A construction site safety compliance low-power ai camera device and method based on edge side risk chain determination

Abstract

The application discloses a construction site safety compliance low-power-consumption AI camera device and method based on edge side risk chain determination, and relates to the technical field of construction site safety visual monitoring.The device comprises an event trigger wake-up module, a dual-light acquisition module, a construction site scene feature filtering module, a lightweight AI inference module, a human-equipment association module, an edge risk chain determination module, a hierarchical early warning module, a network interruption buffer module and a multi-mode power supply module; the adaptive fusion of visible light and infrared images improves the robustness in a dust raising, night or strong light environment; fixed interference such as scaffolding and building materials is removed by a construction site scene mask; a lightweight deep learning model (INT8 quantization improved YOLO-Tiny) is used to realize synchronous detection of personnel, safety helmets, reflective clothing and safety belts; and accurate area matching of personnel and protective equipment is completed based on human skeleton key points; finally, a three-level risk chain of "protective equipment wearing state-operation area attribute-personnel time sequence moving track" is constructed on the edge side, and local real-time hierarchical early warning is realized.The application does not depend on cloud computing throughout the process, can independently operate in a network interruption state, has low power consumption, low false alarm rate and low deployment cost, is suitable for complex construction site scenes such as building construction, municipal engineering and mine operation, and has significant novelty, creativity and practicality.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence visual inspection, edge computing, and construction site safety monitoring technology, specifically to a low-power AI camera device and method for construction site safety compliance based on edge-side risk chain determination. Background Technology

[0002] Construction sites, municipal engineering projects, and other construction sites have stringent safety and compliance requirements, and existing monitoring and AI recognition solutions have the following technical shortcomings: (1) Traditional monitoring only records video and relies on manual inspection and playback. It cannot automatically identify violations in real time, and the early warning is seriously delayed. (2) Existing construction site AI security solutions mostly adopt a cloud-based centralized processing architecture. The construction site network environment is poor and the bandwidth is limited, which can easily lead to delays, packet loss, and warning failures, and cannot meet the real-time requirements. (3) Most detections only identify a single target and do not associate personnel, protective equipment, work area, and temporal behavior. In multi-person scenarios, equipment and personnel mismatch is likely to occur, resulting in high false alarm and false negative rates. (4) The construction site is subject to dust, backlight, nighttime, obstruction, and interference from complex debris, which significantly reduces the accuracy of the general visual model; (5) The camera operates at high power continuously, which cannot meet the low power deployment requirements of solar energy, temporary locations, etc.

[0003] In existing technologies, dual-light cameras, edge AI inference, safety helmet detection, and low-power triggering are all existing methods applied independently, and they have obvious technical bottlenecks: dual-light acquisition alone cannot solve the combined interference of construction site dust and complex backgrounds; edge inference alone can only achieve single-target recognition; low-power triggering alone cannot balance wake-up speed and detection accuracy; and detection of protective equipment alone is prone to mismatch in multi-person scenarios. Currently, there is no solution to deeply integrate these technologies with functions such as construction site-specific semantic filtering, precise human-equipment association, edge-side three-level risk chain determination, and local closed-loop during network outages. Furthermore, a complete edge local closed-loop process of "acquisition-filtering-identification-association-determination-early warning-caching" has not been formed. This results in the inability to simultaneously address the core pain points of construction site monitoring, such as real-time performance, anti-interference, low power consumption, high accuracy, and availability during network outages. The simple superposition of existing single-function modules cannot achieve the aforementioned synergistic effects, nor can it meet the actual safety monitoring needs of complex construction site scenarios. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies, such as reliance on cloud computing, poor environmental adaptability, high false alarm and false negative rates, high overall power consumption, and inability to function properly when the network is offline. It provides a construction site safety compliance AI camera device and method that can operate independently at the edge, has strong anti-interference capabilities, low power consumption, and high recognition accuracy. This invention is not a simple superposition of existing technology modules, but rather a deep collaborative approach that utilizes edge risk chain judgment, dual-light adaptive fusion, construction site-specific semantic masks, low-power event triggering, and precise human skeleton matching to construct a complete edge-local closed loop of "collection-filtering-recognition-association-judgment-early warning-caching." This achieves high accuracy, low latency, low power consumption, and network-off-line usability for safety and compliance monitoring in complex construction site environments. The overall technical effect is significantly better than existing single recognition solutions, effectively solving the technical bottleneck of existing technologies that cannot simultaneously address multiple pain points, demonstrating outstanding creativity and practicality. To achieve the above objectives, this invention adopts the following technical solution:

[0005] An AI-powered construction site safety compliance monitoring device based on edge-side risk chain determination includes: a dual-light acquisition module, an event-triggered wake-up module, a construction site scene feature filtering module, a lightweight AI inference module, a human-equipment association module, an edge risk chain determination module, a tiered early warning module, a network outage caching module, and a multi-mode power supply module. These modules work together to achieve the aforementioned end-to-end edge local closed-loop and safety compliance monitoring functions. The specific structure and functions are as follows:

[0006] The modules are electrically connected and integrated into an IP67-rated protective housing, making them suitable for complex outdoor construction site environments.

[0007] The device is an independent edge-side operating device that can complete the entire process of detection, judgment and early warning without cloud interaction.

[0008] The event-triggered wake-up module is configured to use independent detection elements such as infrared pyroelectric sensors and microwave sensors to sense whether personnel have entered the area. When no personnel are present, the control device enters a low-power sleep mode. After detecting personnel, it wakes up quickly within 100ms and dynamically adjusts the working frame rate, significantly reducing the standby power consumption of the entire device.

[0009] The construction site scene feature filtering module is configured to: adaptively weight and fuse visible light and infrared images to suppress dust and backlight interference, and generate a semantic mask based on a pre-built construction site scene interference library to remove background interference such as scaffolding, steel bars, and building materials, thereby improving target detection accuracy.

[0010] The edge risk chain determination module is configured to execute a three-level risk linkage determination logic locally on the device, and output level one, level two, and level three risk levels by comprehensively considering the wearing status of personal protective equipment, the attributes of the work area, and the personnel movement trajectory information.

[0011] The graded early warning module executes graded audible and visual alarms and pushes information to the safety officer's terminal based on the risk level.

[0012] Furthermore, the dual-light acquisition module includes a 4-megapixel visible light camera and a 2-megapixel infrared camera coaxially mounted, capable of adaptively weighted image fusion based on light intensity and dust levels. Specifically, when the light intensity is <50 lux (nighttime), the infrared image accounts for 80% of the weight, and the visible light image accounts for 20%. When the dust concentration is high, the infrared image accounts for 60% of the weight, and the visible light image accounts for 40%. In backlit scenarios, the visible light image is first subjected to histogram equalization processing before being weighted and fused with the infrared image, effectively suppressing dust, halos, and noise interference. This dual-light fusion strategy differs from existing general dual-light fusion solutions, specifically designed for typical interference scenarios such as construction site dust, backlight, and nighttime. The fusion weights can be adaptively adjusted in real time according to the on-site environment, and the target clarity of the fused image is improved by more than 30% compared to existing technologies, providing a high-quality image foundation for subsequent target detection.

[0013] Furthermore, the event-triggered wake-up module is used to detect people based on changes in screen pixels. When no people are present, the control device enters sleep mode with power consumption not exceeding 1W. When people are detected, the device wakes up within 100ms and dynamically adjusts the frame rate to 5~25FPS, adaptively adjusting the frame rate according to the density of people's activities, thus balancing detection accuracy and power consumption control.

[0014] Furthermore, the construction site scene feature filtering module pre-constructs a construction site interference library, which includes common background clutter on construction sites such as scaffolding, steel bars, building materials, fences, formwork, construction machinery, and temporary facilities. A targeted semantic mask is generated using a semantic segmentation algorithm to accurately filter out these background interferences, highlighting and retaining the target features of personnel and protective equipment, significantly reducing the false detection rate caused by background interference. This interference library is specifically designed for construction site scenes, unlike existing general interference filtering solutions. It can flexibly update interference samples according to different construction site scenes (building construction, municipal engineering, mining operations), achieving a background interference filtering accuracy of over 95%, effectively solving the technical problems of false filtering and missed filtering in existing general filtering modules in construction site scenes.

[0015] Furthermore, the lightweight AI inference module adopts an improved YOLO-Tiny model with INT8 quantization. The model size is less than 5MB, enabling simultaneous detection of personnel, safety helmets, reflective vests, and safety belts on the edge NPU, with an inference speed of at least 60 FPS, meeting real-time detection requirements. The model is trained on a construction site-specific dataset containing over 100,000 samples covering complex scenarios such as construction site dust, nighttime conditions, occlusion, and backlighting. The model's loss function is optimized for the dense workforce and diverse equipment found on construction sites, improving detection accuracy in complex environments. Compared to the existing general-purpose YOLO-Tiny model, this improved model achieves over 15% higher detection accuracy and over 10% faster inference speeds in complex construction site scenarios, while reducing size by 20%, adapting to the low-computing-power and low-storage deployment requirements of edge computing.

[0016] Furthermore, the human-equipment association module divides the head, torso, and waist regions based on 17 key points of the human skeleton. Through IOU matching and ID tracking, it achieves regional matching between personnel and protective equipment, avoiding target mismatch in multi-person scenarios and improving the accuracy of protective equipment wearing status determination. This association mechanism differs from existing simple target binding schemes. It achieves precise matching through skeletal key point region division and continuous binding between personnel and equipment through ID tracking. The mismatch rate in multi-person scenarios is <0.5%, solving the false alarm and missed alarm problems caused by equipment-person mismatch in existing technologies when multiple people are working together.

[0017] Furthermore, the three-level risk assessment rule is as follows: Level 1 Risk: Personnel are not wearing safety helmets or reflective clothing and have not entered the hazardous work area; Level 2 risk: Personnel not wearing safety helmets or reflective vests, or entering ordinary construction areas; Level 3 risk: Personnel are not wearing protective equipment such as safety helmets, reflective vests, or safety belts, and are entering dangerous restricted areas such as foundation pits, tower cranes, high-altitude operations, or power distribution areas.

[0018] Furthermore, the graded early warning module performs the following actions: Level 1 risk is indicated by a buzzer alert; Level 2 risk is indicated by an audible and visual alarm; Level 3 risk is indicated by an audible and visual alarm and a push notification to the safety officer's terminal. The push notification includes the alarm time, risk level, alarm location, and a real-time screenshot of the scene, facilitating rapid response by the safety officer.

[0019] Furthermore, the offline caching module can locally store no less than 7 days of alarm videos and records in 1080P@5fps format. After connecting to the network, it automatically synchronizes to the background management system to ensure that alarm data is not lost and achieve full traceability in the offline state.

[0020] Furthermore, the multi-mode power supply module supports DC12V, POE and 5V solar power supply, adapting to various deployment scenarios on construction sites and solving the problem of inconvenient power supply at temporary construction site locations; the overall operating power consumption is no higher than 3.5W, and the sleep power consumption is no higher than 1W, enabling long-term stable operation at temporary monitoring points powered by solar energy.

[0021] Furthermore, the device uses the HiSilicon Hi3516DV300 as the main control chip, with a built-in NPU computing power of no less than 2 TOPS; it is equipped with 2GB of LPDDR4 memory and 8GB of eMMC storage; the whole machine adopts an IP67 waterproof and dustproof structure, which is suitable for complex outdoor construction site environments.

[0022] A construction site safety compliance identification method based on edge-side risk chain determination, applied to any of the above-mentioned devices, characterized by comprising the following steps: (1) Event triggering: Sleep when no one is present / Wake up when someone is present. This is achieved through an event-triggered wake-up module. When no one is present, the device sleeps with low power consumption and wakes up within 100ms after a person is detected. The working frame rate is dynamically adjusted to 5~25FPS. (2) Dual-light image acquisition and adaptive fusion. Visible light and infrared dual-light images are acquired through a dual-light acquisition module, and adaptive feature fusion is performed according to the light intensity and dust level to suppress interference from dust, backlight, nighttime, etc. (3) Construction site interference mask filtering. A scene semantic mask is generated through the construction site scene feature filtering module to remove background interference such as scaffolding, steel bars, and building materials, while retaining the target features of personnel and protective equipment; (4) Lightweight AI target detection at the edge. A lightweight AI inference module is used to simultaneously detect personnel and protective equipment such as helmets, reflective vests, and safety belts at the edge. (5) Human-equipment association based on skeletal key points. The human-equipment association module divides the head, torso, and waist regions based on 17 human skeletal key points. Through IOU matching + ID tracking mechanism, the region matching between individual personnel and protective equipment is completed; (6) Three-level determination of edge risk chain. Through the edge risk chain determination module, the protective equipment wearing status, the attributes of the work area and the personnel's time-series movement trajectory are combined, and the three-level risk linkage determination is completed and the corresponding risk level is output by comparing with the preset three-level risk rules. The three-level risk determination process is completed independently on the edge end, without relying on cloud computing and network transmission. Risk determination and alarm can still be realized normally under network interruption and bandwidth limitation conditions. The risk chain determination logic is a progressive linkage determination, not a single condition determination. First, the protective equipment wearing status is determined, then the attributes of the work area are combined, and finally the personnel's time-series movement trajectory is associated to form a complete risk determination chain, avoiding misjudgment and omission caused by the existing single condition determination. (7) Tiered early warning output. Through the tiered early warning module, corresponding tiered audible and visual alarms and information push operations are executed according to the risk level. Alarm information is pushed to the safety officer's terminal simultaneously for three levels of risk. (8) Local caching and synchronization of alarm data. The alarm-related video clips and data are cached locally through the offline caching module. They can be stored for no less than 7 days when the network is offline and automatically synchronized to the background management system after the network is connected.

[0023] Compared with the prior art, the present invention has the following beneficial effects: (1) The edge-side independent inference and risk chain judgment are adopted, which can complete the whole process detection and early warning without relying on the cloud server. It can still work stably under poor network, insufficient bandwidth or network outage conditions, which significantly improves the real-time performance and reliability of the system and solves the problems of easy delay and failure of traditional solutions. According to the test, the risk judgment latency of this device is ≤150ms, and it can work stably for more than 7 days in the network outage state without loss of alarm data. Compared with the existing cloud centralized processing solution, the real-time performance is improved by more than 60%, and the network outage adaptability is significantly better than the existing solution, which fully meets the actual needs of unstable network on the construction site. (2) By adaptively fusing visible light and infrared dual-light images and using semantic mask filtering for construction site scenarios, it can effectively suppress interference from dust, nighttime, backlight and complex backgrounds, significantly reduce false alarm rate and false negative rate, and significantly enhance environmental adaptability. According to the test, in complex scenarios such as dust, nighttime and backlight, the false alarm rate is <2% and the false negative rate is <1%, and the detection accuracy is better than the existing general vision model. Compared with the existing single light acquisition or general filtering scheme, the false alarm rate in complex scenarios is reduced by more than 70% and the false negative rate is reduced by more than 80%, which is suitable for various complex working environments on construction sites. (3) The device adopts an event-triggered sleep-wake mechanism. When there are no personnel, the device operates with low power consumption and starts up quickly when there are personnel. The overall power consumption is significantly reduced. It can support solar power supply and temporary site deployment, making it more energy-efficient and flexible. The whole machine sleep power consumption is ≤1W and working power consumption is ≤3.5W. It can work continuously for more than 3 days (sunny days) in solar power supply mode. Compared with the existing continuous high power consumption monitoring equipment, the power consumption is reduced by more than 60%. There is no need for complex power supply wiring, which solves the industry pain point of inconvenient power supply at temporary sites and significantly improves deployment flexibility. (4) Based on the key points of the human skeleton, the precise area matching of personnel and protective equipment is realized. Combined with the IOU matching + ID tracking mechanism, the mismatch rate in multi-person scenarios is <0.5%, which effectively avoids the confusion of target ownership when multiple people are working together and significantly improves the accuracy of risk judgment. At the same time, a three-level graded early warning is adopted. The early warning strategy is more scientific and more in line with the actual needs of construction site safety management, which facilitates the graded handling by safety officers and improves the efficiency of safety management. (5) The device is highly integrated, the model is lightweight, the hardware cost is controllable, and the on-site deployment is simple. It can be widely used in complex construction site scenarios such as building construction, municipal engineering, and mining operations. It is highly practical and has high engineering application and promotion value. Attached Figure Description

[0024] Figure 1 This is a block diagram of the overall structure of the AI ​​camera device of the present invention, showing the connection relationship of each component, wherein: 100 is the main body of the AI ​​camera device; 101 is the event-triggered wake-up module; 102 is the dual-light acquisition module; 103 is the construction site scene feature filtering module; 104 is the lightweight AI inference module; 105 is the human-equipment association module; 106 is the edge risk chain determination module; 107 is the hierarchical early warning module; 108 is the network outage caching module; and 109 is the multi-mode power supply module.

[0025] Figure 2 The flowchart of the construction site safety compliance identification method of the present invention shows the complete steps of the method, wherein: 201 is event triggering: no one is asleep / someone is awakened; 202 is dual-light image acquisition and adaptive fusion; 203 is construction site interference mask filtering; 204 is edge-side lightweight AI target detection; 205 is human-equipment association based on skeletal key points; 206 is edge risk chain three-level judgment; 207 is hierarchical early warning output; 208 is alarm data local caching and synchronization.

[0026] Figure 3 This is a logical diagram of the edge risk chain determination of the present invention, showing the core elements and hierarchical relationship of risk determination, wherein: 301 is the protective equipment wearing status; 302 is the work area attribute; 303 is the personnel temporal movement trajectory; 304 is the edge risk chain comprehensive determination unit; 305 is the first-level risk; 306 is the second-level risk; and 307 is the third-level risk. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to specific embodiments.

[0028] Hardware Implementation Example: This device uses the Hisilicon Hi3516DV300 as the main control chip, with a built-in NPU computing power of no less than 2 TOPS. Image acquisition uses a coaxially deployed 4-megapixel visible light camera and a 2-megapixel infrared camera, with a lens focal length of 4mm and a field of view of 110°. It is equipped with 2GB of LPDDR4 memory and 8GB of eMMC storage, and supports external TF card expansion. The device's sleep power consumption is no more than 1W, and its working power consumption is no more than 3.5W. The power supply supports DC12V power supply (input voltage range 9~15V), POE power supply (compliant with IEEE 802.3af standard), and 5V solar power supply (compatible with 5~6V solar panels). The whole machine adopts an IP67 waterproof and dustproof structure, and the shell is made of aluminum alloy. The protection level meets the needs of complex environments such as wind, rain, and dust in outdoor construction sites. The operating temperature range is -20℃~60℃.

[0029] Algorithm Example: An improved lightweight YOLO-Tiny object detection model is adopted, with a size of 4.7MB after model pruning and INT8 quantization. The inference speed on the edge NPU is no less than 62FPS. The training dataset contains more than 120,000 samples in complex scenarios such as construction site dust, nighttime, occlusion, and backlighting, covering personnel and protective equipment samples of different age groups and different work scenarios. The model achieves detection accuracy of 98.5%, 97.8%, 97.2%, and 96.5% for personnel, safety helmets, reflective clothing, and safety belts, respectively. 17 human skeleton key points are extracted through a lightweight OpenPose network with an extraction speed of no less than 55FPS, achieving accurate segmentation of the head, torso, and waist regions. The IOU matching + ID tracking mechanism completes accurate matching of personnel and protective equipment, with a false detection rate of <0.5% in multi-person scenarios. This improved model is specifically optimized for construction site scenarios and is adapted to the NPU computing power of the HiSilicon Hi3516DV300 main control chip in this device. While ensuring lightweight design, its detection accuracy and speed are superior to existing general-purpose edge AI models. Furthermore, the training dataset closely matches the actual construction site scenario, avoiding the problem of poor adaptability of existing general-purpose models in construction site scenarios.

[0030] Working Example: This device is fixedly installed with brackets at monitoring points around hazardous work areas such as construction site pits, tower cranes, and aerial work platforms. When no personnel enter the monitored area, the device is in a low-power sleep state, consuming approximately 0.8W. When the event-triggered wake-up module detects that the pixel change in the image exceeds a preset threshold (indicating personnel entry), it wakes up the device within 100ms and enters working mode, dynamically adjusting the frame rate to 15FPS (25FPS when personnel are dense, and 5FPS when personnel are sparse). The dual-light acquisition module simultaneously acquires visible light and infrared images, and adaptively weights and fuses the images according to the ambient light intensity (e.g., backlighting at dusk) to suppress halos and noise. The construction site scene feature filtering module calls a pre-built construction site interference library to generate a semantic mask, removing background interference such as scaffolding and rebar, while retaining personnel and protective equipment targets. Lightweight A The I-inference module synchronously detects personnel and protective equipment. If a person is detected not wearing a safety helmet, the human-equipment association module uses 17 key points of the human skeleton to divide the head area, accurately matching the person with the state of not wearing a safety helmet, avoiding mismatch with surrounding personnel. The edge risk chain judgment module further detects the person's movement trajectory, determining that they have entered the dangerous work area of ​​the foundation pit, meeting the three-level risk judgment rules. The graded early warning module immediately triggers on-site audible and visual alarms (audible and visual alarm distance not less than 50 meters) and pushes the alarm time, risk level, alarm location, and real-time on-site screenshot to the safety officer's terminal. At the same time, the network outage caching module caches the corresponding alarm video clips (10 seconds before and after the alarm) and alarm records locally. If the network is down, it can be continuously saved for 8 days. After the network is restored, it will be automatically synchronized to the construction site background management system, realizing full traceability of alarm data. This workflow forms a complete edge-local closed loop, with each module working collaboratively, solving the problem of independent operation and poor collaboration of existing solutions. In practical applications, it can effectively reduce the incidence of construction site safety accidents and improve safety management efficiency.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A low-power AI camera device for construction site safety compliance based on edge-side risk chain determination, characterized in that, include: Dual-light acquisition module, event-triggered wake-up module, construction site scene feature filtering module, lightweight AI inference module, human-equipment association module, edge risk chain determination module, graded early warning module, network outage caching module, and multi-mode power supply module; Each module is electrically connected and integrated within an IP67-rated enclosure; The event-triggered wake-up module is configured to: use independent detection elements such as infrared pyroelectric sensors and microwave sensors to sense whether personnel have entered the area; when no personnel have entered, the control device is in a low-power sleep state; and after detecting personnel entering, it wakes up and enters the working mode within 100ms, effectively reducing the standby power consumption of the device and solving the problem of excessive power consumption during continuous operation of traditional monitoring equipment. The construction site scene feature filtering module is configured to: suppress dust interference through adaptive fusion of visible light and infrared images, and remove background interference such as scaffolding and building materials based on the construction site scene mask, thereby solving the problem of high false alarm rate in image recognition under complex construction site environments; The edge risk chain determination module is configured to complete the three-level risk linkage determination locally on the device. The determination conditions integrate the wearing status of personnel protective equipment, the attributes of the work area and the personnel's temporal movement trajectory. No cloud interaction is required, which solves the problems of delay and network failure caused by the reliance on the cloud in the existing technology. The tiered early warning module executes corresponding tiered audible and visual alarms and pushes information based on the three-level risk level, with the push recipients being the safety officer's terminal. Unlike existing construction site intelligent monitoring patents that only achieve low power consumption, single hazard detection, or cloud-based risk assessment, this device constructs a closed-loop process at the edge through deep collaboration of multiple modules, including dual-light anti-interference, edge risk chain determination, and precise human-equipment association. It simultaneously addresses technical pain points such as construction site dust interference, low power consumption, network outage availability, and multi-person target mismatch. The technical effect is significantly superior to existing technologies, with algorithms and technical features mutually supporting each other to form a unique technical approach.

2. The apparatus according to claim 1, characterized in that, The dual-light acquisition module includes a 4-megapixel visible light camera and a 2-megapixel infrared camera set coaxially, and can adaptively perform image weighted fusion according to light intensity and dust level.

3. The apparatus according to claim 1, characterized in that, The event-triggered wake-up module is used to detect people based on changes in screen pixels. When no people are present, the control device enters sleep mode with a power consumption of no more than 1W. When people are detected, the device wakes up within 100ms and dynamically adjusts the frame rate to 5~25FPS.

4. The apparatus according to claim 1, characterized in that, The construction site scene feature filtering module pre-builds a construction site interference library and can generate semantic masks to filter background interference such as scaffolding, steel bars, building materials, and fences.

5. The apparatus according to claim 1, characterized in that, The lightweight AI inference module uses an improved YOLO-Tiny model with INT8 quantization. The model size is less than 5MB and it is used to simultaneously detect personnel, safety helmets, reflective vests, and safety belts.

6. The apparatus according to claim 1, characterized in that, The human-equipment association module divides the head, torso, and waist regions based on 17 key points of the human skeleton, enabling regional matching between personnel and protective equipment and avoiding target mismatch in multi-person scenarios.

7. The apparatus according to claim 1, characterized in that, The three levels of risk are specifically divided as follows: Level 1 Risk: Not wearing a helmet or reflective vest, and not entering a hazardous area; Level 2 risk: Not wearing a safety helmet and reflective clothing, or entering a regular construction area; Level 3 risk: Entering a hazardous work area without wearing protective equipment.

8. The apparatus according to claim 1, characterized in that, The graded early warning module performs the following actions: Level 1 risk is indicated by a buzzer alert; Level 2 risk is indicated by an audible and visual alarm; and Level 3 risk is indicated by an audible and visual alarm and a notification to the safety officer's terminal.

9. The apparatus according to claim 1, characterized in that, The offline caching module can locally store alarm videos and records for no less than 7 days, and automatically synchronize them after connecting to the network.

10. The apparatus according to claim 1, characterized in that, The multi-mode power supply module supports DC12V, POE and 5V solar power supply, adapting to various deployment scenarios on construction sites and solving the problem of inconvenient power supply at temporary construction sites.

11. A method for identifying construction site safety compliance based on edge-side risk chain determination, applied to the apparatus described in any one of claims 1 to 10, characterized in that, Includes the following steps: (1) The wake-up module is triggered by an event to achieve: low-power sleep when there are no personnel, and wake-up within 100ms after personnel are detected; (2) Acquire visible light and infrared dual-light images through the dual-light acquisition module, and perform adaptive feature fusion to suppress dust interference; (3) Generate a scene semantic mask through the construction site scene feature filtering module to remove background interference such as scaffolding and steel bars, and retain the target features of personnel and protective equipment; (4) Simultaneously detect personnel and protective equipment such as safety helmets, reflective vests, and safety belts on the edge side using a lightweight AI inference module; (5) Through the human body-equipment association module, the head, torso and waist areas are divided based on 17 key points of the human skeleton, and the area matching between the individual and the protective equipment is completed. (6) Through the edge risk chain judgment module, the three-level risk linkage judgment is completed and the corresponding risk level is output by comprehensively considering the wearing status of protective equipment, the attributes of the work area and the personnel movement trajectory; (7) Through the graded early warning module, corresponding graded sound and light alarms and information push operations are executed according to the risk level; (8) The alarm-related video clips and data are cached locally through the offline caching module. They can be stored for a long time when the network is offline and automatically synchronized when the network is connected. Unlike existing construction site safety identification methods that rely on cloud-based judgment and have a high false alarm rate, this method completes the entire process of identification and risk assessment locally at the edge, without the need for cloud interaction. Furthermore, it reduces the probability of mismatch in multi-person scenarios by accurately associating human body with equipment. Combined with an algorithm design optimized for construction site scenarios, it further improves the identification accuracy and scenario adaptability.

12. The method according to claim 11, characterized in that, The three-level risk assessment specifically involves combining the detection results of the lightweight AI inference module, the matching results of the human-equipment association module, and the real-time movement trajectory of personnel, and comparing them with the preset three-level risk rules to complete the risk level assessment. The entire assessment process is completed locally at the edge, without the need for cloud interaction.

Images