Intelligent construction site real-time monitoring method and system

Abstract

The application discloses a smart construction site real-time monitoring method and system, relates to the technical field of artificial intelligence, and comprises the following steps: collecting video data and three-dimensional coordinate data of construction site staff, obtaining posture features and position features, inputting the posture features into a posture recognition model based on a long short-term memory network, outputting the instantaneous posture category probability of the staff, inputting the position features into a regional risk assessment model based on a support vector machine, outputting the risk level of the position where the staff is located, performing weighted summation on the quantitative values of the instantaneous posture category probability and the risk level, outputting a comprehensive risk score, generating an early warning instruction, marking the position where the corresponding staff is located on a three-dimensional electronic map of the construction site, and calling corresponding real-time video data. The application realizes accurate discrimination of staff risks and quantitative assessment of scene danger levels by combining posture features with position features and performing quantitative risk assessment, and significantly improves early warning accuracy and system reliability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and system for real-time monitoring of smart construction sites. Background Technology

[0002] Smart construction sites are an inevitable product of the informatization development of the construction industry. They represent an emerging integrated management model for construction sites, deeply integrating IoT, big data, cloud computing, artificial intelligence, and building information modeling technologies. The aim is to transform the traditional production organization and management methods of construction sites. By deploying various sensors and data acquisition devices on the construction site, smart construction sites can comprehensively perceive, collect, and intelligently analyze key production factors in real time, thereby achieving visualized, intelligent, and refined management of engineering projects. This model effectively improves the quality and safety of projects, production efficiency, and cost control, and is a key pathway to promote the digital and intelligent transformation and upgrading of the construction industry.

[0003] To address the issues of inaccurate identification and untimely warnings of individual risk events involving construction workers, existing technologies rely on single video surveillance analyses of worker behavior and location-based intrusion detection. However, these methods often misjudge normal worker postures as risk events and fail to differentiate between behaviors of varying degrees of danger at the same location. This results in low reliability, high false alarm and missed alarm rates in the monitoring system's warnings. Therefore, a smart construction site real-time monitoring method and system are proposed to solve these problems. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for real-time monitoring of smart construction sites to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: Firstly, a real-time monitoring method for smart construction sites, comprising the following steps S1-S7:

[0006] S1. Collect and preprocess video data and 3D coordinate data of on-site workers to obtain preprocessed video data and 3D coordinate data;

[0007] S2. Obtain pose and position features from the preprocessed video data and 3D coordinate data;

[0008] S3. Input the posture features into the posture recognition model based on the long short-term memory network, and output the instantaneous posture category probability of the worker. The posture categories include standing, walking, squatting and falling.

[0009] S4. Input the location features into the regional risk assessment model based on support vector machine, and output the risk level of the staff's location;

[0010] S5. Using preset weighting coefficients, the quantified values ​​of instantaneous pose category probability and risk level are weighted and summed to output a comprehensive risk score. Using a historical dataset composed of video data and 3D coordinate data, the pose recognition model and regional risk assessment model are trained, and the weighting coefficients are optimized.

[0011] S6. Based on the comprehensive risk score, set the risk assessment logic, output the risk assessment result, and generate early warning instructions;

[0012] S7. Based on the warning instructions, mark the location of the corresponding staff on the three-dimensional electronic map of the construction site and retrieve the corresponding real-time video data.

[0013] A further improvement to the technical solution of this invention lies in: In step S1, the process of collecting and preprocessing video data and three-dimensional coordinate data of on-site workers to obtain preprocessed video data and three-dimensional coordinate data includes:

[0014] S11. Acquire and preprocess video data;

[0015] S111. Deploy closed-circuit television cameras at fixed high points in the predetermined area of ​​the construction site. The closed-circuit television cameras continuously collect video data covering the predetermined area.

[0016] S112. Decompose the acquired video data into an image frame sequence, and perform grayscale processing on each color image in the image frame sequence. Perform Gaussian filtering on each color image after grayscale processing to obtain preprocessed video data.

[0017] S12. Acquire and preprocess three-dimensional coordinate data;

[0018] S121. Deploy positioning base stations at known coordinate locations on the construction site and equip each worker's safety helmet with a positioning tag;

[0019] S122. The positioning tag periodically sends out wireless signals, which are received by the positioning base station. By calculating the time difference of the wireless signals arriving at different positioning base stations, the three-dimensional coordinate data of the positioning tag is calculated.

[0020] S123. Based on the building information model of the construction site, the calculated three-dimensional coordinate data is transformed into a unified construction site coordinate system using coordinate transformation formulas. The calculation process is as follows:

[0021] ;

[0022] in, This represents the converted coordinates of the construction site. This represents the calculated three-dimensional coordinates. Represents the rotation matrix. Represents the translation vector;

[0023] S124. Remove outliers from the converted 3D coordinate data to obtain preprocessed 3D coordinate data.

[0024] A further improvement to the technical solution of this invention lies in: S2, the process of obtaining pose features and position features from the preprocessed video data and three-dimensional coordinate data includes:

[0025] S21. Posture features include the two-dimensional pixel coordinates of the main joints of the worker's body, and the ratio of the height to the width of the minimum bounding rectangle determined based on the two-dimensional pixel coordinates.

[0026] S211. Train a human skeleton detection network model based on a convolutional neural network architecture using a large-scale public dataset containing human pose annotation information.

[0027] S212. Input each frame of the preprocessed video data into the pre-trained human skeleton detection network model to identify and output the two-dimensional pixel coordinates of the staff's body joints in the image.

[0028] S213. Based on the output two-dimensional pixel coordinates, determine the minimum bounding rectangle that surrounds the body joints, and calculate the ratio of the height to the width of the minimum bounding rectangle.

[0029] S22. Location features include the Euclidean distance between the preprocessed three-dimensional coordinate data of the worker at the current moment and the nearest point on the boundary of the preset danger zone, as well as the rate of change of the worker's current coordinate value in the vertical direction compared to the coordinate value at the previous moment.

[0030] A further improvement to the technical solution of this invention lies in: In S3, the posture features are input into a posture recognition model based on a long short-term memory network, and the instantaneous posture category probability of the worker is output. The posture categories include standing, walking, squatting, and falling. The process includes:

[0031] S31. Combine the attitude features in chronological order to form an attitude feature sequence with a preset time step.

[0032] S32. Input the pose feature sequence into the pose recognition model based on the long short-term memory network. The long short-term memory network processes the temporal dependencies in the pose feature sequence through its internal input gate, forget gate and output gate structure, and outputs the hidden state vector.

[0033] S33. Input the hidden state vector into the fully connected layer inside the Long Short-Term Memory network, and use the activation function to calculate the instantaneous posture class probabilities corresponding to standing, walking, squatting, and falling postures, respectively. The pose category corresponding to the highest instantaneous pose category probability is taken as the pose category predicted by the pose recognition model. Represents the fully connected layer for the first Output values ​​for each pose category Represents the total number of pose categories.

[0034] A further improvement to the technical solution of this invention lies in: In S4, the process of inputting location features into a regional risk assessment model based on a support vector machine and outputting the risk level of the worker's location includes:

[0035] S41. Combine location features into a location feature vector. ;

[0036] S42. Using a historical location feature dataset containing different risk level category labels, find the optimal hyperplane that separates the data points in the historical location feature dataset corresponding to different risk level category labels and maximizes the classification margin by solving a convex quadratic programming problem. Then, divide the decision space into decision intervals corresponding to each risk level based on the optimal hyperplane. The risk level category labels include safety, attention, and danger.

[0037] S43, Transfer the position feature vector The input is fed into a support vector machine-based regional risk assessment model, which uses an optimal hyperplane to analyze the location feature vectors. Classify;

[0038] S44. The optimal hyperplane is defined by a decision function, which is calculated as follows:

[0039] ;

[0040] in, The output value of the decision function. The number of support vectors, For Lagrange multipliers, For the first support vectors Risk level category labels For kernel function, For bias terms;

[0041] S45. Based on the output value of the decision function Determine the location feature vector within the decision interval. The corresponding risk level category label is used to output the risk level of the staff's location.

[0042] A further improvement to the technical solution of this invention lies in: In S5, the process of using preset weighting coefficients to perform a weighted summation of the quantified values ​​of the instantaneous attitude category probability and risk level, and outputting a comprehensive risk score, includes:

[0043] S51. Map the safety, caution, and danger risk levels to preset risk values ​​respectively. ;

[0044] S52. Use preset weighting coefficients. and The probability value of falling posture in the instantaneous posture category probability. Risk Values The weighted sum is then used to calculate the overall risk score. .

[0045] A further improvement to the technical solution of this invention lies in: In S5, the process of training the pose recognition model and the regional risk assessment model using a historical dataset composed of video data and three-dimensional coordinate data, and optimizing the weight coefficients, includes:

[0046] S53. Train the pose recognition model using a historical dataset consisting of video data and 3D coordinate data;

[0047] S531. Extract pose feature sequences from historical datasets and label each pose feature sequence with the corresponding true pose category.

[0048] S532. Input the labeled posture feature sequence into the posture recognition model, calculate the error between the posture category predicted by the posture recognition model and the true posture category through the loss function, and use the optimization algorithm to update the internal parameters of the posture recognition model in reverse according to the error until the error converges to the preset range.

[0049] S54. Train the regional risk assessment model using historical datasets;

[0050] S541. Extract location feature vectors from historical datasets and label each location feature vector with the corresponding real risk level category label;

[0051] S542. Using labeled position feature vectors as training data, determine the optimal hyperplane by solving a convex quadratic programming problem.

[0052] S55. Optimize the weighting coefficients;

[0053] S551. Divide the historical dataset into a validation dataset;

[0054] S552. Set performance metrics for evaluating the accuracy of risk event determination;

[0055] S553. Use a grid search method to traverse different combinations of weight coefficients within a preset range of values;

[0056] S554. For each combination of weight coefficients, calculate the comprehensive risk score using the data in the verification dataset, and iterate through the preset risk judgment thresholds. If the comprehensive risk score is greater than the risk judgment threshold, it is judged as a risk event. Then, calculate the value of the performance index under different risk judgment thresholds.

[0057] S555. Select the combination of weighting coefficients that makes the performance index values ​​reach the optimal value and the corresponding risk judgment threshold, and use them as the final weighting coefficients and risk judgment thresholds, respectively.

[0058] A further improvement to the technical solution of this invention lies in: In S6, the process of setting risk judgment logic based on the comprehensive risk score, outputting the risk judgment result, and generating an early warning instruction includes:

[0059] S61. Compare the comprehensive risk score calculated in real time with the risk judgment threshold determined in the process of weight coefficient optimization.

[0060] S62. When the comprehensive risk score is greater than the risk assessment threshold, the output risk assessment result is that a risk event has occurred; when the comprehensive risk score is not greater than the risk assessment threshold, the output risk assessment result is that the status is normal.

[0061] S63. When the risk assessment result indicates that a risk event has occurred, generate an early warning instruction containing the current staff member's identification and three-dimensional coordinate data.

[0062] A further improvement to the technical solution of this invention is that, in S7, the process of marking the location of the corresponding worker on the three-dimensional electronic map of the construction site according to the early warning instruction and retrieving the corresponding real-time video data includes:

[0063] S71. The monitoring terminal receives and parses the warning command;

[0064] S72. In the pre-loaded 3D electronic map of the construction site, locate the spatial position of the workers based on the 3D coordinate data parsed from the early warning instructions;

[0065] S73. At the located spatial position, generate and display a highlighted visual marker;

[0066] S74. Based on the three-dimensional coordinate data, combined with the preset installation location and monitoring field of view information of each closed-circuit television camera, match and determine the closed-circuit television camera corresponding to the coordinate point of the coverage space location.

[0067] S75. Send a retrieval command to the designated closed-circuit television camera to obtain real-time video data synchronized with the time the warning command was generated;

[0068] S76. Display the acquired real-time video data in the visual interface of the monitoring terminal.

[0069] Secondly, the smart construction site real-time monitoring system is used to realize the smart construction site real-time monitoring method, including a construction site data acquisition module, a personnel feature extraction module, a personnel posture analysis module, a regional risk assessment module, a weighted assessment optimization module, a risk judgment and early warning module, and a response execution module, wherein the modules are connected by electrical signals.

[0070] The construction site data acquisition module is used to collect and preprocess video data and 3D coordinate data of on-site workers to obtain preprocessed video data and 3D coordinate data.

[0071] The personnel feature extraction module is used to obtain pose and position features from preprocessed video data and 3D coordinate data;

[0072] The personnel posture analysis module is used to input posture features into a posture recognition model based on a long short-term memory network and output the instantaneous posture category probability of the staff. The posture categories include standing, walking, squatting and falling.

[0073] The regional risk assessment module is used to input location features into a support vector machine-based regional risk assessment model and output the risk level of the staff's location.

[0074] The weighted evaluation and optimization module is used to use preset weight coefficients to weight and sum the quantified values ​​of instantaneous pose category probability and risk level, output a comprehensive risk score, and use a historical dataset composed of video data and 3D coordinate data to train the pose recognition model and the regional risk assessment model, and optimize the weight coefficients.

[0075] The risk assessment and early warning module is used to set risk assessment logic based on the comprehensive risk score, output the risk assessment results, and generate early warning instructions;

[0076] The response execution module is used to mark the location of the corresponding workers on the 3D electronic map of the construction site according to the early warning instructions, and to retrieve the corresponding real-time video data.

[0077] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows:

[0078] 1. This invention provides a method and system for real-time monitoring of smart construction sites. By combining the posture characteristics of workers with high-precision position characteristics for comprehensive analysis, it can accurately distinguish between dangerous behaviors and normal working postures, significantly reducing misjudgments caused by analysis of a single data source.

[0079] 2. This invention provides a method and system for real-time monitoring of smart construction sites, which constructs a dual model of posture recognition and regional risk assessment, and outputs a quantitative comprehensive risk score by using a weighted fusion method. This enables refined judgment of risk levels under different scenarios and solves the problem of not being able to distinguish the degree of danger of the same behavior in different locations.

[0080] 3. This invention provides a method and system for real-time monitoring of smart construction sites. Based on multi-dimensional data fusion and quantitative risk assessment, it establishes a clear risk judgment logic, which effectively improves the overall reliability and timeliness of the real-time monitoring system for smart construction sites and significantly reduces the underreporting and false reporting of risk events. Attached Figure Description

[0081] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0082] Figure 1 The flowchart illustrates the real-time monitoring method for smart construction sites provided by this invention.

[0083] Figure 2 This is a structural block diagram of the smart construction site real-time monitoring system provided by the present invention. Detailed Implementation

[0084] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. 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.

[0085] Example 1, such as Figure 1 As shown, the present invention provides a real-time monitoring method for smart construction sites, including the following steps S1-S7:

[0086] S1. Collect and preprocess video data and 3D coordinate data of on-site workers to obtain preprocessed video data and 3D coordinate data.

[0087] In some embodiments, video data is acquired and preprocessed.

[0088] In some embodiments, closed-circuit television (CCTV) cameras are deployed at fixed high points in a predetermined area of ​​the construction site, and the CCTV cameras continuously collect video data covering the predetermined area.

[0089] In some embodiments, the acquired video data is decomposed into a sequence of image frames, and each color image in the sequence is converted to grayscale. Gaussian filtering is then applied to each grayscale-processed color image to obtain preprocessed video data.

[0090] In some embodiments, three-dimensional coordinate data are acquired and preprocessed.

[0091] In some embodiments, positioning base stations are deployed at known coordinate locations on the construction site, and each worker's safety helmet is equipped with a positioning tag.

[0092] In some embodiments, the positioning tag periodically transmits wireless signals, which are received by the positioning base station. By calculating the time difference between the arrival of the wireless signals at different positioning base stations, the three-dimensional coordinate data of the positioning tag is calculated.

[0093] In some embodiments, based on the building information model of the construction site, the calculated three-dimensional coordinate data is transformed into a unified construction site coordinate system using a coordinate transformation formula. The specific calculation formula is as follows:

[0094]

[0095] in, This represents the converted coordinates of the construction site. This represents the calculated three-dimensional coordinates. Represents the rotation matrix. This represents the translation vector.

[0096] In some embodiments, outlier removal is performed on the converted 3D coordinate data. 3D coordinate points whose spatial distance from the previous valid 3D coordinates exceeds a preset distance threshold are identified as outlier jump points and removed, thus obtaining preprocessed 3D coordinate data.

[0097] S2. Obtain pose and position features from the preprocessed video data and 3D coordinate data.

[0098] In some embodiments, the posture features include two-dimensional pixel coordinates of the main joints of the worker's body, and the ratio of the height to the width of the minimum bounding rectangle determined based on the two-dimensional pixel coordinates.

[0099] In some embodiments, a human skeleton detection network model based on a convolutional neural network architecture is trained using a large-scale public dataset containing human pose annotation information.

[0100] In some embodiments, each frame of the preprocessed video data is input into a pre-trained human skeleton detection network model to identify and output the two-dimensional pixel coordinates of the worker's body joints in the image.

[0101] In some embodiments, the minimum bounding rectangle surrounding the body joints is determined based on the output two-dimensional pixel coordinates, and the ratio of the height to the width of the minimum bounding rectangle is calculated.

[0102] In some embodiments, the location features include the Euclidean distance between the preprocessed three-dimensional coordinate data of the worker at the current moment and the nearest point on the boundary of the preset danger zone, and the rate of change of the worker's current coordinate value in the vertical direction compared to the coordinate value at the previous moment.

[0103] S3. Input the posture features into the posture recognition model based on the long short-term memory network, and output the instantaneous posture category probability of the staff. The posture categories include standing, walking, squatting and falling.

[0104] In some embodiments, the pose features are combined in chronological order to form a pose feature sequence with a preset time step.

[0105] In some embodiments, the pose feature sequence is input into a pose recognition model based on a long short-term memory network. The long short-term memory network processes the temporal dependencies in the pose feature sequence through its internal input gate, forget gate, and output gate structure, and outputs a hidden state vector.

[0106] In some embodiments, the hidden state vector is input into a fully connected layer inside a long short-term memory network, and the instantaneous posture category probabilities corresponding to standing, walking, squatting, and falling postures are calculated using activation functions. The pose category corresponding to the highest instantaneous pose category probability is taken as the pose category predicted by the pose recognition model. Represents the fully connected layer for the first Output values ​​for each pose category Represents the total number of pose categories.

[0107] S4. Input the location features into the regional risk assessment model based on support vector machine, and output the risk level of the staff's location.

[0108] In some embodiments, location features are combined into a location feature vector. .

[0109] In some embodiments, using a historical location feature dataset containing different risk level category labels, a convex quadratic programming problem is solved to find the optimal hyperplane that separates the data points in the historical location feature dataset corresponding to different risk level category labels and maximizes the classification margin. Based on the optimal hyperplane, decision intervals corresponding to each risk level are divided in the decision space. The risk level category labels include safety, attention, and danger.

[0110] In some embodiments, the location feature vector The input is fed into a support vector machine-based regional risk assessment model, which uses an optimal hyperplane to analyze the location feature vectors. Classify them.

[0111] In some embodiments, the optimal hyperplane is defined by a decision function, the specific formula for calculating the decision function is as follows:

[0112]

[0113] in, The output value of the decision function. The number of support vectors, For Lagrange multipliers, For the first support vectors Risk level category labels For kernel function, This is a bias term.

[0114] In some embodiments, based on the output value of the decision function Determine the location feature vector within the decision interval. The corresponding risk level category label is used to output the risk level of the staff's location.

[0115] S5. Using preset weighting coefficients, the instantaneous pose category probability and risk level quantification values ​​are weighted and summed to output a comprehensive risk score. Using a historical dataset composed of video data and 3D coordinate data, the pose recognition model and regional risk assessment model are trained, and the weighting coefficients are optimized.

[0116] In some embodiments, the risk levels of safety, caution, and danger are mapped to preset risk values. .

[0117] In some embodiments, preset weighting coefficients are used. and The probability value of falling posture in the instantaneous posture category probability. Risk Values The weighted sum is then used to calculate the overall risk score. .

[0118] In some embodiments, the pose recognition model is trained using a historical dataset consisting of video data and 3D coordinate data.

[0119] In some embodiments, pose feature sequences are extracted from historical datasets, and each pose feature sequence is labeled with the corresponding true pose category.

[0120] In some embodiments, a sequence of labeled pose features is input into a pose recognition model. The error between the pose category predicted by the pose recognition model and the true pose category is calculated using a loss function. An optimization algorithm is then used to update the internal parameters of the pose recognition model in reverse based on the error until the error converges to a preset range.

[0121] In some embodiments, historical datasets are used to train the regional risk assessment model.

[0122] In some embodiments, location feature vectors are extracted from historical datasets, and each location feature vector is labeled with a corresponding real risk level category label.

[0123] In some embodiments, labeled location feature vectors are used as training data, and the optimal hyperplane is determined by solving a convex quadratic programming problem.

[0124] In some embodiments, the weighting coefficients are optimized.

[0125] In some embodiments, a validation dataset is partitioned from the historical dataset.

[0126] In some embodiments, performance metrics are set to evaluate the accuracy of risk event determination.

[0127] In some embodiments, a grid search method is used to traverse different combinations of weight coefficients within a preset range of values.

[0128] In some embodiments, for each combination of weight coefficients, a comprehensive risk score is calculated using data from the verification dataset, and a preset risk judgment threshold is iterated through. Cases where the comprehensive risk score is greater than the risk judgment threshold are judged as risk events, and then the performance index values ​​under different risk judgment thresholds are calculated.

[0129] In some embodiments, the combination of weighting coefficients that makes the performance index values ​​optimal and the corresponding risk assessment threshold are selected as the final weighting coefficients and risk assessment thresholds, respectively.

[0130] S6. Based on the comprehensive risk score, set the risk judgment logic, output the risk judgment result, and generate early warning instructions.

[0131] In some embodiments, the comprehensive risk score calculated in real time is compared with the risk assessment threshold determined during the weighting coefficient optimization process.

[0132] In some embodiments, when the comprehensive risk score is greater than the risk assessment threshold, the output risk assessment result is that a risk event has occurred; when the comprehensive risk score is not greater than the risk assessment threshold, the output risk assessment result is that the state is normal.

[0133] In some embodiments, when the risk assessment result indicates that a risk event has occurred, an early warning instruction containing the current staff member's identification and three-dimensional coordinate data is generated.

[0134] S7. Based on the warning instructions, mark the location of the corresponding staff on the three-dimensional electronic map of the construction site and retrieve the corresponding real-time video data.

[0135] In some embodiments, the monitoring terminal receives and parses the warning command.

[0136] In some embodiments, the spatial location of workers is determined based on the three-dimensional coordinate data parsed from the warning instructions in a pre-loaded three-dimensional electronic map of the construction site.

[0137] In some embodiments, a highlighted visual marker is generated and displayed at the located spatial position.

[0138] In some embodiments, based on three-dimensional coordinate data and combined with preset installation location and monitoring field of view information of each closed-circuit television camera, the closed-circuit television camera corresponding to the coordinate point of the coverage space location is matched and determined.

[0139] In some embodiments, a retrieval command is sent to a designated closed-circuit television camera to obtain real-time video data synchronized with the time the warning command is generated.

[0140] In some embodiments, the acquired real-time video data is displayed in the visual interface of the monitoring terminal.

[0141] Example 2, as Figure 2 As shown, based on Embodiment 1, the present invention also provides a technical solution: a smart construction site real-time monitoring system, used to realize a smart construction site real-time monitoring method, including a construction site data acquisition module, a personnel feature extraction module, a personnel posture analysis module, a regional risk assessment module, a weighted assessment optimization module, a risk judgment and early warning module, and a response execution module, wherein the modules are electrically connected.

[0142] The construction site data acquisition module is used to collect and preprocess video data and 3D coordinate data of on-site workers to obtain preprocessed video data and 3D coordinate data.

[0143] The personnel feature extraction module is used to obtain posture and position features from preprocessed video data and 3D coordinate data.

[0144] The personnel posture analysis module is used to input posture features into a posture recognition model based on a long short-term memory network and output the instantaneous posture category probability of the staff. The posture categories include standing, walking, squatting and falling.

[0145] The regional risk assessment module is used to input location features into a support vector machine-based regional risk assessment model and output the risk level of the location of the staff.

[0146] The weighted evaluation and optimization module is used to use preset weight coefficients to perform a weighted summation of the quantified values ​​of instantaneous pose category probability and risk level, outputting a comprehensive risk score. It uses a historical dataset composed of video data and 3D coordinate data to train the pose recognition model and the regional risk assessment model, and optimizes the weight coefficients.

[0147] The risk assessment and early warning module is used to set risk assessment logic based on the comprehensive risk score, output the risk assessment results, and generate early warning instructions.

[0148] The response execution module is used to mark the location of the corresponding workers on the 3D electronic map of the construction site according to the early warning instructions, and to retrieve the corresponding real-time video data.

[0149] It should be noted that the aforementioned Figure 1 The explanations and effects of the method embodiments shown are also applicable to the method of this embodiment, and the principle is the same. Therefore, this embodiment will not be limited thereto.

[0150] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for real-time monitoring of a smart construction site, characterized in that, The method comprises the following steps: S1, collecting and preprocessing video data and three-dimensional coordinate data of workers on the construction site to obtain preprocessed video data and three-dimensional coordinate data; S2, obtaining posture features and position features from the preprocessed video data and three-dimensional coordinate data; S21, obtaining posture features from the preprocessed video data, the posture features including two-dimensional pixel coordinates of worker body joint points, and a ratio of height value and width value of a minimum circumscribed rectangle determined according to the two-dimensional pixel coordinates; S22, obtaining position features from the preprocessed three-dimensional coordinate data, the position features including a Euclidean distance between the preprocessed three-dimensional coordinate data of the worker at the current time and a nearest point on the boundary of a preset dangerous area, and a rate of change of the coordinate value of the worker at the current time in the vertical direction compared with the coordinate value at the previous time; S3, inputting the posture features into a posture recognition model based on a long short-term memory network to output an instantaneous posture category probability of the worker, the posture category including standing, walking, squatting and falling; S31, combining the posture features into a posture feature sequence with a preset time step in chronological order; S4, inputting the position features into a regional risk assessment model based on a support vector machine to output a risk level of the position where the worker is located; S5, using a preset weight coefficient to perform weighted summation on the instantaneous posture category probability and a quantitative value of the risk level to output a comprehensive risk score: S51, the risk level category label includes safety, attention and danger, and safety, attention and danger in the risk level are respectively mapped to preset risk values ; S52, using preset weight coefficients with , the fall posture probability value in the instantaneous posture category probability with the risk value weighted sum, calculate the comprehensive risk score ; Using a historical data set composed of the video data and the three-dimensional coordinate data, the posture recognition model and the regional risk assessment model are trained, and the weight coefficient is optimized; S551, dividing a validation data set from the historical data set; S552, setting a performance indicator for evaluating the accuracy of risk event determination; S553, using a grid search method to traverse different combinations of the weight coefficients within a preset value range; S554, for each combination of the weight coefficients, calculating the comprehensive risk score using data in the validation data set, and traversing a preset risk determination threshold, determining a risk event when the comprehensive risk score is greater than the risk determination threshold, and then calculating the value of the performance indicator under different risk determination thresholds; S555, selecting the combination of the weight coefficients and the corresponding risk determination threshold that make the value of the performance indicator optimal as the final weight coefficient and the risk determination threshold, respectively; S6, based on the comprehensive risk score, setting a risk determination logic to output a risk determination result and generate an early warning instruction; S7, according to the early warning instruction, marking the position where the corresponding worker is located on a three-dimensional electronic map of the construction site, and calling the corresponding real-time video data. 2.The smart construction site real-time monitoring method of claim 1, wherein, In the S1, the process of collecting and preprocessing the video data and the three-dimensional coordinate data of the workers on the construction site to obtain the preprocessed video data and the three-dimensional coordinate data comprises: S11, collecting and preprocessing the video data; S111, deploying a closed-circuit television camera at a fixed high point of a predetermined area in a construction site, the closed-circuit television camera continuously collecting the video data covering the predetermined area; S112, decomposing the collected video data into a sequence of image frames, and performing grayscale processing on each color image in the sequence of image frames, and performing Gaussian filtering on each grayscale image to obtain preprocessed video data; S12, collecting and preprocessing three-dimensional coordinate data; S121, arranging a positioning base station at a known coordinate position in the construction site, and providing each worker's safety helmet with a positioning tag; S122, the positioning tag periodically sends wireless signals outward, and the positioning base station receives the wireless signals, calculates the time difference of the wireless signals reaching different positioning base stations, and solves the three-dimensional coordinate data of the positioning tag; S123, according to the building information model of the construction site, the solved three-dimensional coordinate data is converted into a unified construction site coordinate system through a coordinate transformation formula; S124, performing outlier rejection on the converted three-dimensional coordinate data to obtain preprocessed three-dimensional coordinate data. 3.The smart construction site real-time monitoring method of claim 2, wherein, In S2, the process of obtaining posture features from preprocessed video data includes: S211, training a human skeleton detection network model based on a convolutional neural network architecture through a large-scale public data set containing human posture annotation information; S212, inputting each image in the preprocessed video data into the pre-trained human skeleton detection network model to identify and output the two-dimensional pixel coordinates of the body joint nodes in the image; S213, determining the minimum circumscribed rectangle frame surrounding the body joint nodes according to the output two-dimensional pixel coordinates, and calculating the ratio of the height value to the width value of the minimum circumscribed rectangle frame. 4.The smart construction site real-time monitoring method of claim 3, wherein, In S3, the posture features are input into a posture recognition model based on a long short-term memory network to output the instantaneous posture category probability of the worker, and the posture categories include standing, walking, squatting and falling, which includes: S32, inputting the posture feature sequence into a posture recognition model based on a long short-term memory network, the long short-term memory network processes the time series dependency relationship in the posture feature sequence through internal input gate, forget gate and output gate structure, and outputs a hidden state vector; S33, inputting the hidden state vector into the fully connected layer inside the long short-term memory network, and calculating the instantaneous posture category probability corresponding to the standing, walking, squatting and falling postures respectively by using an activation function, and taking the posture category corresponding to the maximum value of the instantaneous posture category probability as the posture category predicted by the posture recognition model. 5.The smart construction site real-time monitoring method of claim 3, wherein, In S4, the position features are input into a region risk assessment model based on a support vector machine to output the risk level of the position where the worker is located, which includes: S41, combining the position features into a position feature vector; S42, using the historical location feature data set containing different risk level category labels, finding an optimal hyperplane that separates the data points corresponding to different risk level category labels in the historical location feature data set and maximizes the classification margin by solving a convex quadratic programming problem, and dividing the decision space into decision intervals corresponding to each risk level according to the optimal hyperplane; S43, inputting the location feature vector into the regional risk assessment model based on the support vector machine, and classifying the location feature vector by the optimal hyperplane; S44, the optimal hyperplane is defined by a decision function; S45、determining the position feature vector according to the decision interval to which the output value of the decision function belongs the risk level category label corresponding to the decision interval, and outputting the risk level category label as the risk level of the position where the staff is located. 6.The smart construction site real-time monitoring method of claim 5, wherein, In S5, the process of training the posture recognition model and the regional risk assessment model using the historical data set composed of the video data and the three-dimensional coordinate data includes: S53, training the posture recognition model using the historical data set composed of the video data and the three-dimensional coordinate data; S531, extracting the posture feature sequence from the historical data set, and labeling the corresponding real posture category for each posture feature sequence; S532, inputting the labeled posture feature sequence into the posture recognition model, calculating the error between the posture category predicted by the posture recognition model and the real posture category by a loss function, and updating the internal parameters of the posture recognition model according to the error by an optimization algorithm until the error converges to a preset range; S54, training the regional risk assessment model using the historical data set; S541, extracting the location feature vector from the historical data set, and labeling the corresponding real risk level category label for each location feature vector; S542, taking the labeled location feature vector as training data, and determining the optimal hyperplane by solving the convex quadratic programming problem. 7.The smart construction site real-time monitoring method of claim 2, wherein, In S6, based on the comprehensive risk score, setting a risk judgment logic, outputting a risk judgment result, and generating a warning instruction include: S61, comparing the comprehensive risk score calculated in real time with the risk judgment threshold determined in the weight coefficient optimization process; S62, when the comprehensive risk score is greater than the risk judgment threshold, the output risk judgment result is a risk event, and when the comprehensive risk score is not greater than the risk judgment threshold, the output risk judgment result is a normal state; S63, when the risk judgment result is a risk event, generating a warning instruction containing the current worker's identity and the three-dimensional coordinate data. 8.The smart construction site real-time monitoring method of claim 7, wherein, In S7, according to the warning instruction, marking the position of the corresponding worker on the three-dimensional electronic map of the construction site, and calling the corresponding real-time video data include: S71, the monitoring terminal receives and analyzes the warning instruction; S72, in the preloaded three-dimensional electronic map of the construction site, according to the three-dimensional coordinate data parsed from the warning instruction, the spatial position of the worker is located; S73, generating and displaying a highlighted visual mark at the located spatial position; S74, matching and determining a CCTV camera covering the coordinate point corresponding to the spatial position according to the three-dimensional coordinate data, in combination with preset installation position and monitoring field range information of each CCTV camera; S75, sending a call instruction to the determined CCTV camera to obtain real-time video data synchronized with the moment of generating the early warning instruction; S76, displaying the obtained real-time video data in the visual interface of the monitoring terminal.

9. The intelligent construction site real-time monitoring system, which is used to realize the intelligent construction site real-time monitoring method in any one of claims 1-8, characterized in that, The construction site data acquisition module, the personnel feature extraction module, the personnel posture analysis module, the regional risk assessment module, the weighted evaluation optimization module, the risk judgment and early warning module, and the response execution module are connected by electrical signals. The construction site data acquisition module is configured to acquire and pre-process video data and three-dimensional coordinate data of workers at a construction site, to obtain pre-processed video data and three-dimensional coordinate data. The personnel feature extraction module is configured to obtain posture features and position features from the pre-processed video data and three-dimensional coordinate data. The personnel posture analysis module is configured to input the posture features into a posture recognition model based on a long short-term memory network, to output instantaneous posture category probabilities of the workers, and the posture categories include standing, walking, squatting, and falling. The regional risk assessment module is configured to input the position features into a regional risk assessment model based on a support vector machine, to output risk levels of the positions where the workers are located. The weighted evaluation optimization module is configured to use preset weight coefficients to perform weighted summation on the instantaneous posture category probabilities and quantized values of the risk levels, to output comprehensive risk scores, and to train the posture recognition model and the regional risk assessment model using a historical data set composed of the video data and the three-dimensional coordinate data, and to optimize the weight coefficients. The risk judgment and early warning module is configured to set risk judgment logic based on the comprehensive risk scores, to output risk judgment results, and to generate early warning instructions. The response execution module is configured to mark positions where corresponding workers are located on a three-dimensional electronic map of the construction site according to the early warning instructions, and to call corresponding real-time video data.

Images