Construction site real-time monitoring system and method based on unmanned aerial vehicle technology

Abstract

The invention relates to the technical field of building construction monitoring, in particular to a construction site real-time monitoring system and method based on an unmanned aerial vehicle technology. The method specifically comprises the steps that an unmanned aerial vehicle carries a high-definition camera, an infrared thermal imager and a laser radar and collects multi-modal data of a construction site in real time; a deep learning algorithm is adopted to carry out multi-modal semantic segmentation, RGB images, infrared thermal imaging and laser point cloud projection data are fused, YOLOv8 target detection and time sequence behavior modeling technologies are combined, and abnormal risks are dynamically identified; by constructing a context-aware comprehensive risk scoring model, the risk level of a construction scene is quantified, and a graded early warning mechanism is triggered; and meanwhile, based on a three-dimensional modeling technology, tank vertex cloud data is analyzed by utilizing an LIO-SLAM algorithm, and the inclination degree of the structure is detected and is linked with emergency response. According to the invention, through multi-sensor fusion, dynamic risk assessment and closed-loop control, intelligentization and precision of construction monitoring are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of construction site monitoring, and particularly to a real-time monitoring system and method for construction sites based on unmanned aerial vehicle (UAV) technology. Background Art

[0002] Safety monitoring of construction sites is a core requirement in construction project management. Traditional manual inspections and single-sensor monitoring have bottlenecks such as limited coverage, poor real-time performance, and static risk assessment; existing technologies mostly rely on fixed cameras or UAVs, and it is difficult to adapt to the dynamic changes of construction scenarios. Especially in night operations, high-altitude areas, or complex equipment layouts, monitoring blind spots are likely to occur, and the risk warning mechanism usually based on preset thresholds lacks comprehensive dynamic analysis of personnel behavior, equipment status, and environmental factors.

[0003] Most existing image processing technologies use single-modal data (such as visible light or infrared), resulting in insufficient target segmentation accuracy and difficulty in accurately identifying violations of construction personnel (such as not wearing safety helmets or seat belts), equipment abnormalities (such as hoisting offsets), and structural hidden dangers (such as the tilt of the tank top blowing up).

[0004] In addition, with the development of UAV platforms and artificial intelligence technologies, there is an urgent need to build a full-link monitoring system covering data collection, intelligent analysis, and active intervention to address the deficiencies of existing technologies in multi-modal fusion, behavioral dynamic modeling, and risk quantification decision-making. Summary of the Invention

[0005] The object of the present invention is to propose a real-time monitoring system and method for construction sites based on UAV technology in view of the problems in the background art.

[0006] The technical solution of the present invention: A real-time monitoring system for construction sites based on UAV technology, comprising: A UAV measurement module, configured with a UAV equipped with a high-definition camera, an infrared thermal imager, and a lidar, for real-time collecting RGB images, infrared thermal imaging data, and lidar point cloud data of the construction site; A cloud platform management module, including a data storage and management unit and a data analysis and processing unit, for storing the collected data and performing semantic segmentation, target detection, and dynamic risk assessment on multi-modal data through deep learning algorithms; The data storage and management unit: for storing the image data of the construction site collected by the UAV; The data analysis and processing unit: performing real-time analysis on the image data of the construction site; An alarm and feedback module: triggering a hierarchical response mechanism according to the risk assessment result.

[0007] Preferably, the data analysis and processing unit fuses RGB images, infrared images and laser point cloud projection maps, uses the DeepLabv3+ network for pixel-level semantic segmentation, and optimizes the segmentation accuracy of high-risk areas through a weighted cross-entropy loss function; identifies and detects targets based on the YOLOv8 network, and combines the DeepSORT algorithm to achieve continuous target tracking; and extracts dynamic behavior features through temporal behavior modeling to output a risk probability value; combines the risk probability, environmental context weight and preset threshold to generate a four-level warning level.

[0008] Preferably, the weighted cross-entropy loss function is: ; where L seg is the loss function; y i is the true class label of pixel i; is the predicted probability output by the network; w c (i) is the class weight of the i-th class pixel.

[0009] Preferably, the temporal behavior modeling includes a TSM network and a Transformer encoder, which fuse target trajectory, pose and scene location information to output a risk probability.

[0010] The technical solution of the present invention: A real-time monitoring method for a construction site based on UAV technology, which is applied to the above-mentioned real-time monitoring system for a construction site based on UAV technology, includes the following specific implementation steps: S51. Initialize the UAV flight path and task parameters, and dynamically match the construction site range and environmental factors; S52. The UAV collects multi-modal data and transmits it; S53. Perform adaptive histogram equalization and non-local mean denoising on the RGB image, fuse the infrared image and the point cloud projection map to generate a 6-channel fusion feature; perform semantic segmentation based on the DeepLabv3+ network to output a segmentation map; combine YOLOv8 and DeepSORT to achieve target detection and tracking, and output the detection result; output a risk probability through temporal behavior modeling, combine the environmental context to generate a comprehensive risk score; trigger a four-level warning according to the comprehensive risk score, and output the response level; at the same time, analyze the inclination degree of the storage tank structure based on the lidar three-dimensional modeling and point cloud analysis method; S54. Start the feedback mechanism according to the response level; S55. Store the current multi-modal data, segmentation map, detection result, comprehensive risk score, response level and corresponding feedback measures.

[0011] Preferably, the comprehensive risk score: ; where is the comprehensive risk score at time t; K is the total number of risk behavior types; β k is the basic weight of the k-th type of risk; is the risk probability of the k-th behavior at time t; is the context weight adjusted according to the current operating environment.

[0012] Preferably, the implementation steps of the method based on lidar three-dimensional modeling and point cloud analysis are as follows: S71. Instruct the drone to obtain the point cloud data and odometer data of the tank wall and tank top in the drone coordinate system; S72. Fusion the point cloud data and odometer data through the LIO-SLAM algorithm, screen out the three-dimensional coordinates of the point cloud at the junction of the tank top and the tank wall in the world coordinate system, and calculate the inclination degree of the tank top: Adopt the LIO-SLAM algorithm for fusion and pose optimization; Select frames with a relatively long time interval and a relatively close distance in the historical key frames as candidate loop closure frames, extract the set of feature points of the current frame, downsample, perform scantomap optimization, and use the iterative closest point algorithm to obtain the optimized pose; Downsample the point cloud image to determine the junction position of the tank top and the tank wall: Fill the point cloud and remove the interference of the tank wall point cloud data to obtain a point cloud image containing only the tank top data; Extract the point cloud on the circumference of the tank top edge, based on the maximum value z of the height of the edge point cloud max and the minimum value z min , obtain the inclination degree ΔH of the tank top = z max -z min ; Output the inclination degree ΔH of the tank top; S73. If ΔH ≥ the set allowable maximum tank top inclination height difference threshold ΔH max , trigger an audible and visual alarm and adjust the blowing speed and intake air volume.

[0013] Preferably, the implementation process of adopting the LIO-SLAM algorithm for fusion and pose optimization is as follows: S81. Calculate the IMU pre-integration: Calculate the velocity, position, and rotation change amount between adjacent frames: ; ; ; where, Δt ij represents the time interval between time t i and t j ; is the estimated value of the velocity; is the estimated value of the position; v iand v j are the velocities at times t i and t j respectively; g is the acceleration due to gravity; is the estimated value of the rotation matrix; R i and R j are the rotation matrices at times t i and t j respectively; S82. Feature matching: By using the distance relationships between points and lines and between points and planes, pose solution equations are established: ; ; ; where k, u, v, w are the corresponding feature possessions; is the 3D coordinate of the k-th line feature point in the (i + 1)-th frame; and are the coordinates of the two endpoints of the same line feature in the i-th frame; d ek represents the perpendicular distance from the point at the current moment to the line feature at the previous moment; d pk is the perpendicular distance from the point at the current moment to the plane at the previous moment; is the 3D coordinate of the k-th plane feature point in the (i + 1)-th frame; , and are three non-collinear points defining the plane in the i-th frame; e represents the line feature; P represents the plane feature; ΔT i,i+1 is the relative pose between the current key frame and the previous key frame; T i and T i+1 are the poses of the frames at times i and i + 1; S83. Taking d ek and d pk to form an objective function, and solving for the optimal ΔT i,i+1 by minimizing the error to perform pose optimization.

[0014] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects: The present invention designs a real-time monitoring system and method for construction sites based on UAV technology. By equipping UAVs with multi-modal sensors (high-definition cameras, infrared thermal imagers, lidar), all-weather three-dimensional data collection of construction scenes is realized. Combining the DeepLabv3+ multi-scale semantic segmentation model and the YOLOv8 object detection algorithm significantly improves the recognition accuracy of personnel, equipment, and dangerous areas in images. Using temporal behavior modeling (TSM + attention network) and DeepSORT object tracking technology, dynamic analysis of construction workers' illegal operations (such as not wearing safety helmets) and equipment abnormal states (such as hoisting offset) is carried out, and a context-aware dynamic risk assessment model is constructed through multi-dimensional feature fusion to achieve quantitative scoring of risk probabilities and four-level early warning classification. Introducing lidar point cloud data and the LIO-SLAM algorithm, accurately calculating the inclination height difference of the tank top, and combining a closed-loop response mechanism (such as automatically pausing operations and manual reset) effectively prevent structure collapse accidents. In addition, the system reduces the cost of manual inspections through automatic flight path planning, real-time fusion of multi-source data, and hierarchical alarm pushing, improves the real-time and intelligent level of construction safety management, and provides a high-precision and full-process active safety monitoring solution for complex construction scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 FIG. is a system architecture diagram of a real-time monitoring system for construction sites based on UAV technology proposed by the present invention; Figure 2 FIG. is a schematic diagram of a point cloud image containing only tank top data. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Embodiment 1, as Figure 1 shown, a real-time monitoring system for construction sites based on UAV technology proposed by the present invention includes: a UAV measurement module, a cloud platform management module, and an alarm and feedback module.

[0017] UAV measurement module: A UAV equipped with a high-definition camera, an infrared thermal imager, and lidar is used to collect image data of the construction site; Cloud platform management module: Used to store and analyze the collected image data of the construction site, including: Data storage and management unit: Used to store the image data of the construction site collected by the UAV to ensure the security and traceability of the data; Data analysis and processing unit: Using big data analysis and machine learning technologies, real-time analysis of the image data of the construction site is carried out, including but not limited to detecting potential safety hazards, personnel positioning, and equipment status; Alarm and feedback module: When detecting potential safety risks (including but not limited to illegal operations and personnel crossing boundaries), an alarm is automatically issued, and the real-time situation is fed back to the construction management personnel through mobile terminal devices.

[0018] Embodiment 2. A real-time monitoring method for a construction site based on UAV technology proposed by the present invention is applied to a real-time monitoring system for a construction site based on UAV technology proposed in Embodiment 1. The specific implementation steps are as follows: S1. System initialization stage: To ensure that the UAV can effectively cover the construction site and complete the monitoring task, initialization is first performed, including but not limited to checking of hardware devices, setting of flight environment, and clarification of monitoring targets. Specifically: The user inputs the basic information of the construction site and monitoring requirements (including but not limited to area range, monitoring frequency, specific targets) through a mobile terminal device; According to the geographical location and scale of the construction site, a flight path is automatically generated and dynamically adjusted considering environmental factors (including but not limited to wind speed, weather conditions); Set the flight mode of the UAV (including but not limited to cruising, fixed-point staying, etc.) and task parameters (including but not limited to image acquisition).

[0019] S2. The UAV in the UAV measurement module executes the task according to the predetermined flight task and starts on-site monitoring: After the UAV takes off, it automatically flies according to the preset path, enters the monitoring area, and starts to synchronously collect real-time data of the construction site based on the equipped high-definition camera, infrared thermal imager, and lidar device; Transmit the construction site data to the cloud platform management module.

[0020] S3. The data analysis and processing unit uses intelligent algorithms to perform real-time processing and analysis on the construction site data collected from the UAV, constructs a full-process analysis mechanism from image data preprocessing to intelligent recognition and risk assessment, and realizes efficient intelligent monitoring and dynamic risk perception of the construction site through deep learning and multi-dimensional information fusion, and identifies abnormal situations, including but not limited to potential safety hazards, equipment failures, and personnel violations. The specific implementation steps are as follows: S31. Image data preprocessing and multi-modal semantic segmentation: Since the image data captured by the UAV has characteristics such as large perspective changes, large differences in target sizes, and a lot of noise, it is necessary to first perform standardized preprocessing on the original image, and then introduce a multi-modal semantic segmentation model to fuse RGB images, infrared thermal images, and point cloud projection maps to accurately segment the core targets (including but not limited to personnel, equipment, warning lines, dangerous areas) within the construction area. Specifically: S3101. Process the RGB image channels using Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance local contrast, and introduce the Non-Local Means (NLM) denoising algorithm to reduce noise while preserving edges. Also, use the attitude parameters of the drone sensor to match the SIFT feature points of the image, and combine the inverse projection model to restore the distorted area: ; where (x, y) represents the original coordinates; (x corr , y corr ) represents the corrected coordinates; k1 and k2 represent the radial distortion coefficients, which are obtained from the calibration of the drone's internal parameters; S3102. Fuse the RGB image, infrared image, and laser point cloud projection map to enhance the expression of depth and thermal information in the image and improve the segmentation accuracy. Specifically: (1) Match the timestamps of the RGB, infrared, and point cloud data, and project the point cloud into a 2D grayscale depth map; (2) Use the image key point detection based on the SURF algorithm to calculate the homography matrix H and achieve image spatial alignment: pʹ = H·p; where p is the pixel point coordinate in the original image; pʹ is the registered coordinate; H is a 3×3 homography matrix; (3) Perform spatial registration and channel-level fusion on the RGB image, infrared image, and depth image to construct a multi-modal feature tensor (i.e., a 6-channel fused image) containing spatial, thermal, and geometric structure information; S3103. Adopt the DeepLabv3+ network structure, and combine the Atrous Spatial Pyramid Pooling (ASPP) module of the DeepLabv3+ network to capture local and global structures. The input is the 6-channel fused image, and the output is a pixel-level semantic segmentation map; Use the improved weighted cross-entropy as the loss function: ; where L seg is the loss function; y i is the true class label of pixel i; is the predicted probability output by the network; w c (i) is the class weight of the i-th class pixel, which is set according to experience; For example: the weight of the high-risk class (under the hoisting equipment): >1.5; the weight of the medium-risk class (stacked objects, machinery): ≈1; the weight of the low-risk class (open space): <0.8; S3104. Output the segmented image; S32. Based on the image regions after semantic segmentation, the system uses YOLOv8 for multi-object detection to accurately identify objects such as construction workers, safety helmets, equipment, and signs. To achieve intelligent recognition of dynamic behaviors, a temporal behavior modeling module (TSM (Temporal Shift Module) + attention fusion network) is introduced to extract action features from consecutive frames. Specifically: S3201. Use the object detection network YOLOv8 (You Only Look Once version 8) as the main detector. Input the image after semantic segmentation output into the YOLOv8 network for object recognition and classification, and output the position, category, and confidence: Any detection box B i =[x i ,y i ,w i ,h i ,c i ,s i ; Among them, (x i ,y i ), w i , h i , c i , s i are the center point coordinates, width and height, category, confidence, and score of the bounding box respectively; S3202. Track the detected objects in the video frame sequence. Use the DeepSORT algorithm, combine appearance features and Kalman filtering to complete the numbering and continuous tracking of the objects, and generate a trajectory sequence for each object; S3203. Build a temporal modeling network: TSM (Temporal Shift Module) + Transformer Encoder. Output the key frame images, trajectories, and pose information of the objects over a period of time to the temporal modeling network and output the recognition results; Accordingly: Introduce multi-dimensional behavior feature fusion: image + motion trajectory + scene position information, bind behaviors to semantic regions, and improve the recognition confidence (for example, the behavior of "not wearing a safety helmet" must be established within the "personnel area"); S3204. Convert the recognition results into quantitative probability values to provide input for subsequent risk assessment: ; Among them, is the risk probability of the k-th behavior at time t; are j behavior influence features (including but not limited to: construction workers not wearing safety helmets, abnormal state of lifting equipment); is the weight coefficient; σ is the Sigmoid activation function, and the output value is normalized to [0,1]; S33. Quantify the identified risk behaviors in multiple dimensions and construct a comprehensive risk scoring model for the construction scenario to reflect the current risk level and potential problems at the operation site in a quantitative form, specifically as follows: S3301. Construct a risk factor database: Define various risk events, including but not limited to "abnormal attitude of lifting equipment" or "obstacles piled up in the construction area"; S3302. Extract the risk intensity corresponding to each type of behavior and weight it based on its behavior probability and environmental context; S3303. Output the overall risk score: ; Among them, is the comprehensive risk score at time t; K is the total number of types of risk behaviors; β k is the basic weight of the k-th type of risk; is the risk probability of the k-th behavior at time t; is the context weight adjusted according to the current operation environment (including but not limited to "night operation", "near the edge", "confined space"); Accordingly: Jointly model the static image information and the dynamic behavior inference result to achieve dynamic risk assessment under context awareness, convert the probability output into a globally available quantitative result, which is the basis for system decision-making and response; S34. According to the risk scoring result, automatically activate the hierarchical response mechanism, specifically as follows: Preset a threshold to map the comprehensive risk score to a four-level response level Level(t): ; S35. Package and upload the current risk scenario image, segmentation map, detection result, score value, and response level to the data storage and management unit, and send the response level to the alarm and feedback module.

[0021] S4. The alarm and feedback module activates the feedback mechanism according to the response level Level(t), including but not limited to: local voice broadcast (drone shouting); push notifications to mobile terminal devices (construction administrators); early warning synchronization on the supervision platform.

[0022] S5. The data storage and management unit uses the built-in database to store the current risk scenario image, segmentation map, detection result, score value, response level, and corresponding feedback measures.

[0023] Embodiment 3. A real-time monitoring method for a construction site based on drone technology proposed by the present invention further includes a lidar three-dimensional modeling and point cloud analysis method, and its specific implementation steps are as follows: A1. The drone measurement module commands the drone to carry a lidar, hover at a height of 15 m, with a scanning frequency of 10 Hz, a vertical viewing angle covering -30° to +10°, and a point cloud density ≥ 10 points / m², ensuring that the edge of the large storage tank roof is clearly visible; Based on this: Taking the position of the drone as the origin, the direction the drone is facing as the x-axis, the transverse axis of the drone as the y-axis, and the vertical axis direction as the z-axis to establish a coordinate system, obtaining the point cloud data and odometer data of the tank wall and tank roof in the drone coordinate system, and transmitting the point cloud data and odometer data to the cloud platform management module.

[0024] A2. The data analysis and processing unit fuses the point cloud data and odometer data through the LIO-SLAM algorithm, and then the positions of these point cloud data in the world coordinate system can be obtained. By screening out the three-dimensional coordinates of the point cloud at the junction of the tank roof and the tank wall in the world coordinate system, the inclination degree of the tank roof can be judged through calculation. Specifically: A21. Using the LIO-SLAM algorithm for fusion and pose optimization, specifically: (1) Calculate the IMU pre-integration: Calculate the velocity, position, and rotation change amount between adjacent frames: ; ; ; Among them, Δt ij represents the time interval between time t i and t j ; is the estimated value of velocity; is the estimated value of position; v i and v j are the velocities at time t i and t j respectively; g is the gravitational acceleration; is the estimated value of the rotation matrix; R i and R j are the rotation matrices at time t i and t j respectively; (2) Feature matching: Using the distance relationship between points and lines and points and planes to establish a pose solution equation: ; ; ; Among them, k, u, v, w are the corresponding features all; is the 3D coordinate of the k-th line feature point in the (i + 1)-th frame (current frame); and are the coordinates of the two endpoints of the same line feature in the i-th frame (previous frame); d ek represents the current moment point is the perpendicular distance from the current moment point to the line feature at the previous moment; d pk is the current moment point is the perpendicular distance from the current moment point to the plane at the previous moment; are the 3D coordinates of the k-th plane feature point in the (i + 1)-th frame; 、 and are three non-collinear points defining the plane in the i-th frame; e represents the line feature; P represents the plane feature; ΔT i,i+1 is the relative pose between the current key frame and the previous key frame; T i and T i+1 are the poses of the frames at the i-th moment and the (i + 1)-th moment; Accordingly: Combine d ek and d pk to form an objective function, and solve for the optimal ΔT by minimizing the error i,i+1 ; It should be noted that LIO-SLAM is a type of 3D lidar SLAM algorithm. The full name of SLAM is Simultaneous Localization and Mapping, which is a technology covering simultaneous mapping and localization. LIO-SAM is a tightly coupled lidar inertial odometry framework based on factor graph optimization; A22. Adopt the method of matching the lidar scan frame with the map, select frames with a relatively large time interval and a relatively close distance in the historical key frames as candidate loop closure frames, extract the set of feature points of the current frame, downsample, perform scantomap optimization, and use the iterative closest point algorithm to obtain the optimized pose; A23. Downsample the point cloud image to reduce the computational processing amount while retaining the point cloud features; determine the intersection position between the tank top and the tank wall. Since the intersection position is relatively far from the lidar and the point cloud is sparse, it is necessary to fill the point cloud; then remove the interference of the tank wall point cloud data to obtain a point cloud image containing only the tank top data (as Figure 2 shown); A24. Based on the tank top point cloud, extract the point cloud on the circumference of the tank top edge. Since the number of sampling points on the edge is sufficient, the maximum value z max of the edge point cloud height can be used to subtract the minimum value z min to obtain the inclination degree ΔH of the tank top = z max -z min ; A25. Transmit the inclination degree ΔH of the tank top to the alarm and feedback module.

[0025] A3. The alarm and feedback module receives the tank top inclination height difference ΔH. If ΔH ≥ ΔHmax (The set maximum allowable threshold of the height difference of the tank top inclination, which meets the requirements of the tank top specification), triggers an audible and visual alarm and transmits the tank top balance data in real time, adjusts the blowing speed and the intake air volume, and then adjusts the tank top attitude.

[0026] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art to which the present invention pertains.

Claims

1. A real-time monitoring system for construction sites based on drone technology, characterized in that, Including: The UAV measurement module is configured with a UAV equipped with a high-definition camera, an infrared thermal imager, and a lidar, and is used to collect RGB images, infrared thermal imaging data, and lidar point cloud data of the construction site; The cloud platform management module includes a data storage and management unit and a data analysis and processing unit, and is used to store the collected data and perform semantic segmentation, target detection, and dynamic risk assessment on the multimodal data through deep learning algorithms; The data storage and management unit: is used to store the image data of the construction site collected by the UAV; The data analysis and processing unit: performs real-time analysis on the image data of the construction site; The alarm and feedback module: triggers a hierarchical response mechanism according to the risk assessment result.

2. The real-time monitoring system for a construction site based on UAV technology according to claim 1, characterized in that The data analysis and processing unit fuses RGB images, infrared images, and lidar point cloud data, performs pixel-level semantic segmentation using the DeepLabv3+ network, and optimizes the segmentation accuracy of high-risk areas through a weighted cross-entropy loss function; identifies and detects targets based on the YOLOv8 network, and combines the DeepSORT algorithm to achieve continuous target tracking; and extracts dynamic behavior features through temporal behavior modeling, outputs a risk probability value; combines the risk probability, environmental context weight, and preset threshold to generate a four-level early warning level.

3. The real-time monitoring system for construction sites based on UAV technology according to claim 2, wherein, The weighted cross-entropy loss function is: ； Among them, L seg is the loss function; y i is the true class label of pixel i; is the predicted probability output by the network; w c (i) is the class weight of the i-th class of pixels.

4. The real-time monitoring system for a construction site based on UAV technology according to claim 2, characterized in that, The temporal behavior modeling includes a TSM network and a Transformer encoder, which fuse target trajectory, pose, and scene location information and output a risk probability.

5. A real-time monitoring method for a construction site based on UAV technology, which is applied to a real-time monitoring system for a construction site based on UAV technology according to any one of claims 1 to 4, characterized in that, Including the following specific implementation steps: S51. Initialize the UAV flight path and task parameters, and dynamically match the construction site range and environmental factors; S52. The UAV collects multimodal data and transmits it; S53. Perform adaptive histogram equalization and non-local mean denoising on the RGB image, and fuse the infrared image and the point cloud projection map to generate a 6-channel fusion feature; Perform semantic segmentation based on the DeepLabv3+ network and output a segmentation map; combine YOLOv8 and DeepSORT to achieve target detection and tracking, and output the detection result; output a risk probability through temporal behavior modeling, and generate a comprehensive risk score in combination with the environmental context; Trigger a four-level early warning according to the comprehensive risk score and output the response level; at the same time, analyze the inclination degree of the storage tank structure based on the lidar three-dimensional modeling and point cloud analysis method; S54. Start the feedback mechanism according to the response level; S55. Store the current multimodal data, segmentation map, detection result, comprehensive risk score, response level, and feedback measures.

6. The real-time monitoring method for a construction site based on drone technology according to claim 5, characterized in that Comprehensive risk score: ; Among them, is the comprehensive risk score at time t; K is the total number of risk behavior types; β k is the basic weight of the k-th type of risk; is the risk probability of the k-th behavior at time t; is the context weight adjusted according to the current working environment.

7. A real-time monitoring method for a construction site based on UAV technology according to claim 5, characterized in that, The implementation steps based on the lidar three-dimensional modeling and point cloud analysis method are: S71. Instruct the UAV to obtain the point cloud data and odometer data of the tank wall and tank top in the UAV coordinate system; S72. Fuse the point cloud data and odometer data through the LIO-SLAM algorithm, screen out the three-dimensional coordinates of the point cloud at the junction of the tank top and the tank wall in the world coordinate system, and calculate the inclination degree of the tank top: Adopt the LIO-SLAM algorithm for fusion and pose optimization; Select frames with a relatively large time interval and a relatively close distance in the historical key frames as candidate closed-loop frames, extract the set of feature points of the current frame, downsample, perform scantomap optimization, and use the iterative closest point algorithm to obtain the optimized pose; Downsample the point cloud image and determine the intersection position between the tank top and the tank wall: fill the point cloud and remove the interference of the tank wall point cloud data to obtain a point cloud image containing only the tank top data; Extract the point cloud on the circumference of the tank top edge, and based on the maximum value z of the height of the edge point cloud max and the minimum value z min , obtain the inclination degree ΔH of the tank top = z max -z min ; Output the inclination degree ΔH of the tank top; S73. If ΔH ≥ the set maximum allowable threshold value ΔH of the tank top inclination height difference max , trigger an audible and visual alarm and adjust the blowing speed and air intake volume.

8. The real-time monitoring method for a construction site based on UAV technology according to claim 7, characterized in that, The implementation process of fusing and optimizing the pose using the LIO-SLAM algorithm is as follows: S81. Calculate the IMU pre-integration: calculate the velocity, position, and rotation change amount between adjacent frames; ； ； ； where, Δt ij represents the time interval between time t i and t j ; is the estimated value of the velocity; is the estimated value of the position; v i and v j are the velocities at time t i and t j respectively; g is the acceleration due to gravity; is the estimated value of the rotation matrix; R i and R j are the rotation matrices at time t i and t j respectively; S82. Feature matching: establish a pose solution equation using the distance relationship between points and lines and between points and planes; ； ； ； where k, u, v, and w belong to the corresponding features; is the 3D coordinate of the k-th line feature point in the (i + 1)-th frame; and are the coordinates of the two endpoints of the same line feature in the i-th frame; d ek represents the current time point to the perpendicular distance from the previous time line feature; d pk is the current time point to the perpendicular distance from the previous time plane; is the 3D coordinate of the k-th plane feature point in the (i + 1)-th frame; , and are three non-collinear points that define the plane in the i-th frame; e represents the line feature; P represents the plane feature; ΔT i,i+1 is the relative pose between the current key frame and the previous key frame; T i and T i+1 are the poses of the frames at time i and time (i + 1); S83. Combine d ek and d pk to form an objective function, and solve for the optimal ΔT by minimizing the error i,i+1 for pose optimization.

Images