Construction area safety hidden danger early warning system and method based on AI visual identification

Abstract

The application relates to the technical field of building construction safety monitoring, and particularly discloses a construction area safety hidden danger early warning system and method based on AI visual identification, which comprises an end-side collection layer, a side-side calculation layer, a cloud-side management layer and an application layer. The method comprises the following steps: image collection and transmission; image preprocessing and enhancement; dividing the construction area into grids and extracting quantifiable parameters; calculating the quantifiable scores of safety hidden dangers and determining the hidden danger grades by using a quantifiable evaluation model; triggering a graded early warning according to the grades; and closed-loop disposal management. The application establishes quantifiable evaluation models for different hidden dangers, calculates comprehensive quantifiable scores by using a multiplication fusion model, and realizes multi-hidden danger superposition risk evaluation. By means of accurate extraction of quantifiable parameters and a graded early warning mechanism, the application solves the technical problems of single identification dimension, lack of quantifiable evaluation and inability to form a closed-loop management in the prior art, and significantly improves the intelligent level of construction site safety management.

Description

Technical Field

[0001] This invention relates to the field of construction safety monitoring technology, specifically to an early warning system and method for safety hazards in construction areas based on AI visual recognition. Background Technology

[0002] As one of the pillar industries of the national economy, the construction industry is characterized by complex working environments, high personnel mobility, numerous high-altitude operations, a high density of large machinery and equipment, and frequent overlapping operations, making it a high-risk industry for work safety accidents. Statistics show that thousands of safety accidents occur annually in the construction sector, with falls from heights accounting for over 50%, falling object accidents for approximately 15%, collapses for about 10%, and machinery-related injuries for about 8%. In-depth analysis of the causes of these accidents reveals that approximately 90% are closely related to unsafe acts by people and unsafe conditions of equipment.

[0003] Traditional construction site safety management mainly relies on two methods: manual inspections by safety management personnel and post-event review via video surveillance systems. However, both methods have significant technical shortcomings. Manual inspections have time and spatial blind spots; safety officers typically inspect every 2-4 hours, making it difficult to achieve 24 / 7 real-time monitoring. Furthermore, the rates of missed inspections and misjudgments are high due to factors such as the safety officer's experience level, sense of responsibility, and physical fatigue. In recent years, some technology vendors have begun to apply artificial intelligence technology to construction safety monitoring, launching basic applications such as safety helmet recognition based on target detection. However, existing technical solutions still have the following prominent problems: 1. Existing systems typically adopt a binary judgment mode of whether a violation has occurred, and the output result is only an alarm or no alarm, lacking a quantitative assessment of the severity of safety hazards. For example, for the violation of not wearing a safety helmet, it is impossible to distinguish whether the person is in a normal area or a high-risk area, whether the person has only temporarily removed the helmet or has not worn it for a long time, or whether there are other overlapping risks around, resulting in a lack of hierarchical basis for alarms and easy alarm fatigue; 2. Weak spatial perception capability. Most existing technologies only perform target detection based on two-dimensional images and lack the ability to map image coordinates to the three-dimensional space of the actual construction scene, making it impossible to accurately determine whether the person is in a truly dangerous area (such as the edge of a foundation pit, near an opening, etc.), resulting in a high false alarm rate for electronic fences.

[0004] Therefore, this invention proposes a construction area safety hazard early warning system and method based on AI visual recognition. Summary of the Invention

[0005] The purpose of this invention is to provide a construction area safety hazard early warning system and method based on AI visual recognition, thereby solving the above-mentioned technical problems: The objective of this invention can be achieved through the following technical solutions: A construction area safety hazard early warning system based on AI visual recognition includes an edge-side data acquisition layer, an edge-side computing layer, a cloud-side management layer, and an application layer. The end-side acquisition layer includes multiple high-definition camera devices and supporting lighting devices deployed in key areas of the construction site to adapt to the complex environment of building construction; The edge computing layer includes an edge computing node cluster deployed at the construction site, which is communicatively connected to the end-side acquisition layer. It is used to receive multi-channel video stream data, perform AI inference, extract quantifiable parameters, conduct quantitative assessment of safety hazards, and trigger graded early warnings. The cloud-side management layer, including the cloud server cluster, is communicatively connected to the edge computing layer and is used to receive alarm data and quantifiable parameters uploaded by edge computing nodes, and to perform data aggregation, storage, and management platform display. The application layer includes management backend terminals, mobile terminals, and on-site audible and visual alarm devices, used to display alarm information and quantitative assessment results, issue handling instructions, and generate safety management reports.

[0006] As a further description of the technical solution of the present invention, the edge computing layer includes a data preprocessing unit, a feature extraction unit, an AI risk level judgment unit, and a local caching unit; The data preprocessing unit decodes, dehazes, denoises, equalizes brightness, compresses wide dynamic range, and corrects geometric distortion on the acquired raw image data, removes interference information, and outputs a standardized image to be identified. The feature extraction unit extracts feature parameters based on the standardized image to be identified and processes the feature parameters. The AI-based hazard classification and judgment unit has a built-in hazard risk level judgment standard, analyzes the extracted feature parameters, and maps the analysis results to the hazard risk level judgment standard. The local caching unit caches hazard images, video clips, and early warning records from the past 7 days. It can store data locally when the network is offline and automatically synchronize it to the cloud after the network is restored.

[0007] A method for early warning of safety hazards in construction areas based on AI visual recognition, the method comprising the following steps: S1: Image acquisition and transmission; S2: Image preprocessing and enhancement; S3: Feature extraction, performing AI multi-dimensional feature extraction on the preprocessed image to obtain quantifiable parameters; S4: Quantitative assessment of safety hazards. Based on the extracted quantifiable parameters, a weighted quantitative assessment model is used to calculate the quantitative score of safety hazards, and the level of safety hazards is determined according to the quantitative score. S5: Tiered early warning, triggering corresponding level of early warning action based on the level of safety hazard; S6: Closed-loop management, generating a hazard work order and pushing it to the responsible person. After the responsible person completes the rectification, they upload feedback, and the system verifies and archives the case. As a further description of the technical solution of this invention, the specific process of S2 includes: S2.1: Video stream decoding, extracting continuous image frames at a frame rate of no less than 25 frames per second; S2.2: The dark channel prior dehazing algorithm is used to dehaze the image, improving the image clarity in foggy and dusty environments; S2.3: Use Gaussian filtering or bilateral filtering algorithms to denoise the image and reduce image noise; S2.4: An adaptive histogram equalization algorithm is used for brightness equalization to improve image quality under backlight and low light conditions; S2.5: Dynamic range compression is applied to images captured by a wide dynamic range camera to enhance detail in both bright and dark areas; S2.6: Perform geometric distortion correction based on the camera installation angle and distortion parameters, and output the corrected standard image.

[0008] As a further description of the technical solution of the present invention, the safety hazard quantitative assessment in S4 includes: safety hazard quantitative assessment of not wearing a safety helmet, safety hazard quantitative assessment of not wearing a safety belt while working at height, safety hazard quantitative assessment of lack of edge protection, safety hazard quantitative assessment of blocked fire lanes, and safety hazard quantitative assessment of open flames and smoke. For scenarios with multiple safety hazards, a comprehensive quantitative assessment of safety hazards should be conducted.

[0009] As a further description of the technical solution of the present invention, the calculation of the safety hazard quantitative score includes: the calculation of the safety hazard quantitative score for not wearing a safety helmet, taking into account the duration of stay, the number of people around who are not wearing safety helmets, and whether it is working at height; The calculation of the quantitative score for the safety hazard of not wearing a safety belt while working at height takes into account the working height, the confidence level of the safety belt hook detection, and the impact of wind load. The calculation of the quantitative score for safety hazards caused by the lack of edge protection takes into account the length of the missing protection, the height of the guardrail, and the frequency of personnel activity in the protected area. The calculation of the quantitative score for the safety hazard of blocked fire lanes takes into account factors such as the volume of the obstructing material, the distance between the obstruction and the exit, the width of the passage, and whether an emergency situation has occurred. The calculation of the quantitative score for safety hazards of open flames and smoke takes into account factors such as flame area, smoke concentration, distance between the fire source and combustibles, whether there are fire extinguishing facilities, whether it is in a hazardous materials area, and the speed of flame spread. The comprehensive safety hazard quantitative assessment adopts a weighted quantitative assessment model, which is as follows: ,in, This represents the overall quantitative score for safety hazards. The maximum possible score for the k-th type of security hazard set by the system. Let represent the quantified score of the k-th type of safety hazard, and m represent the total number of quantifiable safety hazard categories detected.

[0010] As a further description of the technical solution of the present invention, the graded early warning in S5 includes: when ∈[0, When a hazard is identified as a general risk, an early warning record is generated on the management platform and marked with a yellow border. when ∈[ , When a hazard is identified as a major hidden danger, a local voice prompt is triggered, a warning message is pushed to the team leader's mobile terminal, and a hazard work order is generated. when ∈[ , When a major hazard is identified, the local sound and light alarm is triggered, the remote broadcasting system is activated, a warning message is sent to the project manager and safety director, a hazard work order is generated, and video clips are captured. when ∈[ , When a fire is identified as a particularly serious hazard, all audible and visual alarms will be activated, the emergency plan will be initiated, an emergency evacuation order will be sent, a voice alarm call will be automatically dialed, and information will be synchronized with the company headquarters.

[0011] As a further description of the technical solution of the present invention, the specific working process of S6 includes: automatic generation and dispatch of work orders, and determination of responsible persons and responsible leaders based on the type of hidden danger and the quantitative score combined with the responsibility matrix rules; On-site handling and AI verification: The person in charge uploads photos after rectification, and the system uses image similarity comparison algorithms and AI detection models to automatically verify the rectification effect; Review and archive: After the safety management personnel review and confirm the case, archive and close the case, and record the complete timeline of the handling process. An escalation mechanism is in place to automatically send an escalation alert to the next higher level of management if the handling is not completed within the specified time limit.

[0012] As a further description of the technical solution of the present invention, it also includes a key target tracking process, comprising: Extract the bounding box coordinates of the target in the image coordinate system: ,in, The x-coordinate is the leftmost position of the target, which is the left edge of the target. The vertical coordinate is the topmost point of the target, which is the position of the top edge of the target. The x-coordinate is the rightmost position of the target, which is the right edge of the target. The vertical coordinate is the bottommost point, which is the position of the bottom edge of the target; Calculate the coordinates of the center point of the target in the image coordinate system: ; Through the system's preset transformation matrix Map the image coordinates to the construction coordinate system, where, Let the target be located in the construction coordinate system; ; Based on the transformed construction coordinates, determine the target and the danger zone. Shortest Euclidean distance ,in, As a reference point on the boundary of the danger zone, Let Q be the coordinates of the reference point in the construction coordinate system; like =0 indicates that the target is in the danger zone; like >0 indicates that the target is outside the danger zone. The numerical value represents the shortest distance from the target to the boundary of the danger zone.

[0013] The beneficial effects of this invention are: First, this invention constructs a multi-dimensional quantifiable parameter system encompassing target spatial location, posture, trajectory, aggregation degree, protective facility status, and environmental status. Combined with a perspective transformation matrix, it precisely maps image coordinates to the construction coordinate system, solving the technical problems of existing technologies' single-dimensional recognition and weak spatial perception capabilities. This enables accurate determination of the spatial relationship between personnel and hazardous areas. Second, this invention establishes dedicated quantitative assessment models for five typical hazards: not wearing a safety helmet, not wearing a safety belt while working at height, lack of edge protection, blocked fire exits, and open flames and smoke. This transforms vague safety hazard judgments into objective and precise quantitative scores. A multiplicative fusion model is used to comprehensively assess the risks of multiple overlapping hazards, solving the technical problems of existing technologies lacking quantitative assessment mechanisms and being unable to handle the risks of multiple overlapping hazards. Third, this invention establishes a four-level graded early warning mechanism based on quantitative scores, triggering differentiated early warning actions according to the severity of the hazard. This solves the technical problems of lacking graded alarm criteria and easily causing "alarm fatigue." Finally, this invention forms a complete closed-loop management process through mechanisms such as automatic work order dispatch, AI visual verification and rectification, and timeout upgrade push notifications. This solves the technical problems of existing technologies that only issue alarms without taking action and cannot quantitatively assess safety management effectiveness. This invention significantly improves the level of intelligence, accuracy of early warning, and efficiency of hazard rectification in construction site safety management. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a partial structural diagram of the construction area safety hazard early warning system based on AI visual recognition of the present invention. Detailed Implementation

[0016] 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, and 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.

[0017] Please see Figure 1 As shown, this invention provides a construction area safety hazard early warning system based on AI visual recognition, including an edge-side acquisition layer, an edge-side computing layer, a cloud-side management layer, and an application layer. The end-side acquisition layer includes multiple high-definition camera devices and supporting lighting devices deployed in key areas of the construction site to adapt to the complex environment of building construction; The image acquisition device includes: High-definition visible light industrial camera: resolution no less than 4K, frame rate 25fps, used for image acquisition of personnel behavior, facility status, and environmental hazards under normal daylight conditions; Infrared thermal imaging camera: Temperature measurement range -20℃-150℃, used for target detection, open flame identification, and monitoring of overheating hazards of electrical equipment in nighttime, backlight, and dust-covered environments, making up for the environmental adaptability shortcomings of visible light cameras; Panoramic intelligent PTZ camera: Supports 360° rotation and zoom cruise, used for full-area inspection without blind spots, and periodically collects images of non-fixed monitoring points; Lighting unit: It adopts dual modes of infrared and white light supplementary lighting, which automatically switches according to the ambient light intensity to ensure the image acquisition quality at night and in low light environments.

[0018] The edge computing layer includes an edge computing node cluster deployed at the construction site, which is communicatively connected to the end-side acquisition layer. It is used to receive multi-channel video stream data, perform AI inference, extract quantifiable parameters, conduct quantitative assessment of safety hazards, and trigger graded early warnings. The cloud-side management layer, including the cloud server cluster, is communicatively connected to the edge computing layer and is used to receive alarm data and quantifiable parameters uploaded by edge computing nodes, and to perform data aggregation, storage, and management platform display. The application layer includes management backend terminals, mobile terminals, and on-site audible and visual alarm devices, used to display alarm information and quantitative assessment results, issue handling instructions, and generate safety management reports.

[0019] The edge computing layer includes a data preprocessing unit, a feature extraction unit, an AI-based hazard classification and judgment unit, and a local caching unit. The data preprocessing unit decodes, dehazes, denoises, equalizes brightness, compresses wide dynamic range, and corrects geometric distortion on the acquired raw image data, removes interference information, and outputs a standardized image to be identified. The feature extraction unit extracts feature parameters based on the standardized image to be identified and processes the feature parameters. The AI-based hazard classification and judgment unit has a built-in hazard risk level judgment standard, analyzes the extracted feature parameters, and maps the analysis results to the hazard risk level judgment standard. The local caching unit caches hazard images, video clips, and early warning records from the past 7 days. It can store data locally when the network is offline and automatically synchronize it to the cloud after the network is restored.

[0020] A method for early warning of safety hazards in construction areas based on AI visual recognition, the method comprising the following steps: S1: Image acquisition and transmission. Image acquisition devices deployed in key areas of the construction site acquire video streams in real time and transmit the video streams to edge computing nodes via the RTSP protocol. The key areas include, but are not limited to, the main entrance and exit of the construction site, construction passages, the top of the tower crane, the perimeter of the foundation pit, high-altitude work surfaces, material processing areas, hazardous materials warehouses, edge openings, unloading platforms, and construction elevator entrances and exits. S2: Image preprocessing and enhancement; The specific process of S2 includes: S2.1: Video stream decoding, extracting continuous image frames at a frame rate of no less than 25 frames per second; S2.2: The dark channel prior dehazing algorithm is used to dehaze the image, improving the image clarity in foggy and dusty environments; S2.3: Use Gaussian filtering or bilateral filtering algorithms to denoise the image and reduce image noise; S2.4: An adaptive histogram equalization algorithm is used for brightness equalization to improve image quality under backlight and low light conditions; S2.5: Dynamic range compression is applied to images captured by a wide dynamic range camera to enhance detail in both bright and dark areas; S2.6: Perform geometric distortion correction based on the camera installation angle and distortion parameters, and output the corrected standard image.

[0021] S3: Divide the construction area into several grids and extract the quantifiable parameters of each grid. S4: Quantitative assessment of safety hazards. Based on the extracted quantifiable parameters, a quantitative assessment model is used to calculate the quantitative score of safety hazards, and the safety hazard level of each grid is determined according to the quantitative score. The quantitative assessment of safety hazards in S4 includes: quantitative assessment of safety hazards such as not wearing a safety helmet, not wearing a safety belt while working at height, lack of edge protection, blocked fire lanes, and open flame and smoke hazards. For scenarios with multiple safety hazards, a comprehensive quantitative assessment of safety hazards should be conducted.

[0022] The calculation of the quantitative score for safety hazards includes: the calculation of the quantitative score for the safety hazard of not wearing a safety helmet, which takes into account the duration of stay and the number of people not wearing safety helmets. ; In the formula, The average time that people not wearing helmets stayed in the area. The number of people in the area who are not wearing safety helmets. The number of people in the area. and These are the weighting coefficients. Quantify the safety hazard of not wearing a safety helmet into a score; The calculation of the quantitative score for the safety hazard of not wearing a safety belt while working at height takes into account the working height, the confidence level of the safety belt hook detection, and the impact of wind load. ; In the formula, To assess the confidence level of safety belt hook detection for workers at heights, target detection is used to identify whether the safety belt hook is securely fixed to a reliable structure. =1, otherwise =0, For workers operating at height, The maximum height set for the system. The wind load influence factor is determined based on weather data. and These are the weighting coefficients. Quantify the safety hazard of not wearing a safety belt while working at heights into a score; The calculation of the quantitative score for safety hazards caused by the lack of edge protection takes into account the length of the missing protection, the height of the guardrail, and the frequency of personnel activity in the protected area. For areas such as foundation pits and building edges, the quantification score for safety hazards caused by inadequate edge protection is as follows: : ; In the formula, For the missing protection length, Set a missing protection length threshold for the system. For the height of the guardrail, For the standard height of the guardrail, The frequency of human activity in a region is the number of people entering that region per unit of time. Set the highest frequency of personnel activity in the area for the system. , and These are the weighting coefficients. Quantify the safety hazards of inadequate edge protection into numerical scores; The calculation of the quantitative score for the safety hazard of blocked fire lanes takes into account factors such as the volume of the obstructing material, the distance between the obstruction and the exit, the width of the passage, and whether an emergency situation has occurred. ; In the formula, The volume of material blocking the channel. Set a maximum volume threshold for the system. The distance from the obstruction to the nearest safe exit. This is the actual width of the channel. For standard channel width, This is an emergency sign; if it is an emergency... =1, otherwise, =0, , and These are the weighting coefficients. Quantify the safety hazards of blocked fire exits into scores; The calculation of the quantitative score for safety hazards of open flames and smoke takes into account factors such as flame area, smoke concentration, distance between the fire source and combustibles, whether there are fire extinguishing facilities, whether it is in a hazardous materials area, and the speed of flame spread. ; In the formula, The flame area The maximum threshold for flame area. This is an estimate of the smoke concentration. The distance between the ignition source and the combustible material. This is a dangerous goods symbol. If dangerous goods are present, =1, otherwise, =0, For the speed of flame spread, - These are the weighting coefficients.

[0023] The comprehensive safety hazard quantitative assessment adopts a weighted quantitative assessment model, which is as follows: ,in, This represents the overall quantitative score for safety hazards. The maximum possible score for the k-th type of security hazard set by the system. Let represent the quantified score of the k-th type of safety hazard, and m represent the total number of quantifiable safety hazard categories detected.

[0024] S5: Tiered early warning, triggering corresponding level of early warning action based on the level of safety hazard; The graded early warning system in S5 includes: when ∈[0, When a hazard is identified as a general risk, an early warning record is generated on the management platform and marked with a yellow border. when ∈[ , When a hazard is identified as a major hidden danger, a local voice prompt is triggered, a warning message is pushed to the team leader's mobile terminal, and a hazard work order is generated. when ∈[ , When a major hazard is identified, the local sound and light alarm is triggered, the remote broadcasting system is activated, a warning message is sent to the project manager and safety director, a hazard work order is generated, and video clips are captured. when ∈[ , When a fire is identified as a particularly serious hazard, all audible and visual alarms will be activated, the emergency plan will be initiated, an emergency evacuation order will be sent, a voice alarm call will be automatically dialed, and information will be synchronized with the company headquarters.

[0025] S6: Closed-loop management of handling issues, generating a hazard work order and pushing it to the responsible person. After the responsible person completes the rectification, they upload feedback, and the system verifies and archives the case to close it.

[0026] The above technical solutions collectively constitute the core closed-loop mechanism for quantitative assessment and graded early warning of safety hazards in this invention. First, the system uses AI visual recognition technology to establish dedicated quantitative assessment models for five typical safety hazards at construction sites (not wearing a safety helmet, not wearing a safety belt while working at height, lack of edge protection, blocked fire exits, open flames, and smoke). Each model extracts multi-dimensional quantifiable parameters from the images—for example, not wearing a safety helmet requires consideration of factors such as the duration of personnel stay, the number of similar violators in the surrounding area, and whether the work is at height; not wearing a safety belt while working at height requires assessment of the working height, the confidence level of the safety belt hook detection, and the impact of wind load; lack of edge protection considers the length of the missing protection, the height deviation of the guardrail, and the frequency of personnel activity in the area; blocked fire exits are analyzed based on the volume of the obstruction, the distance to the exit, the remaining width of the passage, and the emergency situation; open flames and smoke are assessed by integrating multiple indicators such as flame area, smoke concentration, distance between the fire source and combustibles, the availability of fire extinguishing equipment, the attributes of hazardous materials in the area, and the speed of flame spread—substituting each parameter into the corresponding quantitative model to calculate the independent score for each type of hazard. When multiple potential hazards exist simultaneously at the site, the system employs a multiplicative fusion model. The system calculates a comprehensive quantitative score, scientifically reflecting the cumulative effect of multiple risks through the principle of probability multiplication. Based on the calculated comprehensive score, the system automatically matches a preset four-level risk threshold range: when the score falls into the first range and is determined to be a general hazard, the system generates an early warning record on the management platform and marks it with a yellow border; when it falls into the second range and is determined to be a major hazard, a local voice prompt is triggered and a work order is pushed to the team leader; when it falls into the third range and is determined to be a serious hazard, the system activates the sound and light alarm and remote broadcast system, simultaneously pushes the information to the project manager and safety director, and captures video evidence; when it falls into the fourth range and is determined to be an extremely serious hazard, the system immediately activates the full-area sound and light alarm linkage, automatically triggers the emergency plan, pushes emergency evacuation instructions to all personnel, automatically dials the voice alarm phone, and synchronizes information with the company headquarters, realizing intelligent safety management throughout the entire process from precise quantitative assessment to graded differentiated response.

[0027] The specific working process of S6 includes: automatic generation and dispatch of work orders, and determination of responsible persons and leaders based on the type of hidden danger and quantitative score combined with the responsibility matrix rules; On-site handling and AI verification: The person in charge uploads photos after rectification, and the system uses image similarity comparison algorithms and AI detection models to automatically verify the rectification effect; Review and archive: After the safety management personnel review and confirm the case, archive and close the case, and record the complete timeline of the handling process. An escalation mechanism is in place to automatically send an escalation alert to the next higher level of management if the handling is not completed within the specified time limit.

[0028] It also includes the key target tracking process, including: Extract the bounding box coordinates of the target in the image coordinate system: ,in, The x-coordinate is the leftmost position of the target, which is the left edge of the target. The vertical coordinate is the topmost point of the target, which is the position of the top edge of the target. The x-coordinate is the rightmost position of the target, which is the right edge of the target. The vertical coordinate is the bottommost point, which is the position of the bottom edge of the target; Calculate the coordinates of the center point of the target in the image coordinate system: ; Through the system's preset transformation matrix Map the image coordinates to the construction coordinate system, where, Let the target be located in the construction coordinate system; ; Based on the transformed construction coordinates, determine the target and the danger zone. Shortest Euclidean distance ,in, As a reference point on the boundary of the danger zone, Let Q be the coordinates of the reference point in the construction coordinate system; like =0 indicates that the target is in the danger zone; like >0 indicates that the target is outside the danger zone. The numerical value represents the shortest distance from the target to the boundary of the danger zone.

[0029] The key target tracking process described in the above technical solution is the core technology for accurately determining the spatial relationship between personnel and hazardous areas in construction zones. First, the system extracts the bounding box coordinates of the target from the image. And calculate its center point. This completes the positioning of the target in the image coordinate system. Then, using a pre-calibrated perspective transformation matrix H, the coordinates of the image center point are mapped to the actual construction coordinate system, obtaining the target's true position at the construction site. Based on this, the system calculates the target and the danger zone. Shortest Euclidean distance of the boundary .like =0 indicates that the target is within the danger zone; if >0 indicates that the target is outside the danger zone. The numerical value represents the shortest distance from the target to the boundary of the danger zone. In summary, the working principle of this invention is as follows: This invention proposes a construction area safety hazard early warning system and method based on AI visual recognition. Its core working principle can be summarized as a closed-loop intelligent safety management system covering the entire process under a "device-edge-cloud" collaborative architecture. First, the system uses various types of image acquisition devices (high-definition visible light cameras, infrared thermal imaging cameras, panoramic cameras, and supplementary lighting units) deployed in key areas of the construction site to acquire multiple video streams in real time, which are then transmitted to edge computing nodes via the RTSP protocol. The edge computing nodes perform preprocessing on the original images, including decoding, defogging, denoising, brightness equalization, wide dynamic range compression, and geometric distortion correction. Based on the algorithm, they extract the bounding box coordinates of the target in the image coordinate system. Then calculate the center point. The image coordinates are precisely mapped to the construction coordinate system using a preset perspective transformation matrix H. Calculate the target and the danger zone Shortest Euclidean distance This enables precise determination of the spatial relationship between personnel and hazardous sources. Based on this, the system establishes dedicated quantitative assessment models for five typical hazards (not wearing a safety helmet, not wearing a safety belt while working at height, lack of edge protection, blocked fire exits, open flames, and smoke)—for example, the ratio of the total time spent without a safety helmet (t) to the number of people not wearing helmets, with a quantified score. The confidence level of working at height without wearing a safety belt is determined by a combination of factors including the working height and wind load, and the resulting quantitative score. The assessment of deficiencies in edge protection is based on a comprehensive evaluation of the length of the missing section, the height deviation of the guardrail, and the frequency of personnel activity, resulting in a quantified score. Fire lane obstruction is assessed by a comprehensive evaluation of the obstruction volume, distance to exit, reduction in lane width, and emergency conditions, resulting in a quantified score. The overall score is determined by the combined factors of open flame and smoke, including flame area, smoke concentration, distance between the ignition source and combustible material, hazard markings, and spread rate. —Independent quantitative scores are calculated for each type of hazard. When multiple hazards exist simultaneously on-site, the system employs a multiplicative fusion model. The system calculates a comprehensive quantitative score to scientifically reflect the cumulative effect of multiple risks. Based on this comprehensive score, the system automatically matches a preset four-level risk threshold range and triggers differentiated early warning actions: general hazards are only recorded and marked; larger hazards push work orders to the team leader and trigger local voice prompts; major hazards trigger audible and visual alarms and remote broadcasts, push notifications to the project manager and safety director, and capture video evidence; particularly serious hazards activate full-area audible and visual alarm linkage, automatically push emergency evacuation instructions, make voice alarm calls, and synchronize information with company headquarters. Finally, the system forms a complete closed-loop management system through mechanisms such as automatic work order dispatch, AI visual verification of rectification, and timeout upgrade push, realizing intelligent safety control throughout the entire process from hazard identification, quantitative assessment, graded early warning to handling verification.

[0030] The following are specific examples of the present invention: Taking a large-scale housing construction project as an example, the project has a total construction area of ​​approximately 150,000 square meters, including three high-rise residential buildings (33 floors) and supporting underground parking garages. At peak times, there are approximately 500 workers on the construction site, three tower cranes, four construction elevators, and a total of 86 image acquisition devices deployed on the construction site.

[0031] Five AI edge computing nodes were deployed on the construction site, each equipped with an NVIDIA Jetson AGX Orin module (275 TOPS computing power), and connected to cameras via gigabit switches. The edge nodes are equipped with a Linux operating system and Docker containerized AI inference services. Each node can process 16 video streams, with inference latency controlled within 10 milliseconds and a complete alarm process within 500 milliseconds.

[0032] Example 1: Identification and handling of hazards caused by not wearing a safety helmet Scene description: On *year*month*day at 14:23:17, at the working face on the east side of the foundation pit, worker Wang from the support team temporarily removed his safety helmet during the operation.

[0033] A 4K bullet camera deployed on the east side of the foundation pit captures video streams in real time and transmits them to the edge computing node at 25fps via the RTSP protocol.

[0034] Edge computing nodes perform preprocessing on the original image: Decode and extract image frames (resolution 1920×1080). Dark channel prior dehazing: The day's weather was lightly hazy; dehazing improved image contrast by 15%. Gaussian filtering for noise reduction: reducing image sensor noise Adaptive histogram equalization: Improves details in backlit areas Feature extraction (key target tracking) Personnel target: Bounding box coordinates (540, 380, 620, 520), target center point calculation: =(540+620) / 2=580, =(380+520) / 2=450. The image coordinates are mapped to the construction coordinate system through the pre-calibrated transformation matrix H, and the actual construction coordinates of the personnel are obtained as P=(12.5,23.8).

[0035] The danger zones are rectangles (10.0, 20.0), (10.0, 30.0), and (15.0, 30.0). (15.0, 20.0). Calculate the shortest distance from the target to the danger zone: 0, indicating the target is within the danger zone.

[0036] Extract quantifiable parameters: dwell time t = 2.4 seconds (60 frames in total), number of people not wearing helmets in the area (1), total number of people in the area (3), weighting coefficient. =0.6, =0.4 (system preset), calculate the quantitative score for not wearing a safety helmet: =0.18, at this point there are no other hidden dangers, the comprehensive quantitative score is... = =0.18. Belongs to [0, ) range (preset) =0.3), which is judged as a general hidden danger. The action is as follows: generate an early warning record on the management platform, mark the person with a yellow border on the monitoring screen, do not trigger an immediate push, and wait for the safety officer to pay attention during the inspection.

[0037] Example 2: The following potential hazards exist simultaneously in a certain construction area: Not wearing a helmet =0.7683; Working at height without a safety belt. =1; lack of edge protection. =1; The total number of quantifiable safety hazard categories detected is m=3, and the maximum possible score for each type of hazard is 1. =1.0. Comprehensive quantitative score calculation: It should be noted that all formulas in this application are calculated by removing dimensions and obtaining numerical values. =1.

[0038] Warning judgment: =1.0∈[ , ), preset =0.85, which is considered a particularly serious hidden danger.

[0039] Warning action: Trigger all audible and visual alarms on the work surface. The emergency response plan was activated, and an emergency evacuation order was sent to all project management personnel and workers on that floor. Automatically dials the project manager and safety director to issue voice alerts; Synchronize information with the company headquarters' security management platform; The construction elevator and all entrance and exit gates are automatically opened to facilitate the rapid evacuation of personnel. AI continuously tracks and updates dynamic information on potential hazards in real time.

[0040] Handling process: Safety management personnel used a remote broadcasting system to call out to workers in the adjacent area to immediately fasten their safety belts and evacuate. The safety officer quickly arrived at the scene, supervised the two workers to fasten their safety belts, and organized the evacuation of personnel from the danger zone. The edge protection railings were immediately reinstalled; Safety training was conducted for three workers who were not wearing safety helmets; After rectification is completed, upload photos, pass AI verification, and archive the work order; Total closed-loop time: 8 minutes.

[0041] The system preset thresholds, threshold ranges, and coefficients involved in this application are all empirical values, selected by those skilled in the art based on actual conditions. The formulas are derived from software simulations using collected data, resulting in a formula that most closely approximates the real situation. All parameters involved in this application have undergone uniform normalization to remove dimensions and calculate their numerical values.

[0042] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A construction area safety hazard early warning system based on AI visual recognition, characterized in that, It includes the edge-side acquisition layer, the edge-side computing layer, the cloud-side management layer, and the application layer; The end-side acquisition layer includes multiple high-definition camera devices and supporting lighting devices deployed in key areas of the construction site to adapt to the complex environment of building construction; The edge computing layer includes an edge computing node cluster deployed at the construction site, which is communicatively connected to the end-side acquisition layer. It is used to receive multi-channel video stream data, perform AI inference, extract quantifiable parameters, conduct quantitative assessment of safety hazards, and trigger graded early warnings. The cloud-side management layer, including the cloud server cluster, is communicatively connected to the edge computing layer and is used to receive alarm data and quantifiable parameters uploaded by edge computing nodes, and to perform data aggregation, storage, and management platform display. The application layer includes management backend terminals, mobile terminals, and on-site audible and visual alarm devices, used to display alarm information and quantitative assessment results, issue handling instructions, and generate safety management reports.

2. The construction area safety hazard early warning system based on AI visual recognition as described in claim 1, characterized in that, The edge computing layer includes a data preprocessing unit, a feature extraction unit, an AI-based hazard classification and judgment unit, and a local caching unit. The data preprocessing unit decodes, dehazes, denoises, equalizes brightness, compresses wide dynamic range, and corrects geometric distortion on the acquired raw image data, removes interference information, and outputs a standardized image to be identified. The feature extraction unit extracts feature parameters based on the standardized image to be identified and processes the feature parameters. The AI-based hazard classification and judgment unit has a built-in hazard risk level judgment standard, analyzes the extracted feature parameters, and maps the analysis results to the hazard risk level judgment standard. The local caching unit caches hazard images, video clips, and early warning records from the past 7 days. It can store data locally when the network is offline and automatically synchronize it to the cloud after the network is restored.

3. A method for early warning of safety hazards in construction areas based on the system described in any one of claims 1 to 2, characterized in that, The method includes the following steps: S1: Image acquisition and transmission; S2: Image preprocessing and enhancement; S3: Feature extraction, performing AI multi-dimensional feature extraction on the preprocessed image to obtain quantifiable parameters; S4: Quantitative assessment of safety hazards. Based on the extracted quantifiable parameters, a weighted quantitative assessment model is used to calculate the quantitative score of safety hazards, and the level of safety hazards is determined according to the quantitative score. S5: Tiered early warning, triggering corresponding level of early warning action based on the level of safety hazard; S6: Closed-loop management of handling issues, generating a hazard work order and pushing it to the responsible person. After the responsible person completes the rectification, they upload feedback, and the system verifies and archives the case to close it.

4. The method for early warning of safety hazards in construction areas according to claim 3, characterized in that, The specific process of S2 includes: S2.1: Video stream decoding, extracting continuous image frames at a frame rate of no less than 25 frames per second; S2.2: The dark channel prior dehazing algorithm is used to dehaze the image, improving the image clarity in foggy and dusty environments; S2.3: Use Gaussian filtering or bilateral filtering algorithms to denoise the image and reduce image noise; S2.4: An adaptive histogram equalization algorithm is used for brightness equalization to improve image quality under backlight and low light conditions; S2.5: Dynamic range compression is applied to images captured by a wide dynamic range camera to enhance detail in both bright and dark areas; S2.6: Perform geometric distortion correction based on the camera installation angle and distortion parameters, and output the corrected standard image.

5. The method for early warning of safety hazards in construction areas according to claim 3, characterized in that, The quantitative assessment of safety hazards in S4 includes: quantitative assessment of safety hazards such as not wearing a safety helmet, not wearing a safety belt while working at height, lack of edge protection, blocked fire lanes, and open flame and smoke hazards. For scenarios with multiple safety hazards, a comprehensive quantitative assessment of safety hazards should be conducted.

6. The method for early warning of safety hazards in construction areas according to claim 3, characterized in that, The calculation of the quantitative score for the safety hazard includes: The calculation of the quantitative score for the safety hazard of not wearing a safety helmet takes into account factors such as the duration of the stay, the number of people around who are not wearing safety helmets, and whether the work is being done at height. The calculation of the quantitative score for the safety hazard of not wearing a safety belt while working at height takes into account the working height, the confidence level of the safety belt hook detection, and the impact of wind load. The calculation of the quantitative score for safety hazards caused by the lack of edge protection takes into account the length of the missing protection, the height of the guardrail, and the frequency of personnel activity in the protected area. The calculation of the quantitative score for the safety hazard of blocked fire lanes takes into account factors such as the volume of the obstructing material, the distance between the obstruction and the exit, the width of the passage, and whether an emergency situation has occurred. The calculation of the quantitative score for safety hazards of open flames and smoke takes into account factors such as flame area, smoke concentration, distance between the fire source and combustibles, whether there are fire extinguishing facilities, whether it is in a hazardous materials area, and the speed of flame spread. The comprehensive safety hazard quantitative assessment adopts a weighted quantitative assessment model, which is as follows: ,in, This represents the overall quantitative score for safety hazards. The maximum possible score for the k-th type of security hazard set by the system. Let represent the quantified score of the k-th type of safety hazard, and m represent the total number of quantifiable safety hazard categories detected.

7. The method for early warning of safety hazards in construction areas according to claim 3, characterized in that, The graded early warning system in S5 includes: when ∈[0, When a hazard is identified as a general risk, an early warning record is generated on the management platform and marked with a yellow border. when ∈[ , When a hazard is identified as a major hidden danger, a local voice prompt is triggered, a warning message is pushed to the team leader's mobile terminal, and a hazard work order is generated. when ∈[ , When a major hazard is identified, the local sound and light alarm is triggered, the remote broadcasting system is activated, a warning message is sent to the project manager and safety director, a hazard work order is generated, and video clips are captured. when ∈[ , When a fire is identified as a particularly serious hazard, all audible and visual alarms will be activated, the emergency plan will be initiated, an emergency evacuation order will be sent, a voice alarm call will be automatically dialed, and information will be synchronized with the company headquarters.

8. The method for early warning of safety hazards in construction areas according to claim 3, characterized in that, The specific working process of S6 includes: automatic generation and dispatch of work orders, and determination of responsible persons and leaders based on the type of hidden danger and quantitative score combined with the responsibility matrix rules; On-site handling and AI verification: The person in charge uploads photos after rectification, and the system uses image similarity comparison algorithms and AI detection models to automatically verify the rectification effect; Review and archive: After the safety management personnel review and confirm the case, archive and close the case, and record the complete timeline of the handling process. An escalation mechanism is in place to automatically send an escalation alert to the next higher level of management if the handling is not completed within the specified time limit.

9. The method for early warning of safety hazards in construction areas according to claim 3, characterized in that, It also includes the key target tracking process, including: Extract the bounding box coordinates of the target in the image coordinate system: ,in, The x-coordinate is the leftmost position of the target, which is the left edge of the target. The vertical coordinate is the topmost point of the target, which is the position of the top edge of the target. The x-coordinate is the rightmost position of the target, which is the right edge of the target. The vertical coordinate is the bottommost point, which is the position of the bottom edge of the target; Calculate the coordinates of the center point of the target in the image coordinate system: ; Through the system's preset transformation matrix Map the image coordinates to the construction coordinate system, where, Let the target be located in the construction coordinate system; ； Based on the transformed construction coordinates, determine the target and the danger zone. Shortest Euclidean distance ,in, As a reference point on the boundary of the danger zone, Let Q be the coordinates of the reference point in the construction coordinate system; like =0 indicates that the target is in the danger zone; like >0 indicates that the target is outside the danger zone. The numerical value represents the shortest distance from the target to the boundary of the danger zone.

Images