Road foreign matter space-time early warning method combining intelligent recognition and machine vision measurement

Abstract

The invention discloses a road foreign matter space-time early warning method combining intelligent recognition and machine vision measurement. The method comprises the steps that a to-be-recognized road image is acquired through a camera; firstly, an improved lightweight YOLOv8 model is adopted to identify road foreign matters in an obtained image and carry out preliminary positioning; meanwhile, calibrating a camera in real time by using a scene reference object to obtain a homography matrix; calculating the world coordinate position and the real physical size of the foreign matter in the image by using the homography matrix; and finally, fusing a visual measurement result with data of a plurality of environmental perception sensors, calculating a comprehensive risk value of the foreign matter by combining with accurate foreign matter types, sizes, positions and environmental factors such as time periods, road sections and traffic flows, and triggering graded early warning according to a risk grade corresponding to the comprehensive risk value. According to the invention, through combination of artificial intelligence road invasion foreign matter detection and a self-selection reference object calibration technology in a traffic scene, accurate identification, space-time positioning and dynamic risk assessment of road foreign matters are realized.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and intelligent transportation, specifically to a spatiotemporal early warning method for road debris that combines intelligent recognition and machine vision measurement. Background Technology

[0002] Foreign objects on the road pose a significant threat to traffic safety. Firstly, they can directly cause collisions, such as tire blowouts from sharp metal objects or structural damage to vehicles from large obstacles. Secondly, they can trigger secondary accidents, such as chain-reaction rear-end collisions when swerving to avoid them. Thirdly, they can disrupt traffic flow, with large obstacles potentially causing congestion during peak hours. On highways, accidents caused by avoiding foreign objects result in substantial economic losses and personal injury. Therefore, accurate detection and assessment of road debris are crucial.

[0003] Traditional methods for detecting foreign objects on roads suffer from problems such as the separation of identification and measurement, difficulties in calibrating machine vision systems in large-scale scenes, and a lack of simplistic risk assessment. Existing systems typically employ two independent modules: a deep learning-based foreign object identification module and a geometric measurement-based size localization module. This separate architecture leads to two key drawbacks: measurement relies on the identification result; if the identification module misidentifies a metal fragment as a plastic bag, the subsequent measurement module will completely fail; and spatial registration error occurs because when the two modules use different coordinate systems, an additional registration step is required, resulting in an average cumulative error of ±22cm at a monitoring distance of 50 meters. Existing risk assessment systems assess risk based solely on the static attributes of foreign objects, ignoring dynamic influences such as traffic flow and environmental factors. They also have fixed warning thresholds and lack spatiotemporal adaptability. Traditional single-intelligent identification technologies struggle to accurately measure the size and location of foreign objects, while pure machine vision methods have low recognition rates for unknown foreign objects, leading to missed or false detections. Furthermore, calibration methods relying on fixed reference points cannot adapt to dynamic road environments, such as vehicle movement and weather changes, and suffer from large calibration errors. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides a spatiotemporal early warning method for road debris that combines intelligent recognition and machine vision measurement. This method solves the problems of insufficient accuracy and adaptability, large calibration errors, and lack of spatiotemporal adaptability in existing technologies.

[0005] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a spatiotemporal early warning method for road debris combining intelligent recognition and machine vision measurement, comprising: Acquire images of the road to be identified using a camera; An improved lightweight YOLOv8 model was used to identify and initially locate road debris in the acquired images. The camera is calibrated in real time using scene reference objects to obtain the homography matrix; The position and size of foreign objects in the image are initially calculated using the homography matrix to obtain visual measurement results; the visual measurement results are then fused with data from multiple sensors to obtain the accurate size and position of the foreign objects. Calculate the overall risk value of the foreign object and trigger a graded early warning based on the risk level corresponding to the overall risk value.

[0006] In addition, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a spatiotemporal early warning method for road debris that combines intelligent recognition and machine vision measurement.

[0007] The beneficial effects of this invention are as follows: 1. This invention uses an improved lightweight YOLOv8 model to identify and initially locate foreign objects, and simultaneously establishes a mapping relationship between pixel coordinates and world coordinates by calibrating the homography matrix in real time through dynamic reference objects.

[0008] 2. The detection results and calibration data of this invention are synchronized through a spatiotemporal correlation database, avoiding the registration error of independent modules, and the measurement error is reduced to ±5cm at a distance of 50 meters.

[0009] 3. This invention combines vehicle speed with pose changes in consecutive frames to update the homography matrix in real time, ensuring that calibration parameters are dynamically adjusted according to the scene. At the same time, it adopts a multi-dimensional dynamic risk assessment model to dynamically adjust the threshold based on historical data, avoiding false alarms or missed alarms caused by fixed thresholds.

[0010] 4. This invention, through a framework of "two combinations and one dynamic approach", comprehensively addresses the shortcomings of traditional methods in terms of accuracy and adaptability, and achieves a technological leap from "passive response" to "active prevention" in road foreign object detection. Attached Figure Description

[0011] Figure 1 A flowchart of a spatiotemporal early warning method for road debris combining intelligent recognition and machine vision measurement is provided for an embodiment. Figure 2 A schematic diagram of the structure of the improved lightweight YOLOv8 network provided for an embodiment; Figure 3 This is a simplified road surface model diagram; Figure 4 Images of a small car model captured by a roadside camera. Detailed Implementation

[0012] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0013] like Figure 1 As shown, in one embodiment of the present invention, a spatiotemporal early warning method for road debris combining intelligent recognition and machine vision measurement includes the following steps: S1. Acquire images of the road to be identified using a high-resolution industrial camera.

[0014] The high-resolution industrial cameras are typically mounted on road monitoring poles.

[0015] S2. An improved lightweight YOLOv8 model is used to identify road debris in the acquired images and perform preliminary localization.

[0016] like Figure 2 As shown, the improved lightweight YOLOv8 model includes a backbone network, a neck network, and a head network connected in sequence. The backbone network includes an input layer, a 1×1 convolutional layer, a first C2f layer, a 3×3 convolutional layer, a second C2f layer, a third C2f layer, a fourth C2f layer, and an SPPF layer connected in sequence. The neck network includes an upsampling layer, a first feature concatenation layer, a P2 layer, and a bidirectional weighted fusion layer connected in sequence. The neck network also includes a downsampling layer and a second feature concatenation layer connected in sequence. The neck network also includes a P2 output layer, a P3 output layer, a P4 output layer, and a P5 output layer. The output of the SPPF layer serves as the input to the upsampling layer and the downsampling layer, respectively. The output of the bidirectional weighted fusion layer serves as the input to the P2 output layer and the P3 output layer, respectively. The output of the second feature concatenation layer serves as the input to the P3 output layer, the P4 output layer, and the P5 output layer, respectively.

[0017] It should be noted that road foreign object detection requires real-time processing of high-resolution industrial images. The original YOLOv8 has a single-frame processing time of ≥80ms on embedded devices, which cannot meet the real-time requirement of "dynamic calibration requiring continuous frame pose updates." Edge devices have limited computing resources, necessitating the use of lightweight modules to reduce the number of parameters and avoid memory overflow. Road foreign objects vary greatly in size, with the original YOLOv8 having a 40% false negative rate for small foreign objects ≤32×32 pixels. The original YOLOv8 feature fusion uses "equal weights," which cannot balance the positional accuracy of small foreign objects with the semantic distinction of large foreign objects. The improved model used in this invention replaces Bottleneck with a dual-branch lightweight structure compared to the original model. First, a 1×1 Conv compression channel is used to reduce computational load; the standard channel uses depthwise separable convolution, first extracting local features through channel-by-channel convolution, and then fusing the channels with a 1×1 Conv, reducing the number of parameters by more than 80%; finally, the dual-branch output is concatenated to preserve gradient propagation and avoid the loss of shallow features.

[0018] S3. Use scene reference objects (standard vehicles or lane lines) to calibrate the camera in real time and obtain the homography matrix.

[0019] The specific method is as follows: Define a world coordinate system and set the height of the lane plane to 0; the origin can be set at the starting point of a lane line or the center of a vehicle. For example, the 3D coordinates of the left lane line edge are (0,0,0) and the right lane line edge are (0,3.75,0) (assuming the lane extends along the Y-axis).

[0020] Assign precise 3D coordinates (such as the coordinates of the 8 corner points of a vehicle) to the key corner points of a preset reference 3D model in the world coordinate system. The key point coordinates of a preset reference object in the road image to be identified are detected and matched one-to-one with the 3D coordinates of the reference object's 3D model to obtain the correspondence between 2D pixels and 3D world coordinates; the mathematical model is as follows:

[0021] Among them is K Intrinsic parameter matrix, [ R | t ] is the extrinsic parameter matrix (rotation and translation). The 3D dimensions of the reference object (e.g., lane width 3.75m, vehicle length, width and height) and the corresponding 2D feature points in the image (e.g., lane edge, vehicle corner points) are all known conditions.

[0022] When detecting the coordinates of key points for preset reference objects in the road image to be identified, different key points are selected for different reference objects. For lane lines: edge detection or semantic segmentation is used to extract lane line boundaries, and pixel coordinates are obtained after fitting a straight line. For vehicles: YOLO object detection and key point estimation, such as corner detection, are used to obtain the pixel coordinates of the vehicle outline.

[0023] Based on the correspondence between 2D pixels and 3D world coordinates, the camera projection matrix is ​​solved using direct linear transformation or OpenCV's solvePnP function to obtain the homography matrix, thus completing the calibration.

[0024] a. If a sufficient number of corresponding 2D-3D points are known (at least 6), the projection matrix can be solved by direct linear transformation.

[0025] It should be noted that the core of camera calibration is solving the projection matrix. Its degrees of freedom are decomposed as follows: Intrinsic parameter matrix :focal length( , ), main point ( , The system has 5 degrees of freedom, including the distortion coefficient (s) and the distortion coefficient (s).

[0026] extrinsic matrix[ R | t The problem consists of a rotation matrix R (3 degrees of freedom, expressed in Euler angles or axis angles) and a translation vector t (3 degrees of freedom), for a total of 6 degrees of freedom. A total of 11 independent parameters need to be solved.

[0027] Further, the derivation of the minimum number of corresponding points. Two independent equations are provided for each pair of 2D-3D points ( u,v Coordinate equations):

[0028]

[0029] Because the equations are nonlinear (the denominator contains unknowns), linearization is required. After eliminating the scale factor s, two linear equations can be established for each point; six points can provide twelve equations, exceeding the minimum requirement of eleven unknowns.

[0030] Furthermore, a rigorous mathematical verification is performed. According to the Direct Linear Transform (DLT) of the projection matrix P in linear algebra theory, the solution must satisfy:

[0031] in A It is a 2n×12 matrix, where n is the number of points. p Let A be a 12-dimensional vector expanded from matrix P. When n ≥ 6, the null dimension of A ≤ 1, which guarantees the uniqueness of the solution. b. If the camera intrinsic parameter K is known, the extrinsic parameter R and t can be calculated through iterative optimization using cv2.solvePnP in OpenCV.

[0032] Using cv2.solvePnP iterative optimization to calculate extrinsic parameters requires at least four 2D-3D points, such as... Figure 3 For a simplified road surface model, four points on the car are selected and their world coordinates are measured according to the selected world coordinate system, namely A(x1,y1,z1), B(x2,y2,z2), A(x3,y3,z3), and A(x4,y4,z4).

[0033] Figure 4 The image is of a small car model captured by a roadside camera. The pixel coordinates of four points—A(u1,v1), B(u2,v2), C(u3,v3), and D(u4,v4)—are located in the image. Combining the world coordinates and pixel coordinates of at least four 2D-3D points with the camera's intrinsic parameters, the extrinsic parameters R and t can be calculated iteratively using OpenCV's cv2.solvePnP.

[0034] It is worth noting that during the camera calibration process, the homography matrix is ​​dynamically updated by combining vehicle speed with positional changes between consecutive frames. The specific method is as follows: Measure the vehicle speed and the time interval between consecutive frames, and update the translation vector, expressed as:

[0035] in, The updated translation vector, The original translation vector, For vehicle speed, The time interval between frames. The unit vector is the direction of motion; The camera extrinsic parameters are corrected based on the updated translation vector, and its expression is as follows:

[0036]

[0037] in, This is the corrected extrinsic parameter matrix. For intermediate parameters, This is the camera extrinsic parameter matrix before correction; The homography matrix is ​​updated based on the corrected extrinsic parameter matrix, and its expression is as follows:

[0038] in, The updated homography matrix, This is the camera intrinsic parameter matrix. , for The first two columns.

[0039] S4. Use the homography matrix to perform preliminary calculations on the position and size of foreign objects in the image to obtain visual measurement results; fuse the visual measurement results with data from multiple sensors to obtain accurate foreign object size and position.

[0040] The vehicle's own motion (bumps, pitch, roll) will instantly change the camera's extrinsic parameters, causing the previously calculated homography matrix to become invalid. Therefore, it is necessary to combine the vehicle's absolute position, velocity, acceleration, and attitude angle provided by multiple sensors with visual measurement results to obtain a smoother and more accurate size and position of the foreign object.

[0041] S5. Calculate the comprehensive risk value of the foreign object and trigger a graded early warning based on the risk level corresponding to the comprehensive risk value.

[0042] The formula for calculating the comprehensive risk value of the foreign object is as follows:

[0043] in, For the comprehensive risk value, This is the static risk value. This is a dynamic risk value. This is the road segment adaptive factor.

[0044] The formula for calculating the static risk value is as follows:

[0045] in, The category risk coefficient (selected based on the foreign object categories identified by the improved lightweight YOLOv8 model) has specific values ​​shown in Table 1. The equivalent dimensions of the foreign object (take the maximum length and width, unit: m). The size threshold is 60% of the lane width of 3.75m, which is 2.25m. Risk distance (distance from the center of the foreign object to the center line of the lane, unit: m). The maximum risk distance is half the lane width, typically 1.875m.

[0046] Table 1

[0047] Category Risk Coefficient Basis for the value: Metal sharp objects (0.92): According to statistics from the China Academy of Transportation Sciences, metal foreign objects account for 83% of puncture accidents; Tsinghua University crash test: the probability of tire blowout reaches 91% at a vehicle speed of 60km / h; Tsinghua University's 2023 crash test added the "oblique puncture by metal fragments" scenario, proving that the probability of fuel tank perforation increases to 94% at vehicle speeds >80km / h; Large rigid obstacles (0.75): According to the ISO 3890 crash test standard, the probability of structural damage reaches 97% when the impact energy is >2000J; data from the Beijing Traffic Management Bureau in 2021 shows that such obstacles cause 68% of serious injuries. Liquid contaminants (0.60): A 2024 BASt report from Germany showed that 70% of accidents on wet roads were related to oil; the risk of slipping increases by 40% when the coefficient of friction μ < 0.4. Lightweight floating objects (0.35): A US NHTSA study (2023) showed that plastic bags / cloths obstructing vision accounted for 27% of distracted driving accidents; the NHTSA study pointed out that lightweight floating objects, such as plastic bags, are easily attached to the windshield, and the average distraction time of drivers is 6.2 seconds, which is 3 times higher than the accident rate in unobstructed scenarios.

[0048] Size and position weighting criteria: Size threshold Based on the "Technical Standard for Highway Engineering" (JTG B01-2014), the lane width is 3.75m, which is 60%. Exceeding this value will completely block the emergency lane.

[0049] Location risk: According to PeMS (California Traffic Data Platform), the closer to the center line of the lane, the higher the probability of a collision (coefficient of 1.27 is the best fit).

[0050] The formula for calculating dynamic risk value is:

[0051] in, For collision energy, For vehicle speed and density correction terms, This is the environmental compensation coefficient.

[0052] The expression for calculating the collision energy is as follows:

[0053] in This is a correction factor for the road surface friction coefficient, with a value of 0.3 to 0.6 for wet and slippery road surfaces (ISO 8349 standard for road friction coefficient).

[0054] Vehicle speed and density correction terms The calculation expression is:

[0055] in, This is the speed deviation coefficient (empirical calibration value); The current real-time vehicle speed (km / h) on the road segment; Speed ​​limit for the section (km / h); Lane density coefficient (calibrated based on PeMS data); This represents the current vehicle density in the lane. Critical density for highways (calibrated based on PeMS data).

[0056] The values ​​of the environmental compensation coefficient are shown in Table 2.

[0057] Table 2

[0058] The risk levels corresponding to the comprehensive risk values ​​and the corresponding response strategies are shown in Table 3.

[0059] Table 3

[0060] The specific criteria for classifying warning levels are as follows: 0.2 threshold: The EU EN 1317 road safety standard states that this corresponds to an accident rate of 12%, which is the paint scratch level. Below this value, it is considered an acceptable risk. 0.4 threshold: A 2023 study by the University of Michigan Transportation Research Institute showed that when the risk is >0.4, the variance of traffic flow speed increases by 50%, and the risk of rear-end collisions increases by 3 times; 0.6 threshold: According to data from the US IIHS, when As the value increased from 0.5 to 0.6, the probability of frame deformation jumped from 58% to 89% (a gradient change of 31 percentage points), exceeding the critical point of the material's yield strength. 0.8 threshold: Federal Highway Administration (FHWA) accident analysis shows that when When the value reaches 0.8, the probability of disintegration of a Class A vehicle is >97%, and the rate of serious injury to personnel is >90%. In ISO 26262 ASIL-D, a serious injury rate of >90% requires the highest level of safety response.

[0061] In practical applications, the early warning threshold is dynamically adjusted based on historical data, and its expression is:

[0062] The adjusted threshold The threshold before adjustment, The base false alarm rate is the average false alarm rate over the past 30 days, and the attenuation factor is the baseline false alarm rate.

[0063] This invention achieves dynamic optimization of system parameters through a four-step closed loop of "detection-calibration-evaluation-feedback", the specific architecture of which is as follows.

[0064] a. Detection: The YOLOv8 model outputs foreign object detection results (category, pixel coordinates); b. Calibration: Based on the real-time homography matrix calculation of dynamic reference objects, pixel coordinates are mapped to the 3D world coordinate system; c. Assessment: The dynamic risk assessment model calculates the comprehensive risk value and triggers tiered early warnings; d. Feedback: Data such as false alarms / missed alarms, calibration errors, and risk level deviations are sent back to the training and calibration module.

[0065] When a system alert is confirmed as a false alarm by manual verification or through multi-sensor verification, the following data should be recorded: Image ROI region, detection confidence, and error category for false foreign object detection; current calibration parameters (homography matrix, camera extrinsic parameters) and environmental parameters.

[0066] Based on the false alarm data, the model retraining focuses on optimizing the following scenarios: enhancing the texture differentiation between metal fragments and plastic bags by introducing frequency domain features; and improving the robustness of lane line detection in low light by fusing infrared image data.

[0067] It should be noted that the above threshold adjustment formula matches the ISO 39001 practical scenario as follows.

[0068] Statistical Analysis and Dynamic Adjustment: The comparison between the baseline false alarm rate and the real-time false alarm rate in the formula is similar to the statistical methods recommended by ISO 39001, such as the mean and percentile methods. For example, the image stabilization threshold of a dashcam (i.e., set through similar dynamic logic). Risk Sensitivity Control: Attenuation Coefficient (Default 0.2) can be regarded as an adjustment parameter for system stability, avoiding frequent fluctuations in the threshold, and meeting the requirements of ISO 39001 for "balancing sensitivity and reliability".

[0069] Performance evaluation linkage: The adjusted results of the formula can be linked to the KPIs of ISO 39001, such as accident rate and false alarm rate, to support the quantitative performance evaluation required by the standard.

[0070] Compliance Statement: All parameter designs comply with the following specifications: Domestic: JTG series highway engineering standards, GA / T 994-2018 Road Traffic Incident Monitoring Specification International: ISO 39001 Road Traffic Safety Management System, US MUTCD Unified Traffic Control Equipment Manual.

[0071] In summary, this invention combines artificial intelligence for road intrusion detection with self-selected reference point calibration technology in traffic scenarios to achieve accurate identification, size measurement, spatiotemporal positioning, and dynamic risk assessment of road intrusions, thus overcoming the shortcomings of traditional methods in terms of accuracy, adaptability, and early warning efficiency.

Claims

1. A method for spatiotemporal early warning of road debris combining intelligent recognition and machine vision measurement, characterized in that, include: Acquire images of the road to be identified using a camera; An improved lightweight YOLOv8 model was used to identify and initially locate road debris in the acquired images. The camera is calibrated in real time using scene reference objects to obtain the homography matrix; The position and size of foreign objects in the image are initially calculated using the homography matrix to obtain visual measurement results; the visual measurement results are then fused with data from multiple sensors to obtain the accurate size and position of the foreign objects. Calculate the overall risk value of the foreign object and trigger a graded early warning based on the risk level corresponding to the overall risk value.

2. The method according to claim 1, characterized in that, The improved lightweight YOLOv8 model comprises a backbone network, a neck network, and a head network connected in sequence. The backbone network includes an input layer, a 1×1 convolutional layer, a first C2f layer, a 3×3 convolutional layer, a second C2f layer, a third C2f layer, a fourth C2f layer, and an SPPF layer connected in sequence. The neck network includes an upsampling layer, a first feature concatenation layer, a P2 layer, and a bidirectional weighted fusion layer connected in sequence. The neck network also includes a downsampling layer and a second feature concatenation layer connected in sequence. Furthermore, the neck network includes P2, P3, P4, and P5 output layers. The output of the SPPF layer serves as the input to the upsampling and downsampling layers, respectively. The output of the bidirectional weighted fusion layer serves as the input to the P2 and P3 output layers, respectively, and the output of the second feature concatenation layer serves as the input to the P3, P4, and P5 output layers, respectively.

3. The method according to claim 1, characterized in that, The specific method for real-time camera calibration using scene reference objects is as follows: Define a world coordinate system and set the height of the lane plane to 0; Assign precise 3D coordinates to the key corner points of a preset reference 3D model in the world coordinate system; The key point coordinates of the preset reference object in the road image to be identified are detected and matched one by one with the 3D coordinates of the reference object's 3D model to obtain the correspondence between 2D pixels and 3D world coordinates. The camera projection matrix is ​​solved by calculating the correspondence between 2D pixels and 3D world coordinates, and the homography matrix is ​​obtained to complete the calibration.

4. The method according to claim 3, characterized in that, During camera calibration, the homography matrix is ​​dynamically updated by combining vehicle speed with positional changes between consecutive frames. The specific method is as follows: Measure the vehicle speed and the time interval between consecutive frames, and update the translation vector, expressed as: in, The updated translation vector, The original translation vector, For vehicle speed, The time interval between frames. The unit vector is the direction of motion; The camera extrinsic parameters are corrected based on the updated translation vector, and its expression is as follows: in, This is the corrected camera extrinsic parameter matrix. For intermediate parameters, This is the camera extrinsic parameter matrix before correction; The homography matrix is ​​updated based on the corrected extrinsic parameter matrix, and its expression is as follows: in, The updated homography matrix, This is the camera intrinsic parameter matrix. , for The first two columns.

5. The method according to claim 4, characterized in that, The formula for calculating the comprehensive risk value of the foreign object is as follows: in, For the comprehensive risk value, This is the static risk value. This is a dynamic risk value. This is the road segment adaptive factor.

6. The method according to claim 5, characterized in that, The formula for calculating static risk value is: in, For the type of risk coefficient, For the equivalent size of the foreign object, For size threshold, For risk distance, This represents the maximum risk distance.

7. The method according to claim 6, characterized in that, The formula for calculating dynamic risk value is: in, For collision energy, For vehicle speed and density correction terms, This is the environmental compensation coefficient.

8. The method according to claim 7, characterized in that, Risk values ​​are categorized using different warning thresholds, including the following levels and corresponding response strategies: When the overall risk value is less than 0.2, the level is normal, and the response strategy is to record in the cloud. When the comprehensive risk value is less than 0.4, the level is "attention" and the response strategy is to provide a prompt via the vehicle terminal. When the comprehensive risk value is less than 0.6, the level is a warning, and the response strategy is a variable speed limit sign + a 20% speed limit. When the overall risk value is less than 0.8, the level is dangerous, and the response strategy is automatic speed limit + navigation route replanning. When the overall risk value is ≥0.8, the level is emergency, and the response strategy is to coordinate with traffic police to close roads and use drones for disposal.

9. The method according to claim 8, characterized in that, The warning threshold is dynamically adjusted based on historical data, and its expression is: The adjusted threshold The threshold before adjustment, The base false alarm rate is the average false alarm rate over the past 30 days, and the attenuation factor is the baseline false alarm rate.

10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the method described in any one of claims 1-9.

Images