Personnel state identification method and system based on target identification detection

Abstract

The invention relates to the technical field of target detection, and discloses a personnel state recognition method and system based on target recognition detection. The method comprises the steps of performing parameterization processing on collected original image data to obtain intelligent BIM data; performing multidisciplinary data exchange processing on the intelligent BIM data to obtain target detection data; performing intelligent posture analysis processing on human body information in the target detection data to obtain optimized posture data; behavior recognition processing is carried out on the optimized attitude data to obtain behavior analysis data, and the behavior analysis data comprises violation behavior types, climbing state judgment and risk level evaluation; performing fuzzy logic evaluation processing on the behavior analysis data to obtain safety compliance data; and performing intelligent early warning processing on the safety compliance data and the real-time monitoring data to obtain full-life-cycle management data. According to the invention, the efficiency of personnel state recognition based on target recognition detection is improved.

Description

Technical Field

[0001] The present application relates to the field of target detection, and in particular to a method and system for identifying the status of a person based on target recognition detection. Background Art

[0002] Existing personnel status recognition technologies mainly rely on traditional computer vision algorithms and simple machine learning models. These methods usually use fixed feature extraction methods, such as HOG (Histogram of Oriented Gradients) or SIFT (Scale Invariant Feature Transform), to recognize human posture and behavior. At the same time, existing technologies have also begun to introduce some deep learning models, such as convolutional neural networks (CNN) and recurrent neural networks (RNN), to improve recognition accuracy. In fields such as engineering construction and industrial production, some systems have begun to try to combine personnel status recognition with BIM (Building Information Modeling) technology to achieve accurate positioning and risk assessment of the working environment.

[0003] However, the existing technology still has many shortcomings in practical applications. First, traditional algorithms have poor adaptability to complex environments and are easily affected by factors such as occlusion and lighting changes, resulting in unstable recognition accuracy. Secondly, although deep learning models have been introduced, they are often single models or simple model combinations, which make it difficult to fully utilize the advantages of multi-source heterogeneous data. Furthermore, the existing BIM is not closely integrated with personnel status recognition, making it difficult to achieve real-time and dynamic risk assessment and early warning. In addition, most systems lack intelligent decision support and full life cycle safety management capabilities, and are unable to provide managers with timely and effective intervention recommendations and long-term safety optimization strategies. Summary of the invention

[0004] The present application provides a method and system for personnel status recognition based on target recognition detection, which are used to improve the efficiency of personnel status recognition based on target recognition detection.

[0005] In the first aspect, the present application provides a personnel status recognition method based on target recognition detection, and the personnel status recognition method based on target recognition detection includes: performing parameterization processing on the collected original image data to obtain intelligent BIM data; performing multidisciplinary data exchange processing on the intelligent BIM data to obtain target detection data, wherein the target detection data includes personnel location information, safety equipment recognition results and danger zone markings; performing intelligent posture analysis processing on the human body information in the target detection data to obtain optimized posture data; performing behavior recognition processing on the optimized posture data to obtain behavior analysis data, wherein the behavior analysis data includes the type of illegal behavior, climbing status judgment and risk level assessment; performing fuzzy logic evaluation processing on the behavior analysis data to obtain safety compliance data; performing intelligent early warning processing on the safety compliance data and real-time monitoring data to obtain full life cycle management data, wherein the full life cycle management data includes real-time early warning information, intervention measures recommendations and safety incident reports.

[0006] In a second aspect, the present application provides a device for identifying a person's state based on target recognition detection, the device for identifying a person's state based on target recognition detection comprising: The acquisition module is used to perform parameter processing on the acquired original image data to obtain intelligent BIM data; An exchange module is used to perform multidisciplinary data exchange processing on the intelligent BIM data to obtain target detection data, wherein the target detection data includes personnel location information, safety equipment identification results and dangerous area markings; An analysis module, used for performing intelligent posture analysis on human body information in the target detection data to obtain optimized posture data; An identification module, used to perform behavior identification processing on the optimized posture data to obtain behavior analysis data, wherein the behavior analysis data includes the type of illegal behavior, the judgment of the climbing state and the risk level assessment; An evaluation module, used for performing fuzzy logic evaluation processing on the behavior analysis data to obtain safety compliance data; The early warning module is used to perform intelligent early warning processing on the safety compliance data and real-time monitoring data to obtain full life cycle management data, wherein the full life cycle management data includes real-time early warning information, intervention measures suggestions and safety incident reports.

[0007] In the technical solution provided by the present application, the original image data is parametrically processed to obtain intelligent BIM data. This step realizes the conversion from 2D image to 3D model, lays the foundation for subsequent spatial analysis and target positioning, and also improves the visualization and comprehensibility of the data. Then, the intelligent BIM data is processed by multidisciplinary data exchange to obtain target detection data containing personnel location information, safety equipment identification results and dangerous area markings. This step realizes the comprehensive perception and analysis of the work scene and provides detailed environmental information for safety management. Then, the human body information in the target detection data is processed by intelligent posture analysis to obtain optimized posture data. This step can accurately capture the actions and postures of the staff and help identify potential unsafe behaviors. Subsequently, the optimized posture data is processed by behavior recognition to obtain behavior analysis data including violation behavior types, climbing status judgment and risk level assessment. This step realizes intelligent identification and risk assessment of staff behavior and can timely discover safety hazards. Then, the behavior analysis data is processed by fuzzy logic evaluation to obtain safety compliance data. This step comprehensively evaluates the behavior of the staff through the fuzzy reasoning system, providing a more objective and comprehensive safety compliance judgment. Finally, the safety compliance data and real-time monitoring data are intelligently processed for early warning, and full life cycle management data including real-time early warning information, intervention measures and safety incident reports are obtained. This step realizes the closed loop from data analysis to safety management, which can not only issue early warnings in a timely manner, but also provide specific intervention measures and support the formulation of long-term safety management strategies. The whole method realizes the transformation of the safety management mode from passive response to active prevention through the application of data-driven and intelligent algorithms, and significantly improves the accuracy and real-time performance of safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0009] Figure 1 A schematic diagram of an embodiment of a method for identifying a person's state based on target recognition detection in an embodiment of the present application; Figure 2 This is a schematic diagram of an embodiment of a personnel status recognition device based on target recognition detection in an embodiment of the present application. DETAILED DESCRIPTION

[0010] The embodiment of the present application provides a method and system for identifying the state of a person based on target recognition detection. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than the content illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0011] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of a method for identifying a person's state based on target recognition detection includes: Step S101, performing parameter processing on the collected original image data to obtain intelligent BIM data; Step S102: performing multidisciplinary data exchange processing on the intelligent BIM data to obtain target detection data, wherein the target detection data includes personnel location information, safety equipment identification results and dangerous area markings; Step S103, performing intelligent posture analysis and processing on human body information in the target detection data to obtain optimized posture data; Step S104, performing behavior recognition processing on the optimized posture data to obtain behavior analysis data, wherein the behavior analysis data includes the type of illegal behavior, the judgment of the climbing state and the risk level assessment; Step S105, performing fuzzy logic evaluation processing on the behavior analysis data to obtain safety compliance data; Step S106: Perform intelligent early warning processing on the safety compliance data and real-time monitoring data to obtain full life cycle management data, wherein the full life cycle management data includes real-time early warning information, intervention measures suggestions and safety incident reports.

[0012] It is understandable that the execution subject of the present application may be a personnel status recognition device based on target recognition detection, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0013] Specifically, the collected raw image data is parametrically processed to obtain intelligent BIM data. This process includes image denoising, histogram equalization, geometric correction, illumination compensation, feature extraction, feature mapping, spatial information fusion, visualization processing, semantic segmentation and parametric modeling. For example, a median filter is used for image denoising, a cumulative distribution function is used for histogram equalization, an affine transformation is applied for geometric correction, and a Retinex algorithm is used for illumination compensation. Then, a convolutional neural network is used to extract image features, and these features are associated with the BIM model through feature mapping. Spatial information fusion uses the ICP algorithm to align 2D images with 3D BIM models, and visualization processing uses WebGL technology to render 3D scenes. Semantic segmentation uses the U-Net network, and finally intelligent BIM data is generated through parametric modeling. Multidisciplinary data exchange processing is performed on the intelligent BIM data to obtain target detection data. This step includes data format conversion, data classification, feature extraction, target detection, non-maximum suppression, coordinate mapping, target classification, safety equipment identification and dangerous area division. Data format conversion uses the IFC standard, and data classification is based on a hierarchical clustering algorithm. The ResNet50 network is used for feature extraction, and the YOLOv5 algorithm is used for target detection. Non-maximum suppression selects targets by calculating IoU (intersection over union), and the homography matrix is ​​used to convert 2D coordinates into 3D space coordinates for coordinate mapping. The SVM algorithm is used for target classification, the Faster R-CNN network is used for safety equipment identification, and the dangerous area division is based on the spatial clustering algorithm.

[0014] The human body information in the target detection data is intelligently analyzed and processed to obtain optimized posture data. This includes human body region extraction, multi-scale feature extraction, key point positioning, temporal information fusion, occlusion inference, 3D reconstruction, posture angle calculation, motion trajectory analysis and confidence assessment. The Mask R-CNN algorithm is used for human body region extraction, and the FPN network is used for multi-scale feature extraction. The improved HRNet algorithm is used for key point positioning, and the Kalman filter is used for temporal information fusion. Occlusion inference is based on the graph convolution network, and the monocular depth estimation algorithm is used for 3D reconstruction. The cosine theorem is used for posture angle calculation, the DTW algorithm is used for motion trajectory analysis, and the confidence assessment is based on the Monte Carlo dropout technology. The optimized posture data is processed for behavior recognition to obtain behavior analysis data. This step includes spatiotemporal graph construction, feature extraction, multi-level convolution, attention mechanism, behavior classification, climbing state judgment, height estimation and safety risk assessment. The ST-GCN model is used for spatiotemporal graph construction, and the I3D network is used for feature extraction. The 3D convolutional neural network is used for multi-level convolution, and the self-attention model is used for the attention mechanism. Behavior classification is based on LSTM network, and decision tree algorithm is used to judge the climbing state. Height estimation is achieved through triangulation method, and safety risk assessment adopts Fuzzy AHP method. Fuzzy logic evaluation is performed on behavior analysis data to obtain safety compliance data. This includes fuzzy rule base construction, input variable fuzzification, fuzzy reasoning, dynamic weight adjustment, fuzzy output generation, defuzzification, multi-dimensional evaluation, comprehensive analysis and safety compliance calculation. The fuzzy rule base is built based on expert knowledge, and the input variable fuzzification adopts triangular membership function. Fuzzy reasoning uses improved Mamdani fuzzy reasoning system, and dynamic weight adjustment is based on adaptive neural fuzzy reasoning system (ANFIS). Fuzzy output generation adopts centroid method, and defuzzification uses maximum membership method. Multi-dimensional evaluation includes three dimensions: safety, efficiency and compliance. The comprehensive analysis adopts analytic hierarchy process (AHP), and safety compliance calculation is based on weighted average method.

[0015] Finally, the safety compliance data and real-time monitoring data are intelligently processed for early warning to obtain full life cycle management data. This step includes data fusion, risk level assessment, multi-threshold early warning mechanism setting, early warning trigger judgment, early warning information personalization, early warning information generation, intelligent intervention decision, remote expert support, closed-loop feedback and event analysis. Data fusion uses DS evidence theory, and risk level assessment uses random forest algorithm. The multi-threshold early warning mechanism is based on the dynamic threshold method, and the early warning trigger judgment uses pattern recognition technology. Early warning information personalization is achieved through collaborative filtering algorithm, and early warning information generation uses multimodal output algorithm. Intelligent intervention decision is based on reinforcement learning method, and remote expert support system uses knowledge graph technology. The closed-loop feedback mechanism uses PID control principle, and event analysis uses association rule mining algorithm.

[0016] For example, at a large construction site, the method first processes images captured by the site’s surveillance cameras. Image denoising reduces noise by 30%, and histogram equalization increases image contrast by 40%. Geometric correction corrects the 5° tilt caused by the camera’s installation angle, and illumination compensation improves visibility in shadowed areas. The convolutional neural network extracts 128-dimensional feature vectors, which are mapped to 3,000 components in the BIM model. Spatial information fusion keeps the registration error between the 2D image and the 3D BIM model within 5 cm. The semantic segmentation accuracy reaches 95%, and the resulting intelligent BIM data contains detailed information on 50 work areas. In the target detection phase, the YOLOv5 algorithm identifies 100 workers, 50 pieces of safety equipment, and 10 hazardous areas on the site with an accuracy of 98%. In the human posture analysis phase, the improved HRNet algorithm locates 17 key points on the human body with an average error of less than 2 cm. Temporal information fusion reduces jitter by 30%, and the depth error of 3D reconstruction is controlled within 10 cm. Behavior recognition processing identified five common violations, such as not wearing a helmet and entering an unauthorized area, with an accuracy rate of 92%. The accuracy of the height status judgment was 95%, and the height estimation error was controlled within 20cm. The fuzzy logic evaluation system evaluated the behavior of each worker based on 100 fuzzy rules, and the safety compliance score ranged from 0-100. Finally, the intelligent early warning system successfully warned of three potential safety accidents based on real-time data, provided 10 targeted intervention measures, and generated a safety incident report containing 30 days of data.

[0017] In the embodiment of the present application, the original image data is parameterized to obtain intelligent BIM data. This step realizes the conversion from 2D image to 3D model, lays the foundation for subsequent spatial analysis and target positioning, and also improves the visualization and comprehensibility of data. Then, the intelligent BIM data is processed by multidisciplinary data exchange to obtain target detection data containing personnel location information, safety equipment identification results and dangerous area markings. This step realizes the comprehensive perception and analysis of the work scene and provides detailed environmental information for safety management. Then, the human body information in the target detection data is processed by intelligent posture analysis to obtain optimized posture data. This step can accurately capture the actions and postures of the staff and help identify potential unsafe behaviors. Subsequently, the optimized posture data is processed by behavior recognition to obtain behavior analysis data including violation behavior types, climbing status judgment and risk level assessment. This step realizes intelligent identification and risk assessment of staff behavior and can timely discover safety hazards. Then, the behavior analysis data is processed by fuzzy logic evaluation to obtain safety compliance data. This step comprehensively evaluates the behavior of the staff through the fuzzy reasoning system, providing a more objective and comprehensive safety compliance judgment. Finally, the safety compliance data and real-time monitoring data are intelligently processed for early warning, and full life cycle management data including real-time early warning information, intervention measures and safety incident reports are obtained. This step realizes the closed loop from data analysis to safety management, which can not only issue early warnings in a timely manner, but also provide specific intervention measures and support the formulation of long-term safety management strategies. The whole method realizes the transformation of the safety management mode from passive response to active prevention through the application of data-driven and intelligent algorithms, and significantly improves the accuracy and real-time performance of safety management.

[0018] In a specific embodiment, the process of executing step S101 may specifically include the following steps: (1) performing image denoising on the original image data to obtain denoised image data, and performing histogram equalization on the denoised image data to obtain enhanced image data; (2) performing geometric correction processing on the enhanced image data to obtain corrected image data, and performing illumination compensation processing on the corrected image data to obtain compensated image data; (3) Perform feature extraction processing on the compensated image data through a convolutional neural network to obtain image feature data, and perform feature mapping processing on the image feature data to obtain BIM feature data; (4) Perform spatial information fusion processing on the BIM feature data to obtain fused feature data, and visualize the fused feature data through graphical processing to obtain visualized BIM data; (5) Perform semantic segmentation processing on the visualized BIM data to obtain segmented BIM data, and perform parametric modeling processing on the segmented BIM data through attribute association processing to obtain intelligent BIM data.

[0019] Specifically, the original image data is subjected to image denoising to obtain denoised image data. Image denoising is a technique for eliminating or reducing unnecessary noise in an image. Common methods include median filtering, Gaussian filtering and wavelet transform. The non-local means (NLM) denoising algorithm is used. The algorithm estimates the true value of each pixel by searching for similar pixel blocks in the entire image, thereby effectively retaining image details while reducing noise. The denoised image data is subjected to histogram equalization to obtain enhanced image data. Histogram equalization is a commonly used image enhancement technique that improves the image contrast by redistributing the gray value distribution of the image. This step uses the cumulative distribution function (CDF) to implement histogram equalization, thereby enhancing the overall visual effect and detail visibility of the image. The enhanced image data is subjected to geometric correction to obtain corrected image data. Geometric correction is to correct geometric deformation caused by factors such as camera angle and lens distortion in the image. This method uses affine transformation based on control points for geometric correction, and calculates the transformation matrix by the least squares method to correct the geometric shape of the image.

[0020] The corrected image data is subjected to illumination compensation processing to obtain compensated image data. Illumination compensation aims to eliminate the influence of uneven illumination on image quality. Here, the Retinex algorithm is used for illumination compensation. This algorithm is based on the working principle of the human visual system and can effectively improve the local contrast and color reproduction of the image. Subsequently, the compensated image data is subjected to feature extraction processing by a convolutional neural network to obtain image feature data. Convolutional Neural Network (CNN) is a deep learning model that is particularly suitable for processing data with a grid structure, such as images. In this method, ResNet-50 is used as a feature extraction network. This network solves the gradient vanishing problem of deep networks through residual learning and can extract rich multi-scale features.

[0021] The image feature data is processed by feature mapping to obtain BIM feature data. Feature mapping is the process of associating the extracted image features with BIM (Building Information Model) data. This step uses a fully connected layer to convert the features extracted by CNN into feature representations compatible with the BIM model, realizing the mapping of 2D image features to 3D BIM features. Next, the BIM feature data is processed by spatial information fusion to obtain fused feature data. Spatial information fusion aims to integrate 2D image features with the spatial information of the 3D BIM model. Here, the iterative closest point (ICP) algorithm is used for 2D-3D registration to accurately locate the image features to the corresponding positions of the BIM model.

[0022] The fused feature data is visualized through graphical processing to obtain visualized BIM data. The visualization uses WebGL technology to render the fused feature data into an intuitive 3D scene for subsequent analysis and interaction. Then, the visualized BIM data is semantically segmented to obtain segmented BIM data. Semantic segmentation is a pixel-level classification task that aims to assign a semantic label to each pixel in the image. This method uses the U-Net network for semantic segmentation, which has an encoder-decoder structure that can effectively capture multi-scale contextual information and achieve accurate pixel-level segmentation.

[0023] Finally, the segmented BIM data is parametrically modeled through attribute association processing to obtain intelligent BIM data. Parametric modeling is the process of associating the segmented BIM data with a predefined parametric model. This step uses a rule-based reasoning system to associate the segmentation results with the geometric and semantic attributes of the BIM model to generate intelligent BIM data with rich semantic information.

[0024] For example, in a task of personnel status recognition at a large construction site, a 4000x3000 pixel original image was collected. The NLM denoising algorithm increased the image's signal-to-noise ratio from the original 20dB to 28dB. Histogram equalization processing increased the image contrast by 40%, making the originally blurred outline of the worker clearly discernible. Geometric correction corrected the 5-degree tilt caused by the camera installation angle, ensuring that the vertical lines in the image were truly vertical. Lighting compensation processing effectively brightened the shadow area, increasing the visibility of the dark area from only 20% to 80%. The ResNet-50 network extracted 2048-dimensional feature vectors, which were mapped to the 500-dimensional BIM feature space through a fully connected layer. The ICP algorithm registered the 2D image features with the 3D BIM model, and the average registration error was controlled within 5cm. WebGL rendering generated a 3D scene model with 100,000 vertices. The U-Net network performed semantic segmentation on the scene and accurately identified 95% of the workers, equipment, and building structures. Finally, parametric modeling linked the segmentation results with the BIM model and generated intelligent BIM data containing 500 parametric objects, each of which contained an average of 20 attributes, such as location, size, material, and safety status.

[0025] In a specific embodiment, the process of executing step S102 may specifically include the following steps: (1) Perform data format conversion on intelligent BIM data to obtain standardized BIM data, and perform data classification on standardized BIM data to obtain classified BIM data; (2) Perform feature extraction on the classified BIM data to obtain a BIM feature vector, and perform target detection on the BIM feature vector using the YOLOv5 algorithm to obtain candidate target data; (3) Performing non-maximum suppression processing on the candidate target data to obtain the screened target data, and performing coordinate mapping processing on the screened target data to obtain spatial positioning data; (4) Performing target classification processing on the spatial positioning data to obtain classified target data, and performing security device identification processing on the classified target data through a deep learning algorithm to obtain device identification data; (5) The equipment identification data is processed into dangerous areas to obtain area marking data, and the area marking data is integrated to obtain target detection data.

[0026] Specifically, the intelligent BIM data is converted into standardized BIM data. The Industry Foundation Classes (IFC) standard is used to convert BIM data from different sources and formats into a standardized IFC format. IFC is an international standard for building information modeling that ensures data interoperability and consistency. During the conversion process, the geometric information, attribute information, and relationship information in the original BIM data are mapped to the corresponding entities and attributes in the IFC model.

[0027] The standardized BIM data is processed for data classification to obtain classified BIM data. The data classification uses a hierarchical clustering algorithm to divide BIM objects into different categories according to their attributes and characteristics, such as structural elements, equipment, personnel, etc. The hierarchical clustering algorithm calculates the similarity between objects and gradually merges similar objects to form a tree-like hierarchical structure. This classification method can effectively organize and manage complex BIM data, laying the foundation for subsequent feature extraction and target detection. The classified BIM data is processed for feature extraction to obtain a BIM feature vector. The feature extraction uses the ResNet50 deep convolutional neural network, which solves the gradient vanishing problem in deep network training through residual learning and can extract rich multi-scale features. ResNet50 contains 50 convolutional layers. By extracting and combining features layer by layer, it finally outputs a 2048-dimensional feature vector, which contains the high-level semantic information of the BIM object.

[0028] Next, the BIM feature vector is processed for target detection by the YOLOv5 algorithm to obtain candidate target data. YOLOv5 is an advanced single-stage target detection algorithm that converts the target detection problem into a regression problem and can simultaneously predict the bounding box and category of the target. The YOLOv5 algorithm first divides the input image into grids, then predicts multiple bounding boxes and confidences for each grid, and finally filters out candidate targets according to a preset threshold. Non-Maximum Suppression (NMS) is performed on the candidate target data to obtain the filtered target data. NMS is a post-processing technique used to eliminate overlapping detection boxes and retain the best detection results. NMS first sorts all candidate boxes according to confidence, and then calculates the intersection over union (IoU) between overlapping boxes. If the IoU exceeds the preset threshold, the box with the highest confidence is retained and other overlapping boxes are deleted. This step effectively reduces redundant detection and improves the accuracy of target detection.

[0029] Subsequently, the selected target data is processed by coordinate mapping to obtain spatial positioning data. Coordinate mapping converts the target coordinates in the 2D image space into coordinates in the 3D BIM model space. This process uses the homography matrix (Homography Matrix) for conversion, which describes the projection relationship between two planes. By solving the homography equation, the 2D image coordinates can be mapped to the 3D BIM coordinate system to achieve accurate spatial positioning of the target. The spatial positioning data is processed for target classification to obtain classified target data. The target classification adopts the support vector machine (SVM) algorithm. SVM is an effective binary classifier that separates data points of different categories by finding the optimal hyperplane. In the multi-class classification problem, a one-vs-all strategy is adopted to train an SVM classifier for each category. During the classification process, each SVM calculates the probability that the target belongs to the class, and finally selects the class with the highest probability as the classification result.

[0030] Then, the classified target data is processed for safety equipment identification through a deep learning algorithm to obtain equipment identification data. The Faster R-CNN (Faster Region-based Convolutional Neural Network) algorithm is used here. This algorithm is a two-stage target detection model, including a region proposal network (RPN) and a target classification network. RPN first generates candidate regions that may contain targets, and then the target classification network accurately classifies and regresses these regions. Faster R-CNN is particularly suitable for identifying safety equipment, such as helmets and seat belts, because it can handle targets of different scales and shapes. The equipment identification data is processed for dangerous area division to obtain regional labeling data. The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) spatial clustering algorithm is used for dangerous area division. Based on the concept of density, DBSCAN can discover clusters of any shape and has good robustness to noisy data. By setting appropriate density thresholds and neighborhood radius, DBSCAN can effectively identify potential dangerous areas, such as high-altitude work areas and mechanical operation areas.

[0031] Finally, the area marking data is integrated to obtain the target detection data. During the data integration process, the personnel location information, safety equipment identification results, and dangerous area marking information are integrated into a unified data structure. This data structure uses the JSON (JavaScript Object Notation) format, which has good readability and scalability. The integrated target detection data contains the category, location coordinates, confidence, associated safety equipment information, and area type of each detected target.

[0032] For example, in a task of personnel status recognition at a large construction site, the intelligent BIM data containing 10,000 BIM objects was first converted into IFC format, and the standardized data volume was 500MB. Through the hierarchical clustering algorithm, these objects were divided into 15 main categories, including 4 structural elements, 6 equipment types, and 5 personnel categories. The ResNet50 network extracted 2048-dimensional feature vectors from each BIM object, and a total of 10,000 feature vectors were generated. When processing these feature vectors, the YOLOv5 algorithm set a confidence threshold of 0.5 and initially detected 1,500 candidate targets. After NMS processing, where the IoU threshold was set to 0.5, 1,200 valid targets were retained after screening. Coordinate mapping mapped these 1,200 targets from 2D image space (resolution of 3840×2160) to 3D BIM space (size of 100m×50m×30m). The SVM classifier classified the 1,200 targets with an accuracy of 95%, and correctly identified the categories of 1,140 targets. The Faster R-CNN algorithm identified safety equipment and successfully identified 480 of the 500 safety-related targets with an accuracy of 96%. The DBSCAN algorithm set the neighborhood radius to 5 meters and the minimum number of samples to 3, identifying 10 potential danger areas within the construction site. The final integrated target detection data contains detailed information on 1,140 targets, with an average of 20 attribute fields per target, and a total data volume of approximately 5MB.

[0033] In a specific embodiment, the process of executing step S103 may specifically include the following steps: (1) Perform human region extraction processing on the target detection data to obtain human ROI data, and perform feature extraction processing on the human ROI data through a multi-scale feature pyramid network to obtain multi-scale feature data; (2) The improved HRNet algorithm is used to locate the key points of the multi-scale feature data to obtain two-dimensional key point data, and the two-dimensional key point data is fused with time series information to obtain continuous posture data; (3) Perform occlusion inference processing on the continuous posture data to obtain the completed posture data, and perform three-dimensional reconstruction processing on the completed posture data through a monocular depth estimation algorithm to obtain three-dimensional posture data; (4) Calculate the posture angle of the three-dimensional posture data to obtain angle feature data, and perform motion trajectory analysis on the angle feature data to obtain motion feature data; (5) Perform confidence assessment on the motion feature data to obtain credibility data, and perform data fusion processing on the credibility data to obtain optimized posture data.

[0034] Specifically, the target detection data is processed for human region extraction to obtain human ROI (Region of Interest) data. The Mask R-CNN algorithm is used, which adds a branch network on the basis of Faster R-CNN to generate an accurate segmentation mask for the target. Mask R-CNN can not only detect the bounding box of the human body, but also generate pixel-level human contours, thereby accurately extracting the human body area. The human ROI data is processed for feature extraction through a multi-scale feature pyramid network (FPN) to obtain multi-scale feature data. FPN is a top-down architecture that can use the multi-layer feature map of a convolutional neural network to generate a feature pyramid with strong semantic information. The processing process of FPN can be expressed by the following formula: ; in, Indicates The feature map of the layer, U represents the upsampling operation, represents the convolutional feature map of the corresponding layer, Represents the convolution operation. This structure can effectively capture human features of different scales and provide rich information for subsequent posture estimation. The improved HRNet (High-Resolution Network) algorithm is used to locate the key points of multi-scale feature data to obtain two-dimensional key point data. The improvement of HRNet is that the attention mechanism is introduced to enhance the network's perception of key points. The key point location process can be expressed as: ; Among them, K represents the key point coordinates, A represents the attention weight, and F represents the feature map. represents element-wise multiplication, is a differentiable operation used to transform the heatmap into exact coordinates.

[0035] The two-dimensional key point data is processed by time series information fusion to obtain continuous posture data. This step uses the Kalman filter, which can optimally estimate the posture based on historical observations and current measurements. The state update equation of the Kalman filter is: ; in, represents the current state, F is the state transfer matrix, B is the control input matrix, is the control vector, is process noise. The continuous posture data is processed by occlusion inference to obtain the completed posture data. Occlusion inference uses a graph convolutional network (GCN), which can use the topological structure of the human skeleton to effectively infer the position of the occluded key points. A forward propagation of GCN can be expressed as: ; in, represents the node features of the lth layer, A is the adjacency matrix, D is the degree matrix, is the weight matrix, is the activation function. Then, the completed posture data is 3D reconstructed through the monocular depth estimation algorithm to obtain 3D posture data. Monocular depth estimation uses the DenseDepth network, which estimates depth information from a single image through an encoder-decoder structure and jump connections. The loss function of depth estimation includes: ; in, is the depth reconstruction error, is the gradient consistency error, is the structural similarity error, , , is the weight coefficient. The three-dimensional posture data is processed by posture angle calculation to obtain angle feature data. Angle calculation is based on the cosine theorem, for example, to calculate the elbow angle : ; in, and are the vectors of the upper arm and forearm respectively.

[0036] The angle feature data is processed by motion trajectory analysis to obtain action feature data. The motion trajectory analysis uses the Dynamic Time Warping (DTW) algorithm, which can calculate the similarity of two time series and is not affected by changes in motion speed. The DTW distance is defined as: ; Among them, X and Y are two time series, w is the alignment path, d is the distance function, and K is the length of the alignment path. represents the element in the X sequence at the kth step of the alignment path. Represents the element in the Y sequence at the kth step of the alignment path. Indicates that the path with the smallest total distance is selected from all possible alignment paths. Finally, the action feature data is processed for confidence assessment to obtain credibility data, and the credibility data is processed for data fusion to obtain optimized posture data. The confidence assessment uses Monte Carlo dropout technology to estimate the uncertainty of the model through multiple forward propagations without dropout. The data fusion adopts the weighted average method to integrate according to the confidence of each part.

[0037] For example, in a task of personnel status recognition at a construction site, 100 human ROIs with an average size of 256×256 pixels were first extracted from images with a 4K resolution (3840×2160). The FPN network generated feature maps of 5 scales, namely 32×32, 64×64, 128×128, 256×256 and 512×512, each containing 256 channels. The improved HRNet algorithm located 17 human key points on these feature maps, with an average positioning error of less than 3 pixels. The Kalman filter performed temporal fusion on the key point sequence of 30 frames (1 second), reducing the posture jitter by 40%. The GCN network successfully inferred 15% of the occluded key points with an average error of 5 cm. The monocular depth estimation converted the 2D posture to 3D posture, and the average relative error of the depth estimation was 10%. The posture angle calculation obtained 20 joint angles with an accuracy of 1 degree. The DTW algorithm analyzed 100 frames (about 3 seconds) of action sequences and identified 5 common construction actions, such as carrying, bending, climbing, etc., with an accuracy rate of 92%. Monte Carlo dropout was sampled 50 times to obtain the confidence interval of each posture parameter, with an average confidence of 0.85. The final fused optimized posture data contains each worker's 3D joint position, joint angle, action type and corresponding confidence, with a data volume of about 1KB per person per second. These high-precision and high-reliability posture data provide a solid foundation for subsequent safety supervision and efficiency analysis.

[0038] In a specific embodiment, the process of executing step S104 may specifically include the following steps: (1) Performing spatiotemporal graph construction processing on the optimized posture data to obtain posture graph structure data, and performing feature extraction processing on the posture graph structure data through the spatiotemporal graph convolutional network to obtain spatiotemporal feature data; (2) Perform multi-level convolution processing on the spatiotemporal feature data to obtain high-level action feature data, and then perform weight distribution processing on the high-level action feature data through the attention mechanism to obtain weighted feature data; (3) Performing behavior classification processing on the weighted feature data to obtain violation behavior type data, and performing climbing state judgment processing on the violation behavior type data to obtain climbing state data; (4) Performing height estimation processing on the climbing status data to obtain height information data, and performing safety risk assessment processing on the height information data through the risk assessment model to obtain risk level data; (5) Perform data integration processing on the risk level data to obtain behavior analysis data.

[0039] Specifically, the optimized posture data is processed by spatiotemporal graph construction to obtain posture graph structure data. Spatiotemporal graph construction is the process of representing the human posture sequence as a graph structure, where the nodes of the graph represent the key points of the human body and the edges represent the connections between the key points. This representation method can effectively capture the spatial and temporal characteristics of human posture. The construction of the spatiotemporal graph adopts the sliding window method to combine the posture data of consecutive T frames into a graph structure, and each node contains the spatial coordinates and time information of the key point. The posture graph structure data is processed by spatiotemporal graph convolution network (ST-GCN) to obtain spatiotemporal feature data. ST-GCN is a deep learning model specially designed for processing spatiotemporal graph data, which can consider the information of spatial and temporal dimensions at the same time. The core of ST-GCN is the graph convolution operation, which defines convolution and pooling operations on the graph structure, and can effectively extract the spatiotemporal features of the posture sequence. The ST-GCN network contains multiple ST-GCN layers, each of which performs feature conversion and aggregation on the input graph structure, and finally outputs a high-dimensional spatiotemporal feature vector.

[0040] Then, the spatiotemporal feature data is processed by multi-level convolution to obtain high-level action feature data. Multi-level convolution uses a 3D convolutional neural network (3D CNN), which can perform convolution operations in both spatial and temporal dimensions to capture more complex action patterns. 3D CNN contains multiple 3D convolutional layers, 3D pooling layers, and fully connected layers, each of which gradually extracts higher-level action features. Through this hierarchical feature extraction, the network can learn various features from low-level limb movements to high-level behavior patterns. The high-level action feature data is weighted by the attention mechanism to obtain weighted feature data. The introduction of the attention mechanism enables the model to automatically learn the importance of different spatiotemporal regions, so as to better focus on key action details. The self-attention mechanism is used here, which allows the model to consider the information of all positions when calculating the representation of a certain position and dynamically assign weights based on relevance. The self-attention mechanism assigns weights by calculating the similarity between the query, key, and value to obtain a weighted feature representation.

[0041] Subsequently, the weighted feature data is processed for behavior classification to obtain the violation behavior type data. The behavior classification uses the long short-term memory network (LSTM), which is a special recurrent neural network that can effectively process long sequence data and capture long-term dependencies. LSTM controls the flow of information through a gating mechanism, including input gates, forget gates, and output gates, which enables it to learn complex temporal patterns. In this method, the LSTM network receives the weighted feature sequence as input and outputs the category probability distribution of the violation behavior. The violation behavior type data is processed for climbing state judgment to obtain the climbing state data. The climbing state judgment uses a decision tree algorithm, which classifies the data through a series of conditional judgments. Each internal node of the decision tree represents a test on a feature or attribute, each branch represents a test output, and each leaf node represents a category or decision result. In this method, the decision tree uses the violation behavior type, posture features, and environmental information to determine whether the worker is in a climbing state.

[0042] Then, the altitude estimation data is processed to obtain the altitude information data. The altitude estimation adopts the triangulation method, which calculates the actual altitude of the target based on the known camera parameters and the reference objects in the image. Specifically, by comparing the size of the target in the image with the size of the reference objects of known height (such as building structures or standard equipment), combined with the internal and external parameters of the camera, the actual altitude of the target can be inferred.

[0043] Finally, the risk assessment model is used to process the height information data for safety risk assessment to obtain risk level data, and the risk level data is integrated to obtain behavior analysis data. The risk assessment model adopts the fuzzy comprehensive evaluation method, which combines the ideas of fuzzy mathematics and comprehensive evaluation and can handle the uncertainty and ambiguity in the evaluation process. The fuzzy comprehensive evaluation first establishes the evaluation factor set and the comment set, and then maps the factor set to the comment set through the fuzzy relationship matrix, and finally obtains the membership vector of the risk level. The data integration process combines the risk level with the previous behavior classification results, climbing status and height information to form the final behavior analysis data.

[0044] For example, in a task of personnel status recognition at a large construction site, the optimized posture data of 100 workers are first processed. The spatiotemporal graph construction uses a sliding window of 30 frames (1 second) to generate a graph structure containing 17 nodes (corresponding to 17 key points of the human body). The ST-GCN network contains 8 ST-GCN layers, each of which outputs a 256-dimensional feature vector. The 3D CNN network contains 5 3D convolutional layers, and finally outputs 1024-dimensional high-level action features. The self-attention mechanism calculates 16 attention heads, each with a dimension of 64, to obtain weighted feature data. The LSTM network contains 2 layers, each with 128 hidden units, and finally outputs the probability distribution of 10 common violations. The decision tree algorithm uses 5 key features (such as height, posture angle, tool usage, etc.) to judge the climbing state with an accuracy of 95%. The average error of height estimation is controlled within 10cm. The fuzzy comprehensive evaluation model takes into account five risk factors (height, posture stability, protective measures, environmental conditions, and operational complexity), and divides the risk level into five levels (low, lower, medium, higher, and high). The behavioral analysis data finally generated includes each worker's violation behavior type, climbing status, estimated height, and risk level, with a data volume of approximately 2KB per person per second. Through this comprehensive behavioral analysis, the method successfully identified 15 potential high-risk behaviors, including 8 aerial operations without wearing safety belts and 7 improper operations of heavy equipment, providing timely and accurate safety warning information for construction site managers, effectively reducing the incidence of safety accidents.

[0045] In a specific embodiment, the process of executing step S105 may specifically include the following steps: (1) The behavior analysis data is processed by constructing a fuzzy rule base to obtain fuzzy rule set data, and the input variables of the fuzzy rule set data are fuzzified to obtain input fuzzy set data; (2) Perform fuzzy inference processing on the input fuzzy set data through the improved Mamdani fuzzy inference system to obtain inference result data, and dynamically adjust the weight of the inference result data to obtain weighted result data; (3) Perform fuzzy output generation processing on the weighted result data to obtain fuzzy output set data, and defuzzify the fuzzy output set data using the centroid method to obtain accurate scoring data; (4) Perform multi-dimensional evaluation processing on the precise scoring data to obtain multi-dimensional evaluation data, and perform comprehensive analysis and processing on the multi-dimensional evaluation data to obtain comprehensive evaluation data; (5) Calculate and process the comprehensive evaluation data to obtain safety compliance data.

[0046] Specifically, the behavior analysis data is processed by fuzzy rule base construction to obtain fuzzy rule set data. The fuzzy rule base is a set of IF-THEN statements used to describe the relationship between input variables and output variables. These rules are formulated based on the knowledge and experience of domain experts and cover various possible situations and corresponding evaluation results. For example, a rule may be "IF the height risk is high AND the protective measures are insufficient THEN the safety compliance is low". The rule base construction process includes the steps of determining the input and output variables, defining the fuzzy sets of language variables, and designing the rule set. The fuzzy rule set data is processed by fuzzification of the input variables to obtain the input fuzzy set data. Fuzzification is the process of converting the precise input value into a fuzzy set. This step uses membership functions to describe the membership of the input variables to different fuzzy sets. Commonly used membership functions include triangular functions, trapezoidal functions, and Gaussian functions. For example, for the input variable "height risk", three fuzzy sets of "low", "medium", and "high" can be defined, and the triangular membership function can be used to describe the membership of a specific height value to these three fuzzy sets.

[0047] Then, the input fuzzy set data is processed by fuzzy reasoning through the improved Mamdani fuzzy reasoning system to obtain the reasoning result data. The Mamdani fuzzy reasoning system is a commonly used fuzzy reasoning method, which draws conclusions through four steps: fuzzification of input variables, rule evaluation, rule output aggregation and defuzzification. The improved Mamdani system introduces a dynamic weight adjustment mechanism based on the traditional method, which can dynamically adjust the importance of rules according to the characteristics of the input data. This improvement improves the adaptability and reasoning accuracy of the system to different situations. The weight of the reasoning result data is dynamically adjusted to obtain weighted result data. The dynamic adjustment of weights is based on the idea of ​​adaptive neural fuzzy inference system (ANFIS), which optimizes the weight of rules through feedback learning. ANFIS combines the learning ability of neural networks and the reasoning ability of fuzzy systems, and can continuously adjust and optimize the weights of fuzzy rules based on historical data and real-time feedback. This step enables the system to adapt to different working environments and changes in safety requirements.

[0048] Subsequently, the weighted result data is processed for fuzzy output generation to obtain fuzzy output set data. Fuzzy output generation is the process of aggregating the outputs of each rule. Common aggregation methods include the maximum method, weighted average method, etc. In this method, an improved weighted average method is adopted, which takes into account the weight and activation strength of each rule to obtain a more accurate fuzzy output set. The fuzzy output set data is defuzzified by the centroid method to obtain accurate scoring data. Defuzzification is the process of converting a fuzzy output set into an accurate numerical value. The centroid method is a commonly used defuzzification method, which obtains the final accurate value by calculating the centroid of the fuzzy output set. The advantage of the centroid method is that it can comprehensively consider the entire output distribution and obtain a relatively smooth and stable result.

[0049] The precise scoring data is processed by multi-dimensional evaluation to obtain multi-dimensional evaluation data. Multi-dimensional evaluation takes into account multiple aspects such as safety, efficiency and compliance, and each dimension has its specific evaluation criteria and weight. This step expands the single precise score into a score of multiple dimensions, thereby providing a more comprehensive evaluation result. Next, the multi-dimensional evaluation data is processed by comprehensive analysis to obtain comprehensive evaluation data. The comprehensive analysis adopts the analytic hierarchy process (AHP), which determines the relative importance of each factor by decomposing complex problems into a hierarchical structure and comparing them pairwise. AHP can effectively process qualitative and quantitative factors and obtain objective comprehensive evaluation results.

[0050] Finally, the comprehensive evaluation data is processed for safety compliance to obtain safety compliance data. Safety compliance calculation adopts the weighted average method, and the scores of each dimension are weighted averaged according to their importance to obtain the final safety compliance score. This score reflects the degree of compliance of workers' behavior with safety regulations, and provides a quantitative basis for subsequent safety management and intervention.

[0051] For example, in a personnel status recognition task at a large construction site, a fuzzy rule base containing 100 rules was first constructed, covering multiple aspects such as high-altitude work, equipment operation, and personal protection. The input variables include five variables: height risk (low, medium, and high), protective measures (inadequate, general, and sufficient), and operational norms (low, medium, and high). Each variable uses a triangular membership function to define three fuzzy sets. For example, for a worker working at a height of 10 meters, the height risk input value is 0.8 (high), 0.2 (medium), and 0 (low). The improved Mamdani fuzzy inference system processes these inputs and preliminarily concludes that the safety compliance is 0.6. Dynamic weight adjustment Based on the safety record of the past week, the weight of the rules related to high-altitude work is increased by 10%, and the weighted result is recalculated to be 0.55. The fuzzy output generation stage aggregates the outputs of multiple rules to obtain a composite fuzzy set. Defuzzification is performed using the centroid method, and the final accurate score is 75 points (out of 100 points). The multi-dimensional assessment expanded this score into three dimensions: safety (80 points), efficiency (70 points), and compliance (75 points). The weights given by the AHP comprehensive analysis were 0.5, 0.3, and 0.2, respectively, and the calculated comprehensive assessment score was 76 points. Finally, according to industry standards, the 76 points were converted into a safety compliance of 85%. This result shows that the worker's behavior basically complies with safety regulations, but there is still room for improvement, especially in terms of work efficiency. Through this comprehensive safety compliance assessment, this method not only provides quantitative safety assessment results, but also provides a basis for site managers to formulate targeted safety training and improvement measures, effectively improving the overall safety management level.

[0052] In a specific embodiment, the process of executing step S106 may specifically include the following steps: (1) Perform data fusion processing on safety compliance data and real-time monitoring data to obtain fusion status data, and use the random forest algorithm to perform risk level assessment on the fusion status data to obtain dynamic risk data; (2) Perform multi-threshold warning mechanism setting processing on dynamic risk data to obtain warning threshold data, and perform warning trigger judgment processing on the warning threshold data to obtain warning trigger data; (3) Performing personalized warning information processing on the warning trigger data to obtain personalized warning data, and performing warning information generation processing on the personalized warning data through a multimodal output algorithm to obtain multimodal warning data; (4) Perform intelligent intervention decision processing on multimodal warning data through a decision tree algorithm to obtain intervention recommendation data, and perform remote expert support processing on the intervention recommendation data to obtain expert guidance data; (5) Perform closed-loop feedback processing on the expert guidance data to obtain optimization strategy data, and perform event analysis on the optimization strategy data through data mining algorithms to obtain full life cycle management data.

[0053] Specifically, the safety compliance data and real-time monitoring data are fused to obtain fused state data. Data fusion adopts DS evidence theory, which can effectively deal with the uncertainty and conflict of multi-source information. DS evidence theory integrates information from different sources by calculating basic probability distribution function, trust function and likelihood function, and finally obtains a more reliable and comprehensive state description. The fused state data is processed by random forest algorithm for risk level assessment to obtain dynamic risk data. Random forest is an ensemble learning method composed of multiple decision trees. Each decision tree independently classifies or regresses the input data, and finally obtains the result by voting or averaging. The advantages of random forest are that it can process high-dimensional data, has strong noise resistance, and can evaluate the importance of features. In this method, the random forest model uses fused state data as input and outputs risk level assessment results.

[0054] Then, the dynamic risk data is processed by multi-threshold warning mechanism setting to obtain warning threshold data. The multi-threshold warning mechanism is based on the dynamic threshold method, which can dynamically adjust the warning threshold according to historical data and current trends. The calculation of dynamic thresholds takes into account the time series characteristics of the data, such as seasonality, periodicity and trend, so as to make the warning more sensitive and accurate. The warning threshold data contains the thresholds corresponding to different risk levels, which provides a basis for subsequent warning triggering. The warning threshold data is processed for warning trigger judgment to obtain warning trigger data. The warning trigger judgment adopts pattern recognition technology, which can identify data patterns that match predefined risk patterns. The pattern recognition process includes three steps: feature extraction, pattern matching and decision-making. When the identified risk pattern exceeds the preset threshold, the corresponding level of warning is triggered.

[0055] Subsequently, the warning trigger data is processed for personalized warning information to obtain personalized warning data. The personalized warning information adopts the collaborative filtering algorithm, which recommends the most suitable warning information form and content based on the user's historical behavior and preferences. Collaborative filtering calculates the similarity between users to find similar user groups, and then generates personalized recommendations for target users based on the behavioral characteristics of the groups. The personalized warning data is processed for warning information generation through the multimodal output algorithm to obtain multimodal warning data. The multimodal output algorithm can generate warning information in various forms such as text, images, and sounds according to different scenarios and user needs. This method comprehensively considers the urgency of the information, the user's receiving ability, and environmental factors to select the most effective warning method.

[0056] Then, the multimodal warning data is processed for intelligent intervention decision-making through the decision tree algorithm to obtain intervention recommendation data. The decision tree algorithm builds a decision model through a series of if-then rules, and can automatically generate intervention recommendations based on the characteristics of the warning data. The advantage of the decision tree is that it has a clear structure and is easy to understand and execute. In this method, the decision tree model considers multiple dimensions such as warning level, risk type, and environmental factors, and outputs specific intervention measures. The intervention recommendation data is processed by remote expert support to obtain expert guidance data. The remote expert support system uses knowledge graph technology, which can effectively organize and represent the knowledge and experience of domain experts. The knowledge graph describes complex knowledge structures through entities, relationships, and attributes, and supports intelligent reasoning and query. Remote experts can quickly access relevant knowledge through this system and review and supplement the automatically generated intervention recommendations.

[0057] Closed-loop feedback processing is performed on the expert guidance data to obtain optimization strategy data. The closed-loop feedback mechanism is based on the PID (proportional-integral-differential) control principle, which optimizes the system response by continuously adjusting the control parameters. In security management, this mechanism can continuously optimize the early warning and intervention strategies according to the actual effects of the intervention measures. Event analysis and processing of the optimization strategy data is performed through data mining algorithms to obtain full life cycle management data. Data mining uses association rule mining algorithms, such as the Apriori algorithm, which can discover implicit and valuable association rules from large amounts of data. By analyzing the implementation effects of optimization strategies and historical event data, deep laws and trends in security management are discovered to provide data support for full life cycle management.

[0058] For example, in a personnel status recognition task at a large construction site, the safety compliance data (average 85%) and real-time monitoring data (including location, posture, environmental parameters, etc.) of 100 workers are first fused. The DS evidence theory sets the credibility of these two types of data to 0.7 and 0.8 respectively, and a more comprehensive status description is obtained after fusion. The random forest algorithm uses 1,000 decision trees to conduct risk assessment on the fused data and output a risk probability distribution of 5 levels. The multi-threshold warning mechanism sets dynamic thresholds for each risk level based on the data of the past 3 months. For example, the high risk threshold is adjusted from a fixed 0.8 to 0.75-0.85 according to time. The warning trigger judgment identifies 15 potential risk patterns, of which 3 exceed the threshold to trigger warnings. The personalized processing of warning information customizes the form and content of warning information for each worker according to their position, experience and preferences. The multimodal output algorithm generates warning information in three forms: text messages, voice reminders and visual charts. The decision tree algorithm generates 50 specific intervention suggestions based on 20 key features. The remote expert support system connected 10 security experts and optimized the automatically generated intervention suggestions within 5 minutes through rapid retrieval and reasoning of the knowledge graph. The closed-loop feedback mechanism increased the early warning accuracy from the initial 80% to 95% through 30 consecutive days of data. The data mining algorithm analyzed 500 security incidents in the past year and found 10 key risk factor combinations. These findings were integrated into the full life cycle management data, providing a basis for long-term security strategy formulation.

[0059] The above describes the personnel status recognition method based on target recognition detection in the embodiment of the present application. The following describes the personnel status recognition device based on target recognition detection in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of a personnel status recognition device based on target recognition detection includes: The acquisition module 201 is used to perform parameterization processing on the acquired original image data to obtain intelligent BIM data; An exchange module 202 is used to perform multidisciplinary data exchange processing on the intelligent BIM data to obtain target detection data, wherein the target detection data includes personnel location information, safety equipment identification results and dangerous area markings; The analysis module 203 is used to perform intelligent posture analysis on the human body information in the target detection data to obtain optimized posture data; The recognition module 204 is used to perform behavior recognition processing on the optimized posture data to obtain behavior analysis data, wherein the behavior analysis data includes the type of illegal behavior, the judgment of the climbing state and the risk level assessment; An evaluation module 205 is used to perform fuzzy logic evaluation processing on the behavior analysis data to obtain safety compliance data; The early warning module 206 is used to perform intelligent early warning processing on the safety compliance data and real-time monitoring data to obtain full life cycle management data, wherein the full life cycle management data includes real-time early warning information, intervention measures suggestions and safety event reports.

[0060] Through the collaborative cooperation of the above components, the original image data is parametrically processed to obtain intelligent BIM data. This step realizes the conversion from 2D images to 3D models, laying the foundation for subsequent spatial analysis and target positioning, and also improves the visualization and comprehensibility of the data. Then, the intelligent BIM data is processed by multidisciplinary data exchange to obtain target detection data containing personnel location information, safety equipment identification results and dangerous area markings. This step realizes the comprehensive perception and analysis of the work scene and provides detailed environmental information for safety management. Then, the human body information in the target detection data is processed by intelligent posture analysis to obtain optimized posture data. This step can accurately capture the movements and postures of the staff and help identify potential unsafe behaviors. Subsequently, the optimized posture data is processed by behavior recognition to obtain behavior analysis data including violation behavior types, climbing status judgment and risk level assessment. This step realizes the intelligent identification and risk assessment of staff behavior and can timely discover safety hazards. Then, the behavior analysis data is processed by fuzzy logic evaluation to obtain safety compliance data. This step comprehensively evaluates the behavior of the staff through the fuzzy reasoning system, providing a more objective and comprehensive safety compliance judgment. Finally, the safety compliance data and real-time monitoring data are intelligently processed for early warning, and full life cycle management data including real-time early warning information, intervention measures and safety incident reports are obtained. This step realizes the closed loop from data analysis to safety management, which can not only issue early warnings in a timely manner, but also provide specific intervention measures and support the formulation of long-term safety management strategies. The whole method realizes the transformation of the safety management mode from passive response to active prevention through the application of data-driven and intelligent algorithms, and significantly improves the accuracy and real-time performance of safety management.

[0061] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A personnel status recognition method based on target recognition detection, characterized in that: The personnel status recognition method based on target recognition detection includes: Perform parameter processing on the collected original image data to obtain intelligent BIM data; Performing multidisciplinary data exchange processing on the intelligent BIM data to obtain target detection data, wherein the target detection data includes personnel location information, safety equipment identification results and dangerous area markings; Performing intelligent posture analysis and processing on human body information in the target detection data to obtain optimized posture data; Performing behavior recognition processing on the optimized posture data to obtain behavior analysis data, wherein the behavior analysis data includes the type of illegal behavior, the judgment of the climbing state and the risk level assessment; Performing fuzzy logic evaluation processing on the behavior analysis data to obtain safety compliance data; Intelligent early warning processing is performed on the safety compliance data and real-time monitoring data to obtain full life cycle management data, wherein the full life cycle management data includes real-time early warning information, intervention measure recommendations and safety incident reports.

2. The method for identifying a person's state based on target recognition detection according to claim 1, characterized in that: The parameterization processing of the collected original image data to obtain intelligent BIM data includes: Performing image denoising processing on the original image data to obtain denoised image data, and performing histogram equalization processing on the denoised image data to obtain enhanced image data; Performing geometric correction processing on the enhanced image data to obtain corrected image data, and performing illumination compensation processing on the corrected image data to obtain compensated image data; Performing feature extraction processing on the compensated image data through a convolutional neural network to obtain image feature data, and performing feature mapping processing on the image feature data to obtain BIM feature data; Performing spatial information fusion processing on the BIM feature data to obtain fused feature data, and visualizing the fused feature data through graphical processing to obtain visualized BIM data; The visualized BIM data is subjected to semantic segmentation processing to obtain segmented BIM data, and the segmented BIM data is subjected to parametric modeling processing through attribute association processing to obtain intelligent BIM data.

3. The method for identifying a person's state based on target recognition detection according to claim 1, characterized in that: The multidisciplinary data exchange processing is performed on the intelligent BIM data to obtain target detection data, wherein the target detection data includes personnel location information, safety equipment identification results and dangerous area markings, including: Performing data format conversion processing on the intelligent BIM data to obtain standardized BIM data, and performing data classification processing on the standardized BIM data to obtain classified BIM data; Perform feature extraction processing on the classified BIM data to obtain a BIM feature vector, and perform target detection processing on the BIM feature vector using a YOLOv5 algorithm to obtain candidate target data; Performing non-maximum suppression processing on the candidate target data to obtain screening target data, and performing coordinate mapping processing on the screening target data to obtain spatial positioning data; Performing target classification processing on the spatial positioning data to obtain classified target data, and performing security device identification processing on the classified target data through a deep learning algorithm to obtain device identification data; The device identification data is processed by dividing the device identification data into dangerous areas to obtain area marking data, and the area marking data is processed by data integration to obtain target detection data.

4. The method for identifying a person's state based on target recognition detection according to claim 1, characterized in that: The intelligent posture analysis and processing of the human body information in the target detection data to obtain optimized posture data includes: Performing human region extraction processing on the target detection data to obtain human ROI data, and performing feature extraction processing on the human ROI data through a multi-scale feature pyramid network to obtain multi-scale feature data; Performing key point positioning processing on the multi-scale feature data by using an improved HRNet algorithm to obtain two-dimensional key point data, and performing time series information fusion processing on the two-dimensional key point data to obtain continuous posture data; Performing occlusion inference processing on the continuous posture data to obtain completed posture data, and performing three-dimensional reconstruction processing on the completed posture data through a monocular depth estimation algorithm to obtain three-dimensional posture data; Performing posture angle calculation processing on the three-dimensional posture data to obtain angle feature data, and performing motion trajectory analysis processing on the angle feature data to obtain motion feature data; The action feature data is subjected to confidence evaluation processing to obtain credibility data, and the credibility data is subjected to data fusion processing to obtain optimized posture data.

5. The method for identifying a person's state based on target recognition detection according to claim 1, characterized in that: The optimized posture data is subjected to behavior recognition processing to obtain behavior analysis data, wherein the behavior analysis data includes the type of illegal behavior, the judgment of the climbing state and the risk level assessment, including: Performing a spatiotemporal graph construction process on the optimized posture data to obtain posture graph structure data, and performing feature extraction process on the posture graph structure data through a spatiotemporal graph convolutional network to obtain spatiotemporal feature data; Performing multi-level convolution processing on the spatiotemporal feature data to obtain high-level action feature data, and performing weight distribution processing on the high-level action feature data through an attention mechanism to obtain weighted feature data; Performing behavior classification processing on the weighted feature data to obtain violation behavior type data, and performing climbing state judgment processing on the violation behavior type data to obtain climbing state data; Performing height estimation processing on the climbing status data to obtain height information data, and performing safety risk assessment processing on the height information data through a risk assessment model to obtain risk level data; The risk level data is subjected to data integration processing to obtain behavior analysis data.

6. The method for identifying a person's state based on target recognition detection according to claim 1, characterized in that: The performing fuzzy logic evaluation processing on the behavior analysis data to obtain safety compliance data includes: Performing fuzzy rule base construction processing on the behavior analysis data to obtain fuzzy rule set data, and performing input variable fuzzification processing on the fuzzy rule set data to obtain input fuzzy set data; Performing fuzzy inference processing on the input fuzzy set data through an improved Mamdani fuzzy inference system to obtain inference result data, and dynamically adjusting the weight of the inference result data to obtain weighted result data; Performing fuzzy output generation processing on the weighted result data to obtain fuzzy output set data, and defuzzifying the fuzzy output set data by a centroid method to obtain accurate scoring data; Performing multi-dimensional evaluation processing on the precise scoring data to obtain multi-dimensional evaluation data, and performing comprehensive analysis processing on the multi-dimensional evaluation data to obtain comprehensive evaluation data; The comprehensive evaluation data is subjected to safety compliance calculation processing to obtain safety compliance data.

7. The method for identifying a person's state based on target recognition detection according to claim 1, characterized in that: The intelligent early warning processing of the safety compliance data and the real-time monitoring data is performed to obtain the full life cycle management data, wherein the full life cycle management data includes real-time early warning information, intervention measures suggestions and safety incident reports, including: Performing data fusion processing on the safety compliance data and the real-time monitoring data to obtain fusion status data, and performing risk level assessment processing on the fusion status data through a random forest algorithm to obtain dynamic risk data; Performing a multi-threshold warning mechanism setting process on the dynamic risk data to obtain warning threshold data, and performing a warning trigger judgment process on the warning threshold data to obtain warning trigger data; Performing personalized warning information processing on the warning trigger data to obtain personalized warning data, and performing warning information generation processing on the personalized warning data through a multimodal output algorithm to obtain multimodal warning data; Performing intelligent intervention decision processing on the multimodal warning data through a decision tree algorithm to obtain intervention suggestion data, and performing remote expert support processing on the intervention suggestion data to obtain expert guidance data; The expert guidance data is subjected to closed-loop feedback processing to obtain optimization strategy data, and event analysis processing is performed on the optimization strategy data through a data mining algorithm to obtain full life cycle management data.

8. A device for identifying a person's state based on target recognition detection, used to implement the method for identifying a person's state based on target recognition detection as claimed in any one of claims 1 to 7, characterized in that: The personnel status recognition device based on target recognition detection includes: The acquisition module is used to perform parameter processing on the acquired original image data to obtain intelligent BIM data; An exchange module is used to perform multidisciplinary data exchange processing on the intelligent BIM data to obtain target detection data, wherein the target detection data includes personnel location information, safety equipment identification results and dangerous area markings; An analysis module, used for performing intelligent posture analysis on human body information in the target detection data to obtain optimized posture data; An identification module, used to perform behavior identification processing on the optimized posture data to obtain behavior analysis data, wherein the behavior analysis data includes the type of illegal behavior, the judgment of the climbing state and the risk level assessment; An evaluation module, used for performing fuzzy logic evaluation processing on the behavior analysis data to obtain safety compliance data; The early warning module is used to perform intelligent early warning processing on the safety compliance data and real-time monitoring data to obtain full life cycle management data, wherein the full life cycle management data includes real-time early warning information, intervention measures suggestions and safety incident reports.

Images