An AI image recognition-based construction site safety protection intelligent monitoring system

Abstract

The application provides a kind of construction site safety protection intelligent monitoring system based on AI image recognition, relating to safety protection technical field. Including: real-time acquisition of construction site video image data and historical video image data;Video image data is preprocessed to obtain preprocessed image data;According to safety identification and target trajectory, get illegal behavior and safety hazard;Based on the improved AHP-TOPSIS algorithm, a risk prediction model is constructed, the illegal behavior is taken as the core evaluation index, the index weight is dynamically corrected by entropy weight method, and the early warning information and risk diffusion prediction are output;According to the early warning information and risk diffusion prediction, trigger the hierarchical response mechanism, and optimize the equipment control by using deep reinforcement learning. The application effectively reduces accidents by using improved EfficientDet to construct safety protection identification model and other measures, ensures the safety of personnel and equipment, and improves the overall construction safety management level.

Description

Technical Field

[0001] This invention relates to the field of safety protection technology, and in particular to an intelligent monitoring system for construction site safety protection based on AI image recognition. Background Technology

[0002] As a vital pillar of the national economy, the construction industry is characterized by its work environment, which involves numerous high-altitude operations, frequent overlapping work, dense concentration of large equipment, and high personnel mobility. Safety risks persist throughout the entire construction cycle. According to statistics released by the Ministry of Housing and Urban-Rural Development, direct economic losses due to safety accidents in my country's construction sector exceed 10 billion yuan annually in recent years, with a persistently high proportion of accidents resulting in injuries or fatalities. The main causes of these accidents include not wearing safety protective equipment, unauthorized entry into dangerous areas, and improper operation of machinery. Therefore, establishing an efficient and precise safety monitoring system for construction sites is of significant practical importance in reducing the accident rate.

[0003] Traditional safety management relies on on-site inspections by safety officers, which not only requires a significant investment of manpower but is also limited by the physical strength, energy, and professional judgment of the inspectors. While some construction sites have installed video surveillance cameras, the systems only record and play back videos, essentially remaining "post-incident traceability" tools, unable to correlate behavioral trajectories or assess dynamic risks. Current safety monitoring data is mostly stored in a scattered manner, lacking systematic integration and analysis. Managers cannot use data to understand the distribution patterns of hazards or equipment failure trends, resulting in a lack of targeted safety management measures and difficulty in mitigating risks at their source. When hazards such as personnel not wearing safety helmets, illegally climbing scaffolding, or materials blocking fire lanes are detected in the monitoring footage, automatic identification and warnings are not possible, requiring real-time manual monitoring. Large construction sites often deploy dozens or even hundreds of cameras, making it difficult for manual monitoring to cover all footage, resulting in a high rate of missed detections and failing to meet the needs of dynamic safety management. Summary of the Invention

[0004] This invention provides an intelligent monitoring system for construction site safety protection based on AI image recognition, which solves the shortcomings of existing technologies such as inability to automatically identify and warn, high missed detection rate, and inability to meet dynamic safety management needs.

[0005] This invention provides an intelligent monitoring system for construction site safety protection based on AI image recognition, comprising:

[0006] The image acquisition module is used to collect real-time and historical video image data of the construction site.

[0007] The preprocessing module is used to preprocess video image data to obtain preprocessed image data.

[0008] The AI ​​recognition and analysis module uses an improved EfficientDet to extract image features. It employs a weighted BiFPN structure to assign learnable weight parameters to feature layers at different scales. Through backpropagation, the weight parameters are automatically adjusted to give higher weights to feature layers corresponding to small targets, thereby enhancing the transmission efficiency of small target features during the fusion process. It preserves the semantic and detailed information of multi-scale features and uses an improved YOLOv8 target tracking algorithm to generate target trajectories based on preprocessed image data. Based on the security recognition effect and target trajectory, it identifies violations and security risks.

[0009] The safety early warning module is used to build a risk prediction model based on the improved AHP-TOPSIS algorithm. It takes the degree of danger of violations and the scope of impact of safety hazards as the core evaluation indicators, dynamically adjusts the indicator weights through the entropy weight method, and outputs early warning information and risk diffusion prediction.

[0010] The linkage control module is used to trigger a hierarchical response mechanism based on early warning information and risk spread prediction, and to optimize equipment control using deep reinforcement learning.

[0011] According to the present invention, a construction site safety protection intelligent monitoring system based on AI image recognition includes a data preprocessing module comprising:

[0012] The data denoising unit is used to denoise image data using a hybrid algorithm that combines adaptive median filtering and Gaussian filtering.

[0013] The data augmentation unit is used to enhance features and expand diversity in the denoised image data.

[0014] Data normalization units are used to adjust denoised and enhanced image data to a uniform pixel size.

[0015] According to the present invention, an intelligent monitoring system for construction site safety protection based on AI image recognition is provided, wherein the AI ​​recognition and analysis module includes:

[0016] The image annotation unit is used to annotate historical video image data to obtain annotated historical video image data, which is then divided into training set, validation set and test set.

[0017] The model building unit is used to build a security protection and identification model based on the EfficientDet model architecture.

[0018] The model training unit is used to train the security protection identification model using the training set.

[0019] The model validation unit is used to monitor model performance in real time using the validation set, employs an early stopping strategy to prevent overfitting, and adjusts model hyperparameters based on the average accuracy metric on the validation set.

[0020] The model testing unit is used to evaluate the model using a test set, analyze the model's detection performance under different target categories, adjust the model for targets with detection performance below the threshold, and obtain a security protection recognition model. The preprocessed image data is then input into the security protection recognition model to obtain the security recognition performance.

[0021] According to the intelligent monitoring system for construction site safety protection based on AI image recognition provided by the present invention, the process of generating target trajectories by improving the YOLOv8 target tracking algorithm in the AI ​​recognition and analysis module further includes: inputting the temporal feature matrix of target behavior into a classifier, outputting the category probabilities of violations and safety hazards through a softmax function, and determining the target location by combining the bounding box regression results. Based on the category probabilities, target location, and temporal information in the temporal feature matrix, it is determined whether there are safety hazards in the image data; if so, it is determined that there are violations and safety hazards, and the recognition result is output.

[0022] According to the intelligent monitoring system for construction site safety protection based on AI image recognition provided by the present invention, the process of generating target trajectory using the improved YOLOv8 target tracking algorithm in the AI ​​recognition and analysis module further includes: inputting the temporal feature matrix of target behavior into the classifier, outputting the category probability of violation behavior and safety hazard through the softmax function, and determining the target position by combining the bounding box regression result; judging whether there is a safety hazard in the image data based on the category probability, target position, and temporal information in the temporal feature matrix, and if so, determining that there is a violation behavior and safety hazard, and outputting the recognition result.

[0023] According to the present invention, a construction site safety protection intelligent monitoring system based on AI image recognition includes a safety early warning module comprising:

[0024] The weight calculation unit is used to construct an evaluation index system through the analytic hierarchy process. It uses the action risk coefficient of the violation, the impact range of the safety hazard, and environmental metadata as primary indicators to calculate the initial weight of the primary indicators.

[0025] The comprehensive weight calculation unit is used to calculate the information entropy of the primary index based on historical video image data using the entropy weight method, determine the data-driven weights based on the information entropy values, and then dynamically adjust and merge the data-driven weights with the initial weights to obtain the comprehensive index weights.

[0026] The quantified risk unit is used to process the weights of comprehensive indicators using the TOPSIS algorithm, compare the actual values ​​of primary indicators with the ideal solutions, and calculate the relative closeness to obtain the quantified risk level.

[0027] According to the present invention, a construction site safety protection intelligent monitoring system based on AI image recognition is provided. The safety early warning module further includes a risk diffusion unit, which is used to construct a risk prediction model using an improved AHP-TOPSIS algorithm, inputting environmental metadata and quantitative risk levels into the risk prediction model to obtain the risk diffusion trend.

[0028] The early warning information unit is used to obtain early warning information and risk spread prediction based on the quantified risk level and risk spread trend.

[0029] According to the present invention, a construction site safety protection intelligent monitoring system based on AI image recognition is provided, the linkage control module including:

[0030] The lethal warning unit is used to send hardwired commands in real time through the edge computing terminal to cut off the power to dangerous equipment, call the UWB positioning system to lock the personnel in danger, and activate the sound and light guidance device.

[0031] The high-risk early warning unit is used to plan the optimal inspection path based on the improved Dijkstra algorithm and push it to the AR terminal of the management personnel, and link drones to perform 3D modeling of the violation area.

[0032] The low-to-medium risk early warning unit is used to build a database of violation cases through a knowledge graph, generate intelligent prompts containing rectification steps, and associate them with BIM models to mark the location of potential hazards.

[0033] According to the present invention, an intelligent monitoring system for construction site safety protection based on AI image recognition is provided. In the catastrophic warning unit, after receiving catastrophic warning information, the edge computing terminal sends a forced power-off code to the power control system of the hazardous equipment through a hard-wired interface. It calls the UWB positioning system through an internal protocol and sends the coordinate parameters of the hazardous area. The UWB system starts the real-time positioning engine and calculates the three-dimensional coordinates of the personnel in danger based on the tag signal strength and time difference. The positioning result is synchronously transmitted back to the edge terminal. The edge terminal controls the sound and light guidance device in the warning area to emit warning lights and voice prompts of a specific frequency according to the three-dimensional coordinates, guiding the personnel to move to the safe area.

[0034] By constructing a security protection identification model using an improved EfficientDet and generating target trajectories using an improved YOLOv8 target tracking algorithm, warnings are issued based on security identification and target trajectories to identify violations and security risks. The beneficial effects achieved are as follows:

[0035] The EfficientDet model is used to construct a security protection identification model, leveraging its efficient multi-scale feature extraction and target detection capabilities. In this model, the EfficientNet backbone network can flexibly select the appropriate version to extract high-quality features based on hardware computing power. Combined with the BiFPN bidirectional feature pyramid network for feature fusion, it can accurately process targets of different scales in construction site images. This adaptation and application of advanced models to complex construction scenarios is an innovation. Traditional monitoring systems often simply use general target detection methods, making it difficult to accurately identify multi-scale targets and targets in complex construction environments.

[0036] This brings significant changes to construction safety management from multiple dimensions. In terms of recognition accuracy, the AI ​​recognition and analysis module improves the EfficientDet and YOLOv8 algorithms, constructing a safety protection recognition model with EfficientNetV2-S as the backbone network and a weighted BiFPN structure. This strengthens the transmission of small target features, making the detection of violations and safety hazards more accurate. Small targets such as the wearing of safety helmets and improper operation of small equipment during construction are no longer missed. The target tracking algorithm integrates appearance features and motion trajectories, combines 3D convolution to extract action features, and constructs a temporal feature matrix to accurately capture continuous frame behavior, significantly reducing false positives and improving recognition reliability.

[0037] In terms of risk management efficiency, the safety early warning module relies on an improved AHP-TOPSIS algorithm, constructs an indicator system using the analytic hierarchy process, dynamically adjusts weights using the entropy weight method, and quantifies risk levels using the TOPSIS algorithm. It can also predict risk spread trends, allowing managers to be aware of risk evolution in advance and gain time for decision-making. The linkage control module responds in layers according to different risk levels. For critical risks, the early warning quickly cuts off equipment power and locates and guides personnel, minimizing danger and injury. For high-risk risks, it plans inspection routes and uses drone modeling to facilitate rapid investigation. For medium- and low-risk risks, it uses knowledge graphs and BIM models to intelligently prompt rectification, promoting efficient handling of hazards and forming an efficient closed loop from identification to response.

[0038] The system's standardized and automated processes reduce manual inspections and lower labor costs. Blockchain-based evidence storage ensures traceability of response instructions, facilitating accountability and process optimization. Simultaneously, the accumulated historical data and violation case database support subsequent safety management optimization, promoting the intelligent and refined development of construction safety management, strengthening safety defenses at construction sites, effectively reducing accidents, ensuring personnel and equipment safety, and improving the overall level of construction safety management. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of a construction site safety protection intelligent monitoring system based on AI image recognition provided in an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of the process of generating a target trajectory based on preprocessed image data using the improved YOLOv8 target tracking algorithm provided in an embodiment of the present invention.

[0042] Figure 3 This is a schematic diagram of a safety early warning module of an intelligent monitoring system for construction site safety protection based on AI image recognition, provided in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0044] The following is combined Figures 1-3 This invention describes an intelligent monitoring system for construction site safety protection based on AI image recognition.

[0045] like Figures 1-3 As shown in the figure, an intelligent monitoring system for construction site safety protection based on AI image recognition is provided in this embodiment of the invention. The system includes:

[0046] The image acquisition module is used to collect real-time and historical video image data of the construction site.

[0047] Real-time video image data includes: personnel data, equipment data, material data, and environmental data.

[0048] Personnel data includes: the behavior status of construction workers and the number of workers.

[0049] The behavioral status of construction workers includes: the specific operations they are performing, such as excavation, transportation, and welding; whether they are operating according to regulations or engaging in violations, such as not wearing safety helmets or safety harnesses. The location of the workers also includes whether they are concentrated in a specific construction area or scattered at different points.

[0050] The number of personnel includes: visually showing how many people of each trade are involved in the construction at the current construction site, which helps to determine whether the manpower input is reasonable and whether there are any abnormal situations such as personnel gathering.

[0051] Equipment data includes: the operating status of construction machinery and equipment, and the location where the equipment is parked.

[0052] Operating status of construction machinery and equipment: For cranes, check the boom extension angle and the smoothness of the lifting motion; for excavators, check the boom swing frequency and digging depth. Assess whether the equipment is in normal working order, whether there are any signs of malfunction, and whether the machinery is emitting abnormal smoke or vibrating.

[0053] Equipment parking location: Whether all types of construction machinery are placed in the designated parking areas.

[0054] Material data includes: material stacking and material transportation status.

[0055] Material stacking: Whether building materials are stacked neatly and securely, and whether the stacking height meets safety requirements to avoid the risk of collapse.

[0056] Material transportation status: Whether the driving route of the transport vehicles and the loading and unloading of materials are compliant.

[0057] Environmental data includes: the overall layout of the site and the status of safety protection facilities within the site.

[0058] Overall site layout: Are the construction areas clearly divided? Are the boundaries between the living area, work area, and material storage area clearly defined? Are there any overlapping or confusing situations?

[0059] The status of safety protection facilities on site: whether the fence is intact, whether the guardrail is damaged, whether the warning signs are clearly visible, etc.; the weather conditions on site, whether there is water or snow accumulation that may affect construction safety.

[0060] Historical video image data includes historical construction data, historical personnel activity data, equipment usage data, material management data, accident record data, and construction specification data.

[0061] The preprocessing module is used to preprocess video image data to obtain preprocessed image data.

[0062] The data denoising unit is used to denoise the data using a hybrid algorithm that combines adaptive median filtering and Gaussian filtering.

[0063] The data augmentation unit is used to enhance features and expand diversity in the denoised image.

[0064] The data normalization unit is used to adjust the denoised and enhanced image to a uniform pixel size, so as to make the input data format consistent.

[0065] The AI ​​recognition and analysis module is used to build a security protection recognition model using an improved EfficientDet model to perform security recognition on preprocessed image data. It uses an improved YOLOv8 target tracking algorithm to generate target trajectories based on the preprocessed image data, and obtains violations and security risks based on security recognition and target trajectories.

[0066] Targets refer to the various objects and behaviors that need to be identified in the construction scenario, covering construction personnel, machinery and equipment such as tower cranes and welding machines, safety facilities such as safety helmets and guardrails, hazardous areas such as high-altitude work areas and live areas, as well as related behaviors of these objects such as climbing and improper operation of equipment. These targets are the basic units of model detection, including both static object states and dynamic human actions or equipment operating states.

[0067] The process of building a security protection identification model using the EfficientDet model includes:

[0068] The image annotation unit is used to annotate historical video image data to obtain annotated historical video image data, which is then divided into training set, validation set and test set.

[0069] The model building unit is based on the EfficientDet model architecture. EfficientNetV2-S is selected as the backbone network for image feature extraction. A weighted BiFPN structure is used to assign learnable weight parameters to feature layers at different scales. The weight parameters are automatically adjusted through backpropagation, giving higher weights to feature layers corresponding to small targets. This enhances the transfer efficiency of small target features during the fusion process and preserves the semantic and detailed information of multi-scale features.

[0070] Small targets are objects that are small in size or far away, occupying a small proportion of the image, with few pixels and relatively sparse feature information. Examples include small tools falling from a height, workers on distant scaffolding without safety harnesses, exposed thin wires, and sharp building materials scattered on the ground. Because these targets have inconspicuous visual features, they are easily interfered with by complex backgrounds, making them more difficult to identify. Therefore, mechanisms such as weighted BiFPN are needed to enhance their feature weights to ensure that the model can accurately capture and identify them.

[0071] The model training unit is used to train the constructed model using the training set. It adopts the stochastic gradient descent optimization algorithm, sets appropriate learning rate, batch size and number of iterations, and uses data augmentation techniques during training to improve the robustness of the model to the complex environment of the construction scenario.

[0072] By monitoring model performance in real time using the validation set, adopting an early stopping strategy to prevent overfitting, and adjusting model hyperparameters based on metrics such as average precision on the validation set, the model can achieve high precision and recall when identifying construction targets and violations.

[0073] The model is evaluated using a test set to analyze its detection performance under different target categories and scenarios. For targets with poor recognition performance, the model is fine-tuned. Performance can be improved by increasing the training weights of corresponding samples or optimizing the parameters of the feature fusion module. Finally, a safety protection recognition model that meets the needs of construction safety detection is obtained.

[0074] The process of generating a target trajectory using the improved YOLOv8 target tracking algorithm includes:

[0075] The improved YOLOv8 target tracking algorithm assigns a unique ID to each target based on its appearance features and motion trajectory prediction, and generates a sequence of trajectory coordinates for consecutive frames to obtain the target trajectory features.

[0076] 3D convolution is used to extract motion features from continuous video frames, and the motion features are fused with target trajectory features to construct a temporal feature matrix of target behavior.

[0077] The temporal feature matrix of the target behavior is input into the classifier, and the softmax function outputs the category probability of the violation and the safety hazard. At the same time, the target location is determined by combining the bounding box regression results.

[0078] Based on category probability, target location, and time series information in the time series feature matrix, violations and safety hazards are identified, and the violations and safety hazards are obtained.

[0079] The safety early warning module is used to build a risk prediction model based on the improved AHP-TOPSIS algorithm. It uses the degree of danger of violations and the scope of impact of safety hazards as core evaluation indicators, dynamically adjusts the indicator weights through the entropy weight method, and outputs early warning information and risk diffusion prediction.

[0080] The weight calculation unit is used to construct an evaluation index system using the Analytic Hierarchy Process (AHP). It uses the hazard coefficient of violations, the impact range of safety hazards, and environmental metadata as primary indicators to calculate the initial weights of these primary indicators. The evaluation object is clearly defined as the safety risks of the construction site. The hazard coefficient of violations (e.g., the danger of not wearing safety belts while working at height), the impact range of safety hazards (e.g., the size of the area potentially affected by scaffolding collapse), and environmental metadata (e.g., environmental parameters affecting construction safety such as heavy rain and strong winds) are listed as primary indicators, covering three major risk sources: dynamic behavior, static hazards, and environmental interference. The initial weights are calculated using the AHP algorithm. The importance of the three primary indicators is compared pairwise to construct a judgment matrix. The largest eigenvalue and corresponding eigenvector are solved through matrix operations. After consistency verification, the eigenvector is normalized to obtain the initial weights.

[0081] The comprehensive weight calculation unit is used to calculate the information entropy of the primary index based on historical video image data using the entropy weight method, determine the data-driven weights based on the information entropy values, and then dynamically adjust and merge the data-driven weights with the initial weights to obtain the comprehensive index weights.

[0082] The formula for calculating information entropy is expressed as:

[0083]

[0084] In the formula, where is the standardized index value, and m is the number of samples.

[0085] The smaller the entropy value, the stronger the indicator's ability to distinguish risks in historical data, and the greater the corresponding entropy weight. In other words, data-driven weights are more prominent for indicators that are sensitive to risk.

[0086] The quantified risk unit is used to process the weights of comprehensive indicators using the TOPSIS algorithm, compare the actual values ​​of primary indicators with the ideal solutions, and calculate the relative closeness to obtain the quantified risk level.

[0087] The formula for calculating proximity is expressed as follows:

[0088]

[0089] In the formula, The Euclidean distance between the i-th evaluation object and the negative ideal solution, The Euclidean distance between the i-th evaluation object and the positive ideal solution.

[0090]

[0091] In the formula, For weighted actual values, Let be the positive ideal solution for the j-th indicator, and n be the total number of evaluation indicators.

[0092]

[0093] In the formula, Let be the positive ideal solution for the j-th indicator, and n be the total number of evaluation indicators.

[0094] The risk diffusion unit is used to construct a risk prediction model based on an improved AHP-TOPSIS algorithm. Environmental metadata and quantified risk levels are input into the risk prediction model to obtain the risk diffusion trend. Through the improved AHP-TOPSIS algorithm, time-series risk indicators are correlated with environmental metadata to analyze the evolution of risk over time, ultimately outputting the risk diffusion trend, including dynamic information such as the predicted risk level for the next 1-3 hours, the boundary of the impact range expansion, and the peak risk time.

[0095] The early warning information unit is used to generate early warning information and risk spread predictions based on the quantified risk level and risk spread trend. Early warning levels are determined based on the quantified risk level output by the quantified risk unit. Combined with the risk spread trend output by the risk spread unit, information on future risk development is supplemented, integrating information such as risk level, current risk location, spread path, and affected populations to generate a complete early warning information and risk spread prediction report containing "early warning level, risk description, spread prediction, and response recommendations," providing accurate basis for subsequent coordinated control.

[0096] The early warning information includes: risk level, risk subject information, spatiotemporal characteristics, risk description, and response recommendations.

[0097] Risk level is used to clarify the urgency and severity of the current risk. Based on the quantified risk level, the classification results include catastrophic, high-risk, medium-risk, and low-risk levels. Risk subject information includes the type of violation involved, the specific form of the safety hazard, and related environmental metadata. Spatiotemporal characteristics include the specific location, duration, or time of discovery of the risk, as well as the scope of impact determined by the risk diffusion trend. Risk description is a quantified value of indicators such as the hazard coefficient of the violation, the scope of impact of the safety hazard, and key parameters in the risk diffusion trend. Response recommendations are preliminary response guidelines based on the risk level and diffusion trend.

[0098] The linkage control module is used to trigger a hierarchical response mechanism based on early warning information and risk spread prediction, and to optimize equipment control using deep reinforcement learning.

[0099] The lethal warning system sends a hardwired command in real time from the edge computing terminal to cut off the power to the hazardous equipment, invokes the UWB positioning system to locate the personnel in danger, and activates the audio-visual guidance device. Upon receiving the lethal warning information, the edge computing terminal immediately sends a cut-off command to the power control system of the hazardous equipment via the hardwired interface. The command includes the equipment's unique identifier and a forced power-off code. The edge terminal invokes the UWB positioning system through its internal protocol, sending the coordinate parameters of the danger zone. The UWB system activates its real-time positioning engine, calculating the three-dimensional coordinates of the personnel in danger based on the tag signal strength and time difference, with a positioning accuracy controlled within 30 centimeters. The positioning results are simultaneously transmitted back to the edge terminal. Based on the positioning coordinates, the edge terminal triggers the audio-visual guidance device in the corresponding area, controlling the device to emit warning lights and voice prompts at specific frequencies, guiding the personnel to a safe area.

[0100] The high-risk warning system uses an improved Dijkstra algorithm to plan the optimal inspection path and push it to the management personnel's AR terminal, while simultaneously coordinating with drones to create a 3D model of the violation area. Upon receiving the high-risk warning information, the system extracts the boundary coordinates of the violation area associated with the warning, as well as the distribution data of surrounding roads and obstacles. This data is then input into the improved Dijkstra algorithm module. By incorporating road traffic coefficients and safety redundancy distances, the algorithm plans the optimal inspection path from the management personnel's current location to the violation area within one second. The path data is converted into a spatial coordinate sequence recognizable by the AR terminal. The path data is then pushed to the management personnel's AR terminal via a wireless communication module, where a blue path guide line and key node prompts are overlaid on the real-time screen. Simultaneously, the system sends a 3D modeling task command to the drone dispatch platform, including the latitude and longitude range of the violation area and the modeling accuracy requirements. After receiving the command, the drone autonomously plans its flight path, with its onboard LiDAR and high-definition camera simultaneously collecting data. After the flight, it automatically generates a 3D point cloud model of the violation area and uploads it to the management platform for remote viewing.

[0101] The low-to-medium risk early warning system constructs a violation case database using a knowledge graph, generating intelligent prompts containing rectification steps and linking them to the BIM model to mark the location of potential hazards. Upon receiving low-to-medium risk early warning information, the system extracts hazard feature parameters and performs semantic matching with the violation case database in the knowledge graph. The case database uses ontology construction technology to link hazard types, rectification standards, historical handling solutions, and other knowledge nodes. When the matching degree exceeds 85%, the system automatically retrieves the rectification process of similar cases. Based on the rectification process, structured intelligent prompts are generated, including operation steps, a list of required tools, and safety precautions. Simultaneously, the system converts the hazard location coordinates into component coordinates in the BIM model, highlights the structural component where the hazard is located on the BIM platform, and links it to the intelligent prompt content. All response commands undergo hash value calculation via a blockchain client, generating blocks containing timestamps, operator information, and command content, which are linked to the existing blockchain ledger to ensure the entire lifecycle of the commands is traceable and tamper-proof.

[0102] In summary, this embodiment provides an intelligent monitoring system for construction site safety protection based on AI image recognition. It constructs a safety protection recognition model using an improved EfficientDet algorithm and generates target trajectories using an improved YOLOv8 target tracking algorithm. Based on safety recognition and target trajectories, it issues warnings about violations and safety hazards. The system's processes are standardized and automated, reducing manual inspections and lowering labor costs. Blockchain-based evidence storage ensures traceability of response instructions, facilitating responsibility identification and process optimization. Furthermore, the accumulated historical data and violation case database support subsequent safety management optimization, promoting the intelligent and refined development of construction safety management, strengthening the safety defenses of construction sites, effectively reducing accidents, ensuring the safety of personnel and equipment, and improving the overall level of construction safety management.

[0103] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device to execute the methods of various embodiments or some parts of the embodiments.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart monitoring system for construction site safety protection based on AI image recognition, characterized in that, include: The image acquisition module is used to collect real-time and historical video image data of the construction site. The preprocessing module is used to preprocess the video image data to obtain preprocessed image data; The AI ​​recognition and analysis module is used to extract image features using an improved EfficientDet algorithm. It employs a weighted BiFPN structure to assign weight parameters to feature layers at different scales, automatically adjusting these parameters through backpropagation to give higher weights to feature layers corresponding to small targets. This enhances the transmission efficiency of small target features during the fusion process, preserves the semantic and detailed information of multi-scale features, and constructs a security protection recognition model to perform security recognition on the preprocessed image data. An improved YOLOv8 target tracking algorithm is used to generate target trajectories based on the preprocessed image data. Based on the security recognition effect and the target trajectory, violations and security risks are identified. The safety early warning module is used to build a risk prediction model based on the improved AHP-TOPSIS algorithm. It takes the degree of danger of the violation and the scope of the impact of the safety hazard as the core evaluation indicators, dynamically corrects the indicator weights through the entropy weight method, and outputs early warning information and risk diffusion prediction. The linkage control module is used to trigger a hierarchical response mechanism based on the early warning information and risk spread prediction, and to optimize equipment control using deep reinforcement learning. The lethal warning unit is used to send hardwired commands in real time via an edge computing terminal to cut off the power to dangerous equipment, call the UWB positioning system to lock the personnel in danger, and activate the sound and light guidance device. The high-risk early warning unit is used to plan the optimal inspection path based on the improved Dijkstra algorithm and push it to the AR terminal of the management personnel, and link the drone to perform 3D modeling of the violation area. The low-to-medium risk early warning unit is used to build a database of violation cases through a knowledge graph, generate intelligent prompts containing rectification steps, and associate them with BIM models to mark the location of potential hazards.

2. The intelligent monitoring system for construction site safety protection based on AI image recognition according to claim 1, characterized in that, The data preprocessing module includes: The data denoising unit is used to denoise image data using a hybrid algorithm that combines adaptive median filtering and Gaussian filtering. The data augmentation unit is used to enhance features and expand diversity in the denoised image data. Data normalization units are used to adjust denoised and enhanced image data to a uniform pixel size.

3. The intelligent monitoring system for construction site safety protection based on AI image recognition according to claim 1, characterized in that, The AI ​​recognition and analysis module includes: An image annotation unit is used to annotate the historical video image data to obtain annotated historical video image data, and to divide the annotated historical video image data into a training set, a validation set and a test set; The model building unit builds a security protection and identification model based on the EfficientDet model architecture. The model training unit is used to train the security protection identification model using the training set; The model validation unit is used to monitor model performance in real time using the validation set, adopts an early stopping strategy to prevent overfitting, and adjusts model hyperparameters based on the average accuracy metric on the validation set. The model testing unit is used to evaluate the model using a test set, analyze the model's detection performance under different target categories, adjust the model for targets with detection performance below a threshold, obtain a security protection recognition model, and input the preprocessed image data into the security protection recognition model to obtain the security recognition performance.

4. The intelligent monitoring system for construction site safety protection based on AI image recognition according to claim 1, characterized in that, In the AI ​​recognition and analysis module, the process of generating target trajectories using the improved YOLOv8 target tracking algorithm includes: using the improved YOLOv8 target tracking algorithm to assign a unique ID to each target based on the target appearance features and motion trajectory prediction, and generating a sequence of trajectory coordinates for consecutive frames to obtain target trajectory features; using 3D convolution to extract action features from consecutive video frames, and fusing the action features with the target trajectory features to construct a temporal feature matrix of target behavior.

5. The intelligent monitoring system for construction site safety protection based on AI image recognition according to claim 4, characterized in that, In the AI ​​recognition and analysis module, the process of generating target trajectories using the improved YOLOv8 target tracking algorithm also includes: inputting the temporal feature matrix of the target behavior into a classifier, outputting the category probabilities of violations and safety hazards through the softmax function, and determining the target position by combining the bounding box regression results; judging whether there are safety hazards in the image data based on the category probabilities, target position, and temporal information in the temporal feature matrix, and if so, determining that there are violations and safety hazards, and outputting the recognition result.

6. The intelligent monitoring system for construction site safety protection based on AI image recognition according to claim 1, characterized in that, The security early warning module includes: The weight calculation unit is used to construct an evaluation index system through the analytic hierarchy process. It takes the action risk coefficient of the violation, the impact range of the safety hazard, and environmental metadata as primary indicators and calculates the initial weight of the primary indicators. The comprehensive weight calculation unit is used to calculate the information entropy of the first-level index based on the historical video image data using the entropy weight method, determine the data-driven weight based on the information entropy value, and dynamically adjust and obtain the comprehensive index weight by weighting and fusing the data-driven weight with the initial weight. The quantification risk unit is used to process the weights of the comprehensive indicators using the TOPSIS algorithm, compare the actual values ​​of the primary indicators with the ideal solutions, and calculate the relative closeness to obtain the quantification risk level.

7. The intelligent monitoring system for construction site safety protection based on AI image recognition according to claim 6, characterized in that, The security early warning module also includes: The risk diffusion unit is used to build a risk prediction model using an improved AHP-TOPSIS algorithm. Environmental metadata and quantitative risk levels are input into the risk prediction model to obtain the risk diffusion trend. The early warning information unit is used to obtain early warning information and risk spread prediction based on the quantified risk level and risk spread trend.

8. The intelligent monitoring system for construction site safety protection based on AI image recognition according to claim 1, characterized in that, In the lethal warning unit, after receiving the lethal warning information, the edge computing terminal sends a forced power-off code to the power control system of the hazardous equipment through a hard-wired interface. It also calls the UWB positioning system through an internal protocol and sends the coordinate parameters of the hazardous area. The UWB system starts the real-time positioning engine and calculates the three-dimensional coordinates of the personnel in danger based on the tag signal strength and time difference. The positioning results are synchronously transmitted back to the edge terminal. The edge terminal controls the sound and light guidance device in the warning area to emit warning lights and voice prompts of a specific frequency according to the three-dimensional coordinates, guiding the personnel to move to the safe area.

Images