Unmanned aerial vehicle-based highway ponding detection and early warning method

Abstract

The invention discloses a highway ponding detection and early warning method based on an unmanned aerial vehicle. Belongs to the technical field of pavement ponding detection. The implementation process is as follows: after the quad-rotor unmanned aerial vehicle takes off, the water accumulation condition of the highway pavement is acquired in real time through advanced sensor equipment such as a camera, an infrared thermal imager and a laser radar; then, performing optimal route planning of highway ponding position detection by adopting a PO algorithm; a YOLOv8 algorithm is adopted to classify and identify accumulated water in an acquired expressway image, and an iRMB attention module is used as a feature extractor. The quadrotor unmanned aerial vehicle carries a camera to carry out image data acquisition, and the system combines high maneuverability and quick response capability of the unmanned aerial vehicle to realize quick detection, positioning and early warning of accumulated water on the expressway; the expressway operation management level and the emergency handling capacity of emergencies can be improved, and safety guarantee is provided for expressway operation management units.

Description

Technical Field

[0001] The invention belongs to the technical field of road surface waterlogging detection and relates to a highway waterlogging detection and early warning method based on an unmanned aerial vehicle. Background Art

[0002] Maturity of UAV technology: After years of development, UAV technology has achieved high stability, endurance and intelligence. UAVs can carry various sensors and equipment to perform high-altitude, long-distance, high-definition photography and monitoring.

[0003] Demand for highway management: As an important infrastructure for transportation, the safety and smoothness of highways are of vital importance. However, natural disasters such as rainstorms and floods often lead to waterlogging on highways, seriously affecting driving safety. Therefore, timely and accurate detection and early warning of waterlogging on highways are particularly important.

[0004] Limitations of traditional detection methods: Traditional methods of waterlogging detection, such as manual inspections and ground sensors, have problems such as low efficiency, limited coverage, and susceptibility to interference. Drone technology can overcome these limitations and achieve efficient and accurate waterlogging detection. Summary of the invention

[0005] In view of the above problems, the purpose of the present invention is to propose a highway waterlogging detection and early warning method based on drones.

[0006] The technical solution of the present invention is: a highway water accumulation detection and early warning method based on drones described in the present invention is realized based on a quad-rotor drone equipped with a laser radar and a camera, and its implementation steps are as follows:

[0007] Step 1: UAV selection: Select a UAV suitable for highway inspection and ensure that the UAV has stable flight performance and high-definition image acquisition capabilities;

[0008] Sensor integration: Equip the drone with a high-precision water level sensor to monitor water level changes in real time;

[0009] Map and route planning: Plan the flight routes and inspection points of drones based on the geographical characteristics of the highway and areas prone to waterlogging;

[0010] Step 2: Drone takes off: According to the predetermined flight route and inspection points, the drone takes off autonomously and performs the inspection task;

[0011] Real-time monitoring: The water level sensor carried by the drone monitors the water accumulation on the highway in real time, while the drone's high-definition camera captures images of the waterlogged area; the track planning for waterlogging detection is completed based on the PO algorithm;

[0012] Data transmission: The drone transmits the water level data and image information monitored in real time to the ground control station or data center;

[0013] Step 3: Data reception: The ground control station or data center receives the data transmitted by the drone;

[0014] Data analysis: Analyze the received image information using computer vision and deep learning algorithms to identify the location, scope and depth of the waterlogged area;

[0015] Warning judgment: Based on the analysis results and the preset warning threshold, determine whether it is necessary to issue a waterlogging warning;

[0016] Optimization algorithm: YOLOv8 algorithm is used to classify and identify water accumulation in the collected highway images, and iRMB attention module is used as feature extractor;

[0017] Step 4: Warning release: Once it is determined that a waterlogging warning is necessary, the warning information will be released through SMS, display screens and other channels;

[0018] Emergency response: The driver quickly activates the emergency response mechanism based on the early warning information and takes necessary control measures to ensure his own safety.

[0019] Furthermore, the laser radar in step (1) completes data collection on the flooded area of ​​the highway, and the specific method is: select a laser radar system suitable for road water detection, which should have high sensitivity, high resolution and real-time data processing capabilities; ensure that the laser radar system is equipped with a dual-echo or multi-echo mode to more accurately measure the depth of water accumulation; build a digital model, and then complete the track planning of the highway water accumulation situation according to the PO algorithm.

[0020] Furthermore, regarding the objective function setting of the UAV trajectory planning, the constraints of flight altitude, pitch angle, and turning angle must be taken into account.

[0021] The objective function can be expressed as:

[0022] f=p1f1+p2f2+p3f3

[0023] In the formula, f1, f2, and f3 represent the range cost, threat cost, and altitude cost respectively; the weight of the objective function is adjusted by adjusting p1, p2, and p3;

[0024] Range cost: Considering the fuel limit of the UAV, a shorter range means that the UAV can complete the mission before the fuel is consumed; the range cost function f1 can be expressed as:

[0025]

[0026] In the formula, ||l e l e+1 || represents the distance between the e-th flight node and the e+1-th flight node; S represents the total number of flight nodes;

[0027] Threat cost: The drone must avoid existing obstacles when working; the threat cost function f2 is defined as:

[0028]

[0029] Where, d e is the danger level of the e-th flight node;

[0030] Height cost: To ensure safe flight of the drone and avoid collision with the ground, the drone must maintain a certain safe distance from the ground, and the flight altitude must not exceed the maximum flight altitude limit. The height cost function g3 is defined as:

[0031]

[0032] Where p1 and p2 are proportional functions;

[0033] Flight altitude constraint: When a drone is flying, its flight altitude must be restricted in consideration of safety and relevant airspace regulations. The maximum and minimum flight altitudes are set to h max and h min , the flight altitude should meet the following requirements:

[0034] h min ≤h≤h max

[0035] Pitch angle constraint: Set the maximum pitch angle to γ max , the flight pitch angle should satisfy: -γ max ≤γ≤γ max , the flight angle calculation formula is:

[0036]

[0037] Turning angle constraint: The turning angle of the drone should be as smooth as possible during the turning process and cannot exceed its maximum allowed turning angle β max , the turning angle formula is as follows:

[0038]

[0039] f1=(x e -x e-1 )(x e+1 -x e )+(y e -y e-1 )(y e+1 -ye )

[0040]

[0041] (x e ,y e ,z e ) and (x e+1 ,y e+1 ,z e+1 ) represent the e-th and e+1-th track nodes of the UAV flight aircraft respectively.

[0042] Furthermore, the PO algorithm has four stages:

[0043] Phase 1: In the PO algorithm, in the first three iterations, exploration and exploitation operations are performed simultaneously until initialization is completed in the phase transition phase; that is:

[0044]

[0045] In the formula, Sep Cost The variable value is related to each stage of mining and exploration. Time It is also a variable with a constant value, usually 1; PF1 and PF2 are parameters with fixed values ​​that must be set before the optimization process and are used to prioritize each of the f1 and f2 functions;

[0046] Phase 2: In this phase, three different functions f1, f2, and f3 are used for scoring;

[0047] The first function emphasizes the upgrading part and leads to the prioritization of one of the two phases of exploration and development, which is chosen and executed better than the other; the first function places more emphasis on the exploration phase;

[0048]

[0049] In the formula, and represents the amount of the first function based on the exploration phase or the development phase, where t represents the current iteration number; and is the cost of the best solution before improvement in the current choice; on the other hand, and is the cost of the best solution obtained by improving the current choice; T t explore and T t exploitis the number of unselected iterations from the last selection to the current selection; PF1 is a user-adjustable parameter that must be set to a value between 0 and 1 before the optimization operation; this parameter determines the importance of the first function; and is prioritized as the value of this function increases, and decreases as its priority decreases;

[0050] The second function also emphasizes the resonant components and causes the phase to perform better than other preferred phases. Good performance is checked and measured in sequence; therefore, this function can help in its selection during the development phase.

[0051] Phase 3: The drones randomly jump into the search space to search for water accumulation; first, the entire population is sorted in ascending order, and then its solution is improved using the following formula:

[0052] If rand1>0.5,Z i,G =R Dim *(Ub-Lb)+Lb

[0053] Otherwise

[0054] Z i,G =X a,G +G.(X a,G -X b,G )+G.(((X a,G -X a,G )-(X c,G -X d,G ))+((X c,G -X d,G )-(X e,G -X f,G )))

[0055] G = 2.rand2-1

[0056] Where Ub and Lb are the lower and upper bounds of the problem, R Dim is a randomly generated number between 0 and 1, and in the dimensions of the problem; rand1 is also a randomly generated number between 0 and 1; X aG , X bG , X cG , X dG , X eG and X fG is the solution in the whole population, which are randomly selected faces; G is calculated using the second formula;

[0057] Among them, rand2 is a randomly generated number uniformly distributed between 0 and 1; according to the above formula, one of the two equations is selected to generate a different solution, and then the generated new solution is applied to improve the current solution;

[0058] Phase 4:

[0059] During the development phase, the PO algorithm was improved using two different operators, which were simulated using the following equation for the behavior;

[0060]

[0061] The formula shown above shows the two strategies used in PO; considering that in the mountain lion, in order to hunt in the first mode, in the development phase, for the running and ambush strategies, the situation in the formula is used, the operation is carried out using a drone to simulate the rapid running of the mountain lion towards the prey.

[0062] Furthermore, in step (3), the YOLOv8 algorithm is used to classify and identify the water accumulation in the highway image. Since the water collection area is small and difficult to detect and the shape is irregular, the iRMB has a strong feature extraction capability and can effectively extract the features of the target, while improving the robustness and detection accuracy. The specific steps are as follows:

[0063] Step 1: Collect and annotate highway images of flooded areas and mark the bounding boxes of flooded areas.

[0064] Step 2: Accurately mark the water accumulation area in the image, usually as a rectangular box;

[0065] Step 3: Perform data enhancement on the image to improve the generalization ability of the model;

[0066] Step 4: Select YOLOv8 as the base model and load the pre-trained weights;

[0067] Step 5: Introduce the iRMB module into the feature extraction network of YOLOv8 to enhance the model’s ability to learn waterlogging features.

[0068] Furthermore, the YOLOv8 is an advanced object detection framework, the core of which is a deep convolutional neural network for predicting the category and location of the object directly from the input image;

[0069] The architecture of YOLOv8 can be divided into three main parts: input layer, feature extraction layer and detection layer; the main task of the input layer is to preprocess the input image data for subsequent network processing; this includes scaling the image to meet the requirements of the network input size, and normalization to reduce the impact of illumination changes and color deviations on the model;

[0070] The feature extraction layer is the core of YOLOv8, responsible for extracting useful features from preprocessed images. This layer is usually composed of multiple convolutional layers, each of which applies a set of filters to extract specific types of features, such as edges, textures, or more complex patterns. As the network level deepens, the extracted features gradually transition from simple to complex, forming a feature hierarchy from shallow to deep.

[0071] The detection layer is the last part of YOLOv8, which is responsible for converting the features extracted by the feature extraction layer into specific detection results. This layer uses a series of convolutional layers to predict the bounding box, category, and confidence of the target. YOLOv8 uses an anchor box mechanism to predict the position and size of the bounding box, and uses category-specific confidence scores to evaluate the reliability of the detection results. Finally, non-maximum suppression (NMS) is used to remove overlapping detections to ensure that each target is detected only once.

[0072] Furthermore, the inverse residual attention module (iRMB) is to enhance the model's ability to learn key features, especially in complex backgrounds or multi-target scenes; the design of iRMB combines residual learning and attention mechanisms to improve the network's ability to recognize and respond to aircraft features; the structure of iRMB can be expressed as:

[0073] IRMB(x)=ReLU(x+λ.F(x,{W i}))

[0074] Where x is the input feature, F(x,{W i}) is a series of convolution operations applied to x, w is the set of weights of the convolution layer, and λ is a learnable scaling parameter used to adjust the contribution of the residual part;

[0075] In iRMB, F represents a residual block containing an attention mechanism; the design of this residual block can be further refined as follows:

[0076] F(x,{W i})=W3*σ(W2*ReLU(W1*x))

[0077] Where * represents the convolution operation, W1, W2, and W3 are the weights of different convolutional layers, and σ is the Sigmoid activation function, which is used to implement the attention mechanism.

[0078] In iRMB, the role of the attention mechanism is to emphasize the parts of the input features that are relevant to aircraft detection; this is achieved through the Sigmoid activation function, which generates an attention map between 0 and 1 for weighting the input features;

[0079] Specifically, the attention map is generated by learning the global distribution of input features, which can be expressed as:

[0080] A(x)=σ(W A *x)

[0081] Where A(x) is the generated attention map, W A is the learned attention weight;

[0082] The introduction of iRMB enables the network to focus more effectively on features that are critical to the aircraft detection task; by combining residual learning and attention mechanism, iRMB not only improves the expressiveness of features, but also enhances the network's ability to recognize aircraft features in complex backgrounds and multi-target environments; in addition, the design of iRMB also takes computational efficiency into consideration, ensuring that the model does not excessively increase the computational burden while improving performance.

[0083] The beneficial effects of the present invention are: the system combines the high maneuverability and rapid response capabilities of drones to achieve rapid detection, positioning and early warning of waterlogging on highways; it can improve the level of highway operation management and emergency response capabilities for sudden incidents, and provide safety guarantees for highway operation management units; the drone-based highway waterlogging detection and early warning method improves the efficiency and accuracy of waterlogging detection, enhances the timeliness and effectiveness of early warning, ensures traffic safety, reduces losses, and improves the level of emergency management. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is an operation flow chart of the present invention. DETAILED DESCRIPTION

[0085] The specific technical solutions of the present invention are further described in detail below in conjunction with specific examples.

[0086] As shown in the figure, the highway water accumulation detection and early warning method based on drone of the present invention is realized based on a four-rotor drone equipped with a laser radar and a camera (DJI action4 camera). The implementation process is as follows: after the four-rotor drone takes off, the water accumulation situation on the highway road surface is obtained in real time through advanced sensor equipment such as cameras, infrared thermal imagers, and laser radars; then the PO algorithm is used to perform optimal track planning for highway water accumulation position detection; and the YOLOv8 algorithm is used to classify and identify the water accumulation in the collected highway images, and the iRMB attention module is used as a feature extractor; the camera is carried out to collect highway image data, and then the image data is classified and identified; the four-rotor drone is equipped with a camera to collect image data;

[0087] S1. Select a LiDAR system suitable for road water detection, which should have high sensitivity, high resolution and real-time data processing capabilities; ensure that the LiDAR system is equipped with dual-echo or multi-echo mode to more accurately measure the depth of water accumulation;

[0088] Build a digital model and then complete the trajectory planning for highway flooding based on the PO algorithm;

[0089] S2: Complete the path planning for the flooded location on the highway according to the PO algorithm; the UAV has a total of 4 stages in the PO algorithm:

[0090] Phase 1: In the PO algorithm, in the first three iterations, exploration and development operations are performed simultaneously until initialization is completed in the phase transition phase:

[0091]

[0092] Where, the variable value of SeqCost is related to each stage of mining and exploration, SeqTime is also a constant value variable, which is usually 1; PF1 and PF2 are parameters with fixed values, which must be set before the optimization process and are used to prioritize each of the f1 and f2 functions;

[0093] Phase 2: In this phase, scoring is performed using three different functions f1, f2, and f3; the first function emphasizes the upgrading part and leads to the prioritization of one of the two phases of exploration and exploitation, which is selected and executed better than the other;

[0094] The first function places more emphasis on the exploration phase;

[0095]

[0096] In the formula, and represents the amount of the first function based on the exploration phase or the development phase, where t represents the current iteration number; and is the cost of the best solution before improvement in the current choice; on the other hand, and is the cost of the best solution obtained by improving the current choice; T t explore and T t exploitis the number of unselected iterations from the last selection to the current selection; PF1 is a user-adjustable parameter that must be set to a value between 0 and 1 before the optimization operation; this parameter determines the importance of the first function; and it is prioritized as the value of this function increases, and decreases as its priority decreases; the second function also emphasizes the resonant component and causes the phase to perform better than the other prioritized phases; good performance is checked and measured in sequence; therefore, this function can help to select it during the development phase;

[0097] Phase 3: The drones randomly jump into the search space to search for water accumulation; first, the entire population is sorted in ascending order, and then its solution is improved using the following formula:

[0098] If rand1>0.5,Z i,G =R Dim *(Ub-Lb)+Lb

[0099] Otherwise

[0100] Z i,G =X a,G +G.(X a,G -X b,G )+G.(((X a,G -X a,G )-(X c,G -X d,G ))+((X c,G -X d,G )-(X e,G -X f,G )))

[0101] G = 2.rand2-1

[0102] Where Ub and Lb are the lower and upper bounds of the problem, R Dim is a randomly generated number between 0 and 1, and in the dimensions of the problem; rand1 is also a randomly generated number between 0 and 1; X aG , X bG , X cG , X dG , X eG and X fG is the solution in the whole population, which are randomly selected faces; G is calculated using the second formula; where rand2 is a randomly generated number uniformly distributed between 0 and 1; According to the above formula, one of the two equations is selected to generate a different solution, and then the generated new solution is applied to improve the current solution;

[0103] Phase 4:

[0104] During the development phase, the PO algorithm was improved using two different operators, which were simulated using the following equation for this behavior:

[0105]

[0106] The formula shown above shows the two strategies used in PO; considering that in the case of the mountain lion, for hunting in the first mode, in the development phase, for the running and ambush strategies, the situation in the formula is used, which is carried out so that the drone simulates the rapid running of the mountain lion towards the prey;

[0107] Step 1: Construct highway terrain model;

[0108] Step 2: Determine the constraints of the drone’s trajectory planning based on the drone’s own conditions (battery, load, etc.);

[0109] Step 3: Clean, denoise and correct the collected data to ensure the accuracy of the data;

[0110] Step 4: Use the preprocessed data to train the algorithm model, including parameter tuning, feature selection, etc.

[0111] Step 5: Integrate the trained algorithm model into the drone, such as algorithm loading and sensor calibration;

[0112] Step 6: Find the optimal path and improve or replan the trajectory according to the four stages of the PO algorithm;

[0113] Step 7: Output the final collected data, process and analyze it to ensure the effectiveness of trajectory planning;

[0114] S3. Principle of water accumulation identification:

[0115] From the perspective of plane images, the characteristics of the horizontal plane are that the grayscale is relatively light and the contour is relatively smooth; however, in practical applications, due to the influence of factors such as dust and light, it is difficult to accurately extract the horizontal plane based on grayscale or contour features alone; therefore, these two methods are combined to improve the recognition and extraction accuracy of waterlogging points; MATLAB software is used to perform image grayscale and binarization operations on the waterlogging area, which realizes the automatic recognition of the waterlogging area and can accurately locate the waterlogging area;

[0116] Image grayscale: A grayscale image is composed of a set of data matrices. The values ​​of the matrix represent the depth of the grayscale, usually in [0, 255], with white being 255 and black being 0. In order to better reflect the grayscale characteristics of the water body, the collected color image needs to be grayscaled.

[0117] Image grayscale change: On this basis, a target recognition method based on grayscale transformation is proposed; the imadju function in the MATLAB toolkit can complete this operation, and its general grammar is g=imadjust(f,[lowin,highin],[lowout,highout]); this function sets the grayscale of the original image lower than 255*highout to low output, and the grayscale higher than 255*highout to high output;

[0118] Image binarization: The goal of binarization is to reduce redundant data, so as to obtain clearer image features. It is very important to select a suitable threshold to binarize the image, and the result will directly affect the quality of the image. Among them, the identification of water points requires high speed and accuracy.

[0119] The traditional average method and threshold setting method are not ideal, and the bimodal method and iterative method require a long calculation time. Therefore, when determining the threshold, the following methods should be selected based on experience:

[0120]

[0121] Where T is the threshold value to be sought, MAX is the maximum value of the image matrix to be processed, and MIN is the minimum value of the image matrix to be processed.

[0122] Image denoising: Mean filtering can effectively smooth out impulse noise while protecting sharp edges of the image, making the image clearer and more complete. Mean filtering is a technique that directly performs smoothing on the image space domain. The imdilate function in the MATLAB toolbox can implement this operation. This process reduces sharp changes in the grayscale of the image. In a noisy digital image, noise can be seen as sharp changes in grayscale, so smoothing can reduce noise.

[0123] S4. Determination of water accumulation:

[0124] Statistical connected domain shape features: calculate the ellipticity e, density C and area A of the connected domain;

[0125] Among them, ellipticity and compactness are defined as:

[0126]

[0127] In the formula, I max I represents the length of the major axis of the ellipse with the same standard second-order center distance as the region; min represents the length of the minor axis of the ellipse with the same standard second-order center distance as the region; A represents the area of ​​the connected domain;

[0128] Judgment conditions: According to the shape characteristics of the water accumulation, select appropriate thresholds T for e, C, and Ae , T C , T A , according to the following formula, determine whether the connected domain R is a waterlogged area;

[0129]

[0130] In the formula, A total represents the total area of ​​the image, T total is the threshold of the proportion of the non-integral connected domain to the total image area; the first discrimination condition is to eliminate the influence of road markings. When there are road markings, the road area except the markings will be extracted by binary value, which will increase the amount of calculation, so the area of ​​the connected domain and the total image area A are used to calculate the ratio of the non-integral connected domain to the total image area. total Compare the two, eliminate the influence of marking lines, and improve the operation speed;

[0131] S5. Use the YOLOv8 algorithm to classify and identify waterlogging in highway images. Since the waterlogging area is small and difficult to detect and the shape is irregular, iRMB has strong feature extraction capabilities and can effectively extract the features of the target, while improving robustness and detection accuracy.

[0132] Step 1: Collect and annotate highway images of flooded areas and mark the bounding boxes of flooded areas.

[0133] Step 2: Accurately mark the water accumulation area in the image, usually as a rectangular box;

[0134] Step 3: Perform data enhancement on the image to improve the generalization ability of the model;

[0135] Step 4: Select YOLOv8 as the base model and load the pre-trained weights;

[0136] Step 5: Introduce the iRMB module into the feature extraction network of YOLOv8 to enhance the model’s ability to learn waterlogging features.

[0137] The inverse residual attention module (iRMB) is designed to enhance the model's ability to learn key features, especially in complex backgrounds or multi-target scenes; the design of iRMB combines residual learning and attention mechanisms to improve the network's ability to recognize and respond to aircraft features; the structure of iRMB can be expressed as:

[0138] IRMB(x)=ReLU(x+λ.F(x,{W i}))

[0139] Where x is the input feature, F(x,{W i}) is a series of convolution operations applied to x, w is the set of weights of the convolution layer, and λ is a learnable scaling parameter used to adjust the contribution of the residual part;

[0140] In iRMB, F represents a residual block containing an attention mechanism; the design of this residual block can be further refined as follows:

[0141] F(x,{W i})=W3*σ(W2*ReLU(W1*x))

[0142] In the formula, * represents the convolution operation, W1, W2, and W3 are the weights of different convolutional layers, and σ is the Sigmoid activation function, which is used to implement the attention mechanism;

[0143] In iRMB, the role of the attention mechanism is to emphasize the parts of the input features that are relevant to aircraft detection; this is achieved through the Sigmoid activation function, which generates an attention map between 0 and 1 for weighting the input features;

[0144] Specifically, the attention map is generated by learning the global distribution of input features, which can be expressed as:

[0145] A(x)=σ(W A *x)

[0146] Where A(x) is the generated attention map, W A is the learned attention weight;

[0147] The introduction of iRMB enables the network to focus more effectively on features that are critical to the aircraft detection task; by combining residual learning and attention mechanism, iRMB not only improves the expressiveness of features, but also enhances the network's ability to identify aircraft features in complex backgrounds and multi-target environments; in addition, the design of iRMB also takes computational efficiency into consideration, ensuring that the model will not excessively increase the computational burden while improving performance;

[0148] S6: The specific steps of YOLOv8 algorithm combined with iRMB are:

[0149] Step 1: The UAV infrared thermal imager collects road surface images;

[0150] Step 2: Complete the trajectory planning of the UAV according to the PO algorithm;

[0151] Step 3: Use YOLOv8 as the basic network structure;

[0152] Step 4: Use the YOLOv8 algorithm to classify and identify the water in the collected highway images, and use the iRMB attention module as a feature extractor;

[0153] Step 5: Use the neural network structure of the YOLOv8 algorithm to train the input and output images;

[0154] Step 6: Identify and determine the accumulated water;

[0155] Step 7: Output feature image;

[0156] S7. Early warning method design:

[0157] Electronic map warning: After the remote server completes the highway image collected by the drone, the geographical location information contained in the road image taken by the drone is used to mark the corresponding road section on the remote server, and the road section with accumulated water is marked with a red line, so as to warn the driver to drive carefully;

[0158] Electronic display screen warning: Install electronic display screens at the entrances of highways or places where water accumulates frequently. The electronic display screens will show the water accumulation on the road, warning drivers that water accumulation on the road section may cause danger, and reminding drivers to pay attention to driving safety. Especially after rain, when drivers are not vigilant, using electronic display screens is more intuitive.

[0159] SMS alert:

[0160] "Community SMS" refers to a value-added service that sends specific short messages to specific users in specific areas and specific time periods. The cell division in cellular short messages can be determined by users based on business needs and network conditions. "Community SMS" can be used to notify passengers who have not yet arrived of road flooding to enhance their driving awareness and serve as an early warning.

Claims

1. A highway waterlogging detection and early warning method based on drones, characterized in that: The implementation steps are as follows: Step (1): UAV selection: Select a UAV suitable for highway inspection and ensure that the UAV has stable flight performance and high-definition image acquisition capabilities; Step (2): UAV takes off: According to the predetermined flight route and inspection points, the UAV takes off autonomously and performs the inspection task; Step (3): Data reception: The ground control station or data center receives the data transmitted by the drone; Step (4): Warning issuance: Once it is determined that a waterlogging warning is necessary, the warning information is issued through SMS and display screens.

2. The method for detecting and warning waterlogging on a highway based on an unmanned aerial vehicle according to claim 1, characterized in that: In step (1), the high-definition image acquisition capability is: Sensor integration: Equip the drone with a high-precision water level sensor to monitor water level changes in real time; Map and route planning: Plan the flight routes and inspection points of drones based on the geographical characteristics of the highway and areas prone to waterlogging; In addition, the specific method of completing data collection on the flooded area of ​​the highway is: Choose a LiDAR system suitable for road water detection and ensure that the LiDAR system is equipped with dual-echo or multi-echo mode to measure the depth of water; A digital model is constructed, and then the trajectory planning for highway waterlogging is completed based on the PO algorithm.

3. The method for detecting and warning waterlogging on a highway based on an unmanned aerial vehicle according to claim 1, characterized in that: In step (1), the objective function for the UAV trajectory planning is set according to the constraints of flight altitude, pitch angle and turning angle. The objective function is expressed as: f=p1f1+p2f2+p3f3 In the formula, f1, f2, and f3 represent the range cost, threat cost, and altitude cost respectively; the weight of the objective function is adjusted by adjusting p1, p2, and p3; Voyage cost: According to the fuel limit of the UAV, a shorter range means that the UAV can complete the mission before the fuel is consumed; the range cost function f1 is expressed as: In the formula, ||l e l e+1 || represents the distance between the e-th flight node and the e+1-th flight node; S represents the total number of flight nodes; Threat Cost: The drone needs to avoid obstacles when working, and the threat cost function f2 is defined as: Where, d e is the danger level of the e-th flight node; Height cost: In order to ensure the safety of the UAV flight and the flight altitude cannot exceed the maximum flight altitude limit, the altitude cost function g3 is defined as: Where p1 and p2 are proportional functions; Flight altitude constraints: When a drone is flying, considering safety and relevant airspace regulations, its flight altitude is restricted, with the maximum and minimum flight altitudes set at h and h respectively. max and h min , the flight altitude should meet the following requirements: h min ≤h≤h max Pitch angle constraint: Set the maximum pitch angle to γ max , the flight pitch angle should satisfy: -γ max ≤γ≤γ max , the flight angle calculation formula is: Turn Angle Constraint: The turning angle of the drone should be smooth during the turning process and should not exceed its maximum allowed turning angle β max , the turning angle formula is as follows: f1=(x e -x e-1 )(x e+1 -x e )+(y e -y e-1 )(y e+1 -y e ) In the formula, (x e ,y e ,z e ) and (x e+1 ,y e+1 ,z e+1 ) represent the e-th and e+1-th track nodes of the UAV flight aircraft respectively.

4. The method for detecting and warning waterlogging on a highway based on an unmanned aerial vehicle according to claim 1, characterized in that: In step (2), the drone takes off autonomously and performs the inspection task as follows: Real-time monitoring: The water level sensor carried by the drone monitors the water accumulation on the highway in real time, while the drone's high-definition camera captures images of the waterlogged area; the track planning for waterlogging detection is completed based on the PO algorithm; Data transmission: The drone transmits the monitored water level data and image information to the ground control station or data center in real time.

5. The method for detecting and warning waterlogging on a highway based on an unmanned aerial vehicle according to claim 4 is characterized in that: The PO algorithm has four stages: Phase 1: In the PO algorithm, in the first three iterations, exploration and exploitation operations are performed simultaneously until initialization is completed in the phase transition phase; that is: In the formula, Sep Cost The variable value is related to each stage of mining and exploration. Time is a constant-valued variable, taking the value 1; PF1 and PF2 are parameters with fixed values, which must be set before the optimization process and are used to prioritize each of the f1 and f2 functions; Phase 2: In this phase, scoring is performed using three different functions f1, f2 and f3; the first function emphasizes the upgrading part and leads to the prioritization of one of the two phases of exploration and development; The first function places more emphasis on the exploration phase; Right now: In the formula, and represents the quantity of the first function based on the exploration phase or the development phase, where t represents the current iteration number; and is the cost of the best solution before improvement in the current choice; on the other hand, and is the cost of the best solution obtained by improving the current choice; and is the number of unselected iterations from the previous selection to the current selection; PF1 It is a user-adjustable parameter that must be set to a value between 0 and 1 before performing the optimization operation; This parameter determines the importance of the first function; it is prioritized as the value of this function increases, and decreases as its priority decreases; the second function also emphasizes the resonant component and causes the phase to perform better than the other prioritized phases; good performance is checked and measured in order; Phase 3: The drone randomly jumps into the search space to search for accumulated water; First, the entire population is sorted in ascending order, and then the solution is improved using the following formula: If rand1>0.5,Z i,G =R Dim *(Ub-Lb)+Lb Otherwise Z i,G =X a,G +G.(X a,G -X b,G )+G.(((X a,G -X a,G )-(X c,G -X d,G ))+((X c,G -X d,G )-(X e,G -X f,G ))) G = 2.rand2-1 Where Ub and Lb are the lower and upper bounds of the problem, R Dim is a randomly generated number between 0 and 1, and in the dimension of the problem; rand1 is also a randomly generated number between 0 and 1; X aG , X bG , X cG , X dG , X eG and X fG is the solution in the whole population, which are randomly selected faces; G is calculated using the second formula; where rand2 is a randomly generated number uniformly distributed between 0 and 1; According to the above formula, one of the two equations is selected to produce a different solution, and then the generated new solution is applied to improve the current solution; Phase 4: During the development phase, the PO algorithm was improved using two different operators, which were simulated using the following equation for the behavior; namely: The formula shown above shows the two strategies used in PO; considering that in the mountain lion, for hunting in the first mode, in the development phase, for the running and ambush strategies, using the situation in the formula, the operation is carried out using a drone to simulate the mountain lion's fast running towards the prey.

6. The method for detecting and warning waterlogging on a highway based on a drone according to claim 1 is characterized in that: In step (3), the data received from the drone transmission is specifically: Data analysis: Analyze the received image information using computer vision and deep learning algorithms to identify the location, scope and depth of the waterlogged area; Warning judgment: Based on the analysis results and the preset warning threshold, determine whether it is necessary to issue a waterlogging warning; Optimization algorithm: The YOLOv8 algorithm is used to classify and identify water accumulation in the collected highway images, and the iRMB attention module is used as the feature extractor.

7. The method for detecting and warning waterlogging on a highway based on a drone according to claim 6, characterized in that: The YOLOv8 algorithm is used to classify and identify water accumulation in highway images. Due to the strong feature extraction capability of iRMB, it can improve robustness and detection accuracy while extracting the features of the target; The specific steps are as follows: Step 1: Collect and annotate highway images of flooded areas and mark the bounding boxes of flooded areas. Step 2: Accurately mark the water accumulation area in the image, usually as a rectangular box; Step 3: Perform data enhancement on the image to improve the generalization ability of the model; Step 4: Select YOLOv8 as the base model and load the pre-trained weights; Step 5: Introduce the iRMB module into the feature extraction network of YOLOv8 to enhance the model’s ability to learn waterlogging features.

8. The method for detecting and warning waterlogging on a highway based on an unmanned aerial vehicle according to claim 7, characterized in that: The YOLOv8 is a deep convolutional neural network; Its architecture is divided into three parts: input layer, feature extraction layer and detection layer; The task of the input layer is to preprocess the input image data, including scaling the image to meet the network input size requirements and normalizing it to reduce the impact of illumination changes and color deviations on the model; The feature extraction layer is the core of YOLOv8 and is responsible for extracting useful features from the preprocessed image. This layer consists of multiple convolutional layers, each of which applies a set of filters to extract a specific type of feature. As the network level deepens, the extracted features gradually transition from simple to complex, forming a feature hierarchy from shallow to deep; The detection layer is responsible for converting the features extracted by the feature extraction layer into specific detection results; this layer uses a series of convolutional layers to predict the bounding box, category and confidence of the target; YOLOv8 uses an anchor box mechanism to predict the position and size of the bounding box, and uses category-specific confidence scores to evaluate the reliability of the detection results; finally, non-maximum suppression is used to remove overlapping detections to ensure that each target is detected only once.

9. The method for detecting and warning waterlogging on a highway based on an unmanned aerial vehicle according to claim 7, characterized in that: The iRMB module combines residual learning and attention mechanism to improve the network's ability to recognize and respond to aircraft features; the structure of the iRMB module is expressed as: IRMB(x)=ReLU(x+λ.F(x,{W i })) Where x is the input feature, F(x,{W i }) is a series of convolution operations applied to x, w is the set of weights of the convolution layer, and λ is a learnable scaling parameter used to adjust the contribution of the residual part; In the iRMB module, F represents a residual block containing an attention mechanism; this residual block is refined as follows: F(x,{W i })=W3*σ(W2*ReLU(W1*x)) In the formula, * represents the convolution operation, W1, W2, and W3 are the weights of different convolutional layers, and σ is the Sigmoid activation function, which is used to implement the attention mechanism; In addition, in the iRMB module, the role of the attention mechanism is to emphasize the part of the input features related to aircraft detection; It is implemented through the Sigmoid activation function, which generates an attention map between 0 and 1 to weight the input features; Specifically, the attention map is generated by learning the global distribution of input features, expressed as: A(x)=σ(W A *x) Where A(x) is the generated attention map, W A are the learned attention weights.

10. The method for detecting and warning waterlogging on a highway based on a drone according to claim 1, characterized in that: In step (4), the warning release information is specifically: Emergency response: The driver quickly activates the emergency response mechanism based on the early warning information and takes necessary control measures to ensure his own safety.

Images