Construction site worker behavior intelligent identification and early warning system

Abstract

The invention discloses a construction site worker behavior intelligent identification and early warning system, which collects multi-directional video streams in real time through a front-end video collection module, guarantees timely data updating, and facilitates real-time monitoring of worker behaviors. The early warning and display module displays a field picture, a worker skeleton diagram, a behavior state and duration time in real time, early warning is immediately carried out when abnormity occurs, an instant feedback mechanism of the early warning and display module enables safety management personnel to rapidly know the abnormity and take measures, accidents caused by information lag are avoided, and the defect that the real-time early warning capability is insufficient is overcome; the behavior recognition module adopts an EnhanceST-GCN model, can better capture a space-time relation of skeleton key points, accurately recognize a complex behavior mode, improve recognition precision and solve the problem of low recognition precision, and meanwhile, the model is high in video data processing efficiency, rapid in analysis, feature extraction and classification, reduces unnecessary calculation steps and meets the real-time monitoring requirement of a construction site.

Description

Technical Field

[0001] This invention relates to the field of worker behavior recognition and early warning technology, and in particular to an intelligent recognition and early warning system for worker behavior at construction sites. Background Technology

[0002] In the construction industry, worker safety and efficiency are of paramount importance. With the rapid development of computer vision and deep learning technologies, it has become possible to use video analytics to achieve intelligent identification and early warning of worker behavior on construction sites. Video acquisition equipment can acquire video data from construction sites in real time, providing a rich source of information for subsequent analysis. Deep learning models, such as graphical convolutional networks (GCN) and their improved models, have demonstrated powerful capabilities in processing video stream data and identifying complex behavioral patterns.

[0003] However, although there are some video analytics-based security monitoring systems, there are still shortcomings in dedicated systems for recognizing worker behavior at construction sites: some systems can only achieve simple video recording and playback functions, lacking real-time behavior recognition and early warning capabilities; while some systems have behavior recognition functions, the recognition accuracy and efficiency need to be improved, and they cannot meet the needs of the complex and ever-changing environment at construction sites. Summary of the Invention

[0004] In view of this, the present invention proposes an intelligent identification and early warning system for worker behavior at construction sites, which can effectively solve the shortcomings of existing technologies, such as lack of real-time behavior identification and early warning capabilities, and low identification accuracy and efficiency.

[0005] The technical solution of this invention is implemented as follows:

[0006] A construction site worker behavior intelligent recognition and early warning system includes:

[0007] The front-end video acquisition module is used to collect video stream data in real time, covering all directions and perspectives of the worker's activity range;

[0008] The behavior recognition module is used to identify worker behavior categories based on video stream data using the EnhancedST-GCN model;

[0009] The early warning and display module is used to display real-time on-site video footage, worker skeleton diagrams, identified behavioral states and their durations, and to issue early warnings when abnormal behavior or non-working states are detected and timeouts are exceeded.

[0010] As a further optional solution to the aforementioned intelligent recognition and early warning system for worker behavior at construction sites, the behavior recognition module includes:

[0011] The pose estimation unit is used to extract human skeleton key point sequences, including coordinate information and confidence information, from each worker detected in the video stream data in real time based on the YOLO11-Pose model.

[0012] The behavior recognition unit is used to process the human skeleton key point sequence based on the EnhancedST-GCN model to identify the worker's behavior category;

[0013] A dual-window detection unit is used to output stable behavior category recognition results.

[0014] As a further optional solution to the intelligent recognition and early warning system for worker behavior at the construction site, the pose estimation unit extracts a sequence of human skeletal key points, including coordinate information and confidence information, in real time for each worker detected in the video stream data based on the YOLO11-Pose model. Specifically, this includes:

[0015] The pre-trained YOLO11-Pose model is used to process each frame of the video stream data in real time to detect the position of all workers and their human body regions in the picture.

[0016] For each worker detected, extract key points of their human skeleton;

[0017] For each extracted key point, its coordinate information and confidence information are recorded to form a sequence of human skeletal key points containing coordinate and confidence information.

[0018] As a further optional solution for the intelligent recognition and early warning system for worker behavior at the construction site, the key points of the human skeleton are 14, including both ears, both shoulders, both elbows, both wrists, hip joints, both knees, both ankles and nose.

[0019] As a further optional solution to the intelligent recognition and early warning system for worker behavior at the construction site, the behavior recognition unit processes the human skeletal key point sequence based on the EnhancedST-GCN model to identify the worker's behavior category, specifically including:

[0020] The received keypoint sequence is input into the EnhancedST-GCN model. The EnhancedST-GCN model dynamically adjusts the spatiotemporal feature fusion weights according to the confidence level of the keypoints based on the confidence level gating mechanism, and obtains the confidence-weighted spatiotemporal features.

[0021] By using the axial Transformer encoding mechanism to model the spatiotemporal features after confidence weighting in the time dimension, the temporal dependencies of human motion are captured, and features with enhanced temporal dependencies are obtained.

[0022] Based on the decoupled feature pooling strategy, spatial and temporal features are extracted from the features that enhance temporal dependencies, resulting in spatial saliency features and temporal statistical features;

[0023] Based on the classification head, the characteristics of enhanced temporal dependencies, spatial saliency features, and temporal statistical features are used to classify and obtain the probability distribution of the worker's current behavior category;

[0024] The behavior category with the highest probability is output as the recognition result.

[0025] As a further optional solution to the intelligent recognition and early warning system for worker behavior at the construction site, the dual-window detection unit includes:

[0026] A temporal buffer window is used to accumulate a fixed-length keypoint sequence and the corresponding behavior recognition results;

[0027] A probabilistic smoothing window is used to perform a weighted average of behavior recognition results for a fixed-length keypoint sequence.

[0028] As a further optional solution for the intelligent recognition and early warning system for worker behavior at the construction site, the length of the time-series buffer window is 60 frames, the length of the probability smoothing window is 15 recognition cycles, and an exponential decay weighted strategy is used for averaging.

[0029] A method for intelligent identification and early warning of worker behavior at construction sites, specifically including:

[0030] Real-time acquisition of video stream data covering all directions and perspectives of the worker's activity area;

[0031] The EnhancedST-GCN model was used to identify worker behavior categories based on video stream data;

[0032] It displays real-time video footage, worker skeleton diagrams, identified behavioral states and their durations, and issues warnings when abnormal behavior or non-working states exceed the time limit.

[0033] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described intelligent identification and early warning method for worker behavior at construction sites.

[0034] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described intelligent identification and early warning method for worker behavior at construction sites.

[0035] The beneficial effects of this invention are as follows: The front-end video acquisition module collects video stream data in real time, covering all directions and angles of the worker's activity range. This means the system can acquire real-time images of the construction site, ensuring data timeliness and enabling real-time monitoring of worker behavior. Simultaneously, the early warning and display module displays real-time video footage, worker skeleton diagrams, identified behavioral states, and their durations. It immediately issues warnings when abnormal behavior or non-working status is detected and timeouts occur. This instant feedback mechanism allows safety management personnel to quickly learn of abnormal situations at the construction site and take timely measures, avoiding safety accidents caused by information delays and effectively compensating for the lack of real-time warning capabilities in existing technologies. Furthermore, the behavior recognition module… The module employs the EnhancedST-GCN model to identify worker behavior categories based on video stream data. This model can better capture the spatiotemporal relationships between key points on a worker's skeleton, accurately identifying various complex behavioral patterns, such as normal operations and violations, significantly improving the accuracy of behavior recognition and solving the problem of low recognition accuracy in existing systems. Furthermore, the EnhancedST-GCN model is highly efficient in processing video stream data, quickly analyzing the acquired video data, extracting key features, and classifying behaviors. Compared to some traditional behavior recognition methods, it reduces unnecessary computational steps, enabling the identification of worker behavior in a shorter time, thus meeting the needs of real-time monitoring at construction sites. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of the composition of a construction site worker behavior intelligent recognition and early warning system according to the present invention;

[0038] Figure 2 This is a schematic diagram of the structure of the EnhancedST-GCN model in this invention;

[0039] Figure 3 This is a flowchart illustrating the workflow of the dual-window detection mechanism in this invention.

[0040] Figure 4 A flowchart of a method for intelligent identification and early warning of worker behavior at construction sites;

[0041] Figure 5 This is a schematic diagram of the composition of a computing device according to the present invention. Detailed Implementation

[0042] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] refer to Figures 1 to 5 A construction site worker behavior intelligent recognition and early warning system includes a front-end video acquisition module, a behavior recognition module, and an early warning and display module, wherein:

[0044] The front-end video acquisition module includes surveillance cameras deployed in key areas of the construction site. The surveillance cameras are high-resolution network cameras to adapt to the complex environment of open-air construction sites and achieve effective ranging over a wide range, used to collect video stream data in real time covering all directions of the workers' activity range.

[0045] Specifically, by deploying multiple surveillance cameras in key areas of the construction site, video stream data can be collected from various angles. Images from different angles can provide more comprehensive information on worker behavior. For example, from the front, the worker's operating actions can be clearly seen, and from the side, the worker's body posture and movement trajectory can be observed. These multi-angle data complement each other, which helps the behavior recognition module to analyze and judge the worker's behavior category more accurately, greatly improving the accuracy of behavior recognition.

[0046] The system collects video stream data in real time, covering all directions and perspectives of the worker's activity range, ensuring the real-time nature and continuity of the data. The system can continuously acquire worker behavior information and promptly detect dynamic changes in behavior. For example, it can capture and identify the process of a worker transitioning from normal operation to violation of regulations in real time, providing the possibility for timely early warning.

[0047] A behavior recognition module is used to identify worker behavior categories based on video stream data using an EnhancedST-GCN model. In some embodiments, the behavior recognition module includes:

[0048] The pose estimation unit is used to extract human skeleton key point sequences, including coordinate information and confidence information, from each worker detected in the video stream data in real time based on the YOLO11-Pose model.

[0049] The behavior recognition unit is used to process the human skeleton key point sequence based on the EnhancedST-GCN model to identify the worker's behavior category;

[0050] A dual-window detection unit is used to output stable behavior category recognition results.

[0051] Specifically, based on the YOLO11-Pose model, the pose estimation unit can detect the key points of each worker's skeleton in real time and accurately from video stream data, and extract sequences containing coordinate information and confidence information. The YOLO11-Pose model has high detection accuracy and speed, and can quickly locate the joint positions of workers, such as shoulders, elbows, and knees, in complex construction site environments. The coordinate information accurately describes the position of the key points in the image, while the confidence information reflects the reliability of the detection results, providing high-quality basic data for subsequent behavior recognition. Construction sites often have problems such as worker occlusion, lighting changes, and complex backgrounds. The YOLO11-Pose model has undergone extensive training and optimization, and has good robustness to these complex situations. For example, even if part of a worker's body is occluded by building materials or other workers, the model can still detect visible key points as accurately as possible and reasonably infer the position of the occluded part, ensuring the integrity and accuracy of the key point sequence.

[0052] The behavior recognition unit uses the EnhancedST-GCN model to process the human skeleton keypoint sequence output by the pose estimation unit. The EnhancedST-GCN model has unique advantages in processing spatiotemporal sequence data, effectively capturing the spatiotemporal relationship between keypoints. It can analyze the changes in the skeletal posture of workers at different time points and identify various behavior patterns, such as walking, carrying, and operating machinery. Through learning and training on a large amount of data, the model can accurately map keypoint sequences to corresponding behavior categories, improving the accuracy of behavior recognition. The behaviors of workers at construction sites are diverse and complex. The EnhancedST-GCN model has strong generalization ability and can adapt to the behavioral habits of different workers and their behavior in various construction scenarios. Whether it is the worker's routine work behavior or some special and uncommon behaviors, the model can effectively identify them, providing comprehensive behavioral information for the safety management of construction sites.

[0053] The dual-window detection unit further processes the behavior recognition results to output stable behavior category recognition results. It can comprehensively consider short-term and long-term behavior recognition information, reducing fluctuations in recognition results caused by misidentification of individual frames or noise interference. For example, at a certain moment, behavior recognition errors may occur due to video blurring or key point detection errors, but the dual-window detection unit can filter out outliers by comprehensively analyzing the results of multiple time points, outputting more stable and reliable behavior categories. The existence of this unit enhances the reliability of the entire behavior recognition module. In practical applications at construction sites, the system needs to work continuously and stably to provide accurate information for safety management personnel. The dual-window detection unit can effectively cope with various uncertainties, ensuring the accuracy and stability of behavior recognition results, thereby improving the reliability of the entire intelligent recognition and early warning system for worker behavior at construction sites.

[0054] In some embodiments, the pose estimation unit extracts a sequence of human skeletal key points, including coordinate information and confidence information, for each worker detected in the video stream data based on the YOLO11-Pose model in real time. Specifically, this includes:

[0055] The pre-trained YOLO11-Pose model is used to process each frame of the video stream data in real time to detect the position of all workers and their human body regions in the picture.

[0056] For each worker detected, extract key points of their human skeleton;

[0057] For each extracted key point, its coordinate information and confidence information are recorded to form a sequence of human skeletal key points containing coordinate and confidence information.

[0058] Specifically, the pre-trained YOLO11-Pose model processes each frame of the video stream data in real time, enabling rapid detection of the positions and human body regions of all workers in the scene. This means the system can instantly acquire dynamic information about workers at the construction site without significant delays. For example, when a worker begins a new operation, the system can detect their positional changes in a short time, providing timely data support for subsequent behavior analysis and meeting the needs of real-time monitoring at the construction site. For each detected worker, the model can quickly extract their skeletal key points and record coordinate and confidence information to form a key point sequence. This efficient processing capability allows the system to process a large amount of video frame data in a short time, ensuring that even in complex construction site environments, it can still quickly and accurately acquire worker posture information, improving the efficiency of the entire behavior recognition module.

[0059] The YOLO11-Pose model, pre-trained, possesses powerful object detection capabilities. It can accurately detect the positions of all workers and their body areas in video footage. Even in situations where workers occlude each other or the background is complex at a construction site, the model can still perform detection tasks well. For example, in scenarios where multiple workers are working simultaneously and their bodies partially occlude each other, the model can still accurately identify the approximate position and body area of ​​each worker. For each detected worker, the model can accurately extract key points of the human skeleton. These key points cover important positions such as major joints. The coordinate information can accurately describe the position of the key points in the image, while the confidence information reflects the reliability of the detection results. Accurate extraction of skeletal key points is crucial for subsequent behavior recognition. It provides precise data for analyzing workers' postures and movements, helping to improve the accuracy of behavior recognition.

[0060] In some embodiments, the key points of the human skeleton are 14, including both ears, both shoulders, both elbows, both wrists, hip joints, both knees, both ankles, and nose.

[0061] Specifically, these 14 key points cover the major joints and prominent parts of the human body from top to bottom: the ears and nose help determine the position and orientation of the head; the shoulders, elbows, and wrists clearly reflect the movements and postures of the upper limbs; the hip joint is a key part connecting the upper and lower limbs, while the knees and ankles play a decisive role in the movement of the lower limbs; through these key points, the posture of the human body in space can be comprehensively and accurately described, whether it is standing, walking, bending over, or other complex movements, all of which can be effectively captured and represented; for various behaviors of workers on construction sites, such as carrying heavy objects and operating tools, these key points can accurately reflect the detailed changes in movements. For example, when a worker carries an item, the degree of bending of the arms and the angle of body tilt can be reflected through the coordinate changes of the key points, providing rich information for subsequent behavior recognition;

[0062] In the process of behavior recognition, the features extracted based on these 14 key points can more accurately reflect the worker's behavioral characteristics. Compared with the use of fewer key points, more key points can provide more detailed motion information, which helps the model better understand the worker's behavior patterns. For example, by analyzing the relative positions and movement trajectories of the key points of the shoulders and elbows, it is possible to more accurately determine whether the worker is pushing, pulling, or lifting. Comprehensive and accurate key point information can reduce the misjudgment rate of behavior recognition. In the complex environment of the construction site, there may be various interference factors, such as the movement of other objects and changes in light. The setting of 14 key points can provide more stable and reliable features, enabling the behavior recognition model to better distinguish different behavior categories and reduce misjudgments caused by insufficient or ambiguous features.

[0063] In some embodiments, the behavior recognition unit processes the human skeletal keypoint sequence based on the EnhancedST-GCN model to identify the worker's behavior category, specifically including:

[0064] The received keypoint sequence is input into the EnhancedST-GCN model. The EnhancedST-GCN model dynamically adjusts the spatiotemporal feature fusion weights according to the confidence level of the keypoints based on the confidence level gating mechanism, and obtains the confidence-weighted spatiotemporal features.

[0065] By using the axial Transformer encoding mechanism to model the spatiotemporal features after confidence weighting in the time dimension, the temporal dependencies of human motion are captured, and features with enhanced temporal dependencies are obtained.

[0066] Based on the decoupled feature pooling strategy, spatial and temporal features are extracted from the features that enhance temporal dependencies, resulting in spatial saliency features and temporal statistical features;

[0067] Based on the classification head, the characteristics of enhanced temporal dependencies, spatial saliency features, and temporal statistical features are used to classify and obtain the probability distribution of the worker's current behavior category;

[0068] The behavior category with the highest probability is output as the recognition result.

[0069] Specifically, the EnhancedST-GCN model uses a confidence gating mechanism to dynamically adjust the spatiotemporal feature fusion weights based on the confidence of key points. In construction site video data, due to issues such as occlusion and motion blur, the confidence of different key points will vary. By dynamically adjusting the weights, the model can pay more attention to the feature information provided by key points with high confidence and reduce noise interference from key points with low confidence, thereby obtaining more accurate and reliable confidence-weighted spatiotemporal features, laying a good foundation for subsequent behavior recognition.

[0070] By using the axial Transformer encoding mechanism to model the spatiotemporal features after confidence weighting in the time dimension, it is possible to capture the temporal dependencies of human movement. Worker behavior is a continuous process, and temporal dependencies are crucial for accurate behavior identification. For example, the process of a worker from preparation to actual operation can be well reflected by time dimension modeling.

[0071] Based on the decoupled feature pooling strategy, spatial and temporal features are extracted from the features that enhance temporal dependencies, resulting in spatial saliency features and temporal statistical features. Spatial saliency features can reflect the worker's posture and position changes in space, while temporal statistical features can reflect information such as the duration and frequency of behavior. Multi-dimensional feature extraction comprehensively characterizes the features of worker behavior.

[0072] Classifying workers based on features that enhance temporal dependencies, spatial saliency features, and temporal statistical features allows for a comprehensive consideration of behavioral characteristics across different dimensions. These features describe worker behavior from various perspectives, and by analyzing them comprehensively, the classification head can more accurately categorize worker behavior into the appropriate categories. For example, for two similar behaviors, "carrying" and "operating," spatial saliency features may reflect differences in body posture, while temporal statistical features may reflect differences in the duration and rhythm of the action. Classifying by integrating these features can improve classification accuracy and reduce misclassification.

[0073] The design of the EnhancedST-GCN model and its related mechanisms ensures recognition accuracy while also considering computational efficiency. The axial Transformer encoding mechanism and decoupled feature pooling strategy can complete feature extraction and classification tasks within a reasonable time, meeting the needs of real-time behavior recognition at construction sites and improving the overall efficiency of the behavior recognition system.

[0074] In some embodiments, the dual-window detection unit includes:

[0075] The temporal buffer window is used to accumulate a continuous sequence of human skeleton key points of fixed length T (T≥10) and the corresponding behavior recognition results. When the accumulated number of frames reaches T, the window data is updated according to the first-in-first-out principle to ensure temporal continuity.

[0076] The probability smoothing window runs synchronously with the temporal buffer window, and performs weighted averaging on the behavior recognition results of the recent K (3≤K≤T) frames.

[0077] Specifically, the temporal buffer window is used to accumulate a fixed length T (T≥10) of continuous human skeletal key point sequences and corresponding behavior recognition results. On the construction site, the worker's behavior is a continuous process. By accumulating a certain length of data, the dynamic changes of behavior can be captured more comprehensively. For example, a worker may need multiple consecutive frames to complete a complete operation. The temporal buffer window can save the data of these frames to provide complete information for subsequent analysis. When the accumulated number of frames reaches T, the window data is updated according to the first-in-first-out principle. This mechanism ensures that the data in the window always maintains the latest continuous behavior information, avoiding data gaps and chaos. During real-time monitoring, even if new data keeps coming in, the continuity of the temporal sequence can be guaranteed, enabling the system to accurately track changes in worker behavior and provide a guarantee for stable behavior recognition.

[0078] The probabilistic smoothing window operates synchronously with the temporal buffer window, performing weighted averaging on the recent K (3≤K≤T) frame behavior recognition results. Construction site environments are complex, and video data may contain noise and fluctuations, leading to unstable behavior recognition results for single frames. Weighted averaging reduces the impact of these fluctuations, making the behavior recognition results smoother and more stable. For example, a frame might be misidentified due to image blur, but the probabilistic smoothing window's comprehensive processing of multiple frame results can reduce the impact of such misidentifications on the final result. As worker behavior gradually changes, the probabilistic smoothing window adapts better. It doesn't rely solely on single-frame recognition results but comprehensively considers recent multi-frame results, thus more accurately reflecting the true state of behavior. For instance, as a worker transitions from normal operation to violation, the probabilistic smoothing window smoothly reflects this behavioral trend, providing a more reliable basis for timely warnings.

[0079] In some embodiments, the length of the temporal buffer window is 60 frames, the length of the probability smoothing window is 15 recognition periods, and an exponential decay weighted strategy is used for averaging.

[0080] Specifically, many behaviors of workers at construction sites have a certain degree of continuity and coherence. The length of 60 frames can cover a relatively long time period. For example, some complex operations may take several seconds to complete. According to the common video frame rate (such as 30 frames / second), 60 frames correspond to about 2 seconds, which is enough to capture the start, development and end of a worker's complete behavior. This provides comprehensive temporal information for behavior recognition. The longer temporal buffer window can better maintain the continuity of the time sequence when updating data according to the first-in-first-out principle. Even during the data update process, it can ensure that a sufficiently long continuous behavioral data is always retained within the window, making the system's tracking of worker behavior more stable and reducing the gaps in behavioral information caused by data updates.

[0081] The 15-cycle recognition period can, to some extent, eliminate fluctuations and noise in single-cycle recognition results. Construction site environments are complex, and video data may be affected by factors such as changes in lighting and occlusion, leading to inaccuracies in the results of a single recognition cycle. By applying an exponentially decaying weighted average to the results of 15 cycles, the impact of outliers can be reduced, making the behavior recognition results smoother and more stable, and improving the reliability of the recognition results. This length of probability smoothing window can comprehensively consider recent recognition results and reflect dynamic changes in behavior in a timely manner. The exponentially decaying weighted strategy gives greater weight to recognition results closer to the current time and smaller weight to results further away. This ensures the stability of the results while also enabling rapid response to sudden changes in worker behavior, such as when normal behavior quickly turns into illegal behavior, the system can promptly capture this trend.

[0082] The exponential decay weighting strategy assigns different weights to results from different recognition periods based on time sequence, making recent recognition results have a greater impact on the final average result. This aligns with the actual needs of worker behavior recognition at construction sites, as the current behavioral state better reflects the actual situation, while the behavioral state over a longer period has a relatively smaller impact on the current judgment. This weighting method can more accurately reflect the current characteristics of behavior and improve the accuracy of behavior recognition. When smoothing probability results, this strategy can perform weighting based on the temporal correlation of the data, avoiding the information ambiguity that may result from simple averaging. It can better preserve the trend information of behavioral changes while reducing noise interference, making the smoothed result closer to the real worker behavior state and providing a more accurate basis for system early warning and decision-making.

[0083] The early warning and display module, including an LED display screen and audio equipment, is used to display real-time video footage, worker skeleton diagrams, identified behavioral states and their durations, and to issue an audible alarm when abnormal behavior or non-working status is detected for an extended period.

[0084] Specifically, the LED display screen can display real-time video footage, worker skeleton diagrams, identified behavioral states, and their duration. This allows safety managers to intuitively understand the real-time situation at the construction site, as if they were there in person. For example, the video footage allows them to observe the overall environment of the construction site and the distribution of workers; the worker skeleton diagrams clearly show the workers' body postures and movement details; and the behavioral state and duration information allows managers to clearly know the behavior each worker is currently performing and how long that behavior has lasted, providing comprehensive and detailed information for timely monitoring of construction dynamics.

[0085] When abnormal behavior or non-working status timeout is detected, the audio equipment can promptly issue an audible alarm. The audible alarm has the characteristics of wide range and easy attention-grabbing. Even in noisy construction site environments, it can quickly alert safety management personnel, enabling them to take measures as soon as possible to avoid potential safety accidents.

[0086] In some embodiments, the system further includes a data storage module, which may be integrated into or external to a workstation, for storing pre-trained model parameters, configuration information, and historical data of recorded and identified behavioral states.

[0087] A method for intelligent identification and early warning of worker behavior at construction sites, specifically including:

[0088] Real-time acquisition of video stream data covering all directions and perspectives of the worker's activity area;

[0089] The EnhancedST-GCN model was used to identify worker behavior categories based on video stream data;

[0090] It displays real-time video footage, worker skeleton diagrams, identified behavioral states and their durations, and issues warnings when abnormal behavior or non-working states exceed the time limit.

[0091] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described intelligent identification and early warning method for worker behavior at construction sites.

[0092] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described intelligent identification and early warning method for worker behavior at construction sites.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A construction site worker behavior intelligent recognition and early warning system, characterized in that, include: The front-end video acquisition module is used to collect video stream data in real time, covering all directions and perspectives of the worker's activity range; The behavior recognition module is used to identify worker behavior categories based on video stream data using the EnhancedST-GCN model; The early warning and display module is used to display real-time on-site video footage, worker skeleton diagrams, identified behavioral states and their durations, and to issue early warnings when abnormal behavior or non-working states are detected and timeouts are exceeded.

2. The intelligent identification and early warning system for worker behavior at construction sites according to claim 1, characterized in that, The behavior recognition module includes: The pose estimation unit is used to extract human skeleton key point sequences, including coordinate information and confidence information, from each worker detected in the video stream data in real time based on the YOLO11-Pose model. The behavior recognition unit is used to process the human skeleton key point sequence based on the EnhancedST-GCN model to identify the worker's behavior category; A dual-window detection unit is used to output stable behavior category recognition results.

3. The intelligent identification and early warning system for worker behavior at construction sites according to claim 2, characterized in that, The pose estimation unit, based on the YOLO11-Pose model, extracts in real-time sequences of human skeletal key points, including coordinate and confidence information, for each worker detected in the video stream data. Specifically, this includes: The pre-trained YOLO11-Pose model is used to process each frame of the video stream data in real time to detect the position of all workers and their human body regions in the picture. For each worker detected, extract key points of their human skeleton; For each extracted key point, its coordinate information and confidence information are recorded to form a sequence of human skeletal key points containing coordinate and confidence information.

4. The intelligent identification and early warning system for worker behavior at construction sites according to claim 3, characterized in that, The key points of the human skeleton are 14, including both ears, both shoulders, both elbows, both wrists, hip joints, both knees, both ankles, and nose.

5. The intelligent identification and early warning system for worker behavior at construction sites according to claim 4, characterized in that, The behavior recognition unit processes the human skeletal key point sequence based on the EnhancedST-GCN model to identify the worker's behavior categories, specifically including: The received keypoint sequence is input into the EnhancedST-GCN model. The EnhancedST-GCN model dynamically adjusts the spatiotemporal feature fusion weights according to the confidence level of the keypoints based on the confidence level gating mechanism, and obtains the confidence-weighted spatiotemporal features. By using the axial Transformer encoding mechanism to model the spatiotemporal features after confidence weighting in the time dimension, the temporal dependencies of human motion are captured, and features with enhanced temporal dependencies are obtained. Based on the decoupled feature pooling strategy, spatial and temporal features are extracted from the features that enhance temporal dependencies, resulting in spatial saliency features and temporal statistical features; Based on the classification head, the characteristics of enhanced temporal dependencies, spatial saliency features, and temporal statistical features are used to classify and obtain the probability distribution of the worker's current behavior category; The behavior category with the highest probability is output as the recognition result.

6. The intelligent identification and early warning system for worker behavior at construction sites according to claim 5, characterized in that, The dual-window detection unit includes: A temporal buffer window is used to accumulate a fixed-length keypoint sequence and the corresponding behavior recognition results; A probabilistic smoothing window is used to perform a weighted average of behavior recognition results for a fixed-length keypoint sequence.

7. The intelligent identification and early warning system for worker behavior at construction sites according to claim 6, characterized in that, The temporal buffer window has a length of 60 frames, and the probability smoothing window has a length of 15 recognition periods, and an exponential decay weighted strategy is used for averaging.

8. A method for intelligent identification and early warning of worker behavior at construction sites, characterized in that, Specifically, it includes: Real-time acquisition of video stream data covering all directions and perspectives of the worker's activity area; The EnhancedST-GCN model was used to identify worker behavior categories based on video stream data; It displays real-time video footage, worker skeleton diagrams, identified behavioral states and their durations, and issues warnings when abnormal behavior or non-working states are detected and timeouts are triggered.

9. A computing device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the intelligent identification and early warning method for worker behavior at the construction site as described in claim 8.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the intelligent identification and early warning method for worker behavior at the construction site as described in claim 8.

Images