Construction site safety supervision unmanned aerial vehicle system based on AI identification and real-time shouting

Abstract

The embodiment of the application provides a construction site safety supervision unmanned aerial vehicle system based on AI identification and real-time shouting, which comprises: an unmanned aerial vehicle platform, wherein the unmanned aerial vehicle platform is provided with a multi-modal sensor, the multi-modal sensor comprises a visible light camera, a thermal imager and a laser radar; an edge computing module integrated with an identification model, wherein the identification model is trained by combining an improved YOLOv5 algorithm with an attention mechanism, the identification model is used for real-time analysis on collected sensor data, identifies non-wearing of protective equipment, intrusion in a dangerous area and abnormal state of equipment, and generates structured data of a hidden danger type, a position and a risk level based on an identification result; a communication module, which adopts 5G and WiFi6 double-link transmission, synchronizes the structured data and real-time video stream to a cloud management platform, and receives a task instruction issued by the platform; and a directional audio device, which is triggered to project a voice alarm to a hidden danger occurrence area in response to the task instruction.

Description

Technical Field

[0001] This invention belongs to the field of construction site safety supervision technology, and in particular relates to a construction site safety supervision drone system based on AI recognition and real-time announcement. Background Technology

[0002] In existing technologies, construction site safety monitoring drone systems typically use drones equipped with multimodal sensors such as visible light cameras and thermal imagers to collect data, and transmit the data to a cloud server via 4G / 5G networks. They then use target detection algorithms such as YOLO to identify safety hazards such as not wearing protective equipment or intrusion into dangerous areas.

[0003] Some systems integrate a voice alarm module, which broadcasts warning information to the construction site through ordinary loudspeakers.

[0004] However, existing technologies still have the following technical problems: 1. Insufficient real-time performance: Cloud processing results in high end-to-end inference latency, which cannot meet the needs of immediate early warning for sudden safety incidents at construction sites; 2. Weak environmental adaptability: Single sensors are prone to false detections and missed detections at night, in strong light, or in complex obstruction scenarios, and their multimodal data fusion and analysis capabilities are insufficient; 3. Poor targeting of early warnings: Ordinary loudspeaker broadcasts have a wide coverage area but weak directionality, which can easily interfere with personnel in non-hazardous areas. Summary of the Invention

[0005] The purpose of this invention is to provide a construction site safety monitoring drone system based on AI recognition and real-time announcements, so as to at least solve the problems of insufficient real-time performance, weak environmental adaptability and poor early warning targeting in the existing technologies.

[0006] To achieve the above objectives, the embodiments of this application provide the following technical solutions.

[0007] According to one embodiment of this application, a construction site safety monitoring drone system based on AI recognition and real-time announcement is provided; Includes the following modules: The drone platform is equipped with multimodal sensors, including a visible light camera, a thermal imager, and a lidar, which are used to collect images, temperature distribution, and three-dimensional point cloud data of the construction site scene, respectively. An edge computing module integrating an identification model, which is trained by combining an improved YOLOv5 algorithm with an attention mechanism, is used to analyze the collected sensor data in real time, identify unprotected personnel, intrusion into dangerous areas, and abnormal equipment status, and generate structured data on hazard type, location, and risk level based on the identification results. The communication module uses dual-link transmission of 5G and WiFi 6 to synchronize structured data and real-time video streams to the cloud management platform and receive task instructions issued by the platform. A directional audio device, in response to the task instruction, is triggered to project a voice alarm onto the area where the hazard occurs. The effective broadcast distance of the directional audio device in noisy environments is no less than 300 meters.

[0008] Preferably, the visible light camera is used to capture images of the construction site scene. The visible light camera is equipped with a wide dynamic range imaging module and a low-light enhancement unit to capture clear images in strong light, backlight and low light environments at night. The image acquisition parameters of the visible light camera are: resolution not less than 4K, frame rate not less than 30fps, and the lens has an optical zoom capability of more than 10x, which is used for adaptive focusing and shooting of key monitoring areas.

[0009] Preferably, the thermal imager is equipped with an infrared focal plane array detector, with a temperature measurement range covering -20℃ to 500℃ and a resolution of not less than 640×512. It marks high-temperature areas or low-temperature anomalies that exceed the preset temperature threshold and generates a temperature heat map. The thermal imager also integrates an ambient temperature compensation module to eliminate the impact of ambient temperature changes on temperature measurement accuracy.

[0010] Preferably, the lidar is equipped with a laser emission array of at least 16 lines, a point cloud density of not less than 200 points / m², a ranging range covering 0.5m to 200m, and a scanning frequency in the range of 10Hz-20Hz. The lidar distinguishes between static and dynamic targets through point cloud motion vector analysis and generates a three-dimensional point cloud model containing height information.

[0011] Preferably, the step of the lidar distinguishing between static and dynamic targets through point cloud motion vector analysis includes: N frames of 3D point cloud data are continuously collected within a preset time window. The motion error of the UAV itself is eliminated by the spatiotemporal registration algorithm that integrates IMU and GPS, and a point cloud sequence normalized to the same coordinate system is obtained. After voxel mesh downsampling of point clouds in adjacent frames, the improved Iterative Closest Point (ICP) algorithm is used to calculate the motion vector of each point cloud, which includes three-dimensional spatial displacement and velocity vectors; an adaptive dynamic threshold model is established based on the standard deviation of the motion vector of the point cloud, and the threshold is dynamically adjusted according to the complexity of the construction site environment. Multi-dimensional verification is performed by combining the reflection intensity and cluster volume characteristics of point clouds: when the motion vector of a point cloud cluster exceeds the corresponding threshold, and the reflection intensity conforms to the material characteristics of the human body or equipment, and the cluster volume is within the preset range, it is determined to be a dynamic target; otherwise, it is a static target. The motion trajectory of the dynamic target is predicted by Kalman filtering to generate a position estimate for a future time period, which is used to trigger the warning action of the directional audio device in advance.

[0012] Preferably, in the recognition model: The backbone network is connected in series with a hybrid attention module after the C3 module of YOLOv5. The hybrid attention module includes a channel attention branch and a spatial attention branch in sequence. The neck network adds a coordinate attention module between the lateral and vertical connections of the FPN-PAN structure to capture the positional information and channel dependencies of the target in the feature map, thereby optimizing the bounding box regression accuracy of dangerous areas. The output end adopts an improved detection head structure, and a lightweight self-attention submodule is added to each scale branch.

[0013] Preferably, the training phase of the recognition model includes the following steps: Constructing a multimodal fusion training dataset: Integrating visible light images, thermal imaging temperature maps, and LiDAR point cloud projection maps to form an associated sample library containing spatial coordinates, temperature attributes, and semantic labels. In the phased transfer training, the first phase uses ImageNet pre-trained weights to initialize the backbone network, freezes parameters other than the attention module, and performs feature adaptation training for 10 epochs; in the second phase, all parameters are unfrozen, and a construction site-specific dataset is used for 50 epochs of refined training, with the learning rate initially set to 0.01 and decayed to 1e-5 using a cosine annealing strategy. In dynamic loss optimization, the recognition accuracy of each hazard category is statistically analyzed in real time during training. For categories with a false negative rate >30%, their weight coefficient in Focal Loss is automatically increased. At the same time, dynamic CIoU is used to calculate the bounding box regression loss, and the position error penalty weight is increased according to the target scale.

[0014] Preferably, the training phase of the recognition model further includes an incremental model iteration step, specifically including: when accumulating a target number of newly collected construction site scene data, starting the incremental training mode, freezing the model's bottom feature extraction layer, updating only the attention module and detection head parameters, and retaining historical training knowledge through knowledge distillation.

[0015] Preferably, the step of generating structured data on hazard type, location, and risk level based on the identification results includes: Based on the target category output by the identification model, the types of hidden dangers are classified; The image coordinates output by the recognition model are converted into UAV body coordinates through camera calibration parameters. Then, combined with the three-dimensional spatial coordinates of the LiDAR point cloud and GPS positioning information, the location of the hidden danger in the global coordinate system of the construction site is calculated through the coordinate transformation matrix. A multi-factor risk matrix is ​​constructed, with the probability of occurrence of hidden dangers, the scope of impact, and the severity of consequences as input parameters, and a weighted summation algorithm is used to generate multi-level risk levels. Generate JSON format data containing a unique hazard ID, timestamp, category label, three-dimensional coordinates in the global coordinate system, risk level, index of original data from associated sensors, and target identification confidence level. When synchronizing to the cloud via the communication module, attach a data verification code.

[0016] Compared with existing technologies, the beneficial effects of the construction site safety supervision drone system based on AI recognition and real-time announcements in this application are: First, this application improves the accuracy of identifying safety hazards in complex construction site environments by coordinating multimodal data from visible light cameras, thermal imagers, and lidar, combined with an improved YOLOv5 algorithm and attention mechanism. Specifically, wide dynamic range imaging and low-light enhancement ensure all-weather image quality, while automatic temperature anomaly marking in thermal imaging and three-dimensional boundary delineation by lidar significantly reduce the false detection and false negative rates of single sensors. Furthermore, this application optimizes the edge computing architecture and model, controlling end-to-end inference latency to within 200ms, thus solving the real-time performance limitations of traditional cloud processing.

[0017] Secondly, the dual-link communication of 5G and WiFi 6 in this application ensures the stability of data transmission, and the accurate shouting of the directional audio device within a range of 300 meters avoids invalid interference. In addition, dynamic risk level assessment and structured data generation enable quantitative management and rapid response to hidden dangers, while the dynamic target separation and trajectory prediction technology of lidar can trigger the early warning mechanism in advance, significantly improving the initiative and effectiveness of construction site safety supervision. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] In the attached diagram: Figure 1 This is a structural block diagram of the construction site safety monitoring drone system based on AI recognition and real-time announcement, as described in this invention. Figure 2 This is a structural block diagram of the recognition model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the training process of the recognition model provided in an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0022] In this invention, unless otherwise stated, directional terms such as "upper," "lower," "top," and "bottom" are generally used in relation to the direction shown in the accompanying drawings, or in relation to the vertical, perpendicular, or gravitational direction of the component itself; similarly, for ease of understanding and description, "inner" and "outer" refer to the inner and outer contours of each component itself, but the above directional terms are not intended to limit this invention.

[0023] like Figure 1 The diagram shown is a structural block diagram of a construction site safety monitoring drone system based on AI recognition and real-time announcement, provided in an embodiment of the present invention. like Figure 1 As shown in one embodiment of this application, the construction site safety monitoring drone system based on AI recognition and real-time announcement includes the following modules: The drone platform 101 is equipped with multimodal sensors, including a visible light camera, a thermal imager, and a lidar, which are used to collect images, temperature distribution, and three-dimensional point cloud data of the construction site scene, respectively. The edge computing module 102 integrates the recognition model, which is trained by combining the improved YOLOv5 algorithm with the attention mechanism. The recognition model is used to analyze the collected sensor data in real time, identify the lack of protective equipment, intrusion into dangerous areas and abnormal equipment status, and generate structured data on the type, location and risk level of the hazard based on the recognition results. The communication module 103 uses 5G and WiFi 6 dual-link transmission to synchronize structured data and real-time video streams to the cloud management platform and receive task instructions issued by the platform. The directional audio device 104, in response to the task instruction, is triggered to project a voice alarm to the area where the hazard occurs. The effective broadcast distance of the directional audio device in noisy environments is not less than 300 meters.

[0024] This application's dual-link 5G and WiFi 6 communication ensures the stability of data transmission, and the directional audio device's accurate shouting within a 300-meter range avoids unnecessary interference.

[0025] Furthermore, in some embodiments of this application, the visible light camera is used to capture images of the construction site scene. The visible light camera is equipped with a wide dynamic range imaging module and a low-light enhancement unit to capture clear images in strong light, backlight and low light environments at night. The image acquisition parameters of the visible light camera are: resolution not less than 4K, frame rate not less than 30fps, and the lens has an optical zoom capability of more than 10x, which is used for adaptive focusing and shooting of key monitoring areas.

[0026] The thermal imager is equipped with an infrared focal plane array detector, with a temperature measurement range covering -20℃ to 500℃ and a resolution of not less than 640×512. It marks high-temperature areas or low-temperature anomalies that exceed the preset temperature threshold and generates a temperature heat map. The thermal imager also integrates an ambient temperature compensation module to eliminate the impact of ambient temperature changes on temperature measurement accuracy.

[0027] The lidar is equipped with a laser emission array of at least 16 lines, a point cloud density of not less than 200 points / ㎡, a ranging range covering 0.5m to 200m, and a scanning frequency in the range of 10Hz-20Hz. The lidar distinguishes between static and dynamic targets through point cloud motion vector analysis and generates a three-dimensional point cloud model containing height information.

[0028] Preferably, the step of the lidar distinguishing between static and dynamic targets through point cloud motion vector analysis includes: N frames of 3D point cloud data are continuously collected within a preset time window. The motion error of the UAV itself is eliminated by the spatiotemporal registration algorithm that integrates IMU and GPS, and a point cloud sequence normalized to the same coordinate system is obtained. After voxel mesh downsampling of point clouds in adjacent frames, the improved Iterative Closest Point (ICP) algorithm is used to calculate the motion vector of each point cloud, which includes three-dimensional spatial displacement and velocity vectors; an adaptive dynamic threshold model is established based on the standard deviation of the motion vector of the point cloud, and the threshold is dynamically adjusted according to the complexity of the construction site environment. Multi-dimensional verification is performed by combining the reflection intensity and cluster volume characteristics of point clouds: when the motion vector of a point cloud cluster exceeds the corresponding threshold, and the reflection intensity conforms to the material characteristics of the human body or equipment, and the cluster volume is within the preset range, it is determined to be a dynamic target; otherwise, it is a static target. The motion trajectory of the dynamic target is predicted by Kalman filtering to generate a position estimate for a future time period, which is used to trigger the warning action of the directional audio device in advance.

[0029] This application improves the accuracy of safety hazard identification in complex construction site environments by coordinating multimodal data from visible light cameras, thermal imagers, and LiDAR, combined with an improved YOLOv5 algorithm and attention mechanism. Wide dynamic range imaging and low-light enhancement ensure all-weather image quality, while automatic temperature anomaly marking in thermal imaging and 3D boundary delineation by LiDAR significantly reduce the false positive and false negative rates of single sensors. Furthermore, this application optimizes the edge computing architecture and model, controlling end-to-end inference latency to within 200ms, thus solving the real-time performance limitations of traditional cloud processing.

[0030] like Figure 2 As shown, in one implementation of this embodiment, the recognition model includes: an input terminal 201, a backbone network 202, a neck network 203, and an output terminal 204; wherein: Input terminal 201 uses Mosaic data augmentation and adaptive anchor box calculation to dynamically adjust for construction site scene samples, thereby improving the feature recognition of small targets such as safety helmets and safety belts; The backbone network 202 provided in this embodiment is connected in series with a hybrid attention module after the C3 module of YOLOv5. The hybrid attention module includes a channel attention branch and a spatial attention branch in sequence. The spatial attention branch generates a spatial weight map through convolutional layers and Sigmoid activation. In this embodiment, the channel attention branch and the spatial attention branch are used to enhance the features of workers' limbs, key parts of equipment, etc. The neck network 203 adds a coordinate attention module between the lateral and vertical connections of the FPN-PAN structure to capture the positional information and channel dependencies of the target in the feature map, thereby optimizing the bounding box regression accuracy of dangerous areas; for example, dangerous areas can be the edge of a deep foundation pit, the rotation radius of a tower crane, etc. The output terminal 204 adopts an improved detection head structure, adding a lightweight self-attention submodule on each scale branch. By calculating the correlation of feature map pixels, it improves the target separation capability in occluded scenarios such as workers behind scaffolding or equipment next to stacked materials.

[0031] like Figure 3 As shown in the embodiments of this application, the training phase of the recognition model includes the following steps: Step S301: Construct a multimodal fusion training dataset: Integrate visible light images, thermal imaging temperature maps, and lidar point cloud projection maps to form an associated sample library containing spatial coordinates, temperature attributes, and semantic labels; Step S302: Phased transfer training: In the first phase, the backbone network is initialized with ImageNet pre-trained weights, and parameters other than the attention module are frozen for 10 epochs of feature adaptation training; in the second phase, all parameters are unfrozen, and 50 epochs of fine-tuning training are performed using a construction site-specific dataset, with the learning rate initially set to 0.01 and decayed to 1e-5 using a cosine annealing strategy. Step S303: Dynamic Loss Optimization: During training, the recognition accuracy of each hazard category is statistically analyzed in real time. For categories with a false negative rate > 30%, the weight coefficient in Focal Loss is increased. At the same time, dynamic CIoU is used to calculate the bounding box regression loss, and the position error penalty weight is increased according to the target scale.

[0032] Furthermore, in the embodiments of this application, the training phase of the recognition model also includes an incremental model iteration step, specifically including: when accumulating a target number of newly collected construction site scene data, starting the incremental training mode, freezing the model's bottom feature extraction layer, updating only the attention module and detection head parameters, and retaining historical training knowledge through knowledge distillation.

[0033] Preferably, the step of generating structured data on hazard type, location, and risk level based on the identification results includes: Based on the target category output by the recognition model, the types of hidden dangers are classified; the image coordinates output by the recognition model are converted into UAV body coordinates through camera calibration parameters, and then combined with the three-dimensional spatial coordinates of the LiDAR point cloud and GPS positioning information, the location of the hidden danger in the global coordinate system of the construction site is calculated through the coordinate transformation matrix; a multi-factor risk matrix is ​​constructed, using the probability of occurrence of hidden danger, the scope of impact, and the severity of consequences as input parameters, and a weighted summation algorithm is used to generate multi-level risk levels; JSON format data containing a unique hidden danger ID, timestamp, category label, three-dimensional coordinates of the global coordinate system, risk level, original data index of associated sensors, and target recognition confidence is generated, and a data verification code is attached when synchronizing to the cloud through the communication module.

[0034] The dynamic risk level assessment and structured data generation in this application embodiment enable quantitative management and rapid response to potential hazards. Furthermore, the dynamic target separation and trajectory prediction technology of lidar can trigger early warning mechanisms in advance, significantly improving the initiative and effectiveness of construction site safety supervision.

[0035] The above solutions are merely illustrative examples of preferred embodiments and are not intended to limit the scope of the invention. Appropriate substitutions and / or modifications can be made according to user needs when implementing this invention.

[0036] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0037] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A construction site safety monitoring drone system based on AI recognition and real-time announcements, characterized in that: Includes the following modules: The drone platform is equipped with multimodal sensors, including a visible light camera, a thermal imager, and a lidar, which are used to collect images, temperature distribution, and three-dimensional point cloud data of the construction site scene, respectively. An edge computing module integrating an identification model, which is trained by combining an improved YOLOv5 algorithm with an attention mechanism, is used to analyze the collected sensor data in real time, identify unprotected personnel, intrusion into dangerous areas, and abnormal equipment status, and generate structured data on hazard type, location, and risk level based on the identification results. The communication module uses dual-link transmission of 5G and WiFi 6 to synchronize structured data and real-time video streams to the cloud management platform and receive task instructions issued by the platform. A directional audio device, in response to the task instruction, is triggered to project a voice alarm onto the area where the hazard occurs. The effective broadcast distance of the directional audio device in noisy environments is no less than 300 meters.

2. The construction site safety supervision drone system based on AI recognition and real-time announcement as described in claim 1, characterized in that, The visible light camera is used to capture images of the construction site scene. The visible light camera is equipped with a wide dynamic range imaging module and a low light enhancement unit to capture clear images in strong light, backlight and low light environments at night. The image acquisition parameters of the visible light camera are: resolution not less than 4K, frame rate not less than 30fps, used for adaptive focusing shooting of key monitoring areas.

3. The construction site safety supervision drone system based on AI recognition and real-time announcement as described in claim 2, characterized in that, The thermal imager is equipped with an infrared focal plane array detector, with a temperature measurement range covering -20℃ to 500℃ and a resolution of not less than 640×512. It marks high-temperature areas or low-temperature anomalies that exceed the preset temperature threshold and generates a temperature heat map. The thermal imager also integrates an ambient temperature compensation module to eliminate the impact of ambient temperature changes on temperature measurement accuracy.

4. The construction site safety supervision drone system based on AI recognition and real-time announcement as described in claim 3, characterized in that, The lidar is equipped with a laser emission array of at least 16 lines, a point cloud density of not less than 200 points / ㎡, a ranging range covering 0.5m to 200m, and a scanning frequency in the range of 10Hz-20Hz. The lidar distinguishes between static and dynamic targets through point cloud motion vector analysis and generates a three-dimensional point cloud model containing height information.

5. The construction site safety supervision drone system based on AI recognition and real-time announcement as described in claim 4, characterized in that, The steps involved in using point cloud motion vector analysis to distinguish between static and dynamic targets by lidar include: N frames of 3D point cloud data are continuously collected within a preset time window. The motion error of the UAV itself is eliminated by the spatiotemporal registration algorithm that integrates IMU and GPS, and a point cloud sequence normalized to the same coordinate system is obtained. After voxel mesh downsampling of point clouds in adjacent frames, the improved Iterative Closest Point (ICP) algorithm is used to calculate the motion vector of each point cloud, which includes three-dimensional spatial displacement and velocity vectors; an adaptive dynamic threshold model is established based on the standard deviation of the motion vector of the point cloud, and the threshold is dynamically adjusted according to the complexity of the construction site environment. Multi-dimensional verification is performed by combining the reflection intensity and cluster volume characteristics of point clouds: when the motion vector of a point cloud cluster exceeds the corresponding threshold, and the reflection intensity conforms to the material characteristics of the human body or equipment, and the cluster volume is within the preset range, it is determined to be a dynamic target; otherwise, it is a static target. The motion trajectory of the dynamic target is predicted by Kalman filtering to generate a position estimate for a future time period, which is used to trigger the warning action of the directional audio device in advance.

6. The construction site safety supervision drone system based on AI recognition and real-time announcement as described in claim 5, characterized in that, In the recognition model: The backbone network is connected in series with a hybrid attention module after the C3 module of YOLOv5. The hybrid attention module includes a channel attention branch and a spatial attention branch in sequence. The neck network adds a coordinate attention module between the lateral and vertical connections of the FPN-PAN structure to capture the positional information and channel dependencies of the target in the feature map, thereby optimizing the bounding box regression accuracy of dangerous areas. The output end adopts an improved detection head structure, and a lightweight self-attention submodule is added to each scale branch.

7. The construction site safety supervision drone system based on AI recognition and real-time announcement as described in claim 6, characterized in that, The training phase of the recognition model includes the following steps: Construct a multimodal fusion training dataset: integrate visible light images, thermal imaging temperature maps, and lidar point cloud projection maps to form an associated sample library containing spatial coordinates, temperature attributes, and semantic labels; In the phased transfer training, the first phase uses ImageNet pre-trained weights to initialize the backbone network, freezes parameters other than the attention module, and performs feature adaptation training for 10 epochs; the second phase unfreezes all parameters and uses the construction site dataset for 50 epochs of fine-tuning training, with the learning rate initially set to 0.01 and decayed to 1e-5 using a cosine annealing strategy. In dynamic loss optimization, the recognition accuracy of each hazard category is statistically analyzed in real time during training. For categories with a false negative rate >30%, their weight coefficient in Focal Loss is automatically increased. At the same time, dynamic CIoU is used to calculate the bounding box regression loss, and the position error penalty weight is increased according to the target scale.

8. The construction site safety supervision drone system based on AI recognition and real-time announcement as described in claim 7, characterized in that, The training phase of the recognition model also includes an incremental model iteration step, specifically including: When a target number of new construction site scene data are collected, the incremental training mode is activated, the underlying feature extraction layer of the model is frozen, and only the attention module and detection head parameters are updated. Historical training knowledge is retained through knowledge distillation.

9. The construction site safety supervision drone system based on AI recognition and real-time announcement as described in claim 8, characterized in that, The steps for generating structured data on hazard type, location, and risk level based on the identification results include: Based on the target category output by the identification model, the types of hidden dangers are classified; The image coordinates output by the recognition model are converted into UAV body coordinates through camera calibration parameters. Then, combined with the three-dimensional spatial coordinates of the LiDAR point cloud and GPS positioning information, the location of the hidden danger in the global coordinate system of the construction site is calculated through the coordinate transformation matrix. A multi-factor risk matrix is ​​constructed, with the probability of occurrence of hidden dangers, the scope of impact, and the severity of consequences as input parameters, and a weighted summation algorithm is used to generate multi-level risk levels. Generate JSON format data containing a unique hazard ID, timestamp, category label, three-dimensional coordinates in the global coordinate system, risk level, index of original data from associated sensors, and target identification confidence level. When synchronizing to the cloud via the communication module, attach a data verification code.

Images