monitoring data and the target components in the three-dimensional dynamic fence with the sensor network node constraints as a priori, so as to obtain the global monitoring data point set. The noise filtering unit 230 is used to perform noise filtering and load reduction processing on the global monitoring data point set to obtain the behavioral feature dataset. The behavior feature recognition unit 240 is used to acquire a behavior feature dataset, extract behavior features from the behavior feature dataset, and output a preprocessed behavior fusion feature set.

[0048] The boundary crossing recognition module 300 is used to load the behavior fusion feature set, pre-build a deep learning-based 3D target tracking model, perform 3D target detection on the behavior fusion feature set based on the 3D target tracking model, and output the 3D fence boundary crossing recognition result.

[0049] In summary, this invention provides a 3D fence boundary crossing recognition system and method based on Building Information Modeling (BIM). In this embodiment, the BIM model is imported into edge computing nodes for parsing and recognition. Combining component type attributes, geometric coordinate information, and multi-dimensional indicators such as the number of adjacent components, spatial influence range, process criticality, and hidden risk attributes, order relationship analysis and weighted filtering are performed to construct a 3D dynamic fence containing height layer constraints, horizontal projection constraints, component association constraints, and sensor network node constraints. This achieves deep integration of the fence boundary with BIM components and dynamic adaptation with construction progress, solving the problems of existing 2D fences lacking spatial depth information, being unable to adapt to complex 3D construction environments, and having statically fixed fence boundaries that are inconsistent with actual construction conditions. The technical problem of state disconnection is addressed; and the 3D target tracking model based on deep learning extracts and aggregates global features through a 3D sparse convolutional network to generate 3D candidate boxes, and combines comprehensive confidence prediction to screen valid targets. Finally, it is fused and verified with a 3D dynamic fence through a constraint index layer, a geometric fence judgment layer, and an attribute verification layer, and outputs a structured boundary crossing recognition result that includes whether the boundary has been crossed, the boundary crossing type recognition result, and the associated target component attributes. This not only realizes the leap from 2D plane boundary crossing detection to 3D spatial occupation determination, but also achieves accurate detection and spatial occupation determination of moving targets in complex 3D space. Thus, it significantly improves the intelligence level and risk prevention and control capabilities of construction site safety management, ensuring the accuracy of construction site safety management.

[0050] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions are also within the scope of protection of the present invention.

Claims

1. A three-dimensional fence boundary crossing recognition method based on Building Information Modeling, characterized in that, The method includes: The building information model of the target construction site is obtained and imported into the edge computing node. The building information model is parsed and identified based on the edge computing node to obtain the target component set. The target component set is then combined to generate a three-dimensional dynamic fence. Based on the real-time acquisition of real-time monitoring data within the target construction site by distributed sensor network nodes, edge-side preprocessing of the real-time monitoring data is performed, and the preprocessed behavior fusion feature set is output. Load the behavior fusion feature set, pre-build a deep learning-based 3D target tracking model, perform 3D target detection on the behavior fusion feature set based on the 3D target tracking model, and output the 3D fence crossing recognition result.

2. The three-dimensional fence boundary crossing recognition method based on building information model as described in claim 1, characterized in that: Methods for parsing and recognizing Building Information Models based on edge computing nodes include: Load the target construction site building information model, identify and analyze the target components in the target construction site building information model, read the type attributes and geometric coordinate information of the target components, combine the type attributes and geometric coordinate information of the target components, and perform preliminary screening of the target components based on the preset dangerous area screening rules to generate a preliminary screening set of components; The target components in the building information model of the target construction site are traversed twice to identify the number of adjacent components associated with the target components, the spatial influence range, the process criticality, and the hidden risk attributes. The number of adjacent components associated with the target components, the spatial influence range, the process criticality, and the hidden risk attributes are normalized and weighted using the order relation analysis method to obtain the component screening weight of the target components. Determine whether the component screening weight of the target component exceeds the preset screening threshold; If the component screening weight of the target component exceeds the preset screening threshold, the target component will be merged into the initial screening set of components to obtain the target component set.

3. The three-dimensional fence boundary crossing recognition method based on building information model as described in claim 2, characterized in that: Methods for generating 3D dynamic fences by combining a set of target components include: Load the target component set, and generate the target component spatial composition matrix based on the geometric coordinate information and type attributes of the target components in the target component set; The three-dimensional initial boundary of the target component is generated based on the spatial composition matrix of the target component, and the spatial influence degree, process criticality degree, and hidden risk degree of the target component are identified. Based on the normalized spatial influence degree, process criticality degree, and hidden risk degree of the target component, the three-dimensional initial boundary of the target component is subjected to spatial expansion redundancy processing to generate the three-dimensional redundant boundary of the target component. Based on the edge computing node, perform Boolean union operation on the three-dimensional redundant boundary of the target component to eliminate the overlapping boundary of the three-dimensional redundant boundary of the target component, and obtain the three-dimensional redundant boundary of the target component after the edge overlap region is eliminated. The three-dimensional redundant boundary of the target component is stretched vertically and fused in height to obtain the merged three-dimensional fence of the component. Obtain the merged 3D fence of the components, establish a component attribute mapping table for the 3D fence of the components, and establish the spatial topological relationship between the target components. Based on the component attribute mapping table and the spatial topological relationship, construct a dynamic fence constraint set, embed the dynamic fence constraint set into the 3D fence of the components, and obtain the 3D dynamic fence.

4. The three-dimensional fence boundary crossing recognition method based on building information model as described in claim 1, characterized in that: Methods for edge-side preprocessing of real-time monitoring data include: It receives asynchronous real-time monitoring data from distributed sensor network nodes and timestamps the real-time monitoring data based on the local clock of the edge computing node. The real-time monitoring data is transformed into the component coordinate system of the dynamic fence constraint set, and the mapping relationship between the real-time monitoring data and the target components in the three-dimensional dynamic fence is established with the sensor network node constraints as a priori, so as to obtain the global monitoring data point set. Noise filtering and load reduction processing are performed on the global monitoring data point set to obtain the behavioral feature dataset; Obtain the behavioral feature dataset, extract behavioral features from the dataset, and output the preprocessed behavioral fusion feature set.

5. The three-dimensional fence boundary crossing recognition method based on building information model as described in claim 4, characterized in that: The method for noise filtering and load reduction of the global monitoring data point set includes: Obtain the global monitoring data point set, perform spatial distribution density statistics on the global monitoring data point set, identify and remove global monitoring data points whose number of neighboring points is lower than the preset neighboring point threshold, and output the global monitoring data point set after outlier removal; Based on the mapping relationship between real-time monitoring data and target components in the 3D dynamic fence, global monitoring data points that exceed the 3D redundant boundary of the target component are removed from the global monitoring data point set, resulting in a global monitoring data point set after component spatial constraint filtering. The filter confidence weight of the global monitoring data points is calculated based on the sensor confidence score and the component screening weight of the target components associated with the global monitoring data points. It is then determined whether the filter confidence weight of the global monitoring data points exceeds the preset filter confidence threshold. If the filtering confidence weight of the global monitoring data points exceeds the preset filtering confidence threshold, mean filtering is applied to the global monitoring data points. If the filter confidence weight of the global monitoring data points does not exceed the preset filter confidence threshold, then the global monitoring data points are subjected to extended Kalman filtering. The filtered global monitoring data point set is obtained, and spatial voxel downsampling is performed on the global monitoring data point set. The uniform voxel point group where the global monitoring data point is located is replaced with the voxel center point to obtain the behavior feature dataset after spatial voxel downsampling.

6. The three-dimensional fence boundary crossing recognition method based on building information model as described in claim 5, characterized in that: The method for extracting behavioral features from a behavioral feature dataset includes: Obtain the behavioral feature dataset, extract spatial geometric features from the behavioral feature data in the dataset, construct a geometric feature covariance matrix based on the spatial geometric features of the behavioral feature data, and perform feature decomposition on the geometric feature covariance matrix. The minimum eigenvalue of the geometric feature covariance matrix is ​​used as the behavioral feature normal vector. The mapping relationship of the target components in the 3D dynamic fence associated with the behavioral feature normal vector is used as a priori to perform height-level labeling on the behavioral feature normal vector and output the height-level behavioral feature normal vector. Based on the optical flow estimation network, temporal motion features are extracted from the behavioral feature data in the behavioral feature dataset, and the temporal motion features of the moving target are output. Based on the mapping relationship index of target components in the 3D dynamic fence associated with behavioral feature normal vectors, the component attribute mapping table is queried to obtain the BIM semantic tags associated with behavioral feature normal vectors, the BIM semantic tags within the field of view of the target component are counted, and scene-level semantic context features are generated. The behavioral feature normal vector, temporal motion feature, and semantic context feature are standardized and processed. The correlation features between behavioral feature data are calculated based on the cross-modal attention mechanism. The correlation features are residually connected with the behavioral feature normal vector, temporal motion feature, and semantic context feature to obtain the fused behavioral fusion feature. A behavioral fusion feature set is generated based on the behavioral fusion feature.

7. The three-dimensional fence boundary crossing recognition method based on building information model as described in claim 1, characterized in that: The 3D target tracking model uses a deep neural network model as the initial model and also includes an input layer and an output layer. A feature decoder is introduced between the input layer and the deep neural network model, and a 3D target detection head and a target association tracking block are introduced between the deep neural network model and the output layer.

8. The three-dimensional fence boundary crossing recognition method based on building information model as described in claim 7, characterized in that: Methods for 3D object detection based on behavior fusion feature sets using 3D object tracking models include: The behavior fusion features are subjected to nonlinear transformation to extract high-dimensional hidden layer features. The high-dimensional hidden layer features are then subjected to nonlinear transformation and dimensionality reduction to output the dimensionality-reduced three-dimensional feature tensor. Obtain a three-dimensional feature tensor, divide the three-dimensional feature tensor into a regular three-dimensional voxel grid, and perform sparse convolution processing on the regular three-dimensional voxel grid through a three-dimensional sparse convolution layer to extract local spatial features. Then, downsample the local spatial features to generate aggregated global features. Based on the spatial distribution density of aggregated global features in the component coordinate system of the dynamic fence constraint set, the three-dimensional search space is divided by voxelization, and three-dimensional candidate boxes of moving targets are generated based on the DBSCAN clustering analysis method. The aggregated global features of the three-dimensional candidate boxes are used to classify the existence of targets, and three-dimensional bounding box regression analysis is performed on the effective targets of the three-dimensional candidate boxes. The three-dimensional center coordinates, target size, target category and motion descriptor of the moving target analysis results are output. Obtain the moving target analysis results, index the associated 3D dynamic fence based on the moving target analysis results, and align the coordinates with the component coordinate system of the dynamic fence constraint set in the 3D dynamic fence to determine the associated target components in the 3D dynamic fence. The moving target analysis results are projected onto the three-dimensional dynamic fence of the associated target components, and the spatial occupancy of the moving target analysis results is determined to generate a set of fence occupancy voxels. Traverse the dynamic fence constraint set of the 3D dynamic fence, generate boundary violation verification results, boundary violation type identification results, and associated target component attributes, and output the 3D fence boundary violation identification results containing the boundary violation verification results, boundary violation type identification results, and associated target component attributes.

9. A 3D fence boundary crossing recognition system based on Building Information Modeling (BIM), used to implement the 3D fence boundary crossing recognition method based on BIM as described in any one of claims 1-8, characterized in that: The three-dimensional fence boundary crossing recognition system based on building information modeling includes: The dynamic fence generation module is used to acquire the building information model of the target construction site, import the building information model into the edge computing node, parse and identify the building information model based on the edge computing node to obtain the target component set, and generate a three-dimensional dynamic fence by combining the target component set. The edge data processing module collects real-time monitoring data from the target construction site based on distributed sensor network nodes, performs edge-side preprocessing on the real-time monitoring data, and outputs a preprocessed behavior fusion feature set. The boundary crossing detection module is used to load the behavior fusion feature set, pre-build a deep learning-based 3D target tracking model, perform 3D target detection on the behavior fusion feature set based on the 3D target tracking model, and output the 3D fence boundary crossing detection result.

10. The three-dimensional fence boundary crossing recognition system based on building information model as described in claim 9, characterized in that: The edge data processing module includes: The timestamp alignment unit is used to receive asynchronous real-time monitoring data from distributed sensor network nodes and perform timestamp alignment on the real-time monitoring data based on the local clock of the edge computing node. The coordinate transformation unit is used to acquire real-time monitoring data after timestamp alignment, transform the real-time monitoring data to the component coordinate system of the dynamic fence constraint set, and establish the mapping relationship between the real-time monitoring data and the target components in the three-dimensional dynamic fence with the sensor network node constraints as a priori, so as to obtain the global monitoring data point set. The noise filtering unit is used to filter out noise and reduce load on the global monitoring data point set to obtain the behavioral feature dataset. The behavior feature recognition unit is used to acquire a behavior feature dataset, extract behavior features from the behavior feature dataset, and output a preprocessed behavior fusion feature set.

Images