Safety monitoring method and system for building construction

Abstract

The embodiment of the invention discloses a safety monitoring method and system for building construction, and the method comprises the steps: collecting the original video data of a construction site, carrying out the denoising, illumination correction and frame rate adjustment, and outputting a standardized video data stream; extracting attitude features of the constructors and representing the attitude features as a feature matrix to form an attitude feature matrix set containing spatio-temporal information; the sensitivity and correlation of the feature matrix are analyzed, calibration noise is added after dimension reduction, and feature data conforming to differential privacy are generated; and extracting data advanced representation by utilizing a pre-training model, and completing behavior classification, dangerous area judgment and safety violation detection. And the violation risk is evaluated in combination with the risk level of the construction area, graded early warning is generated, and meanwhile violation information is recorded to form a traceable management mechanism. According to the embodiment of the invention, accurate and efficient safety violation behavior detection is realized, and intelligent technical support is provided for safety management of building construction.

Description

Technical Field

[0001] This application relates to the field of building construction safety monitoring technology, specifically to a safety monitoring method and system for building construction. Background Technology

[0002] Construction safety monitoring is a crucial aspect of construction project management. Its purpose is to promptly detect and prevent potential safety accidents during construction, ensuring the safety of construction workers. With the development of computer vision and artificial intelligence technologies, intelligent video surveillance systems have been widely applied to safety management at construction sites.

[0003] Currently, common construction safety monitoring technologies mainly include manual inspection methods based on traditional video surveillance and automatic detection methods based on computer vision. The former relies on manual real-time viewing of surveillance footage for safety hazard identification, which is inefficient and easily affected by human factors; the latter uses image recognition algorithms to automatically detect dangerous behaviors, such as not wearing a safety helmet or not wearing a safety belt while working at heights. More advanced construction safety monitoring technologies use deep learning methods to analyze surveillance videos in real time, automatically detecting dangerous behaviors through human posture estimation and behavior recognition algorithms. This technology extracts complete image information of personnel from the video for analysis and processing. Although it shows high accuracy in safety incident detection, it poses a serious risk of privacy leakage in practical applications.

[0004] However, while these technologies improve the efficiency of security monitoring, they also raise issues regarding the privacy of construction workers. Because they require capturing and processing complete image data containing personal features such as facial features and body posture, the leakage of personal privacy information is easily possible. Existing privacy protection methods, such as image blurring or region occlusion, can protect privacy to some extent, but they often significantly reduce the accuracy of security incident detection, failing to provide quantifiable privacy guarantees and making it difficult to achieve an effective balance between privacy protection and security supervision. Summary of the Invention

[0005] To address the risks of privacy breaches and the difficulty in balancing privacy protection with detection accuracy in existing construction safety monitoring technologies, this application proposes a safety monitoring method and system for construction.

[0006] The technical solution of this application embodiment is as follows: In a first aspect, embodiments of this application provide a safety monitoring method for building construction, including: Collect raw video data from the construction site, perform noise reduction, illumination correction, and frame rate adjustment on the raw video data, and output a standardized video data stream; Receive the standardized video data stream, use the human key point detection algorithm to extract the posture features of the construction workers, represent the posture of each detected person as a feature matrix, and output a set of posture feature matrices containing spatiotemporal information; The sensitivity and intrinsic correlation of the feature matrices in the attitude feature matrix set are analyzed. Principal component analysis is applied to the attitude feature matrix set to reduce the dimensionality and obtain the dimensionality-reduced feature principal components. The optimal privacy budget allocation strategy is determined. By adding calibrated Laplace noise or Gaussian noise to the dimensionality-reduced feature principal components, privacy-preserving feature data that satisfies ε-differential privacy guarantee is generated. Based on the privacy-preserving feature data, a high-level representation of the spatiotemporal features is extracted through a pre-trained spatiotemporal graph convolutional network model, and behavior classification and dangerous area determination are performed to output the detection results of security violation events. Based on the detection results of the safety violations, the risk level of each violation is assessed in conjunction with the pre-configured risk level information of the construction area, and a graded early warning signal is generated. The violation is recorded and stored in the safety management database to form a traceable safety management mechanism.

[0007] Secondly, embodiments of this application also provide a safety monitoring system for building construction, comprising: The video acquisition and preprocessing module is used to acquire raw video data from the construction site, perform noise reduction, illumination correction, and frame rate adjustment on the raw video data, and output a standardized video data stream. The posture feature extraction module is used to receive the standardized video data stream, extract the posture features of the construction workers using the human key point detection algorithm, represent the posture of each detected person as a feature matrix, and output a set of posture feature matrices containing spatiotemporal information. The differential privacy protection module is used to analyze the sensitivity and intrinsic correlation of the feature matrices in the attitude feature matrix set, apply principal component analysis to reduce the dimensionality of the attitude feature matrix set to obtain the dimensionality-reduced feature principal components, determine the optimal privacy budget allocation strategy, and generate privacy-protected feature data that satisfies ε-differential privacy guarantee by adding calibrated Laplacian noise or Gaussian noise to the dimensionality-reduced feature principal components. The violation detection module is used to extract high-level representations of spatiotemporal features based on the privacy-preserving feature data through a pre-trained spatiotemporal graph convolutional network model, perform behavior classification and dangerous area determination, and output the detection results of safety violation events. The safety management module is used to assess the risk level of each violation event based on the detection results of the safety violation events and in combination with the pre-configured risk level information of the construction area, and generate graded early warning signals. The violation behavior is recorded and stored in the safety management database to form a traceable safety management mechanism.

[0008] Compared with the prior art, the beneficial effects of the embodiments of this application are as follows: 1. This application's embodiment introduces a differential privacy protection mechanism for the first time in building construction safety monitoring. It directly performs privacy processing on the attitude feature matrix at the data acquisition end, achieving quantifiable privacy guarantees. At the same time, it innovatively designs an optimal privacy budget allocation strategy based on the intrinsic correlation analysis of the feature matrix, ensuring that key behavioral features can still retain the main structural information after adding random noise. 2. The embodiments of this application design a risk-level-based hierarchical early warning system, which, combined with a traceable safety management database, achieves efficient safety control at construction sites. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the safety monitoring method for building construction in this embodiment; Figure 2 This is a structural schematic diagram of the safety monitoring system for building construction according to this application. Detailed Implementation

[0011] The preferred embodiments of the present application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only for illustration and explanation of the embodiments of the present application and are not intended to limit the embodiments of the present application.

[0012] Example 1: like Figure 1 As shown in the figure, this application provides a safety monitoring method for building construction, including: Step S1: Collect raw video data from the construction site, perform noise reduction, illumination correction, and frame rate adjustment on the raw video data, and output a standardized video data stream; Step S2: Receive the standardized video data stream, extract the posture features of the construction workers using the human key point detection algorithm, represent the posture of each detected person as a feature matrix, and output a set of posture feature matrices containing spatiotemporal information. Step S3: Analyze the sensitivity and intrinsic correlation of the feature matrices in the attitude feature matrix set, apply principal component analysis to reduce the dimensionality of the attitude feature matrix set to obtain the dimensionality-reduced feature principal components, determine the optimal privacy budget allocation strategy, and generate privacy-preserving feature data that satisfies ε-differential privacy guarantee by adding calibrated Laplace noise or Gaussian noise to the dimensionality-reduced feature principal components. Step S4: Based on the privacy-preserving feature data, extract high-level representations of spatiotemporal features through a pre-trained spatiotemporal graph convolutional network model, perform behavior classification and dangerous area determination, and output the detection results of security violation events; Step S5: Based on the detection results of the safety violation events, the risk level of each violation event is assessed in conjunction with the pre-configured risk level information of the construction area, and a graded early warning signal is generated. The violation is recorded and stored in the safety management database to form a traceable safety management mechanism.

[0013] Specifically, by deploying high-definition network cameras in key areas of the construction site (such as high-altitude work areas, hazardous machinery operation areas, and material hoisting areas), a multi-angle coverage monitoring network is formed to collect real-time video data streams. The collected raw video data contains complete visual information of the construction site, providing a data foundation for subsequent intelligent analysis.

[0014] The acquired raw video data undergoes preprocessing operations such as denoising, illumination correction, and frame rate adjustment. First, time synchronization is performed to ensure the consistency of timestamps across multiple video sources, facilitating subsequent multi-angle behavioral analysis. Then, adaptive image enhancement algorithms are applied to optimize the video data quality, including denoising, illumination unevenness correction, and contrast adjustment, thereby improving image quality. Finally, the processed video undergoes frame rate adjustment and resolution normalization, outputting a standardized video data stream that meets the requirements of subsequent analysis.

[0015] The system receives standardized video data streams and extracts the pose features of construction workers using human keypoint detection algorithms. First, an improved YOLOv5 algorithm is applied to detect and locate each worker in the video. Then, the HRNet pose estimation model is applied to each detected worker region to extract the spatial coordinates of 17 keypoints (such as head, shoulder, elbow, wrist, hip, knee, and ankle). The pose of each detected worker is represented as a feature matrix. Based on the extracted keypoint coordinates, a human pose feature matrix is ​​constructed, with each matrix containing the 2D / 3D coordinates of the keypoints and their confidence scores, forming a mathematical representation of the worker's pose. Combining the temporal information of consecutive video frames, a pose sequence containing spatiotemporal features is constructed to capture dynamic behavioral features, outputting a set of pose feature matrices containing spatiotemporal information.

[0016] The sensitivity and intrinsic correlation of the pose feature matrix set are analyzed. Upon receiving the pose feature matrix set, the sensitivity of the feature matrix is ​​analyzed to determine sensitive and non-sensitive parameters for privacy protection. Sensitive parameters typically include features that can directly or indirectly identify an individual, such as the precise location of facial key points; non-sensitive parameters include features that play an important role in behavior recognition but are not directly related to individual identity, such as the relative positional relationships of limbs.

[0017] Principal component analysis (PCA) is applied to reduce the dimensionality of the pose feature matrix set, yielding the dimensionality-reduced principal components. PCA is used to extract the principal components most critical for behavior recognition, reducing data dimensionality while retaining key information. By calculating the feature covariance matrix and extracting eigenvalues ​​and eigenvectors, the top K principal components with a cumulative contribution rate reaching a preset threshold (e.g., 95%) are selected, and the original high-dimensional pose features are projected into a low-dimensional principal component space.

[0018] Determining the optimal privacy budget allocation strategy involves designing a differential privacy noise addition mechanism based on the sensitivity and inherent correlation of the feature matrix, and then identifying the optimal privacy budget allocation strategy and noise scale. The allocation of the privacy budget ε needs to consider the impact of different feature dimensions on privacy protection and behavior recognition accuracy, and the best budget allocation scheme is found through optimization algorithms.

[0019] This embodiment generates privacy-preserving feature data that satisfies ε-differential privacy by adding calibrated Laplacian or Gaussian noise to the dimensionality-reduced principal components. The scale of the noise added to the dimensionality-reduced principal components is determined based on the privacy budget ε and the feature sensitivity, ensuring ε-differential privacy is met. Then, an inverse transform is applied to reconstruct the noisy principal components into a representation of the original feature space, outputting feature data that satisfies differential privacy.

[0020] Based on privacy-preserving feature data, a pre-trained spatiotemporal graph convolutional network (ST-GCN) model is used to extract high-level representations of spatiotemporal features and perform behavior classification and hazard area determination. The system receives privacy-preserving feature data and inputs it into a pre-trained ST-GCN model, specifically adapted to noisy pose feature data. The ST-GCN model extracts high-level representations of spatiotemporal features, capturing the temporal patterns and spatial structure information of human movements. It outputs safety violation event detection results. A multi-task learning framework is used to simultaneously perform behavior classification and hazard area determination, identifying various safety violations including not wearing seat belts, entering hazardous areas, and improper operation. Confidence thresholding and temporal consistency checks are applied to filter out possible false detections, improving recognition accuracy and outputting safety violation event detection results.

[0021] Next, based on the safety violation detection results and pre-configured construction area risk level information, the risk level of each violation is assessed, and tiered early warning signals are generated. The system receives safety violation detection results and, in conjunction with pre-configured construction area risk level information, assesses the risk level of each violation. Based on the risk level classification, different levels of early warning signals are generated: red (emergency), yellow (warning), and blue (alert), and the warning information is pushed to on-site management personnel via edge devices. Violations are recorded and stored in the safety management database, forming a traceable safety management mechanism. Details of detected violations (including violation type, time, location, risk level, etc.) are recorded in the safety management database, but original images or personally identifiable information are not stored. Based on the safety management database, construction safety analysis reports and trend charts are built to support safety management personnel in conducting safety situation analysis and formulating preventative safety measures.

[0022] In this embodiment, cameras are deployed in key areas of the construction site to form a multi-angle coverage monitoring network, collecting raw video data from the construction site to obtain raw multi-source video data; the raw multi-source video data is time-synchronized to output multi-source video data with consistent timestamps; an adaptive image enhancement algorithm is applied to the multi-source video data with consistent timestamps for denoising, illumination unevenness correction, and contrast adjustment to output enhanced video data; the enhanced video data is then subjected to frame rate adjustment and resolution standardization to obtain the standardized video data stream.

[0023] Specifically, high-definition network cameras are deployed in key areas of the construction site to form a multi-angle monitoring network. The camera deployment plan needs to comprehensively consider the actual layout of the construction site, the characteristics of the work, and the distribution of safety risks. Typically, monitoring cameras are deployed in key areas such as high-altitude work areas, hazardous machinery operation areas, material hoisting areas, tower crane operation areas, foundation pit edges, and temporary power supply areas. The installation height, tilt angle, and horizontal viewing angle of each camera are carefully calculated to ensure coverage of the target work area while obtaining clear images of personnel posture. Multiple cameras are configured with complementary viewing angles, effectively reducing blind spots and improving the completeness and reliability of personnel detection through cross-coverage of different angles. High-definition network cameras supporting PoE power supply are selected, with a resolution of no less than 1920×1080, a frame rate of 25fps or higher, good low-light performance, and a wide dynamic range to adapt to the complex and variable lighting conditions of the construction site. All cameras are connected to an edge computing server via wired or wireless networks to transmit video data streams in real time.

[0024] The raw video data collected in this step contains complete dynamic information about the construction site, including various operational scenarios such as personnel activities, equipment operation, and material handling, providing a rich data foundation for subsequent multi-angle collaborative behavior analysis and safety monitoring.

[0025] Next, the acquired raw multi-source video data undergoes time synchronization processing. Since the surveillance network contains multiple cameras, each with its own internal clock, clock drift and deviation are inevitable during long-term operation. Without time synchronization, inconsistent timestamps will exist between video frames acquired by different cameras, severely impacting subsequent functions such as cross-camera personnel trajectory correlation and multi-view behavior fusion analysis.

[0026] To address this issue, the system employs Network Time Protocol (NTP) or Precise Time Protocol (PTP) for time calibration of all cameras. Specifically, a standard time server is deployed on the edge computing server, and each camera periodically synchronizes its clock with the time server to correct local clock deviations. For critical application scenarios, time synchronization accuracy must reach millisecond or even sub-millisecond levels to ensure that video frames captured by different cameras at the same time correspond precisely. Time synchronization processing also includes timestamping the video stream, attaching a uniform absolute timestamp to each frame to facilitate subsequent processing modules' time-series analysis and event correlation. The consistent timestamps of the output multi-source video data lay a solid foundation for advanced functions such as multi-angle collaborative analysis, personnel behavior trajectory tracking, and spatiotemporal correlation of abnormal events.

[0027] Adaptive image enhancement algorithms are applied to multi-source video data with consistent timestamps for denoising, illumination unevenness correction, and contrast adjustment. Construction sites present extremely complex environmental conditions with numerous interfering factors affecting video image quality. For example, dust and water mist generated during construction reduce image clarity; dense objects such as steel bars, formwork, and scaffolding cause frequent obstructions; outdoor construction is affected by natural lighting, with drastic changes in light intensity and direction at different times and under different weather conditions, easily leading to localized overexposure or underexposure; and uneven artificial lighting during nighttime construction creates shadow areas with strong contrast between light and dark.

[0028] The factors mentioned above can significantly impact the accuracy of subsequent personnel detection and pose estimation. To improve image quality, the system employs a multi-level adaptive image enhancement algorithm to preprocess the video data. The denoising process utilizes a deep learning-based video denoising network, specifically trained on a large-scale construction scene dataset. This network effectively identifies and removes noise interference while preserving important details such as personnel contours and edges in the image.

[0029] First, the illumination non-uniformity correction employs an improved Retinex algorithm combined with adaptive histogram equalization. The image is first decomposed into reflection and illumination components. Then, the illumination component is smoothed to eliminate the influence of local illumination variations. Finally, the image is resynthesized to obtain a uniformly illuminated image. This method effectively improves the visual effect of local shadow and bright areas, making the overall brightness distribution of the image more uniform and reasonable. Contrast adjustment uses adaptive limited contrast histogram equalization (CLAHE), dynamically adjusting the contrast enhancement parameters according to the characteristics of different regions of the image. This improves the distinction between the target and background while avoiding noise amplification caused by over-enhancement. The output video data after these three levels of enhancement processing has higher image quality, better visual clarity, and stronger feature representation capabilities, providing high-quality input data for subsequent intelligent analysis algorithms.

[0030] Secondly, the enhanced video data undergoes frame rate adjustment and resolution standardization. Since the surveillance network may contain cameras of different models and manufacturers, their factory configurations and acquisition parameters often differ. Some cameras use a frame rate of 25fps, while others use 30fps or higher; some output 1080p resolution, while others output 720p or 4K resolution. These parameter differences can complicate subsequent batch processing and feature extraction, and also hinder the fusion analysis of multiple video streams. To address this issue, the system standardizes all video streams, unifying them to preset standard parameters.

[0031] In this embodiment, frame rate adjustment employs different strategies based on the relationship between the original video's frame rate and the target frame rate: when the original frame rate is higher than the target frame rate, uniform frame extraction is used, selecting video frames at fixed intervals; when the original frame rate is lower than the target frame rate, motion-compensated frame interpolation is used, inserting intermediate frames on the timeline by analyzing the motion vectors between adjacent frames to ensure the video's temporal continuity and smooth motion. This system typically sets the target frame rate to 25fps or 30fps, which satisfies the temporal resolution requirements of human motion capture without incurring excessive data processing burden. Resolution normalization uses high-quality bicubic interpolation or Lanczos interpolation algorithms to scale the image. When resolution reduction is required, anti-aliasing filtering is performed before downsampling to avoid moiré patterns and jagged edges.

[0032] When resolution needs to be increased, deep learning-based super-resolution reconstruction techniques are employed to increase pixel density while maintaining image clarity. This system typically sets the target resolution to a uniform 1920×1080, which provides sufficient image detail for accurate human keypoint detection while maintaining reasonable computational and storage overhead. After frame rate adjustment and resolution standardization, the output standardized video data stream has uniform spatiotemporal parameters. All video streams are updated at the same frame rate and output with the same image size, facilitating batch and parallel feature extraction and behavior analysis by subsequent processing modules, significantly improving the overall system's processing efficiency and analytical performance.

[0033] In this embodiment, the standardized video data stream is received, and the YOLOv5 algorithm is applied for personnel detection to output personnel location information. Based on each personnel region corresponding to the personnel location information, the HRNet pose estimation model is applied to extract the spatial coordinates of human key points and output the human key point coordinates. A human pose feature matrix containing key point coordinates and corresponding confidence scores is constructed based on the human key point coordinates, and a single-frame pose feature matrix is ​​output. Based on the single-frame pose feature matrix, a pose sequence is constructed by combining the temporal information of continuous video frames to obtain the pose feature matrix set containing spatiotemporal information.

[0034] Specifically, the system receives standardized video data streams and applies the YOLOv5 algorithm for personnel detection. YOLOv5 is one of the most widely used real-time object detection algorithms. Its core advantage lies in unifying the object detection task into a single-stage regression problem, predicting both the target's location and category in a single forward propagation. Compared to traditional two-stage detection algorithms, it offers faster detection speeds and lower computational latency, making it particularly suitable for real-time video surveillance scenarios. The YOLOv5 algorithm uses CSPDarknet53 as its backbone network for feature extraction. This network effectively reduces computation through a cross-stage partial connection (CSP) structure while maintaining strong feature representation capabilities. During feature extraction, the network extracts image features at multiple scales, forming a feature pyramid, which can effectively handle personnel targets of varying distances and sizes in construction sites.

[0035] The detection head employs a decoupled design, separating classification and regression tasks to improve detection accuracy. For example, in construction scenarios, the YOLOv5 algorithm, after specialized domain-adaptive training, is fine-tuned using a dataset of construction site images containing various work postures, different lighting conditions, and complex background interference, significantly improving the detection accuracy of construction workers wearing safety helmets and reflective vests. During detection, the algorithm analyzes each frame of the input video image and outputs the bounding box information for all detected personnel targets. Each bounding box is represented by four parameters (x, y, w, h), where (x, y) is the position of the bounding box center point in the image coordinate system, and w and h are the width and height of the bounding box, respectively. These four parameters fully define the spatial location and size range of the personnel in the image. In addition to the position parameters, the algorithm also outputs a confidence score for each detection box, which reflects the model's estimate of the reliability of the detection result. The score ranges from 0 to 1, with a higher score indicating a greater probability that the area indeed contains a personnel target. The system sets a confidence threshold (usually 0.5 or 0.6), retaining only detection results with a confidence level exceeding the threshold and filtering out unreliable false detections. Therefore, the output personnel location information provides accurate personnel localization for the subsequent pose estimation module. The pose estimation algorithm only needs to analyze within these personnel areas, without processing the entire image, greatly improving processing efficiency and accuracy.

[0036] Furthermore, based on the personnel location information corresponding to each personnel region, the HRNet pose estimation model is applied to extract the spatial coordinates of human keypoints. HRNet (High-ResolutionNet) is a deep learning model that has made groundbreaking progress in the field of human pose estimation in recent years. Its design philosophy is fundamentally different from the traditional encoder-decoder structure. Traditional pose estimation networks typically downsample the input image to a low resolution for feature extraction, and then upsample it to restore a high-resolution output heatmap. This "downsampling then upsampling" approach leads to the loss of spatial detail information, affecting the accuracy of keypoint localization. HRNet innovatively adopts a parallel multi-resolution sub-network structure, maintaining high-resolution feature representation throughout the forward propagation of the entire network, while processing feature maps of different resolutions in parallel. It achieves information interaction between high and low resolution features through a multi-scale feature fusion module. This design allows the network to retain rich spatial details for accurate localization and obtain sufficient semantic information for keypoint recognition, significantly improving the accuracy of pose estimation, especially for small-scale targets and complex poses.

[0037] In practical applications, for each bounding box detected by YOLOv5, the corresponding person region image is first cropped from the original image based on the bounding box coordinates. Then, the bounding box range is appropriately expanded (usually by 10%-20%) to ensure the complete human body is included, avoiding keypoint loss due to boundary cropping. The cropped person image is then resized to the standard input size of the HRNet model (e.g., 256×192 or 384×288), maintaining the aspect ratio to avoid distortion. The resized image is input into the HRNet model for forward inference, and the model outputs a heatmap of 17 human body keypoints. Each keypoint corresponds to a heatmap, and the values ​​in the heatmap represent the probability that the location is a keypoint. These 17 keypoints are defined according to the standard COCO dataset, including the tip of the nose, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, and right ankle, covering the main joints and important parts of the human body, and can completely describe the human body's posture structure. By finding the location of maximum response on the heatmap of each keypoint, the two-dimensional coordinates of the keypoint in the cropped image coordinate system are obtained. Then, according to the inverse transformation of cropping and scaling, the coordinates are mapped back to the original image coordinate system. The output human keypoint coordinates are represented in pixel coordinates, with each keypoint containing (x, y) coordinate values, accurately pinpointing the location of the keypoint in the original video image.

[0038] A human pose feature matrix is ​​constructed based on the coordinates of human keypoints, containing the keypoint coordinates and corresponding confidence scores. The human pose feature matrix is ​​a structured mathematical representation of human pose information, organizing scattered keypoint coordinates into a unified matrix form, facilitating subsequent mathematical operations and machine learning algorithm processing. For the pose of a single person at a single moment, the feature matrix is ​​defined as 17×3, where 17 rows correspond to 17 human keypoints, and 3 columns correspond to the three attributes of each keypoint: x-coordinate, y-coordinate, and confidence score. The x-coordinate and y-coordinate are directly derived from the output of the HRNet pose estimation model, representing the pixel position of the keypoint in the image. The confidence score is an estimate of the accuracy of the HRNet model in locating the keypoint. It is calculated by extracting the maximum response value from the heatmap corresponding to the keypoint, or by calculating the weighted average of the area surrounding the peak in the heatmap. The confidence score ranges from [0, 1]. A value closer to 1 indicates greater confidence in the model's location of the keypoint, resulting in a more reliable location; a value closer to 0 indicates that the keypoint may be occluded, outside the image range, or that the model cannot accurately locate it.

[0039] In real-world construction scenarios, due to diverse personnel postures and frequent occlusion, some key points often exhibit low confidence levels. For example, when working in a squatting position, lower limb key points may be obscured by tools, and during high-altitude operations, certain key points may be difficult to locate accurately due to perspective issues. Confidence scores provide a quantitative indicator of data quality for subsequent processing. Different processing strategies can be adopted for different key points based on their confidence levels, allowing for special handling or removal of low-confidence key points to prevent unreliable data from interfering with behavior recognition.

[0040] For example, the matrix construction process is as follows: First, create an empty matrix of 17 rows and 3 columns. Then, following the standard order of key points (from the tip of the nose to the right ankle), fill in the x-coordinate of each key point into the corresponding row of the first column, the y-coordinate into the corresponding row of the second column, and the confidence score into the corresponding row of the third column, thus forming a complete posture feature matrix. This matrix fully encodes the body posture state of a single person at a single moment, containing the spatial location information and location reliability information of key points, providing a rich spatial feature foundation for subsequent temporal analysis and behavior recognition.

[0041] Based on the single-frame pose feature matrix, a pose sequence is constructed by combining the temporal information of consecutive video frames. Human actions and behaviors have obvious temporal characteristics, and static pose information at a single moment is often insufficient to accurately determine the type of behavior. For example, a person raising their arm while standing could be waving, performing work at height, or carrying an object; this cannot be distinguished from a single-frame pose alone. Only by combining pose changes over time and analyzing the temporal evolution pattern of the action can the behavior be accurately identified. To capture these dynamic behavioral characteristics, the system employs a sliding time window mechanism to organize the pose feature matrices of multiple consecutive frames into a temporal sequence.

[0042] The specific implementation process is as follows: First, determine the length T of the time window. This parameter needs to be set according to the time scale of the target behavior. For fast actions (such as falling or climbing), a shorter window (such as 30 frames, corresponding to 1.2 seconds in a 25fps video) is usually selected, while for slow actions (such as improper operation or prolonged illegal stay), a longer window (such as 60 frames or more, corresponding to 2-3 seconds) is selected. Then, starting from the first frame of the video, extract the posture feature matrix of T frames continuously. Stack these T matrices in chronological order into a three-dimensional tensor with dimensions T×17×3. The first dimension represents the time series, the second dimension represents key points, and the third dimension represents coordinates and confidence. This three-dimensional tensor completely describes the spatiotemporal evolution of personnel posture within a time period. Through a sliding window mechanism, move backward by a certain step size (such as 5 or 10 frames) each time to extract the posture sequence within the new time window. This allows for the continuous generation of a large number of overlapping posture sequences, covering the entire video time period, and achieving continuous real-time monitoring of personnel behavior at the construction site.

[0043] Understandably, for multiple construction workers present simultaneously in a video, each worker's posture sequence is maintained independently. Personnel detection and tracking algorithms are used to associate the same worker across different frames, ensuring the temporal continuity of the posture sequence. The output posture feature matrix set, containing spatiotemporal information, is a four-dimensional data structure (number of samples × time length × number of keypoints × feature dimension). It simultaneously includes spatial structural information of human posture (relative positional relationships of keypoints, limb angles, etc.) and temporal evolution information of movement (trajectory, velocity, acceleration, etc. of keypoints over time), providing extremely rich feature representations for subsequent behavior recognition algorithms. This enables the system to accurately distinguish between various complex construction operations and safety violations.

[0044] In this embodiment, the key points of the human body include the head, shoulders, elbows, wrists, hips, knees, and ankles.

[0045] Specifically, the definition of key points for the human body includes 17 standard key points, arranged from top to bottom and from center to periphery: tip of nose, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, and right ankle. These key points cover the major joints and important parts of the human body, and can comprehensively describe the posture and movements of the human body. Head key points (tip of nose, eyes, ears) are used to determine the direction and posture of the head; shoulder, elbow, and wrist key points describe the movement of the upper limbs; and hip, knee, and ankle key points describe the position of the lower limbs and the body's center of gravity. Through the spatial relationship of these 17 key points, the body posture and movement type of construction workers can be accurately inferred, providing basic data for the identification of safety violations.

[0046] In this embodiment, the attitude feature matrix set is received, the sensitivity of the feature matrices in the attitude feature matrix set is analyzed, and the partitioning results of sensitive and insensitive parameters are output. Based on the partitioning results, principal component analysis is applied to the attitude feature matrix set to reduce its dimensionality, and the dimensionality-reduced principal component representation is output. Based on the dimensionality-reduced principal component representation, the intrinsic correlation of the feature matrices is analyzed, and a differential privacy noise addition mechanism is designed to determine the privacy budget allocation strategy and noise scale. According to the privacy budget allocation strategy and the noise scale, Laplace noise or Gaussian noise that satisfies ε-differential privacy is added to the dimensionality-reduced principal component representation, and the noisy principal components are output. The noisy principal components are reconstructed into a representation of the original feature space by applying an inverse transformation, and the privacy-preserving feature data is obtained.

[0047] Specifically, the process involves receiving a set of posture feature matrices and analyzing their sensitivity. A key aspect of privacy protection is accurately identifying which parts of the data contain sensitive personally identifiable information and which only reflect behavioral patterns. Sensitivity analysis is a systematic evaluation process that requires quantitative assessment of each feature dimension of the posture feature matrix from multiple dimensions. First, the information entropy of each feature dimension is calculated. Information entropy reflects the richness of information and individual variability of the feature. A higher entropy value indicates greater variation among different individuals, making it more likely to be used for individual differentiation and identity recognition. Information entropy is calculated based on the distribution of feature values ​​in the sample set, quantified by statistically analyzing the probability of occurrence in different numerical ranges and applying the Shannon entropy formula. Second, the correlation between feature dimensions and personal identity is analyzed, which is evaluated by training a specialized identity recognition model. Specifically, an identity classifier is trained for each feature dimension or feature combination using a training dataset containing known identity labels, and the contribution of that feature to the accuracy of identity recognition is measured. If a feature or feature combination can effectively predict personal identity, then that feature has high identity sensitivity.

[0048] For example, the precise combination of head landmarks can reflect the geometric features of a face and may be used for identity recognition, thus exhibiting high identity-relatedness. The third evaluation dimension is the contribution to behavior recognition, quantifying the impact of each feature on behavior classification accuracy through feature importance analysis. Common methods include gradient-based feature attribution techniques, which calculate the gradient of the behavior recognition model's classification result with respect to each input feature; features with larger absolute gradient values ​​have a greater impact on classification decisions and are more important. Another method is permutation importance assessment, which randomly shuffles the values ​​of a feature dimension and observes the degree of performance degradation; a more significant performance degradation indicates a more important feature. Combining the analysis results of these three dimensions, pose features are divided into two main categories: sensitive parameters and insensitive parameters.

[0049] The sensitive parameters primarily include key head-related information, specifically the precise coordinates of five key points: the tip of the nose, left eye, right eye, left ear, and right ear. The spatial distribution patterns of these key points are closely related to the geometric features of the face. Attackers may use this information to reconstruct the facial geometry and combine it with other information from the video for identity verification. The precise locations of these head key points may also reveal an individual's biometric information, such as head size and the distance between facial features. This information is unique to each individual and can be used for cross-scene identity tracking.

[0050] In contrast, non-sensitive parameters primarily include information on key points of the torso and limbs, such as the shoulders, elbows, wrists, hips, knees, and ankles. While these key points also exhibit individual differences, their main value lies in describing human posture and movement patterns, rather than personal identity characteristics. More importantly, non-sensitive parameters contain a wealth of relative positional relationships and geometric information, such as the ratio of shoulder width to arm length, the angle between the upper arm and forearm, and the tilt of the torso. These relative features have limited use in identity recognition but are crucial for behavior recognition.

[0051] In summary, this embodiment, through this refined sensitivity classification, allows for differentiated strategies in subsequent privacy protection processing. Stronger privacy protection is applied to sensitive parameters, while weaker protection is applied to non-sensitive parameters, thus achieving an optimal balance between privacy protection and data utility. The output results of the sensitive and non-sensitive parameter classification are stored in the form of structured metadata, with each feature dimension labeled with a sensitivity tag and sensitivity score, providing precise guidance for subsequent differentiated privacy protection strategies.

[0052] Based on the partitioning results, principal component analysis (PCA) is applied to reduce the dimensionality of the pose feature matrix set. PCA is one of the most classic dimensionality reduction techniques in statistics and machine learning, and its mathematical principle is based on eigenvalue decomposition in linear algebra. The core idea of ​​PCA is to transform potentially correlated high-dimensional features into a set of linearly independent low-dimensional features through orthogonal linear transformations. These new features are called principal components, and they are arranged in order of the amount of information they contain.

[0053] In this model, the first principal component (PCI) is a linear combination of the original features, exhibiting the largest variance among all possible linear combinations, meaning it captures the most prevalent patterns of change in the data. The second PCI is also a linear combination of the original features, but it is orthogonal to the first PCI and has the largest variance among all directions orthogonal to the first PCI. Subsequent PCIs follow the same pattern, each orthogonal to all preceding PCIs and exhibiting the largest variance among the remaining orthogonal directions. In privacy-preserving scenarios involving pose features, applying PCI for dimensionality reduction has multiple important values.

[0054] First, dimensionality reduction can significantly reduce the feature dimension. The original pose feature matrix is ​​high-dimensional, with each human body containing seventeen keypoints, each keypoint containing three attributes, resulting in a 51-dimensional feature matrix. After principal component analysis (PCA), the dimensionality can typically be reduced to between 20 and 30 dimensions, significantly reducing data storage and computational overhead while retaining over 95% of the original information. Second, PCA transforms highly correlated original features into independent principal components, eliminating redundancy between features and making subsequent differential privacy noise addition more efficient. If noise is directly added to the original high-dimensional feature space, more noise needs to be added to achieve the same level of privacy protection due to the correlation between features.

[0055] In principal component space, since principal components are independent of each other, noise addition is more efficient, achieving the same privacy protection effect with less noise, thus providing a better trade-off between noise and utility. Third, principal component analysis (PCA) dimensionality reduction itself also has a certain privacy protection effect because principal components are linear combinations of the original features. Even if an attacker obtains the principal component data, it is difficult to accurately deduce the original keypoint coordinates. Especially when only some principal components are retained and smaller ones are discarded, some information from the original features is permanently lost, increasing the strength of privacy protection. The specific steps of PCA processing include data centering, covariance matrix calculation, eigenvalue decomposition, and feature projection. Data centering is a preprocessing step that calculates the mean of each feature dimension across all samples, and then subtracts the mean of the corresponding dimension from the feature values ​​of each sample, making the mean of the centered data a zero vector. This is the standard preprocessing step for PCA algorithms, ensuring that the direction of the principal components is not affected by the position of the feature mean.

[0056] Next, the covariance matrix of the centered data is calculated. The covariance matrix is ​​a symmetric matrix; its diagonal elements represent the variance of each feature, and its off-diagonal elements reflect the degree of linear correlation between features. Eigenvalue decomposition is performed on the covariance matrix to obtain several eigenvalues ​​and corresponding eigenvectors. The magnitude of the eigenvalues ​​directly reflects the variance contribution of the corresponding principal components. The eigenvalues ​​are sorted from largest to smallest, and the eigenvectors are arranged accordingly. The eigenvectors corresponding to the first few largest eigenvalues ​​are selected as the basis vectors of the principal components. The number of principal components is chosen based on a cumulative variance contribution rate threshold, typically requiring the sum of the variances of the first few principal components to account for 95% or 98% or more of the total variance. In practical applications, for 51-dimensional pose features, typically 25 to 30 principal components are selected to achieve a cumulative contribution rate of 98%, realizing nearly half the dimensionality reduction.

[0057] Finally, the original feature matrix is ​​multiplied by the transformation matrix composed of the selected principal component vectors to complete the projection transformation from the high-dimensional original space to the low-dimensional principal component space. The dimensionality-reduced principal component representation of the output is significantly reduced, but retains most of the information of the original features, laying the foundation for subsequent efficient privacy-preserving processing.

[0058] Based on the principal component representation after dimensionality reduction, the intrinsic correlation of the feature matrix is ​​analyzed, and a differential privacy noise addition mechanism is designed to determine the privacy budget allocation strategy and noise scale. Differential privacy is currently the most stringent and theoretically sound definition of privacy in the field of privacy protection. Its core idea is to add carefully calibrated random noise to the data or query results so that the impact of whether or not the data of any individual is included in the dataset on the final output result is statistically indistinguishable, thereby protecting individual privacy from leakage. The strength of differential privacy is controlled by the privacy budget parameter; the smaller the parameter value, the stronger the privacy protection, but at the same time, the greater the loss of data utility.

[0059] In the application scenario of this embodiment, it is necessary to find the optimal balance between strict privacy protection and effective behavior recognition. This requires designing a refined privacy budget allocation strategy and a noise scale calibration mechanism. First, the intrinsic correlation between the principal components after dimensionality reduction is analyzed. Although principal component analysis theoretically generates orthogonal principal components, in actual data, due to the limited sample size and the complexity of data distribution, nonlinear correlations or statistical dependencies may still exist between principal components. By calculating nonlinear correlation metrics such as mutual information and distance correlation coefficients between principal components, it is possible to identify which principal components encode similar information and which capture independent feature patterns. Another important purpose of correlation analysis is to identify feature redundancy. If some principal components are highly correlated, further merging or removal can be considered to achieve more thorough dimensionality reduction and redundancy removal.

[0060] Next, an adaptive differential privacy noise addition mechanism is designed. Unlike traditional uniform noise addition schemes, this embodiment proposes an adaptive mechanism that dynamically adjusts the noise intensity based on the characteristics of each principal component. The privacy budget allocation strategy comprehensively considers three factors. First, the sensitivity of the principal component is evaluated by analyzing the correlation strength between the principal component and the original sensitive features. If a principal component is mainly contributed by sensitive features, i.e., the coefficient of the principal component vector in the sensitive feature dimension is large, then the principal component has high privacy sensitivity and requires a larger privacy budget, meaning more noise is added. Second, the importance of the principal component is measured by the magnitude of its corresponding eigenvalue and its contribution to the behavior recognition task. Principal components with larger eigenvalues ​​or a greater impact on behavior recognition accuracy are considered more important. To maintain data utility, smaller privacy budgets are allocated to these important principal components, meaning less noise is added. Third, the stability of the principal component is evaluated through cross-validation, analyzing the consistency of the principal component across different data subsets. Principal components with high stability have a stronger tolerance to noise disturbances, and the amount of noise added can be appropriately increased.

[0061] Based on a comprehensive score of these three factors, a personalized privacy budget is assigned to each principal component, typically ranging from 0.1 to 2.0, while simultaneously satisfying the overall privacy budget constraint. The determination of the noise scale strictly follows differential privacy theory. For the Laplace mechanism, the noise scale parameter equals the function sensitivity divided by the privacy budget, where the function sensitivity is the maximum difference between any two adjacent datasets on that principal component. The calculation of the sensitivity requires theoretical derivation or empirical estimation combining the principal component analysis transformation matrix and the value range of the original data. For the Gaussian mechanism, the noise standard deviation is calculated based on the privacy budget and the privacy failure probability parameter, which is typically set to a very small value.

[0062] In practical applications, the choice between the Laplace and Gaussian mechanisms depends on specific privacy requirements and computational efficiency considerations. The Laplace mechanism is simple to implement and has a clear theoretical analysis, while the Gaussian mechanism offers a better noise-to-utility tradeoff in certain situations, particularly when combining multiple privacy-preserving operations, resulting in better privacy amplification. The output privacy budget allocation strategy is represented in vector form, as is the noise scale, providing precise parameter configuration for the subsequent noise addition operation.

[0063] Laplace or Gaussian noise, satisfying differential privacy guarantees, is added to the dimensionality-reduced principal component representation based on the privacy budget allocation strategy and noise scale. Noise addition is the core execution step of the differential privacy protection mechanism. By injecting carefully calibrated random noise into the data, the true value of individual data points is statistically masked, making it impossible for attackers to accurately infer the original information of any individual even if they obtain the noisy data. For the dimensionality-reduced principal component representation matrix, the noise addition operation is performed independently for each principal component dimension. Specifically, for each principal component, several independent and identically distributed random noise values ​​are generated based on the noise type and noise scale parameters determined in the previous step. If Laplace noise is used, the noise values ​​are sampled from a Laplace distribution, which has zero mean and a specific variance. Its peaked and thick-tailed shape means that most noise values ​​are concentrated near zero, while a small number of noise values ​​are distributed over a larger range. This characteristic helps to reduce the impact on the overall statistical properties of the data while protecting privacy. If Gaussian noise is used, the noise values ​​are sampled from a Gaussian distribution, which also has zero mean and a specific variance. Its symmetrical bell-shaped curve makes the noise distribution more uniform and smooth.

[0064] Understandably, the process of generating noise values ​​uses a cryptographically secure random number generator to ensure the true randomness and unpredictability of the noise, which is a crucial prerequisite for differential privacy security. After generating a sufficient number of noise values, these noise values ​​are added one by one to the sample values ​​of the corresponding principal components to obtain noisy principal components. This operation is repeated for all principal components, ultimately resulting in a fully noisy principal component matrix. The noise addition process strictly follows the combinatorial theorem of differential privacy; if differential privacy protection is applied to multiple principal components separately, the overall data satisfies the corresponding differential privacy guarantees.

[0065] In practical implementation, the privacy budget consumed by each noise addition operation is recorded, and the total privacy budget consumption is calculated cumulatively to ensure that it does not exceed the preset privacy protection strength requirements. The output noisy principal component matrix is ​​a feature representation that has undergone privacy protection processing, satisfying strict mathematical privacy guarantees. Any attack attempting to infer the original pose information or personal identity from it will be strongly protected by the differential privacy mechanism.

[0066] The noisy principal components are reconstructed using an inverse transformation to represent the original feature space. While adding noise to the principal component space efficiently achieves differential privacy protection, subsequent behavior recognition models are typically trained in the original feature space. Therefore, the noisy principal component data needs to be transformed back to the original feature space before it can be processed and utilized by existing behavior recognition algorithms. The inverse transformation is the mathematical inverse operation of the dimensionality reduction process in principal component analysis, implemented based on the fundamental principles of linear algebra.

[0067] Specifically, the noisy principal component matrix is ​​multiplied by the transpose of the principal component analysis transformation matrix to obtain the centered reconstructed features. Then, the mean vector of the original features is added to restore the uncentered feature space. The dimensions of the reconstructed feature matrix are exactly the same as those of the original pose feature matrix. Each row in the matrix represents the pose feature of a sample, and each column represents a feature dimension.

[0068] However, it is important to emphasize that although the reconstructed feature matrix is ​​structurally identical to the original matrix, its numerical values ​​have changed significantly. This change stems from two aspects. First, the principal component analysis (PCA) dimensionality reduction process discards smaller principal components, resulting in irrecoverable details encoded by these smaller principal components in the original features, leading to irreversible information loss. Second, the random noise added to the principal component space is propagated to all dimensions of the original feature space during the inverse transformation, causing noise perturbation to each eigenvalue. This dual mechanism results in controlled distortion of the reconstructed features relative to the original features. The degree of distortion is determined by the number of principal components retained by PCA and the intensity of the differential privacy noise.

[0069] In this embodiment, the distortion pattern is carefully designed. For sensitive features, because the corresponding principal components are allocated a large privacy budget and more noise is added, the reconstructed feature values ​​deviate significantly from the original values, making it difficult for attackers to accurately reconstruct facial geometry or perform identity recognition from the distorted head key point locations. For non-sensitive features, because the corresponding principal components are allocated a small privacy budget and less noise is added, the reconstructed feature values ​​are closer to the original values, preserving sufficient behavioral pattern information to support subsequent behavior recognition tasks.

[0070] This embodiment, through this differentiated distortion mechanism, effectively protects personal privacy while maximizing the data's utility for the behavior recognition task. Post-processing techniques can be applied during reconstruction to further optimize the results. For example, for coordinate features, geometric constraints can be applied to ensure that the reconstructed key point locations conform to reasonable ranges in human anatomy; for confidence features, values ​​can be truncated to a reasonable range to ensure semantic correctness. The output privacy-preserving feature data, while meeting the strict mathematical guarantees of differential privacy, maintains its effectiveness for the secure behavior recognition task. Subsequent behavior recognition models can directly use this data for inference without needing to understand the technical details of privacy protection, achieving transparent integration of the privacy protection mechanism.

[0071] In this embodiment, the attitude feature matrix set is labeled based on the partitioning information of sensitive and non-sensitive parameters in the partitioning result, and the labeled attitude feature matrix set is output. Principal component analysis is applied to the labeled attitude feature matrix set to calculate the feature covariance matrix and extract eigenvalues ​​and eigenvectors, and the eigenvalue ranking result and the corresponding eigenvectors are output. Based on the eigenvalue ranking result, the top K eigenvectors with a cumulative contribution rate reaching a preset threshold are selected as principal components, and K principal component vectors are output. Based on the K principal component vectors, a linear transformation is performed on the attitude feature matrix set to project the original high-dimensional attitude feature matrix set onto a low-dimensional principal component space, thereby obtaining the dimensionality-reduced principal component representation.

[0072] Specifically, the pose feature matrix set is labeled based on the partitioning information of sensitive and insensitive parameters in the partitioning results. Feature labeling is a metadata enhancement process that aims to attach semantic labels and attribute information to each feature dimension in the dataset, enabling subsequent data processing algorithms to adopt differentiated processing strategies based on the different attributes of the features. The original form of the pose feature matrix is ​​a pure numerical matrix, where each element only contains numerical information and lacks a description of the meaning and sensitivity of that value. The labeling process first needs to establish a mapping relationship between feature dimensions and sensitivity classification results.

[0073] Taking the pose feature of seventeen key points, each with three attributes, as an example, the expanded feature matrix has fifty-one feature dimensions, which can be organized according to a two-dimensional index of key point number and attribute type. Based on the sensitivity classification results obtained in step one of Example 5, the five head-related key points are classified as sensitive parameters, and their corresponding feature dimensions are labeled as sensitive. The twelve key points of the torso and limbs are classified as non-sensitive parameters, and their corresponding feature dimensions are labeled as non-sensitive.

[0074] The annotation information is stored using a combination of data structures, including lightweight bitmap markers and detailed metadata records. The annotated pose feature matrix remains physically stored in its original numerical matrix form; the values ​​themselves are unchanged. However, logically, each feature dimension is associated with a set of metadata attributes. This metadata is stored and transmitted synchronously with the feature matrix in the form of a database or configuration file. When the data flows through processing modules such as principal component analysis for dimensionality reduction and noise addition, the processing algorithm can query the feature annotation information and adjust processing parameters based on sensitivity labels to achieve intelligent, differentiated processing.

[0075] Understandably, the output labeled pose feature matrix set provides the necessary semantic support for subsequent refined privacy protection, enabling the entire system to perform precise privacy control at the feature level.

[0076] Principal component analysis (PCA) is applied to the labeled pose feature matrix to calculate the feature covariance matrix and extract eigenvalues ​​and eigenvectors. The mathematical foundation of PCA is eigenvalue decomposition, which deals with the covariance matrix of the data. Therefore, calculating the covariance matrix and eigenvalue decomposition are the core steps of the PCA algorithm. First, the pose feature matrix is ​​preprocessed by centering, calculating the mean of each feature dimension across all samples to obtain the mean vector. Then, a centering matrix is ​​constructed where each element equals the original eigenvalue minus the mean of that feature. That is, the mean of each feature is subtracted from the mean of each eigenvalue in each sample, ensuring that the mean of the centered data is zero across all feature dimensions.

[0077] In this embodiment, the centering operation eliminates the influence of the data mean position, so that the direction of the principal components extracted by principal component analysis depends only on the degree of data dispersion, and is not affected by the overall data translation. This is an important prerequisite for ensuring the interpretability and stability of principal component analysis results.

[0078] Next, we calculate the covariance matrix of the centered data. This is a symmetric matrix, where the diagonal elements represent the variance of each feature dimension, reflecting the dispersion of that feature across all samples. The off-diagonal elements reflect the strength of the linear correlation between two features; positive values ​​indicate a positive correlation, negative values ​​indicate a negative correlation, and larger absolute values ​​indicate a stronger correlation. The symmetry and positive semi-definiteness of the covariance matrix are important mathematical properties that ensure the feasibility and numerical stability of eigenvalue decomposition. We then perform eigenvalue decomposition on the covariance matrix, solve the characteristic equation, and find all eigenvalues ​​and their corresponding eigenvectors that satisfy the equation.

[0079] From a geometric perspective, an eigenvector defines a direction in the original feature space, and an eigenvalue represents the variance of the data in that direction. A larger eigenvalue indicates a greater dispersion of the data in that direction, meaning that direction contains more information. Therefore, principal component analysis selects the direction corresponding to the largest eigenvalue as the first principal component, the direction corresponding to the second largest eigenvalue as the second principal component, and so on. Eigenvalue decomposition is typically implemented using standard algorithms from numerical linear algebra. These algorithms have highly optimized implementations in numerical computing libraries, capable of efficiently handling large-scale covariance matrices.

[0080] Next, after eigenvalue decomposition, several eigenvalues ​​and corresponding eigenvectors are obtained. These eigenvectors form an orthogonal basis of the original feature space, where any two distinct eigenvectors are orthogonal to each other. The eigenvalues ​​are sorted in descending order, and the order of the eigenvectors is adjusted accordingly. The sorted eigenvalues ​​and corresponding eigenvectors are output. The numerical sequence of eigenvalues ​​reflects the information distribution of each principal component, and the eigenvector matrix defines the linear transformation relationship from the original feature space to the principal component space. This information provides the mathematical basis for subsequent principal component selection and dimensionality reduction projection.

[0081] Based on the eigenvalue ranking results, the top K eigenvectors whose cumulative contribution rate reaches a preset threshold are selected as principal components. Principal component selection is a crucial decision step in the dimensionality reduction process of principal component analysis, requiring a trade-off between dimensionality reduction effectiveness and information preservation. The cumulative contribution rate is a standard indicator for quantifying this trade-off; it is defined as the proportion of the sum of the top few largest eigenvalues ​​to the sum of all eigenvalues. The cumulative contribution rate ranges from zero to one; the closer the cumulative contribution rate is to one, the more variance information of the original data is preserved by the top few principal components, and the less data distortion there is.

[0082] In practical applications, a preset threshold for the cumulative contribution rate needs to be set according to the accuracy requirements of the specific task. Common values ​​include 90%, 95%, and 98%. The choice of threshold reflects different preferences for the degree of dimensionality reduction and data fidelity. Lower thresholds pursue more aggressive dimensionality reduction, which can significantly compress the feature dimensions, but will lose more detailed information. Higher thresholds pursue more conservative dimensionality reduction, retaining most of the original information, but the dimensionality reduction effect is relatively limited.

[0083] In the application scenario of building construction safety behavior recognition, it is necessary to protect privacy while maintaining the accuracy of behavior recognition. Therefore, a cumulative contribution rate threshold of 95% or 98% is usually selected to achieve a good balance between privacy protection and recognition accuracy.

[0084] The determination of the number of principal components employs an iterative accumulation algorithm. Starting with the sorted sequence of feature values, the cumulative sum and cumulative count are initialized to zero. Then, feature values ​​are sequentially added to the cumulative sum, while simultaneously calculating the current cumulative contribution rate. Iteration stops when the cumulative contribution rate first reaches or exceeds a preset threshold. The cumulative count at this point represents the number of principal components to be retained. Taking a 51-dimensional pose feature as an example, assuming the feature value distribution exhibits a typical rapid decay pattern, and setting the threshold to 95%, the cumulative contribution rate of the first 20 to 25 principal components typically reaches 95%. This means that retaining these principal components allows the original data to represent 95% of its information with less than half the dimensionality, achieving significant dimensionality reduction. After selecting the number of principal components, the first few corresponding feature vectors are extracted and arranged column-wise to form the principal component transformation matrix.

[0085] The output principal component vectors not only define the dimensionality-reduced feature space, but also implicitly encode the correlation structure and information importance ranking among the original features, providing a transformation basis for subsequent feature projection and inverse reconstruction.

[0086] A linear transformation is performed on the attitude feature matrix set using K principal component vectors, projecting the original high-dimensional attitude feature matrix set onto a low-dimensional principal component space. Feature projection is the step in principal component analysis for dimensionality reduction, achieving the mapping from high to low dimensions through matrix multiplication in linear algebra. Multiplying the centered original feature matrix with the principal component transformation matrix yields the dimensionality-reduced principal component representation matrix.

[0087] From a matrix operation perspective, each element of the dimensionality-reduced matrix represents the projection coefficient of the corresponding sample along the corresponding principal component direction. This coefficient quantitatively describes the magnitude of the sample's feature component along that principal component direction. Geometrically, the projection process is equivalent to orthogonally decomposing the original high-dimensional feature vector along several principal component directions, retaining its components in these directions and discarding its components in the remaining directions. Since the principal component directions are selected according to their variance contribution, they capture the main variation patterns of the original data. Therefore, although some directional information is discarded, the projection coefficients of the retained principal component directions can still well represent the main features of the original data. The projected principal component representation has several important mathematical properties.

[0088] First, the different principal components are independent of each other because the principal component vectors are orthogonal eigenvectors of the covariance matrix, and their defined directions are mutually orthogonal. Second, the variance of each principal component is equal to its corresponding eigenvalue, which directly reflects the amount of information contained in that principal component. Third, if all principal components are used for projection, the projection is completely invertible, and the original features can be reconstructed without loss. However, when only the first few principal components are used, the projection is not invertible, and information loss occurs during reconstruction.

[0089] In practical computation, matrix multiplication can be accelerated using techniques such as block matrix multiplication, sparse matrix optimization, and parallel computing. Furthermore, since Principal Component Analysis (PCA) is an unsupervised learning method, its transformation matrix is ​​entirely determined by the statistical properties of the data and is unrelated to the label information for behavior recognition. Therefore, PCA dimensionality reduction may lose some features that are useful for classification but have low variance; this is an inherent limitation of PCA.

[0090] To alleviate this problem, we can consider introducing supervised information based on principal component analysis, or consider both variance contribution and classification discriminant power when selecting principal components. The dimensionality-reduced principal component representation of the output compresses the original high-dimensional feature matrix into a low-dimensional space, with a dimensionality reduction ratio typically between 40% and 50%. While significantly reducing the data dimensionality, it retains 95% to 98% of the variance information of the original features, providing an optimized data representation foundation for subsequent efficient differential privacy protection and efficient behavior recognition.

[0091] In this embodiment, the privacy-preserving feature data is input into a pre-trained spatiotemporal graph convolutional network model for spatiotemporal feature extraction, and a high-level representation of the spatiotemporal features is output. Based on the high-level representation of the spatiotemporal features, a multi-task learning framework is used to simultaneously perform behavior classification and dangerous area determination, and a preliminary violation behavior identification result is output. The preliminary violation behavior identification result is filtered by applying a confidence threshold and combined with a temporal consistency test to filter out false detections, thereby obtaining the security violation event detection result.

[0092] Specifically, privacy-preserving feature data is input into a pre-trained spatiotemporal graph convolutional network model for spatiotemporal feature extraction. The spatiotemporal graph convolutional network is a deep learning model specifically designed for skeleton sequence behavior recognition. It extends traditional convolutional neural networks from regular grid data to graph-structured data, making it particularly suitable for processing naturally graph-structured data such as human poses.

[0093] The human skeleton can be naturally modeled as a graph, where nodes are key points and edges are bones connecting adjacent key points. This graph structure accurately depicts the topological connections and spatial geometry of the human body, better reflecting the inherent structural characteristics of human posture than simply arranging key point coordinates into vectors. The core innovation of the spatiotemporal graph convolutional network model lies in defining graph convolution operations, enabling feature aggregation and propagation on the graph structure.

[0094] In the spatial dimension, graph convolution captures local spatial structural patterns of the human body by aggregating features of each keypoint and its neighboring keypoints. For example, arm posture depends not only on the position of the wrist keypoint but also on the positions of the elbow and shoulder keypoints; spatial graph convolution can effectively aggregate information from these related keypoints. In the temporal dimension, the model applies temporal convolution to capture the evolutionary patterns of movements along the time axis. For example, a waving motion temporally manifests as periodic changes in the positions of arm keypoints; temporal convolution can identify this temporal pattern. Spatiotemporal graph convolutional networks alternately stack spatial graph convolution and temporal convolution to construct deep networks, extracting increasingly abstract spatiotemporal feature representations layer by layer. The spatiotemporal graph convolutional network model used in this embodiment is specially designed and trained to adapt to posture feature data with differential privacy noise.

[0095] Standard spatiotemporal graph convolutional network models assume that the input keypoint coordinates are precise and noise-free. When the input data contains noise, the model's recognition performance will significantly decrease. To address this issue, this embodiment employs noise robustness enhancement techniques during the model training phase. Specifically, this involves artificially injecting synthetic noise with statistical characteristics similar to differential privacy noise into the training data, enabling the model to learn tolerance to noise during training. An attention mechanism is introduced into the model architecture, dynamically adjusting the weights of different keypoint features based on their confidence scores. Low-confidence keypoints are assigned smaller weights, while high-confidence keypoints are assigned larger weights, reducing the impact of noise on the overall feature representation. A robust regularization term is added to the loss function design to penalize the model's excessive sensitivity to input noise and encourage the model to learn stable feature representations.

[0096] Through these techniques, the trained spatiotemporal graph convolutional network model exhibits strong robustness to differential privacy noise. Even after privacy-preserving processing of the input data, the model can still extract effective behavioral features. The model's pre-training utilizes a large-scale construction safety behavior dataset, containing tens of thousands of video clips from construction sites, covering normal work behaviors and various violations. Each video clip is manually labeled with a behavioral category. By pre-training on this domain-specific dataset, the model learns feature representations of typical behavioral patterns in construction scenarios. These pre-trained weights serve as the starting point for transfer learning, enabling the model to adapt more quickly to real-world applications at specific construction sites and achieve high recognition accuracy even with limited on-site labeled data.

[0097] The input privacy-preserving feature data undergoes forward propagation through a spatiotemporal graph convolutional network model. First, multiple spatiotemporal graph convolutional layers extract the spatial structural features of the human posture at each moment. These features encode geometric information such as the relative positional relationships between keypoints and the angles and length ratios of limbs. Then, multiple temporal convolutional layers aggregate the spatial features from consecutive moments to extract the temporal evolution patterns of the action, such as the motion trajectory, velocity, and acceleration of keypoints. Finally, a global spatiotemporal pooling layer aggregates the features of the entire temporal segment into a fixed-dimensional feature vector. The high-level representation of the output spatiotemporal features is a high-dimensional vector, a compact encoding of the input posture sequence, containing the key discriminative information required for behavior recognition. Furthermore, due to the multiple nonlinear transformations, robustness to input noise and the ability to distinguish behavior categories are further enhanced.

[0098] High-level representations based on spatiotemporal features utilize a multi-task learning framework to simultaneously perform behavior classification and hazard zone determination. Multi-task learning is an important learning paradigm in machine learning. Its core idea is to train multiple related tasks simultaneously, leveraging the correlation and shared information between tasks to improve the learning performance of each task. In construction safety monitoring scenarios, behavior classification and hazard zone determination are two closely related but distinct tasks. Behavior classification focuses on what actions construction workers are performing, which is a multi-classification problem. Hazard zone determination focuses on whether the current location of construction workers belongs to a hazard zone, which is a binary or multi-label classification problem. These two tasks are naturally related. The determination of certain violations requires considering both action type and location information simultaneously. For example, the violation of not wearing a safety belt while working at height requires that the action be both a high-altitude work action and the location be within a high-altitude work area. Jointly training these two tasks through a multi-task learning framework can achieve several benefits.

[0099] First, sharing the underlying feature representation reduces the number of model parameters, improves parameter utilization efficiency, and reduces the risk of overfitting, which is particularly important in practical applications with limited labeled data. Second, complementary information between tasks can mutually enhance each other. The human motion features learned in the behavior classification task can help the danger zone determination more accurately understand the intentions of personnel, while the spatial location features learned in the danger zone determination can provide contextual information for behavior classification, improving the overall performance of both tasks. Third, joint training enables the model to output richer structured prediction results, providing more comprehensive information for subsequent violation determination and risk assessment. The network architecture of the multi-task learning framework adopts a design pattern that combines shared and private features. The underlying layer uses the spatiotemporal feature representation extracted by the spatiotemporal graph convolutional network model as shared features. This shared feature is useful for both tasks, encoding general information about behavior and location. The top layer is divided into two parallel task-specific branches, each with an independent fully connected layer and output layer.

[0100] Understandably, the behavior classification branch is a multi-classifier containing several fully connected layers, ultimately outputting the probability distribution of each behavior category through an activation function. The danger zone determination branch is a binary or multi-label classifier, outputting the determination probability of danger zones through an activation function. Both branches share underlying features, and the loss functions of both tasks are simultaneously optimized using backpropagation. The total loss is defined as a weighted sum, and the parameters of the shared feature extractor and the two task-specific branches are updated using gradient descent to minimize the total loss, achieving joint optimization of the two tasks. The initial violation behavior identification result integrates the prediction information from both tasks, including the probability distribution vector of the behavior category, indicating the most likely behavior type and confidence level, the determination result and confidence level of the danger zone, and the violation event determination based on the fusion of both. The initial identification result also includes a confidence score for each determination, providing a quantitative basis for subsequent result filtering and risk assessment.

[0101] Next, the preliminary violation identification results are filtered using a confidence threshold and combined with a temporal consistency test to remove false detections. Although the spatiotemporal graph convolutional network model is specially trained and optimized, in practical applications, due to the complexity of the construction site environment and noise disturbances in the input data, the model's prediction results may still contain uncertainties and misjudgments. The preliminary identification results may contain a large number of predictions with low confidence. These low-confidence predictions often correspond to situations where the model is uncertain. Directly adopting them may lead to a high false alarm rate, affecting the system's practicality and reliability. Therefore, post-processing techniques are needed to filter and optimize the preliminary results, removing unreliable predictions and retaining high-quality detection results.

[0102] Confidence thresholding is the most direct and effective post-processing method. Its basic principle is to set a minimum confidence threshold for the model's output probability, retaining only detection results with confidence scores exceeding this threshold and discarding uncertain predictions with confidence scores below the threshold. For behavior classification tasks, confidence is defined as the probability value of the most probable class. Only when the confidence score reaches the preset threshold is the behavior classification result accepted; otherwise, it is marked as uncertain or requires manual review.

[0103] For hazardous area identification tasks, the confidence level directly uses the output probability value, which also needs to meet the threshold requirement to be accepted. The confidence threshold setting needs to be adjusted according to the fault tolerance requirements of the actual application. In scenarios with extremely high security requirements, a lower threshold can be set to improve the detection recall rate and minimize the omission of real violations, at the cost of potentially introducing more false alarms. In scenarios where the goal is to reduce false alarms, a higher threshold can be set to improve detection accuracy and ensure the reliability of alarms, at the cost of potentially missing some less obvious violations.

[0104] Besides confidence-based screening, temporal consistency testing is another important post-processing technique that leverages the continuity and consistency of human behavior over time. Genuine violations typically last for a period of time and should be detectable across multiple consecutive video frames, while false detections are often instantaneous and accidental, lacking consistency across adjacent frames.

[0105] There are several implementation strategies for temporal consistency testing. One is the voting method, which counts the number of frames in which a certain behavior category appears within a time window and calculates the frequency of occurrence. Only when the frequency exceeds a preset threshold is the occurrence of the behavior confirmed. This method can effectively filter out isolated false detection frames. Another is the temporal smoothing method, which uses moving averages or other temporal smoothing algorithms to smooth the detection results of consecutive frames, eliminating instantaneous jitter and jumps. The smoothing confidence of the current frame is defined as the weighted average of the confidence of the current frame and the confidence of the previous several frames. In this way, even if an abnormally low confidence or misclassification occurs in a certain frame, it will be smoothed by the stable results of the preceding and following frames. A third is the state transition constraint method, which uses hidden Markov models or finite state machines to model the transition rules of behavior states. Some transitions between behavior states are unreasonable or impossible. State transition constraints can correct detection results that do not conform to the behavior evolution logic. Temporal consistency testing can also be combined with confidence information. For high-confidence results detected in multiple consecutive frames, their credibility is further improved, while low-confidence results that only appear in a few frames are suppressed or deleted.

[0106] This embodiment significantly improves the reliability and stability of violation detection results through a dual filtering mechanism of confidence level screening and temporal consistency verification, greatly reducing the false detection rate while maintaining a high detection rate of true violations. The output safety violation detection results are high-quality detection information that has undergone rigorous verification and optimization. Each detection result contains detailed structured data, including the type of violation, the start and end times of the occurrence, the duration, the location of the personnel involved in the video footage, the construction area where the personnel are located, the overall confidence score, and relevant contextual information. This rich detection result information provides reliable and detailed data support for subsequent risk assessment, early warning decision-making, event logging, and accountability tracing, and is a key link in achieving closed-loop management of the entire intelligent safety monitoring system.

[0107] In this embodiment, the violations identified in the preliminary violation identification results include not wearing a seat belt, entering a dangerous area, and improper operation.

[0108] Specifically, the specific types of violations are defined and their identification characteristics are explained below. Not wearing a safety belt is the most common safety violation in high-altitude operations. Its identification characteristics include: the worker is on a work platform or scaffolding at a height exceeding 2 meters, but lacks safety belt connection points in the waist and shoulder areas, and their body posture is unrestrained. Entering a hazardous area refers to workers entering a pre-designated hazardous area without authorization, such as the area below hoisting operations, the area where large machinery operates, or the edge of a foundation pit. Its identification characteristics include: determining the worker's spatial location based on the key points of their feet, combined with the site's zoning information, to determine whether they have entered a hazardous area. Non-standard operations include various specific situations, such as incorrect climbing postures, unauthorized use of tools, and violations of operating procedures. Their identification characteristics vary depending on the specific operation type, typically manifesting as abnormal body movement patterns or body states that do not conform to standard operating postures. The multi-task learning framework has trained identification models for these different types of violations, enabling simultaneous detection of multiple violations and improving the comprehensiveness and real-time nature of safety monitoring.

[0109] In this embodiment, the risk level of each violation event is assessed based on the detection results of the safety violation event and the pre-configured risk level information of the construction area, and the violation event information with risk level is output. A graded early warning signal is generated based on the risk-level violation event information and pushed to the site management personnel, outputting graded early warning information. The violation type, time, location, and risk level in the risk-level violation event information are recorded in the safety management database to establish a safety record archive. Based on the safety record archive, construction safety analysis reports and trend charts are constructed to obtain the traceable safety management mechanism.

[0110] Specifically, the risk level of each violation is assessed based on the detection results of safety violations and pre-configured risk level information for the construction area. Risk level assessment is a crucial step in transforming detected violations from simple behavioral identification results into risk information with safety management significance. It requires comprehensive consideration of the inherent danger of the violation and the risk context of the environment in which the violation occurred, to provide objective and quantitative risk assessment results.

[0111] Understandably, construction sites are highly dynamic and heterogeneous environments, with significant differences in risk levels across different areas and work phases. The same violation can pose entirely different levels of danger in different environments, necessitating a context-aware risk assessment approach. Risk level information for construction areas is a crucial input for risk assessment. This information is pre-configured by the project's safety management personnel during the construction preparation phase, based on professional safety management documents such as construction organization design, safety briefings, and hazard identification. The construction site is first divided into multiple functional areas, each assigned a risk level label based on its function, work content, and environmental conditions.

[0112] High-risk areas typically include: high-altitude work areas (such as work platforms on scaffolding, building rooftops, tower crane operating areas, etc.), which are classified as high-risk due to the risk of falls from heights; heavy machinery work areas (such as within the swing radius of tower cranes, concrete pump truck operating areas, excavator working faces, etc.), which pose risks of falling objects and mechanical injuries; temporary high-voltage power supply areas (such as distribution boxes, cable laying areas, etc.), which pose a risk of electric shock; the edge areas of foundation pits, which pose a risk of collapse and falls; and material hoisting areas and lifting operation areas, which pose a risk of falling objects and mechanical injuries. Medium-risk areas include material processing areas, rebar tying areas, and formwork removal areas; while these areas also pose some danger, they are relatively controllable. Low-risk areas include on-site office areas, tool warehouses, and rest areas, where personnel activities are relatively safe.

[0113] In addition to the inherent risk level of a region, the system also considers the impact of time on risk, and the risk level of certain areas will be dynamically adjusted at specific times. For example, when hoisting operations are underway, the risk level of the hoisting area and its surroundings will be temporarily increased; during nighttime construction, the risk level of some areas will be increased accordingly due to reduced visibility.

[0114] The risk level assessment algorithm employs a multi-factor comprehensive scoring model. This model quantifies multiple risk-influencing factors into numerical indicators and calculates a comprehensive risk score through weighted summation or multiplicative combination. Key factors considered include: the inherent danger of the violation type; different violations have different baseline risk scores. For example, the baseline risk score for "not wearing a safety belt while working at height" is set at 90 points (out of 100) because the consequences of such a violation are extremely serious if an accident occurs. The baseline score for "entering a prohibited area" might be 70 points, indicating a relatively lower risk. The baseline score for "non-standard operation" might be 50 points. The risk level coefficient for the area where the violation occurred is also considered: 1.5 for high-risk areas, 1.0 for medium-risk areas, and 0.5 for low-risk areas. This coefficient is multiplied by the baseline score of the violation to obtain a contextualized risk score. For example, the risk score for "not wearing a safety belt" is 90 × 1.5 = 135 in a high-altitude work area (high risk) and 90 × 1.0 = 90 in a ground construction area (medium risk). The duration of the violation has an impact: the longer the violation lasts, the greater the accumulated risk. For violations exceeding a threshold (e.g., 30 seconds), the risk score is adjusted linearly or logarithmically based on the duration. For example, a violation lasting 1 minute increases the risk score by 20%, and a violation lasting 5 minutes increases it by 50%. The number of personnel involved also affects the risk score. If multiple people violate the rules simultaneously in the same area, or if the violation may affect other workers in the vicinity, the risk score increases accordingly. For each additional affected person, the risk score increases by 10%-15%.

[0115] Regarding environmental factors, severe weather (such as strong winds, heavy rain, and high temperatures) increases operational risks. The system can access meteorological sensor data and dynamically adjust the risk score under severe weather conditions. For example, during winds of level six or higher, the risk score for high-altitude operations increases by 30%. Regarding equipment status factors, if a violation occurs during equipment malfunction or maintenance, the risk score will also increase. This embodiment calculates a quantified risk score (range 0-200 points) for each violation event through a comprehensive calculation of these factors, and then maps it to a predefined risk level based on the score range.

[0116] The system uses a three-tiered risk classification: Level 1 risk (red / urgent), corresponding to a risk score ≥120, indicates that the violation is extremely dangerous and may lead to major casualties or property damage, requiring immediate emergency measures; Level 2 risk (yellow / warning), corresponding to a risk score of 80-119, indicates that the violation is relatively dangerous and may lead to a general accident, requiring prompt handling; Level 3 risk (blue / hint), corresponding to a risk score of 50-79, indicates that the violation is relatively low-risk, but still requires attention and correction; for minor violations with a risk score below 50, the system records them but does not trigger real-time alerts, and reminds the violators during daily inspections and regular safety meetings.

[0117] The risk assessment process is fully automated, determining the risk level within milliseconds of detecting a violation, ensuring real-time warnings. The output of violation information with risk levels is rich, structured data based on the original detection results. Each violation, in addition to basic information such as violation type, time, and location, includes a risk level label (Level 1 / 2 / 3), a risk score (specific numerical value), a description of risk influencing factors (listing major risk contributing factors, such as "high-altitude work area + not wearing a safety belt + long duration"), and recommended handling priority fields. This comprehensive risk information provides a scientific basis for subsequent tiered warnings and differentiated responses, enabling safety management resources to be prioritized for the most dangerous events, thus improving the efficiency and effectiveness of safety management.

[0118] Based on information on violations with risk levels, tiered early warning signals are generated and pushed to on-site management personnel. The early warning mechanism is a crucial link in translating risk information into management action. It needs to immediately alert relevant personnel upon detecting a risk, prompting them to take measures to eliminate the danger and prevent accidents. The core concept of the tiered early warning mechanism is to adopt differentiated early warning methods, notification scopes, and response requirements according to different risk levels. This ensures that high-risk events receive the fastest and most effective response while avoiding the "crying wolf" effect and management fatigue caused by frequent alarms for low-risk events.

[0119] For Level 1 risk (red / emergency) violations, the system triggers the highest level emergency alert. The alert response employs a multi-channel concurrent push strategy to ensure relevant personnel are informed of the hazard in the shortest possible time. On-site audible and visual alarms immediately activate, emitting high-decibel alarms and flashing red warning lights in and around the violation area, directly alerting on-site workers and nearby personnel to the danger. SMS alerts are sent in batches via the SMS gateway to pre-set emergency contacts (including on-site safety officers, area managers, project safety directors, project managers, etc.). The SMS content is concise and clear, including the words "Emergency Alert," the type of violation, a specific location description (e.g., "East side of the scaffolding on the 5th floor of Building 3"), the time of occurrence, and recommended actions (e.g., "Immediately proceed to the site, instruct workers to stop work and fasten their safety belts"). Emergency notifications are also pushed to all online management personnel via the project safety management app. The notification uses a strong alert mode combining vibration, sound, and pop-up windows to ensure it is noticed even when the phone is in your pocket. In addition to text descriptions, the notification includes screenshots or short video clips of the violation (with privacy-protected processing to blur facial features), the location of the violation on the construction site map, and a one-click navigation function to the site. The PC-based management system provides alerts, displaying a prominent red warning window on the large screen in the safety monitoring center, showing detailed event information, real-time video footage, and records of similar historical events, facilitating coordination and handling by monitoring center personnel. Voice call alerts automatically dial the relevant responsible person's number and play a pre-recorded voice warning message in particularly serious situations or when no response is received through the aforementioned channels.

[0120] The response time requirements for Level 1 warnings are extremely strict. Managers receiving the warning must confirm receipt within 2 minutes (by replying with a confirmation code via APP or SMS), arrive at the site or verify the situation remotely via video within 5 minutes, and take effective risk control measures (such as requiring workers to stop work, evacuating people from dangerous areas, and isolating sources of danger) within 10 minutes. The system records the entire response time and handling process, which serves as the basis for safety management performance evaluation.

[0121] For Level 2 risk (yellow / warning) violations, the system triggers a warning level alert. The alert method is relatively mild but still timely. Warning information is primarily sent via mobile app push notifications and the on-site broadcast system. App push notifications use a standard reminder mode (sound alert but no vibration), targeting those directly responsible for the area where the violation occurred (such as area safety officers and team leaders) and project safety management personnel. The notification content includes the violation type, location, time, risk score, and suggested remedial measures. The on-site broadcast system plays standardized warning voice messages (e.g., "Warning: Personnel have been detected entering a dangerous area in the material storage area of ​​Building 3. Relevant management personnel are requested to verify immediately"). The broadcast range is limited to the area where the violation occurred to avoid unnecessary panic caused by broadcasting throughout the site. The response time limit for Level 2 alerts is 15 minutes for confirmation, and on-site verification and handling within 30 minutes. The handling results must be reported back via the app or management system.

[0122] For Level 3 risk (blue / warning) violations, the system records and periodically reminds users, without triggering real-time alarms. Instead, it summarizes the event information and displays it in the pending items list on the safety management interface for managers to monitor during daily inspections. It is also raised as a discussion topic at daily or shift-specific safety debriefing meetings for safety education and reminders. Although Level 3 events may not pose a high immediate danger, frequent occurrences or prolonged neglect can accumulate into greater risks. Therefore, the system statistically analyzes Level 3 events. When a certain type of Level 3 event accumulates to a threshold (e.g., 10 times) within a certain period (e.g., one week), it automatically escalates to a Level 2 warning, alerting managers to the systemic risks associated with this type of issue.

[0123] In this embodiment, the entire process of pushing early warning signals is completed under an edge computing architecture. Early warning decisions, signal generation, message sending, and other processing are all executed on edge servers deployed at the construction site, without relying on cloud servers or external networks. This architecture design has multiple advantages: First, the response latency is extremely low. The entire process from violation detection to early warning push can be controlled within one second. For emergencies, every second of delay could mean the occurrence of an accident, making the low-latency characteristic of edge computing crucial. Second, it offers high reliability. Even if the internet connection at the construction site is interrupted or unstable, the edge computing system can still operate independently, continuously monitoring and issuing early warnings, unaffected by external network failures. Third, it provides strong privacy protection. Video data and sensitive information are processed locally, eliminating the need to upload to the cloud and reducing the risk of privacy leaks. Fourth, it offers controllable costs, reducing the expenses of cloud computing and network bandwidth. The content design of the early warning information follows the principle of operability, not only informing users "what happened," but more importantly, providing guidance on "what should be done." Each early warning message includes suggested handling measures based on best practices in security management and the company's security operating procedures.

[0124] For example, for an early warning of "working at height without a safety belt," the recommended action is to "immediately require workers to stop work, check and correctly wear their safety belts, and only resume work after confirming safety." For an early warning of "entering a dangerous area," the recommended action is to "evacuate irrelevant personnel from the dangerous area, check whether the safety isolation measures in the area are in place, and strengthen on-site patrols." These specific operational guidelines help on-site managers make quick and correct responses. In particular, for inexperienced managers, the guidance function of the early warning information significantly improves their emergency response capabilities. The output of tiered early warning information realizes a rapid closed loop from risk identification to management response, deeply integrating intelligent detection technology with safety management processes. This greatly improves the construction site's ability to perceive and respond to safety risks. According to statistics from actual applications, the tiered early warning mechanism reduces the average response time for high-risk violations from more than 30 minutes in the traditional manual patrol mode to less than 5 minutes.

[0125] In this embodiment, the violation type, time, location, and risk level from the violation event information with risk levels are recorded in the security management database to establish a security record archive. The security record archive is the data foundation for building a traceable security management mechanism. It persistently stores complete information about each violation event, providing reliable historical data support for subsequent security analysis, accountability, and management improvement. The security management database is designed using a relational database (such as MySQL or PostgreSQL) or a time-series database (such as InfluxDB or TimescaleDB). Relational databases are suitable for storing structured event records and performing relational queries, while time-series databases are particularly suitable for processing large amounts of time-stamped event data, supporting efficient time-range queries and time-series analysis.

[0126] The core table of the database is the violation event table. This table creates a record for each detected violation event, and the table's fields are designed to comprehensively cover the key information of the event.

[0127] For example, the event unique identifier field (event id Use UUID or auto-incrementing integer as the primary key to ensure that each event has a unique identifier for easy referencing and association; violation type field (violation) type Use an enumeration type or string to store the categories of violations, such as "not wearing a seatbelt," "entering a dangerous area," "improper operation," and "illegal smoking." This field is indexed to support fast querying and statistics by type. The occurrence field... time Using a high-precision timestamp type (accurate to milliseconds), the occurrence time of violations is accurately recorded. This field serves as the benchmark for time-series analysis and is also a primary field for querying and sorting; the end time field (end...)time Record how long the violation lasts. For momentary events, this may be the same as the time of occurrence. For persistent violations, it records the time when the violation is corrected or disappears. The duration field records the length of time the violation lasts in seconds, which is an important parameter for risk assessment.

[0128] Location field zone Using predefined area codes at the construction site (e.g., "A-3F-01" represents zone 01, floor 3, area A) facilitates area-based statistics and analysis. This field is linked to the construction area configuration table, providing detailed information such as the area's name, function, and risk level; the spatial coordinate field (location) coordinates The precise spatial location of the violation can be stored using geometric types or JSON format, such as the bounding box coordinates (x, y, w, h) of the person in the picture or the two-dimensional coordinates in the construction site plan. This coordinate information supports spatial analysis and visualization.

[0129] Risk level field level The risk level (Level 1 / Level 2 / Level 3 / Low Risk) obtained from the assessment is stored using an enumeration type. This field is the basis for graded early warning and priority ranking; the risk score field (risk...) score It stores quantified risk scores (0-200 points), providing a more granular risk measure than ratings; it also detects confidence fields. score The confidence score (0-1) output by the behavior recognition model is recorded, reflecting the reliability of the detection results; the handling status field... status Use an enumeration type to mark the processing progress of an event (pending / processing / processed / false alarm). This field supports closed-loop management of events and workflow tracking.

[0130] Handling action Use text type to record specific measures taken by on-site management personnel, such as "Workers have been required to wear safety belts and have been checked and confirmed," "Safety warning lines have been set up and personnel have been evacuated," and "Safety training has been provided to relevant personnel"; The handler field... name Record the name or employee number of the manager responsible for handling the incident; the handling time field. time Records the time when the event is processed; the response duration field (response) duration Calculate the time difference between the occurrence of an event and its completion, which is a key indicator for evaluating response efficiency; associate the camera field (camera) idThe event capture field records the camera ID that detected the event, facilitating tracing back to the original video source; the event capture path field (snapshot) path Stores the file path of a screenshot taken at the moment the event occurred. The screenshots are processed for privacy protection (facial features are blurred) for post-event verification and evidence preservation. The remarks field allows administrators to add additional explanatory information.

[0131] Understandably, the database design strictly adheres to privacy protection principles, a key feature of this implementation. The database absolutely does not store raw video images or information that could directly identify an individual, such as high-resolution facial images, names, employee numbers, or ID card numbers. All stored spatial location information is anonymized, retaining only region codes and relative coordinates, without being linked to specific personal identities. Even event screenshots must undergo facial blurring, mosaic processing, or only retain the outline of the human body to ensure that facial features cannot be identified from the screenshot. This privacy-preserving design follows the principle of "data minimization," collecting and storing only the minimum dataset necessary to achieve security management objectives, avoiding excessive collection and misuse of personal information.

[0132] The "Personnel Handling" field records the management personnel who handled the incident, not the workers who violated the rules. The system's purpose is to improve safety management and prevent accidents, not to monitor and punish individuals. Database access control uses a Role-Based Access Control (RBAC) mechanism, with different user roles having different data access permissions. The project safety director and safety management department have full query permissions and can view all violation incident records; regional safety officers can only view incident records for their assigned areas; ordinary project management personnel can only view statistical reports and cannot access raw incident details; the system administrator is responsible for database maintenance and backups but cannot access business data content.

[0133] In this embodiment, all data access operations are recorded in audit logs, including access time, accessing user, accessed data range, and type of operation performed (query / modify / delete). These audit logs are periodically reviewed by an independent information security department to prevent unauthorized access or misuse of data. The physical storage of the database employs encryption measures; both data files and backup files are encrypted using the AES-256 encryption algorithm, ensuring that even if the storage medium is stolen or illegally obtained, the data content cannot be directly read. Data transmission uses TLS / SSL encryption protocols to ensure that data is not eavesdropped on or tampered with during network transmission.

[0134] The established safety record archives have accumulated valuable safety management data assets for construction projects. Over time, the event records in the database have become increasingly rich, and this historical data can be used for various purposes: safety trend analysis, identifying patterns and weaknesses in safety management by analyzing the time, type, and regional distribution of violations; safety performance evaluation, quantifying the effectiveness of safety management measures by statistically analyzing the number and severity of violations in different periods, regions, and construction stages; accountability tracing, allowing for the retrieval of violation records prior to accidents to analyze precursors and root causes and clarify management responsibilities when a safety accident occurs; safety training, using typical violation cases (after de-identification) for safety education and training to improve workers' safety awareness; and management improvement, optimizing construction organization plans, improving safety facility configurations, and adjusting safety management processes based on the statistical analysis results of violations.

[0135] Construction safety analysis reports and trend charts are built based on safety record archives. Data analysis and visualization are important means to transform massive amounts of safety event records into management insights and decision support. Through in-depth mining of historical data, hidden safety patterns and problems can be discovered, providing a basis for scientific safety management. The safety analysis reports adopt a multi-dimensional and multi-layered design to comprehensively display the safety status of the construction site from different perspectives.

[0136] Understandably, time-based statistical analysis is the most fundamental and crucial analytical perspective. This involves statistically analyzing the number of violations at different time granularities, such as daily, weekly, monthly, quarterly, and annual, generating time-series reports. Daily reports summarize all violations that occurred that day, including key indicators such as total number, distribution by type, distribution by risk level, and average response time. These reports are automatically generated at the end of each workday, allowing safety managers to review the day's safety management efforts and identify key issues. Weekly reports are generated at the end of each week, summarizing safety data for the week, focusing on weekly trends, comparing the distribution of violations across different workdays (typically, violation rates are higher on Mondays and Fridays because Mondays see insufficient adjustment at the start of work, while Fridays see decreased attention as the weekend approaches), and analyzing the effectiveness of safety management during the week. Monthly reports provide a comprehensive monthly summary, including not only quantitative statistics but also year-on-year (compared to the same period last year) and month-on-month (compared to the previous month) analyses to assess the improvement or deterioration of the safety situation. Monthly reports are typically an important topic in project management meetings and serve as a basis for management decisions.

[0137] Quarterly and annual reports provide a longer-term, macro-level perspective, identifying seasonal patterns (such as different violation characteristics during hot summer and cold winter periods) and overall annual trends. Annual reports serve as the foundational material for project safety management performance evaluation and summarization. Regional-level statistical analysis uses different areas of the construction site as analytical units, statistically analyzing the frequency, type distribution, and risk level distribution of violation events in each area, generating regional comparison reports.

[0138] In this embodiment, regional comparison clearly identifies areas with weak links in safety management and areas with better management. For example, if the number of violations in a specific construction area (such as the 4th floor of Building 3) is consistently higher than in other areas, this area needs to be given special attention. The reasons should be analyzed (such as high operational difficulty, insufficient management personnel, inadequate safety facilities, etc.), and targeted improvement measures should be taken (increasing the frequency of safety inspections, strengthening safety education, and improving safety protection facilities). Regional analysis can also be combined with construction progress information to analyze the safety characteristics of each area at different construction stages (such as foundation construction, main structure construction, and decoration and finishing construction), providing differentiated strategies for safety management at different stages.

[0139] The statistical analysis of violation types categorizes and summarizes violations by behavior type, generating a type distribution report that displays the frequency ranking, percentage, and time trend of each type of violation. Type analysis identifies the most common and prominent violations. For example, if statistics show that "not wearing a safety belt while working at height" is the most frequent violation, accounting for 40% of all violations, then safety belt use should be prioritized in safety management. This can be achieved through measures such as strengthening education and training, improving the safety belt distribution and inspection system, and setting up prominent warning signs in high-altitude work areas to specifically reduce this type of violation. Type analysis can also assess the effectiveness of safety education and management measures. If the number of violations of a certain type significantly decreases after conducting a specific safety education activity (such as "safety belt wearing standard training"), it indicates that the education measures are effective; conversely, if a certain type of violation continues to occur frequently, it indicates that existing measures are insufficient and strategies need to be adjusted. The statistical analysis of risk levels focuses on the number and percentage of events at different risk levels, with a focus on monitoring the occurrence of high-risk (Level 1) and relatively high-risk (Level 2) events, as these events are most likely to lead to serious accidents and are the core focus of safety management.

[0140] Ideally, a safe environment involves very few or no high-risk incidents, with low-risk incidents dominating. If statistics show an increase in the number or proportion of high-risk incidents, it indicates a serious security situation requiring immediate and strengthened measures. Risk level analysis can also assess improvements in the overall security posture by comparing the number and proportion of high-risk incidents at different times, quantifying progress in security management.

[0141] Trend charts use data visualization technology to present the results of statistical analysis in an intuitive and easy-to-understand graphical way, helping managers to quickly grasp the safety status and changing trends.

[0142] For example, time-series line charts are the most commonly used method for trend visualization. With time as the horizontal axis and the number of violations or other indicators as the vertical axis, the line chart shows the trajectory of the indicator over time. The trend of the line intuitively reflects the improvement (line decreases) or deterioration (line increases) of the safety situation. Multiple curves can be superimposed on the line chart to compare the time changes of different types, regions, or risk levels of events, making it easy to identify correlations and differences. Bar charts are used to compare the quantity or frequency of different categories. For example, the horizontal axis represents different construction areas or violation types, and the vertical axis represents the corresponding number of violations. By comparing the height of the bars, the category with the most violations can be identified. Bar charts can also use stacked or grouped bars to display multi-dimensional information, such as stacking bars within each region according to risk level, simultaneously displaying the total number of violations and the risk structure of the region. Pie charts are used to show the composition ratio, such as the percentage of each type of violation event in the total number of violations. The size of the sectors intuitively shows the relative importance of each part. Pie charts are suitable for showing the proportional relationships of a small number of categories (usually no more than 6-8 categories). Heat maps are an advanced form of spatial visualization. Using a site plan as a background, different shades of color represent the frequency or density of violations at different locations on the plan. Darker colors (e.g., red) indicate more violations at that location, creating an "accident black spot," while lighter colors (e.g., green) indicate a better safety situation. Heat maps visually demonstrate the spatial distribution of violations, helping managers to target their inspections and focus on high-frequency violation areas. Radar charts are used for multi-dimensional comprehensive evaluation. For example, multiple safety indicators for a region (total number of violations, number of high-risk events, average response time, rectification completion rate, etc.) can be plotted on different axes of the radar chart. The shape of the radar chart allows for a comprehensive assessment of the region's safety management level. Radar charts are particularly suitable for comparing the overall performance of different regions or at different times.

[0143] The aforementioned analytical reports and trend charts are not only generated periodically in the form of static reports, but also dynamically displayed in real time through the visual dashboard of the security management information system. After logging into the system, administrators can see the real-time updates of various charts on a comprehensive interface and delve into the details of data of interest through interactive operations (such as clicking, filtering, and zooming). The dashboard adopts a responsive design and supports access on different devices such as PCs, tablets, and mobile phones, allowing administrators to keep abreast of security developments anytime, anywhere. It also supports a custom report function, allowing administrators to set query conditions (such as time range, region, violation type, etc.) according to specific needs to generate customized analytical reports, meeting the needs of specialized analysis and in-depth research.

[0144] Understandably, the traceable safety management mechanism organically integrates violation detection, risk assessment, tiered early warning, data recording, statistical analysis, and visualization, forming a complete closed-loop management process. From real-time detection and early warning of violations, to rapid response and handling of incidents, to systematic recording and in-depth analysis of data, and finally to management improvement and optimization of preventive measures based on the analysis results, the entire process is interconnected, achieving intelligent, data-driven, and scientific construction safety management. Traceability is reflected in multiple levels: every violation event has a complete record, including detailed information such as the time, location, type, risk level, and handling process, allowing traceability to the entire event process; the response behavior and handling measures of management personnel are recorded, enabling accountability and performance evaluation; historical safety management data is stored long-term, allowing for the analysis of long-term safety trends and management effectiveness.

[0145] Therefore, this traceable safety management mechanism not only improves the level of daily safety management, but more importantly, when a safety accident occurs, it can fully restore the safety management status before the accident through data tracing, analyze the causes and consequences of the accident, clarify management responsibilities, learn lessons and improve management, and establish a scientific, efficient and sustainable safety management system for construction companies. It has important practical value and demonstration significance for improving the overall safety management level of the construction industry.

[0146] Example 2: like Figure 2 As shown, this embodiment also provides a safety monitoring system for building construction, including: The video acquisition and preprocessing module 10 is used to acquire raw video data from the construction site, perform noise reduction, illumination correction, and frame rate adjustment on the raw video data, and output a standardized video data stream. This module is deployed on an edge computing device at the construction site and integrates video acquisition, image enhancement, and data standardization functions to ensure the output of a high-quality standardized video data stream for use by subsequent modules.

[0147] The pose feature extraction module 20 receives the standardized video data stream, extracts the pose features of construction workers using a human keypoint detection algorithm, represents the pose of each detected worker as a feature matrix, and outputs a set of pose feature matrices containing spatiotemporal information. This module integrates the YOLOv5 worker detection algorithm and the HRNet pose estimation model, and achieves real-time processing through GPU acceleration. It can simultaneously process multiple construction workers from multiple video streams and output structured pose feature representations.

[0148] The differential privacy protection module 30 is used to analyze the sensitivity and intrinsic correlation of the feature matrices in the attitude feature matrix set, apply principal component analysis to reduce the dimensionality of the attitude feature matrix set to obtain dimensionality-reduced eigenvalues, determine the optimal privacy budget allocation strategy, and generate privacy-preserving feature data that satisfies ε-differential privacy guarantees by adding calibrated Laplacian noise or Gaussian noise to the dimensionality-reduced eigenvalues. This module implements an innovative differential privacy protection mechanism, privatizing attitude features in the early stages of data processing, ensuring that subsequent processed and stored data meet strict privacy protection requirements, and fundamentally preventing the leakage of personal privacy information.

[0149] The violation detection module 40 is used to extract high-level representations of spatiotemporal features based on the privacy-preserving feature data using a pre-trained spatiotemporal graph convolutional network model, and to perform behavior classification and dangerous area determination, outputting the detection results of safety violation events. This module deploys a specially adapted ST-GCN model for noisy feature data, simultaneously completing behavior recognition and region determination through a multi-task learning framework, and employs confidence filtering and temporal consistency checks to improve detection accuracy, achieving high-precision violation detection under the premise of privacy protection.

[0150] The safety management module 50 is used to assess the risk level of each violation event based on the detection results of the safety violation events and in conjunction with pre-configured risk level information of the construction area, and generate graded early warning signals. Violations are recorded and stored in the safety management database, forming a traceable safety management mechanism. This module integrates risk assessment algorithms, a graded early warning system, a database management system, and visualization analysis tools, achieving fully automated management from violation detection to risk response. It provides intelligent and systematic technical support for safety management at construction sites, while strictly adhering to privacy protection principles and not storing any personally identifiable original images or sensitive information.

[0151] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present application should be included within the protection scope of the present application.

Claims

1. A safety monitoring method for building construction, characterized in that, include: Collect raw video data from the construction site, perform noise reduction, illumination correction, and frame rate adjustment on the raw video data, and output a standardized video data stream; Receive the standardized video data stream, use the human key point detection algorithm to extract the posture features of the construction workers, represent the posture of each detected person as a feature matrix, and output a set of posture feature matrices containing spatiotemporal information; The sensitivity and intrinsic correlation of the feature matrices in the attitude feature matrix set are analyzed. Principal component analysis is applied to the attitude feature matrix set to reduce the dimensionality and obtain the dimensionality-reduced feature principal components. The optimal privacy budget allocation strategy is determined. By adding calibrated Laplace noise or Gaussian noise to the dimensionality-reduced feature principal components, privacy-preserving feature data that satisfies ε-differential privacy guarantee is generated. Based on the privacy-preserving feature data, a high-level representation of the spatiotemporal features is extracted through a pre-trained spatiotemporal graph convolutional network model, and behavior classification and dangerous area determination are performed to output the detection results of security violation events. Based on the detection results of the safety violations, the risk level of each violation is assessed in conjunction with the pre-configured risk level information of the construction area, and a graded early warning signal is generated. The violation is recorded and stored in the safety management database to form a traceable safety management mechanism.

2. The method according to claim 1, characterized in that, The raw video data collected from the construction site is then subjected to noise reduction, illumination correction, and frame rate adjustment to output a standardized video data stream, including: Cameras are deployed in key areas of the construction site to form a multi-angle coverage monitoring network, collect raw video data from the construction site, and obtain raw multi-source video data. The original multi-source video data is time-synchronized to output multi-source video data with consistent timestamps; An adaptive image enhancement algorithm is applied to the multi-source video data with consistent timestamps to perform noise reduction, illumination non-uniformity correction and contrast adjustment, and the enhanced video data is output. The enhanced video data is subjected to frame rate adjustment and resolution normalization to obtain the standardized video data stream.

3. The method according to claim 1, characterized in that, The method utilizes a human keypoint detection algorithm to extract the posture features of construction workers and represents each detected worker's posture as a feature matrix, outputting a set of posture feature matrices containing spatiotemporal information, including: Receive the standardized video data stream, apply the YOLOv5 algorithm to perform personnel detection, and output personnel location information; Based on the personnel location information, for each personnel region, the HRNet pose estimation model is applied to extract the spatial coordinates of human key points and output the coordinates of human key points. Based on the coordinates of the human body key points, construct a human posture feature matrix containing the key point coordinates and corresponding confidence scores, and output a single-frame posture feature matrix; Based on the single-frame pose feature matrix, a pose sequence is constructed by combining the temporal information of consecutive video frames, thus obtaining the pose feature matrix set containing spatiotemporal information.

4. The method according to claim 3, characterized in that, The application of the HRNet pose estimation model to extract the spatial coordinates of human key points and output the coordinates of human key points includes: The key points of the human body include the head, shoulders, elbows, wrists, hips, knees, and ankles.

5. The method according to claim 1, characterized in that, The process involves analyzing the sensitivity and intrinsic correlation of the posture feature matrix set, applying principal component analysis to reduce the dimensionality of the posture feature matrix set to obtain dimensionality-reduced principal components, determining the optimal privacy budget allocation strategy, and generating privacy-preserving feature data that satisfies ε-differential privacy guarantees by adding calibrated Laplacian noise or Gaussian noise to the dimensionality-reduced principal components. Receive the attitude feature matrix set, analyze the sensitivity of the feature matrices in the attitude feature matrix set, and output the division results of sensitive parameters and insensitive parameters; Based on the partitioning results, principal component analysis is applied to reduce the dimensionality of the pose feature matrix set, and the reduced principal component representation is output. Based on the dimensionality-reduced principal component representation, the intrinsic correlation of the feature matrix is ​​analyzed and a differential privacy noise addition mechanism is designed to determine the privacy budget allocation strategy and noise scale. Based on the privacy budget allocation strategy and the noise scale, add Laplace noise or Gaussian noise that satisfies the ε-difference privacy guarantee to the dimensionality-reduced principal component representation, and output the noisy principal component; The noisy principal components are reconstructed into a representation of the original feature space by applying an inverse transformation, thus obtaining the privacy-preserving feature data.

6. The method according to claim 5, characterized in that, The application of principal component analysis to reduce the dimensionality of the pose feature matrix set outputs a dimensionality-reduced principal component representation, including: The attitude feature matrix set is labeled based on the partitioning information of sensitive and non-sensitive parameters in the partitioning result, and the labeled attitude feature matrix set is output. Principal component analysis is applied to the labeled attitude feature matrix set to calculate the feature covariance matrix and extract eigenvalues ​​and eigenvectors, outputting the eigenvalue sorting results and the corresponding eigenvectors; Based on the eigenvalue sorting results, the top K eigenvectors whose cumulative contribution rate reaches a preset threshold are selected as principal components, and K principal component vectors are output. A linear transformation is performed on the attitude feature matrix set based on the K principal component vectors, projecting the original high-dimensional attitude feature matrix set onto the low-dimensional principal component space to obtain the dimensionality-reduced principal component representation.

7. The method according to claim 1, characterized in that, The high-level representation of spatiotemporal features is extracted through a pre-trained spatiotemporal graph convolutional network model, and behavior classification and hazard area determination are performed to output safety violation event detection results, including: The privacy-preserving feature data is input into a pre-trained spatiotemporal graph convolutional network model for spatiotemporal feature extraction, and a high-level representation of the spatiotemporal features is output. The high-level representation based on the aforementioned spatiotemporal features utilizes a multi-task learning framework to simultaneously perform behavior classification and dangerous area determination, outputting preliminary results of violation behavior identification. The preliminary violation identification results are filtered using a confidence threshold and combined with a temporal consistency test to filter out false detections, thus obtaining the security violation event detection results.

8. The method according to claim 7, characterized in that, The high-level representation based on the aforementioned spatiotemporal features utilizes a multi-task learning framework to simultaneously perform behavior classification and danger zone determination, outputting preliminary violation behavior identification results, wherein: The preliminary violation identification results include violations such as not wearing a seat belt, entering a dangerous area, and improper operation.

9. The method according to claim 1, characterized in that, The process involves assessing the risk level of each violation based on the detection results of the safety violations, combined with pre-configured risk level information for the construction area, generating tiered early warning signals, and storing violation records in a safety management database to form a traceable safety management mechanism, including: Based on the detection results of the safety violations and the pre-configured risk level information of the construction area, the risk level of each violation is assessed, and violation information with risk level is output. Based on the information on violations with risk levels, a graded early warning signal is generated and pushed to on-site management personnel, and graded early warning information is output. The violation type, time, location, and risk level of the violation event information with risk level are recorded in the security management database to establish a security record archive; Based on the safety record archive, construction safety analysis reports and trend charts are constructed to obtain the traceable safety management mechanism.

10. A safety monitoring system for building construction, characterized in that, include: The video acquisition and preprocessing module is used to acquire raw video data from the construction site, perform noise reduction, illumination correction, and frame rate adjustment on the raw video data, and output a standardized video data stream. The posture feature extraction module is used to receive the standardized video data stream, extract the posture features of the construction workers using the human key point detection algorithm, represent the posture of each detected person as a feature matrix, and output a set of posture feature matrices containing spatiotemporal information. The differential privacy protection module is used to analyze the sensitivity and intrinsic correlation of the feature matrices in the attitude feature matrix set, apply principal component analysis to reduce the dimensionality of the attitude feature matrix set to obtain the dimensionality-reduced feature principal components, determine the optimal privacy budget allocation strategy, and generate privacy-protected feature data that satisfies ε-differential privacy guarantee by adding calibrated Laplacian noise or Gaussian noise to the dimensionality-reduced feature principal components. The violation detection module is used to extract high-level representations of spatiotemporal features based on the privacy-preserving feature data through a pre-trained spatiotemporal graph convolutional network model, perform behavior classification and dangerous area determination, and output the detection results of safety violation events. The safety management module is used to assess the risk level of each violation event based on the detection results of the safety violation events and in combination with the pre-configured risk level information of the construction area, and generate graded early warning signals. The violation behavior is recorded and stored in the safety management database to form a traceable safety management mechanism.

Images