A road construction safety early warning management and control method, device, equipment and medium

Abstract

The application relates to the field of road-related construction safety early warning, and specifically discloses a road-related construction safety early warning management and control method, device, equipment and medium, which comprises the following steps: collecting image sequences of a construction area through a fixed camera or a drone; identifying personnel, vehicles and facilities by using a target detection network after pretreatment; mapping a detection frame to a world coordinate system and grouping spatial positions by using a clustering algorithm; calculating the distance and instantaneous speed of a target to a dangerous boundary based on an electronic fence, determining a dangerous behavior when the distance is less than a first threshold value or the speed direction points to a dangerous area and the speed rate exceeds a limit; outputting a graded early warning signal and connecting a sound and light alarm or a mechanical emergency stop; and simultaneously dynamically updating distance threshold values and speed threshold values according to early warning feedback. The application solves the problem that a fixed threshold value is prone to false positives or false negatives in different construction environments by self-adaptive threshold value adjustment, significantly improves early warning accuracy and environmental adaptability, and guarantees road-related construction safety.

Description

Technical Field

[0001] This invention belongs to the field of road construction safety early warning, specifically a road construction safety early warning and control method, device, equipment and medium. Background Technology

[0002] Road-related construction refers to maintenance, expansion, pipeline laying, and other work activities carried out in or near road traffic areas. Because construction areas intersect with vehicle traffic areas and pedestrian activity areas, safety accidents such as vehicles straying into or people accidentally entering are prone to occur. Therefore, it is necessary to deploy a safety early warning system to monitor potential dangerous behaviors in real time.

[0003] In existing technologies, fixed cameras are typically used to capture images of the site. Target detection algorithms are then used to identify targets such as people and vehicles within the construction area, and a fixed distance threshold for the electronic fence is preset. When a target is detected entering the fence boundary, an audible and visual alarm is triggered.

[0004] However, safety risks vary significantly across different construction scenarios. For example, changes in daylight intensity during nighttime construction, vehicle speed distribution on different road sections, and personnel activity density at different construction stages all affect the appropriate triggering conditions for early warnings. Using fixed thresholds is prone to sensitivity mismatches when environmental conditions change: thresholds set too strictly lead to frequent false alarms, disrupting normal construction operations; thresholds set too leniently result in missed alarms, rendering the early warning function ineffective. Therefore, how to achieve dynamic adaptive adjustment of early warning thresholds has become a pressing technical problem to be solved in the field of road construction safety management. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method, device, equipment, and medium for early warning and control of road construction safety, thereby resolving the issues of false alarms or missed alarms caused by the use of fixed early warning thresholds in the prior art under different construction environments.

[0006] This invention provides a method for early warning and control of road construction safety, comprising the following steps:

[0007] Step S1: Collect image sequences of the road construction area in real time using a fixed camera or an image sensor mounted on a drone;

[0008] Step S2: Preprocess the image sequence, including median filtering for noise reduction, histogram equalization for contrast enhancement, and affine transformation for normalization;

[0009] Step S3: Use a pre-trained object detection network to identify people, vehicles, construction equipment, and temporary facilities in the image, and output the bounding box coordinates of each object. And category confidence level, and filter out categories with confidence levels below a set threshold. The test results;

[0010] Step S4: Map the center of the bounding box of the detected target to the world coordinate system to obtain the 3D coordinates. The spatial locations were grouped using a clustering algorithm to obtain the centroid of each cluster and the number of targets within each cluster.

[0011] Step S5: Based on the predefined electronic fence area Calculate the shortest Euclidean distance from each target to the boundary of the region. and the instantaneous velocity of the target between consecutive frames. ;when Less than the first distance threshold Or the angle between the velocity direction and the normal to the boundary of the danger zone is less than a preset angle threshold and the speed exceeds a velocity threshold. At that time, it was determined to be a dangerous act;

[0012] Step S6: Output graded early warning signals according to the severity of the dangerous behavior. The graded early warning includes a first-level early warning, a second-level early warning, and a third-level early warning, and link the on-site audible and visual alarm or the emergency stop device of the construction machinery.

[0013] Step S7: Dynamically update the first distance threshold based on the actual warning events and manual feedback tags. Second distance threshold and speed threshold .

[0014] Preferably, the target detection network in step S3 is an RTDETR network, and a spatial attention mask is introduced in the Transformer encoding layer of the network. The spatial attention mask is generated by the semantic segmentation map of the construction area and is used to make the network focus on the road surface and shoulder area.

[0015] Preferably, before step S4, a sub-step of target temporal correlation and trajectory prediction is included:

[0016] The Hungarian algorithm is used to combine the intersection-union ratio and the cosine distance of appearance features for inter-frame target matching, and a unique identifier ID is assigned to each target.

[0017] For each matched target, a Kalman filter prediction model is used to predict its state vector in the next frame. , where the state vector , This is the state transition matrix;

[0018] When the predicted distance from the location to the danger zone At that time, a level-two warning was triggered in advance, among which This is a threshold for early warning.

[0019] Preferably, the determination of dangerous behavior in step S5 is replaced by the following method:

[0020] Calculate the dimensionless risk scoring function

[0021]

[0022] in , , These are the weighting coefficients. As the attenuation factor, Clustering indicator variable (when the number of targets in the cluster containing the target is greater than the clustering threshold). (Take 1 if it is true, otherwise take 0).

[0023] when When a Level 3 warning is triggered, A level-two warning was triggered.

[0024] Preferably, the clustering algorithm described in step S4 is K-means clustering, and the number of clusters is... Based on the total number of detection targets Dynamically set to Cluster center ,in For the first There are several clusters.

[0025] Preferably, the graded early warning in step S6 is executed according to the following rules:

[0026] Level 1 warning: Meets the requirements And no other higher-level conditions were triggered to activate the low-frequency sound and light alert;

[0027] Level 2 warning: Meets the requirements or Furthermore, if the angle between the speed direction and the normal direction of the danger zone boundary is less than 30°, a high-frequency audible and visual alarm will be triggered and information will be pushed to the safety officer's handheld terminal.

[0028] Level 3 warning: If the same target triggers a Level 2 warning more than twice within 10 seconds and the distance to the boundary of the danger zone does not increase, the automatic emergency stop device of the construction machinery will be activated or a stop command will be issued through the drone broadcast system.

[0029] Preferably, when the length of the construction area exceeds a preset length threshold or the field of view coverage of the fixed camera is lower than a preset threshold, an unmanned aerial vehicle (UAV) is used to autonomously cruise and collect images. The UAV flies along a preset route, with the route points automatically generated by the BIM model of the construction area. The flight speed is 3m / s to 5m / s, the image overlap rate is not less than 60%, and the images are transmitted back in real time via a 5G network.

[0030] Preferably, the dynamic update in step S7 employs an online gradient descent method to update the distance threshold with the goal of minimizing the early warning error. , and speed threshold Simultaneously, the process noise covariance matrix in the Kalman filter prediction model is updated using exponential moving average. The formula for exponential moving average is as follows: ,in This is a smoothing factor.

[0031] The present invention also provides a road construction safety early warning and control device, comprising:

[0032] The data acquisition module, used to perform the above step S1, includes at least one fixed network camera or drone-borne camera;

[0033] The edge computing module, with a built-in graphics processor or tensor processor, is used to execute the above steps S2 to S5 and output the dangerous behavior judgment result.

[0034] The early warning execution module includes an audible and visual alarm, a wireless communication unit, and a linkage control interface, which are used to execute the graded early warning signal output and mechanical emergency stop control in step S6 above.

[0035] The storage and update module is used to store electronic fence maps, historical early warning logs and model parameters, and supports dynamic updates of thresholds.

[0036] The present invention also provides a road construction safety early warning and control device, comprising: at least one processor, and a memory coupled to the processor, the memory storing instructions executable by the processor; the instructions, when executed by the processor, implement the above-described method.

[0037] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] By integrating dual-mode image acquisition from fixed cameras and drones, and combining preprocessing enhancement with a deep learning target detection network, accurate identification and positioning of personnel, vehicles, and equipment within the construction area were achieved. Based on this, spatial clustering algorithms were used to group and analyze the locations of multiple targets, and based on predefined electronic fence areas, the Euclidean distance from the target to the danger boundary and the instantaneous velocity direction were calculated, thus constructing a multi-level dangerous behavior judgment system ranging from distance exceeding limits to speed intrusion.

[0040] Furthermore, this invention introduces a temporal correlation and Kalman filter prediction mechanism to trigger an early warning before the target actually enters the danger zone, effectively compensating for the lag in pure real-time detection. To address the differences in various construction scenarios, the system uses an online gradient descent method to dynamically and adaptively update distance and velocity thresholds, and optimizes the Kalman filter parameters in real time through exponential moving averages, ensuring that the early warning model always adapts to changes in the on-site environment.

[0041] Furthermore, this invention provides an alternative judgment method based on a risk scoring function and an extended data collection scheme for autonomous drone navigation, significantly improving adaptability under complex working conditions. Ultimately, through closed-loop execution of tiered audible and visual alarms, handheld terminal push notifications, and mechanical emergency stop linkage, intelligent safety management and control of the entire process from perception and analysis to intervention is achieved, effectively reducing the risk of road construction accidents. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the road construction safety early warning and control method in Embodiment 1 of the present invention;

[0043] Figure 2 This is a schematic diagram of the frame structure of the device in Embodiment 2 of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1: This example provides a method for early warning and control of road construction safety, which includes the following steps.

[0046] Step S1: Multi-source image acquisition

[0047] In road construction areas (such as highway reconstruction and expansion sections, urban road excavation sites, etc.), deploy several fixed network cameras (such as Hikvision DS-2CD series, 1920×1080 resolution, 25fps frame rate), and mount visible light cameras on drones (such as DJI Matrice 300 RTK); the drones take off as needed, and both cameras jointly collect real-time image sequences of the construction area; during the collection process, ensure that the camera's field of view covers the entire electronic fence area, and maintain the drone's flight altitude at 30~50 meters to ensure that the target pixel size meets the detection requirements.

[0048] Step S2: Image Preprocessing

[0049] Perform the following steps sequentially on each captured image frame:

[0050] Median filtering for noise reduction: A 3×3 sliding window is used, and the median of the gray values ​​of the pixels within the window is taken as the new value of the center pixel to eliminate salt-and-pepper noise.

[0051] Histogram equalization contrast enhancement: Calculate the image gray-level histogram and cumulatively map the distribution function to make the gray-level distribution of the output image uniform and enhance the details of low-contrast areas.

[0052] Affine transformation normalization: Based on pre-calibrated camera parameters, the image is transformed to a uniform scale (e.g., 640×640 pixels) to eliminate lens distortion and perspective differences.

[0053] The image quality is improved after preprocessing, which facilitates subsequent target detection.

[0054] Step S3: Target Detection and Recognition

[0055] A pre-trained RTDETR (Real-Time Detection TRansformer) network was used as the target detector. This network was pre-trained on the COCO dataset and fine-tuned on a road construction scenario dataset (containing 10,000 manually annotated images labeled with categories such as personnel, vehicles, construction equipment, cones, and fences).

[0056] The RTDETR network structure includes a Transformer encoder and a decoder. This embodiment introduces a spatial attention mask in the encoding layer: first, a lightweight semantic segmentation network (such as U-Net, which takes a pre-processed image as input and outputs pixel-level classification) is used to generate a semantic segmentation map of the construction area, segmenting regions such as the road surface, shoulder, and background; then, this segmentation... Figure 2 Value-based: The mask value for the road surface and shoulder area is 1, while the mask value for other areas is 0. In the self-attention calculation of the Transformer encoder, the mask is multiplied element by element with the attention weight, so that the model only focuses on the area with a mask value of 1, thereby ignoring irrelevant areas such as background vegetation and sky, improving detection accuracy and speed.

[0057] The network outputs the bounding box coordinates of each target. (in These are the pixel coordinates of the center point of the bounding box. (for width and height) and category confidence Set confidence threshold Filter out The test results retain high-confidence targets.

[0058] Time-series tracking and Kalman filter prediction

[0059] Before step S4, a target temporal correlation and trajectory prediction sub-step is added, specifically implemented as follows:

[0060] Inter-frame matching: The Hungarian algorithm is used, and the cost matrix is ​​based on the weighted sum of the intersection-to-union ratio (IoU) and the cosine distance of the appearance features; the appearance features are provided by the 128-dimensional feature vector output by the penultimate layer of the RTDETR network; the IoU weight is 0.7 and the cosine distance weight is 0.3; the matching threshold is set to 0.5.

[0061] Kalman filtering: defining the state vector State transition matrix:

[0062]

[0063] Observation matrix Process noise covariance matrix The initial value is Measure the noise covariance matrix Prediction formula:

[0064]

[0065]

[0066] in To estimate the error covariance matrix; when predicting the distance from the location to the danger zone... When the early warning threshold is set to 1.5 meters, a secondary early warning (i.e., high-frequency audible and visual alarm and push notification) is triggered in advance, the system can issue an alarm about 0.5 to 1 second before the target actually enters the danger zone.

[0067] Step S4: Spatial Location Mapping and Clustering

[0068] (1) Coordinate mapping

[0069] Using camera calibration parameters (including intrinsic parameter matrix) and extrinsic rotation matrix Translation vector ), the center pixel coordinates of the bounding box in the image Transform to the world coordinate system (with a fixed point in the construction area as the origin, and the Z-axis pointing vertically upwards); assume the ground is a plane ( Based on the pinhole camera model:

[0070]

[0071] in The scale factor is obtained by solving for ground constraints. Finally, the three-dimensional coordinates of each target are obtained. ,in The height is considered to be 0 (target on the ground). The height of personnel or vehicles can be ignored. If height information is required, it can be extracted separately.

[0072] (2) K-means spatial clustering

[0073] For all detected target locations in the current frame Perform K-means clustering; number of clusters Based on the total number of targets Dynamically set to (That is, each cluster contains an average of 5 targets); initialize cluster centers. (random selection) (target points), iterative updates:

[0074]

[0075] in For the first 1 cluster. Iterate until the change in cluster centers is less than a threshold (e.g., ...). (meters). Output the centroid of each cluster. and the number of targets within the cluster This is used for subsequent risk assessment.

[0076] Step S5: Determining Dangerous Behavior

[0077] (1) Definition of electronic fence

[0078] Based on the construction design drawings, the electronic fence area is predefined in the system. For example, for excavating a foundation pit, the electronic fence is a rectangular area offset 1 meter outward from the edge of the pit; for hoisting operation areas, it is a circular area with the boom rotation radius plus 3 meters; the electronic fence is stored in the world coordinate system as a list of polygon vertex coordinates.

[0079] (2) Distance and speed calculation

[0080] For each objective Calculate its distance to the electronic fence boundary Shortest Euclidean distance:

[0081]

[0082] in This represents the Euclidean norm.

[0083] Calculate consecutive frame intervals (time intervals) The instantaneous velocity of the target (in seconds, corresponding to 25fps):

[0084]

[0085] (3) Basic judgment rules

[0086] A behavior is considered dangerous if any of the following conditions are met:

[0087] Condition ①: (First distance threshold, initial value set to 2.0 meters, subsequently updated adaptively);

[0088] Condition ②: The angle between the velocity direction and the normal to the boundary of the danger zone (Preset angle threshold, in this embodiment, is taken as...) ), and rate (Speed ​​threshold, initial value set to 1.5 m / s, subsequently updated adaptively).

[0089] The angle between the velocity direction and the normal is calculated using the vector dot product: , The direction of the inner normal of the boundary of the danger zone at the nearest point (pointing towards the interior of the danger zone).

[0090] Step S6: Tiered Early Warning and Execution

[0091] A three-level warning system is issued based on the severity of the dangerous behavior, with the specific rules as follows:

[0092] Level 1 warning: When (Second distance threshold) When the initial value is set to 1.0 meter and no other higher-level conditions are triggered, a low-frequency sound and light prompt is triggered (flash frequency 1Hz, volume 70dB).

[0093] Level 2 warning: When or When the angle between the speed direction and the normal of the boundary of the danger zone is less than 30°, a high-frequency audible and visual alarm (flash frequency 5Hz, volume 90dB) is triggered, and alarm information (including target ID, location, and on-site screenshot) is pushed to the safety officer's handheld terminal (such as a smartphone or walkie-talkie) via the 4G / 5G network.

[0094] Level 3 warning: When the same target triggers a Level 2 warning more than twice within 10 seconds, and the distance to the boundary of the danger zone does not increase (i.e., When danger occurs, the automatic emergency stop device of the construction machinery (such as road rollers, pavers, tower cranes) or the drone broadcast system will issue a stop command such as "Danger! Stop immediately!"

[0095] "Continuous triggering" judgment: The system maintains a secondary warning timestamp queue for each target. If the number of times the target triggers a secondary warning is ≥ 2 within a 10-second time window (sliding window), and the distance between the current frame and the previous frame has not decreased, then a tertiary warning is triggered.

[0096] Step S7: Dynamic Threshold Update

[0097] Based on actual warning events and human feedback (safety officers marking "false alarm" or "missed alarm" on their handheld terminals), the system adaptively updates the first distance threshold using an online gradient descent method. Second distance threshold and speed threshold .

[0098] Define the early warning error function:

[0099]

[0100] in For the first The true labels of each sample (1 indicates danger, 0 indicates safety). The system outputs (1 or 0) based on the current threshold; gradient descent update rule:

[0101]

[0102]

[0103]

[0104] Learning rate An update is performed every 10 new feedback samples received; the threshold values ​​range as follows:

[0105] rice;

[0106] rice;

[0107] m / s.

[0108] gradient , , Approximate calculation using the finite difference method: for example , Take 0.01 meters.

[0109] Furthermore, for the Kalman filter prediction model used before step S4 (see the "Time Series Tracking and Kalman Filter Prediction" section above), its process noise covariance matrix... It also needs to be updated online; this embodiment uses the exponential moving average method:

[0110]

[0111] in As a smoothing factor, It is a covariance matrix estimated in real time based on the most recent batch of prediction errors; through this update, the Kalman filter can adapt to changes in the target motion pattern in the construction environment (such as sudden running of personnel).

[0112] Alternative determination method: based on risk scoring function

[0113] As an alternative to the basic judgment rule in step S5, the system can adopt a dimensionless risk scoring function. Perform continuous risk quantification; this embodiment provides recommended parameter values:

[0114] Weighting coefficients: , , (This can be determined based on the actual scenario using AHP (Analog-Philippines Hierarchy Process) or machine learning regression.)

[0115] Attenuation factor: (so that when) Rice time, );

[0116] Aggregation threshold: (A group of more than 3 people is considered a gathering);

[0117] Risk threshold: , .

[0118] Calculate the risk score:

[0119]

[0120] in Clustering indicator variable (when the number of targets in the cluster containing the target is greater than...) If the value is 1, then the value is 0.

[0121] Judgment rule: When A level 3 warning is triggered at any time; when A level 2 warning is triggered when the warning is triggered; otherwise, it is considered safe or a level 1 warning is triggered (level 1 warnings can be mapped to other warnings as needed). However, there are still situations where the distance is close (but there are still cases where the distance is close). The system can select to use either the basic judgment rule or the risk scoring rule through a configuration switch.

[0122] Autonomous drone cruise data collection

[0123] When the length of the construction area exceeds a preset length threshold (200 meters in this embodiment) or the field of view coverage of the fixed camera (the ratio of the camera coverage area to the total area of ​​the construction area) is lower than a preset threshold (80% in this embodiment), the system automatically switches to drone cruise mode. The drone flies along the route generated by the BIM model, and the route points include the boundary inflection points of the construction area, above key equipment, etc. The flight speed is set to 3~5 m / s, the forward overlap rate (the proportion of overlapping images of adjacent routes) is not less than 60%, and the lateral overlap rate is not less than 30%. The drone transmits images back to the edge computing server in real time through the onboard 5G module, with a latency of less than 100 milliseconds.

[0124] Example 2: This example provides an apparatus for implementing the above method, such as... Figure 2 The framework diagram shows the following:

[0125] Data acquisition module: It consists of multiple fixed network cameras (supporting PoE power supply and RTSP stream output) and a drone (equipped with a 30x zoom camera and a 5G image transmission module), and performs image acquisition according to step S1.

[0126] Edge computing module: It adopts the NVIDIA Jetson AGX Orin embedded platform with a built-in GPU (32GB of video memory), and comes pre-installed with Ubuntu 20.04 and PyTorch and TensorRT inference frameworks; this module loads the trained RTDETR model and tracker, executes steps S2 to S5, and outputs the dangerous behavior judgment result.

[0127] Early warning execution module: includes an audible and visual alarm (S100L type, IP65 protection rating), a 4G DTU wireless communication unit (supports MQTT protocol), and a relay linkage control interface (8 channels, each with a maximum load of 10A / 220VAC); used to execute graded early warning signal output and mechanical emergency stop control.

[0128] Storage and Update Module: Uses 256GB SSD to store the geofence map (GeoJSON format), historical early warning logs (SQLite database), and model parameters (JSON file); supports remote updates of thresholds and algorithm models via a cloud management platform.

[0129] Example 3: This example provides a road construction safety early warning and control device, including at least one processor (such as an Intel Core i7-12700K) and a coupled memory (DDR4 32GB). The memory stores a computer program, which, when executed by the processor, implements the method described above.

[0130] In addition, this embodiment also provides a computer-readable storage medium, such as a USB flash drive, a solid-state drive, or an SD card, on which a computer program is stored, which, when executed by a processor, implements the above-described method.

[0131] The embodiments of the present invention are given for the purposes of illustration and description. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Any changes, modifications, substitutions and variations made by those skilled in the art to the above embodiments within the scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early warning and control of road construction safety, characterized in that, Includes the following steps: Step S1: Collect image sequences of the road construction area in real time using a fixed camera or an image sensor mounted on a drone; Step S2: Preprocess the image sequence, including median filtering for noise reduction, histogram equalization for contrast enhancement, and affine transformation for normalization; Step S3: Use a pre-trained object detection network to identify people, vehicles, construction equipment, and temporary facilities in the image, and output the bounding box coordinates of each object. And category confidence level, and filter out categories with confidence levels below a set threshold. The test results; Step S4: Map the center of the bounding box of the detected target to the world coordinate system to obtain the 3D coordinates. The spatial locations were grouped using a clustering algorithm to obtain the centroid of each cluster and the number of targets within each cluster. Step S5: Based on the predefined electronic fence area Calculate the shortest Euclidean distance from each target to the boundary of the region. and the instantaneous velocity of the target between consecutive frames. ;when Less than the first distance threshold Or the angle between the velocity direction and the normal to the boundary of the danger zone is less than a preset angle threshold and the speed exceeds a velocity threshold. When the preset angle threshold is less than 90°, it is determined to be a dangerous behavior; Step S6: Output graded early warning signals according to the severity of the dangerous behavior. The graded early warning includes a first-level early warning, a second-level early warning, and a third-level early warning, and link the on-site audible and visual alarm or the emergency stop device of the construction machinery. Step S7: Dynamically update the first distance threshold based on the actual warning events and manual feedback tags. and speed threshold .

2. The method according to claim 1, characterized in that, The target detection network mentioned in step S3 is the RTDETR network, and a spatial attention mask is introduced in the Transformer encoding layer of the network. The spatial attention mask is generated by the semantic segmentation map of the construction area and is used to make the network focus on the road surface and shoulder area.

3. The method according to claim 1, characterized in that, Step S4 is preceded by a sub-step involving target temporal correlation and trajectory prediction: The Hungarian algorithm is used to combine the intersection-union ratio and the cosine distance of appearance features for inter-frame target matching, and a unique identifier ID is assigned to each target. For each matched target, a Kalman filter prediction model is used to predict its state vector in the next frame. , where the state vector , This is the state transition matrix; When the predicted distance from the location to the danger zone At that time, a level-two warning was triggered in advance, among which This is a threshold for early warning.

4. The method according to claim 1, characterized in that, The determination of dangerous behavior in step S5 is replaced by the following method: Calculate the dimensionless risk scoring function in , , These are the weighting coefficients. As the attenuation factor, Clustering indicator variable (when the number of targets in the cluster containing the target is greater than the clustering threshold). (Take 1 if it is true, otherwise take 0). when When a Level 3 warning is triggered, A level-two warning was triggered.

5. The method according to claim 1, characterized in that, The clustering algorithm described in step S4 is K-means clustering, and the number of clusters is... Based on the total number of detection targets Dynamically set to Cluster center ,in For the first There are several clusters.

6. The method according to claim 1, characterized in that, The graded early warning described in step S6 shall be executed according to the following rules: Level 1 warning: Meets the requirements And no other higher-level conditions were triggered to activate the low-frequency sound and light alert; Level 2 warning: Meets the requirements or Furthermore, if the angle between the speed direction and the normal direction of the danger zone boundary is less than 30°, a high-frequency audible and visual alarm will be triggered and information will be pushed to the safety officer's handheld terminal. Level 3 warning: If the same target triggers a Level 2 warning more than twice within 10 seconds and the distance to the boundary of the danger zone does not increase, the automatic emergency stop device of the construction machinery will be activated or a stop command will be issued through the drone broadcast system.

7. The method according to claim 1, characterized in that, When the length of the construction area exceeds the preset length threshold or the field of view coverage of the fixed camera is lower than the preset threshold, the drone will autonomously cruise to collect images. The drone will fly along a preset route, with the route points automatically generated by the BIM model of the construction area. The flight speed will be 3m / s to 5m / s, the image overlap rate will be no less than 60%, and the images will be transmitted back in real time via the 5G network. The dynamic update in step S7 uses the online gradient descent method to update the distance threshold with the goal of minimizing the early warning error. , and speed threshold Simultaneously, the process noise covariance matrix in the Kalman filter prediction model is updated using exponential moving average. The formula for exponential moving average is as follows: ,in This is a smoothing factor.

8. A road construction safety early warning and control device, characterized in that, include: The data acquisition module is used to perform step S1 as described in claim 1, and includes at least one fixed network camera or drone-borne camera; An edge computing module, with a built-in graphics processor or tensor processor, is used to execute steps S2 to S5 as described in claim 1 and output the dangerous behavior determination result; The early warning execution module includes an audible and visual alarm, a wireless communication unit, and a linkage control interface, and is used to execute the graded early warning signal output and mechanical emergency stop control in step S6 of claim 1. The storage and update module is used to store electronic fence maps, historical early warning logs and model parameters, and supports dynamic updates of thresholds.

9. A road construction safety early warning and control device, characterized in that, include: At least one processor, and a memory coupled to the processor, the memory storing instructions executable by the processor; When the instructions are executed by the processor, they implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method of any one of claims 1 to 7.

Images