Compound wing unmanned aerial vehicle highway low disturbance inspection method and system based on traffic flow prediction

Abstract

The application discloses a composite wing unmanned aerial vehicle highway low-disturbance inspection method and system based on traffic flow prediction, and aims to solve the problem that the existing unmanned aerial vehicle inspection is easily affected by traffic flow disturbance and may interfere with normal traffic. The method comprises the following steps: acquiring real-time traffic flow data of multi-source fusion; constructing a three-dimensional traffic flow disturbance field prediction model based on a space-time graph convolutional neural network; combining the performance constraints of the unmanned aerial vehicle, and planning a low-disturbance dynamic flight path by using an improved rapid expansion random tree algorithm; performing adaptive flight and triggering local re-planning; and calculating a traffic disturbance index according to a ground vehicle behavior index, and adjusting the flight height and / or lateral offset amount of the composite wing unmanned aerial vehicle in a closed loop. Through the above technical scheme, the application realizes prospective disturbance avoidance, reduces the disturbance to traffic flow, improves imaging stability and inspection safety, and supports multi-machine cooperative and efficient operation.

Description

Technical Field

[0001] This invention relates to the field of collaborative control of intelligent transportation and unmanned systems, specifically to a low-disturbance highway inspection method and system based on traffic flow prediction using a compound-wing unmanned aerial vehicle. Background Technology

[0002] With the deep integration of intelligent transportation systems and drone technology, using drones for highway inspections has become an important direction for improving the level of intelligent road maintenance. Traditional drone inspections are mostly geared towards static infrastructure such as power lines and oil and gas pipelines. Their operating environment is relatively closed and has few interference factors, making it difficult to adapt to the complex flight requirements of highways, which are open, dynamic, and have high-density traffic flow. Especially when performing missions over highways, drones not only face natural wind disturbances but also need to cope with wake turbulence generated by high-speed vehicles, sudden airflow changes caused by traffic jams, and temporary traffic fluctuations caused by emergencies. These factors significantly weaken flight stability, reduce image acquisition quality, and may cause visual interference to drivers on the ground or even pose safety hazards.

[0003] However, existing inspection systems generally lack the ability to perceive and model ground traffic flow. For example, while some solutions incorporate meteorological information to optimize flight paths, they fail to integrate key traffic parameters such as traffic volume, speed distribution, and lane occupancy rates into the path planning decision-making system, making it impossible to predict and avoid high-disturbance areas. Another approach relies on infrastructure such as fixed robotic arms for stationary hovering, which extends monitoring time but is severely limited by the linear, non-fixed-attachment-point physical characteristics of highways and completely ignores the potential psychological and behavioral disturbances caused to traffic participants by drones during low-altitude flight. Furthermore, current path planning is mostly based on static or historical traffic data, making it difficult to respond to the rapid evolution of real-time traffic conditions, resulting in flight strategies lagging behind actual disturbance sources and failing to achieve truly "low-disturbance" inspections.

[0004] Furthermore, the impact of vehicle wake and sudden changes in traffic flow on UAVs is not limited to the road plane but attenuates and spreads with lateral distance and flight altitude. Existing solutions, even those utilizing traffic flow data for path planning, typically treat traffic conditions as two-dimensional congestion or risk heatmaps, lacking a mechanism to map lane interactions, vehicle types, speed differences, and flight altitude together into a three-dimensional disturbance field. This makes it difficult to provide executable three-dimensional low-disturbance flight paths for compound-wing UAVs. Simultaneously, existing UAV inspection path planning often prioritizes UAV safety or imaging quality, lacking a closed-loop control mechanism for traffic disturbances based on ground vehicle behavior feedback. This prevents real-time adjustments to flight altitude and route deviation based on abnormal lane changes, lateral acceleration fluctuations, or emergency braking during inspections. Therefore, it is necessary to provide a low-disturbance highway inspection method that combines traffic disturbance prediction, UAV path planning, and traffic disturbance feedback control. Summary of the Invention

[0005] Based on the above description, the present invention provides a highway inspection method and system that integrates dynamic traffic flow prediction and the high maneuverability of compound-wing UAVs. By proactively modeling the traffic flow disturbance field, it dynamically generates anti-interference and low-impact flight trajectories, thereby ensuring inspection efficiency while minimizing interference to the ground traffic system.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A low-disturbance highway inspection method using a compound-wing UAV based on traffic flow prediction includes the following steps: Step 1: Acquire real-time traffic flow data on highways. Collect traffic volume, average speed, lane occupancy rate, vehicle type distribution, vehicle trajectory and emergency information of the target road segment through roadside sensing units, vehicle communication terminals and traffic management center interfaces to form a multi-source fusion traffic status dataset. Step 2: Construct a traffic flow disturbance field prediction model. Discretize the target road segment into three-dimensional grid cells along the longitudinal direction, transverse direction, and flight altitude direction of the road. Use the three-dimensional grid cells as graph nodes and the spatial adjacency relationship and traffic flow interaction relationship between adjacent lanes as graph edges. The traffic flow interaction relationship is obtained by normalizing the vehicle lane-changing frequency and relative speed within the sliding time window. Based on the traffic state dataset, predict the disturbance intensity of each three-dimensional grid cell within the future time window to generate a three-dimensional disturbance field prediction map. Step 3: Generate a low-disturbance flight path. Adjust the random sampling probability of the fast expanding random tree algorithm according to the three-dimensional disturbance field prediction map, and construct the path cost function with the flight disturbance index, path length, altitude change smoothing term and curvature smoothing term. Under the conditions of satisfying the minimum turning radius, maximum climb rate, minimum safe altitude and inspection coverage constraints of the compound wing UAV, generate a dynamic flight path that avoids the high-disturbance area. Step 4: Execute adaptive inspection flight, control the compound wing UAV to fly along the dynamic flight path, and receive updated traffic flow data and disturbance field prediction map in real time. When the change in disturbance intensity within the preset range ahead of the current flight path exceeds the preset change threshold, trigger local path replanning, and continuously splice the replanned path with the original path before switching to execution. Step 5: Assess the level of traffic interference. During flight, monitor the frequency of lane changes, lateral acceleration fluctuations, frequency of emergency braking events, and the percentage of time the pilot's head is raised (optionally anonymous) using the airborne vision system. Calculate the traffic interference index of the UAV on the traffic flow. When the traffic interference index exceeds a preset safety threshold, control the compound-wing UAV to increase its flight altitude or shift away from the traffic flow centerline until the traffic interference index falls back to a safe range.

[0007] As a preferred embodiment, the traffic flow disturbance field prediction model adopts any one of spatiotemporal graph convolutional neural network, temporal convolutional network or Transformer temporal prediction network; Among them, the node features of each graph node include at least the traffic flow, vehicle speed standard deviation, proportion of large vehicles, historical acceleration change rate mean, lane occupancy rate, emergency event signs, height layer coding, and lateral distance from the lane centerline within the road area mapped by the corresponding three-dimensional grid cell. The edge weight of any graph edge is determined by the spatial distance attenuation term and the traffic flow interaction intensity. The traffic flow interaction intensity is obtained by normalizing the effective lane-changing frequency between adjacent lanes and the relative speed between lane-changing vehicles and vehicles in the target lane within the sliding time window. The traffic flow disturbance field prediction model outputs the disturbance intensity prediction value of each three-dimensional grid cell within the future time window, and the disturbance intensity prediction value is used for subsequent low-disturbance flight path planning.

[0008] As a preferred approach: the instantaneous interaction intensity between adjacent lanes is defined as:

[0009] in, This refers to the normalized lane-changing frequency. The relative velocity after normalization; To avoid abrupt changes in graph edge weights caused by a single abnormal lane change, an exponential moving average method is used to obtain the final traffic flow interaction intensity.

[0010] in, This represents the traffic flow interaction intensity used to construct the graph edge weights at the current moment; Δt represents the traffic flow interaction intensity of the previous fusion cycle; Δt represents the interaction intensity update cycle. This represents the smoothing coefficient, which can be between 0.6 and 0.8.

[0011] As a preferred option, the flight disturbance index in step three is defined as follows:

[0012] in, Representing a path The flight disturbance index; This indicates the perturbation intensity of the three-dimensional perturbation field output in step two at the corresponding spatial location and time. Indicates the high sensitivity coefficient; The high sensitivity coefficient is defined as:

[0013] in, Indicates the minimum safe flight altitude. Indicates the high attenuation constant; due to The closer , The larger the value, the more sensitive the drone is to vehicle wakes and near-ground airflow disturbances when flying at low altitudes.

[0014] As a preferred option: the fast expanding random tree algorithm described in step three suppresses the sampling probability of high-disturbance areas and increases the sampling probability of low-disturbance areas during the random sampling phase. The path cost function includes a disturbance exponent, a path length penalty term, an altitude change smoothing term, and a curvature smoothing term. The projection line of the center line of the highway median onto the ground is used as the lateral offset reference. The dynamic flight path is controlled to maintain a preset offset distance in the lateral direction of the road. The initial offset distance is 15 meters. When the density of large vehicles in the adjacent lane exceeds 0.4 vehicles / 100 meters, the preset offset distance is increased to 25 meters.

[0015] As a preferred option: the local path replanning in step four is limited to a range of 1000 meters before and after the current flight segment. A rolling window optimization strategy is adopted. The new path is continuously spliced ​​with the original path at the connection point by a fourth-order Bezier curve using C2. The heading angle change rate does not exceed 15 degrees / second. The flight control system dynamically adjusts the flight attitude and image acquisition parameters according to the estimated disturbance intensity of the new path, including increasing the pitch damping gain and activating the gimbal active compensation mode.

[0016] As a preferred embodiment: the traffic interference index is composed of the vehicle lateral acceleration variance, the frequency of emergency braking events, and the proportion of optional anonymized driver head tilt duration, all normalized and weighted. The proportion of optional anonymized driver head tilt duration is obtained by estimating the deviation between the driver's head tilt angle and the theoretical gaze angle of the UAV relative to the vehicle. Furthermore, the airborne vision system does not extract, compare, or store driver identity information or facial feature templates. When the traffic interference index exceeds a safety threshold, the flight control system first increases the flight altitude. If the altitude does not drop after a preset time, the flight path centerline is further shifted away from the traffic flow.

[0017] As a preferred option, it also includes establishing a disturbance field-flight stability mapping database to record the attitude angle fluctuations, image blurring, and positioning drift of the UAV under different combinations of traffic flow density and vehicle speed. During the online operation phase, the actual observed stability indicators are compared with the expected values ​​of the corresponding working conditions in the database. If the deviation exceeds 15%, the online correction of the disturbance field prediction model is triggered. The deviation pattern is learned through a lightweight LSTM network and a compensation term is applied to the model output.

[0018] As a preferred solution: when the coverage of a single drone is insufficient, at least two hybrid-wing drones are scheduled to take off from the take-off and landing platform. Each hybrid-wing drone coordinates its flight altitude and longitudinal spacing through a distributed consensus algorithm. Each hybrid-wing drone periodically broadcasts its own position, speed, disturbance field perception results, and remaining battery power. After receiving information from other hybrid-wing drones, it solves a collaborative optimization problem with the flight disturbance index of each drone, the peak value of the multi-drone disturbance field superposition, and the longitudinal spacing deviation as the objective terms, so as to update the flight altitude and longitudinal spacing, so that the peak value of the multi-drone disturbance field superposition does not exceed a preset multiple of the peak value of a single drone's independent operation, and the longitudinal spacing between adjacent drones is kept within a preset safe distance range.

[0019] A low-disturbance highway inspection system based on traffic flow prediction using a hybrid wing UAV includes a roadside sensing unit, an onboard communication receiving unit, a traffic management center interface, an edge computing node, a hybrid wing UAV, a flight control module, a path planning module, and a traffic interference assessment module. The roadside sensing unit, vehicle-mounted communication receiving unit, and traffic management center interface are used to acquire real-time traffic flow data of the highway and form a multi-source fusion traffic status dataset. The edge computing node is used to build a traffic flow disturbance field prediction model based on the traffic state dataset and output a three-dimensional disturbance field prediction map within a future time window. The path planning module is used to generate a low-disturbance dynamic flight path based on the three-dimensional disturbance field prediction map and the flight performance constraints of the compound wing UAV. The flight control module is used to control the compound wing UAV to perform inspection flight along the low-disturbance dynamic flight path, and to perform local path replanning when the disturbance field changes significantly. The traffic interference assessment module is used to calculate the traffic interference index based on the frequency of lane changes of ground vehicles, lateral acceleration fluctuations, frequency of emergency braking events, and the optional percentage of anonymized driver head-up time, and to send altitude adjustment or route deviation commands to the flight control module based on the traffic interference index.

[0020] Compared with the prior art, the technical solution of this application has the following beneficial technical effects: By constructing a traffic flow disturbance field prediction model based on spatiotemporal graph convolutional neural network, the traditional path planning that relies solely on meteorological data is upgraded to a multi-dimensional disturbance perception system that integrates traffic flow dynamics. This system can identify high-disturbance areas in advance within a specific time period, enabling drones to proactively avoid vehicle wakes and sudden airflow changes caused by congestion. This helps reduce the amplitude of flight attitude angle fluctuations and improves image acquisition clarity and geolocation stability.

[0021] Breaking free from dependence on fixed infrastructure: Abandoning fixed-point hovering solutions that require pre-set hardware, such as robotic arm mounting, the system relies entirely on the high mobility and wind resistance of the compound-wing UAV itself to achieve autonomous inspection without attachment points on open linear highways. It is suitable for highway sections without dedicated UAV infrastructure, which helps reduce system deployment costs and maintenance complexity.

[0022] Actively suppressing interference with traffic flow: By introducing a traffic interference index to quantify the impact of drones on traffic flow and driving behavior, and establishing a closed-loop feedback mechanism, the drone can automatically adjust its flight altitude or lateral offset when interference exceeds the standard. This helps reduce the risk of abnormal driving behavior of ground vehicles, reduce the potential impact of drone inspections on traffic operations, and meet the core social need of "low-disturbance" inspections.

[0023] Supports real-time response in highly dynamic environments: Through local path replanning at a predetermined time period and rolling window optimization strategy, it ensures that the flight strategy is always synchronized with the latest traffic conditions. The response delay to scenarios such as sudden accidents and temporary construction is less than the preset delay limit. Compared with planning methods based on static historical data, the timeliness is significantly improved, ensuring the continuity and integrity of inspection tasks.

[0024] Enhancing the efficiency of multi-aircraft collaborative operations: In the multi-aircraft collaborative mode, flight parameters are coordinated through a distributed consensus algorithm to avoid the superposition and amplification of disturbance fields. This is conducive to improving the coverage efficiency of multi-aircraft joint inspections, while the overall disturbance impact is controlled within a preset multiple of the single-aircraft level, providing a scalable technical path for high-frequency inspections of large-scale road networks. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the traffic flow disturbance field prediction model in this invention; Figure 3 This is a logical flowchart of the low-disturbance flight path generation and adaptive inspection flight in this invention. Detailed Implementation

[0026] Example 1: Reference Figure 1 , Figure 2 and Figure 3 A low-disturbance highway inspection method using a compound-wing UAV based on traffic flow prediction, characterized by the following steps: Step 1: Acquire real-time traffic flow data on highways. Collect traffic volume, average speed, lane occupancy rate, vehicle type distribution, and emergency information of the target road segment through roadside sensing units, vehicle communication terminals, and traffic management center interfaces to form a multi-source fusion traffic status dataset.

[0027] Specifically, in step 1, the roadside sensing units are deployed on the guardrails or gantry structures along the highway, with a physical deployment spacing of no more than 500 meters to ensure continuous coverage of the entire road section. Each roadside sensing unit integrates a millimeter-wave radar and a high-definition video camera. The millimeter-wave radar operates at 77 GHz, with a detection range of no less than 200 meters, a lateral resolution better than 0.5 meters, and a longitudinal velocity measurement error of no more than ±0.3 m / s. The high-definition video camera uses a global shutter CMOS sensor with a resolution of no less than 4096×2160 pixels, a frame rate of no less than 30 frames per second, and supports H.265 encoding compression. Data from the two types of sensors are synchronized through a timestamp alignment mechanism, with the synchronization error controlled within ±5 milliseconds. All raw sensing data is uploaded to the edge computing node deployed on the roadside via a built-in 5G communication module. This node uses the NVIDIA Jetson AGX Orin platform and has an AI computing power of no less than 32 TOPS. The data fusion process employs a multi-source fusion algorithm assisted by Kalman filtering and deep learning. First, data from millimeter-wave radar, high-definition video cameras, and vehicle-mounted communication terminals are synchronized to a unified fusion time using timestamp alignment, and spatial unification is achieved using sensor extrinsic parameters and the road coordinate system. Then, vehicle candidate clusters are clustered in the radar point cloud, and road coordinate mapping is performed on the visual detection boxes. A matching cost is constructed based on spatial distance, image box intersection-over-union ratio, vehicle type consistency, and depth appearance features. The Hungarian algorithm is used to complete the registration of radar and visual targets. Subsequently, using vehicle position, speed, and acceleration as state variables, Kalman filtering is used for state prediction and multi-source measurement updates. Finally, a structured traffic state vector containing each vehicle ID, position coordinates, instantaneous velocity vector, acceleration, vehicle type, driving lane, and fusion confidence is generated as input to the subsequent traffic flow disturbance field prediction model. The fusion latency is strictly controlled within 100 milliseconds to meet the real-time requirements of subsequent disturbance field prediction. Simultaneously, the vehicle-mounted communication terminal broadcasts its own state information, including vehicle position, speed, acceleration, and emergency braking signals, via a V2X protocol (such as the C-V2X PC5 interface). This information is received by the roadside unit and injected into the fusion process as a supplementary data source. The traffic management center interface provides macro-level traffic parameters, such as average vehicle speed at the road segment level, congestion index, accident reporting records, and construction area notices, with a data update frequency of no less than 1Hz. The final traffic status dataset is stored in a time-series format in a circular buffer of the edge nodes, retaining historical data from the most recent 300 seconds for model training and online inference.

[0028] S111, Time Synchronization and Coordinate Unification Let the time of unified fusion of edge computing nodes be . Millimeter-wave radar, high-definition video cameras, and vehicle-mounted communication terminals each output time-stamped data. For any sensor data, when its timestamp... satisfy:

[0029] At that time, this data will be included in the current fusion cycle; among which, To determine the maximum permissible synchronization error, in this embodiment, we take... For data that is not perfectly aligned, linear interpolation is used to convert the target position and velocity to a unified fusion time. .

[0030] Establish road coordinate system ,in The longitudinal direction of the road. The horizontal direction of the road. The vertical direction is perpendicular to the road surface. The millimeter-wave radar point cloud coordinates are transformed to the road coordinate system using an extrinsic parameter matrix:

[0031] in, This represents the point cloud coordinates in the millimeter-wave radar coordinate system. This represents the point cloud coordinates in the road coordinate system. This is the rotation matrix from the radar coordinate system to the road coordinate system. It is a translation vector.

[0032] The vehicle detection frame output by the high-definition video camera is as follows:

[0033] in, , Indicates the first The pixel coordinates of the center point of each detection box , This indicates the width and height of the detection frame. Indicates the confidence level of the detection. Indicates the vehicle category. The center point of the bottom edge of the detection box is taken as the pixel coordinates of the vehicle's grounding point.

[0034] Homography matrix from camera to road surface Map it to the road coordinate system:

[0035] in, As a scale factor, , These represent the longitudinal and lateral positions of the visually detected target in the road coordinate system, respectively.

[0036] S112, Radar Point Cloud Clustering and Radar Measurement Generation The millimeter-wave radar output point cloud set is as follows:

[0037] in, For the first The location of each radar point in the road coordinate system The radial velocity at that point, The radar echo intensity is used. A clustering method based on range and velocity consistency is employed to divide the point cloud into several vehicle candidate clusters:

[0038] in, This is the spatial clustering threshold. This is the speed consistency threshold. For the... Each radar cluster calculates the position and velocity of its centroid:

[0039] in, , , This indicates the position of the radar cluster in the road coordinate system. , This represents the longitudinal and lateral velocities estimated from radar velocity measurements and position difference estimation across consecutive frames.

[0040] S113, Spatial registration of visual inspection frame and radar cluster Projecting the center point of the radar cluster onto the camera image plane yields the predicted image box corresponding to the radar cluster. For radar clusters With visual inspection box Construct the matching cost:

[0041] in, For spatial location distance, The intersection-union ratio (IoU) of the radar-predicted image boxes and the visually detected boxes. Penalty for inconsistency in vehicle type For depth feature appearance differences, , , , These are the weighting coefficients.

[0042] The spatial location distance term is defined as follows:

[0043] in, This indicates the measurement position after the visual detection results are mapped to the road coordinate system. This represents the covariance matrix of the joint radar and vision measurements.

[0044] The deep learning-assisted module extracts appearance feature vectors from the visual detection boxes. And save the appearance templates of historical trajectory targets. The appearance difference item is defined as:

[0045] The Hungarian matching algorithm is used to solve the minimum total cost matching problem.

[0046] in, This represents the candidate matching relationship between a group of radar clusters and visual detection boxes. This represents the optimal matching result. When... When the value is less than the preset matching threshold, the radar cluster is considered to be... With visual inspection box They belong to the same vehicle target.

[0047] S114. Vehicle state prediction based on Kalman filtering For each tracked vehicle Establish the vehicle motion state vector:

[0048] in, This indicates the vehicle's position in the road coordinate system. Represents the velocity component. It represents the longitudinal and lateral acceleration components.

[0049] State prediction is performed using a constant acceleration motion model:

[0050]

[0051] in, To predict the state, To predict the covariance matrix, The process noise covariance matrix is... Let be the state transition matrix.

[0052] At the sampling interval In this case, the state transition relations include:

[0053]

[0054]

[0055]

[0056] S115, Multi-source Measurement Update For vehicles The millimeter-wave radar measurement model is as follows:

[0057] The visual measurement model is as follows:

[0058] The measurement model for vehicle-mounted communication terminals is as follows:

[0059] in, , , These represent the measurement vectors from radar, vision, and vehicle communication terminals, respectively. , , These represent the corresponding measurement matrices; , , These represent measurement noise, respectively.

[0060] To reflect the reliability differences of different sensors under different environments, a measurement reliability coefficient is introduced:

[0061] in, Indicates millimeter-wave radar, Indicates camera view. This refers to the vehicle-mounted communication terminal. The reliability of millimeter-wave radar can be determined based on echo intensity and point cloud quantity; the reliability of visual communication can be determined based on detection confidence, occlusion ratio, and image clarity; and the reliability of vehicle-mounted communication can be determined based on communication latency and positioning accuracy. The corresponding effective measurement noise covariance is:

[0062] in, For sensors The basic measurement noise covariance To prevent extremely small constants with a denominator of zero.

[0063] Perform a Kalman update sequentially for each type of valid measurement. (Based on the sensor...) Taking the measurement of innovation as an example, the innovation quantity is:

[0064] The innovation covariance is:

[0065] The Kalman gain is:

[0066] The updated vehicle status is as follows:

[0067] The updated covariance matrix is:

[0068] When multiple sensors have valid measurements within the same fusion cycle, sequential updates are performed in the order of millimeter-wave radar, visual inspection, and vehicle communication terminal; when a sensor measurement is missing, only the measurements from the other sensors are used for updating.

[0069] S116. Vehicle ID Maintenance and Track Confirmation For existing trajectories, if consecutive matches are successful, the original vehicle ID remains unchanged; if consecutive matches are not successful... If no valid measurement is matched within a fusion cycle, the trajectory is deleted. For newly emerging candidate targets, if their continuity... If a fusion cycle is detected, a new vehicle ID is generated. This forms a stable vehicle trajectory.

[0070] in, Indicates the vehicle number. Indicates vehicle type, This indicates the lane number.

[0071] S117. Determination of Driving Lane and Vehicle Type Based on the vehicle center point The lateral distance between the vehicle and the center line of the lane determines the driving lane:

[0072] in, Indicates the longitudinal position of the vehicle First The lateral coordinates of the center line of each lane.

[0073] The vehicle type is determined jointly by the visual classification result, radar target size, and historical categories. Let the probability of the vehicle type output by the visual model be... The radar point cloud estimates the vehicle model probability as follows: The probability of the vehicle model based on historical trajectory is The merged vehicle category is:

[0074] in, , , The weights are respectively for visual, radar, and historical categories, and satisfy the following conditions:

[0075] S118, Output structured traffic state vector For vehicles The final generated structured traffic state vector is:

[0076] in, This represents the confidence level of the vehicle fusion result. The confidence level can be defined as:

[0077] The structured traffic state vectors of all vehicles together constitute the traffic state dataset for the current fusion cycle:

[0078] in, This indicates the number of vehicles that are being tracked stably during the current fusion cycle.

[0079] S119. Generating subsequent disturbance prediction input from structured traffic state vectors. For any grid cell Based on whether a vehicle's location falls within a given grid cell, the following statistics are calculated: traffic flow, speed standard deviation, proportion of large vehicles, lane occupancy rate, and mean rate of acceleration change.

[0080]

[0081]

[0082] in, Represents grid cells The converted traffic flow within the area Indicates the standard deviation of velocity. This indicates the proportion of large vehicles. Represents grid cells Number of vehicles inside, Represents a set of large vehicle categories. This is an indicator function.

[0083] Therefore, the multi-source fusion algorithm integrates the advantages of millimeter-wave radar in distance and speed measurement, high-definition video cameras in vehicle type and lane recognition, and vehicle communication terminals in status broadcasting to generate a structured traffic state vector containing each vehicle ID, location coordinates, instantaneous velocity vector, acceleration, vehicle type, driving lane, and fusion confidence. This provides calculable node features and edge weights for subsequent traffic flow disturbance field prediction models.

[0084] Step 2: Construct a traffic flow disturbance field prediction model. Based on the traffic state dataset, a spatiotemporal graph convolutional neural network is used to model the intensity of wake turbulence, airflow disturbance range, and disturbance duration generated by vehicle motion, and generate a three-dimensional disturbance field prediction map that is distributed longitudinally and laterally along the road in a specific future time period.

[0085] Specifically, in step 2, a traffic flow disturbance field prediction model is constructed: based on the traffic state dataset, a spatiotemporal graph convolutional neural network is used to model the intensity of wake turbulence, airflow disturbance range, and disturbance duration generated by vehicle motion, and to generate a three-dimensional disturbance field prediction map that is distributed longitudinally and laterally along the road in a specific future time period.

[0086] In step 2, the specific calculation process of the traffic flow disturbance field prediction model includes the following steps: S21. Establish a road-flight space coordinate system. The direction of road extension for the target highway segment is used as the reference point. Axis, with the road's transverse direction as The axis is defined by its height relative to the road surface. Axis, establish a local three-dimensional coordinate system The target road segment is divided longitudinally into sections of length [length missing]. The vertical grid divides the road laterally into sections with a width of [missing information]. The horizontal grid is divided into multiple altitude layers, thus forming a three-dimensional grid unit.

[0087] The diagram structure is as follows:

[0088] in, Represents the set of graph nodes. Represents the set of graph edges. Indicates time The dynamic adjacency matrix. Each graph node Corresponding to a three-dimensional mesh element:

[0089] in, Indicates the first The center position of each vertical grid, Indicates the first The center position of each horizontal grid Indicates the first One flight altitude level.

[0090] S22, Construct node features. For time... nodes Construct node feature vectors:

[0091] in, Represents a node Traffic flow within the corresponding road area; Indicates the standard deviation of vehicle speed; This indicates the proportion of large vehicles; Representing history Mean rate of change of vehicle acceleration; Indicates lane occupancy rate; This indicates a sudden event flag. It is set to 1 when there is an accident, construction, sudden braking, or sudden congestion in the corresponding area, and 0 otherwise.

[0092] S23. Calculate the traffic flow interaction intensity. For road grid segments... Adjacent lanes and In length of The number of lane changes in both directions is counted within the sliding time window:

[0093] in, Indicates that the vehicle is in the lane Change to lane Number of times, Indicates that the vehicle is in the lane Change to lane The number of times.

[0094] Adjacent lane changing frequency is defined as:

[0095] in, The length of the road grid segment is in meters; The unit is times / .

[0096] The average relative speed between adjacent lanes is defined as:

[0097] in, This indicates the number of vehicles that can effectively change lanes within the sliding window. Indicates the first The moment a vehicle changes lanes; This indicates the longitudinal speed of the vehicle at the moment of lane change; Indicates the target lane In road grid sections The average longitudinal speed of adjacent vehicles.

[0098] Instantaneous traffic flow interaction intensity is defined as:

[0099] in, This is a reference value for lane changing frequency. This is a relative speed reference value. This represents the truncation function.

[0100] To reduce fluctuations caused by occasional lane changes, an exponential moving average is used to obtain the final traffic flow interaction intensity:

[0101] in, For smoothing coefficients, To update the step size.

[0102] S24. Construct a dynamic adjacency matrix. For graph nodes... and The adjacency matrix elements are defined as follows:

[0103] in, Represents a node and The spatial center distance between them Indicates the spatial attenuation scale. Indicates the vehicle flow interaction enhancement coefficient. Represents a node and The intensity of traffic flow interaction between the corresponding road areas.

[0104] Normalize the adjacency matrix:

[0105] in, for The degree matrix.

[0106] S25. Establish disturbance intensity labels. Synchronous wind field measurement data are acquired by an ultrasonic anemometer array or an airborne wind speed estimation module. Let the instantaneous wind speed vector be... The average wind speed vector within the sliding window is Then the wind speed disturbance is:

[0107] Convert the wind speed disturbance into the equivalent airflow pressure fluctuation amplitude:

[0108] in, Represents a node At any moment The measured disturbance intensity label value, in units of ; Indicates air density; This represents the L2 norm.

[0109] S26. Perform spatiotemporal graph convolution calculation. Let the input time window be... The input sequence is:

[0110] in, For a moment A matrix composed of the features of all nodes.

[0111] No. The output of the layer graph convolution is:

[0112] in, ; Indicates the activation function; Denotes the order of the Chebyshev polynomial; Indicates the first Chebyshev polynomial; Represents the normalized graph Laplacian matrix; Indicates the first Layer Graph convolution parameters; This indicates the bias term.

[0113] The temporal convolution output is:

[0114] in, Indicates the temporal convolution kernel length. Indicates the first Layer expansion factor, Represents the temporal convolution parameters. This indicates the bias term.

[0115] S27. Output the predicted future perturbation field. The model outputs the predicted perturbation intensity values ​​for each node within the future time window:

[0116] in, Represents a node In the future The predicted value of the disturbance intensity; ; and These are the output layer parameters; This indicates the total number of network layers.

[0117] The model training loss function is:

[0118] in, Indicates the number of training samples; Indicates the first The predicted perturbation intensity for each sample; Indicates the first The measured perturbation intensity of each sample; Indicates sample weights; Represents the spatial smoothing constraint weights.

[0119] Through the above steps S21 to S27, the spatiotemporal graph convolutional neural network can output a three-dimensional disturbance field prediction map with a step size of 5 seconds for the next 60 seconds based on the traffic state dataset; the disturbance field at each time step is represented as a tensor isomorphic to the input three-dimensional grid, and its element value represents the disturbance intensity at the corresponding spatial location, with the unit being Pa.

[0120] The detailed process for obtaining the traffic flow interaction intensity between adjacent lanes is as follows: First, based on the structured traffic state vector output in step 1, the vehicle number, road coordinates, longitudinal velocity, lateral velocity, vehicle type, and lane number for each vehicle at consecutive time points are obtained. For vehicle k, its trajectory state at time t can be represented as:

[0121] Wherein, ID_k represents the vehicle number; x_k(t) represents the vehicle's longitudinal position coordinates along the road; y_k(t) represents the vehicle's lateral position coordinates along the road; v_x,k(t) represents the vehicle's longitudinal speed; v_y,k(t) represents the vehicle's lateral speed; c_k(t) represents the vehicle type; and l_k(t) represents the lane number where the vehicle is located.

[0122] For adjacent lanes a and b, when vehicle k is within the sliding time window [t-T_c,t], the following conditions are met:

[0123]

[0124]

[0125] Furthermore, if the duration of the vehicle's position in the original lane a before the lane change is not less than τ_pre, and the duration of its position in the target lane b after the lane change is not less than τ_post, then vehicle k is determined to have committed a valid lane change event from lane a to lane b within road grid segment p. Here, t_a is the stable driving time before the lane change, t_b is the stable driving time after the lane change, T_c is the interaction intensity calculation window (10s in this embodiment), and τ_pre and τ_post can be 0.5s.

[0126] Within road grid segment p, count the number of effective lane changes in both directions between adjacent lanes a and b in the sliding window [t-T_c,t].

[0127] Where N_a→b,p(t) represents the number of valid lane changes from lane a to lane b, and N_b→a,p(t) represents the number of valid lane changes from lane b to lane a.

[0128] Adjacent lane changing frequency is defined as:

[0129] Where F_ab,p(t) represents the lane-changing frequency between adjacent lanes a and b within road grid segment p, in times / (s·100m); L_p represents the length of road grid segment p, in meters.

[0130] For the r-th valid lane change event within the sliding window, let its lane change completion time be t_r^c, the longitudinal velocity of the lane-changing vehicle be v_x,r(t_r^c), and the average longitudinal velocity of the nearest vehicles in the target lane is... Given x, b, p(t_r^c), the relative velocity corresponding to this lane-changing event is:

[0131] If there are no vehicles in front or behind that meet the distance threshold within the target lane, then the average longitudinal velocity of all vehicles in the target lane within road grid segment p is used as the distance threshold. _x,b,p(t_r^c). Therefore, the average relative speed between adjacent lanes a and b within the sliding window is:

[0132] Where ε is a minimal constant to prevent the denominator from being zero. When N_ab,p(t)=0, let ΔV_ab,p(t)=0.

[0133] To eliminate the influence of different dimensions on the calculation of interaction intensity, the lane-changing frequency and relative speed are normalized respectively:

[0134]

[0135] Where F_0 represents the lane change frequency reference value, V_0 represents the relative speed reference value, and clip(·) represents the cutoff function, which is used to restrict the normalization result to the interval [0,1].

[0136] The instantaneous interaction intensity between adjacent lanes is defined as:

[0137] To avoid abrupt changes in graph edge weights caused by a single abnormal lane change, an exponential moving average method is used to obtain the final traffic flow interaction intensity.

[0138] in, This represents the traffic flow interaction intensity used to construct the graph edge weights at the current moment; Δt represents the traffic flow interaction intensity of the previous fusion cycle; Δt represents the interaction intensity update cycle. The value represents the smoothing coefficient, which can be between 0.6 and 0.8 in this embodiment.

[0139] For adjacent nodes v_i and v_j in the graph structure, if they correspond to adjacent lanes a and b within road grid segment p respectively, then:

[0140] The edge weights represent the traffic flow interaction intensity of the graph. Furthermore, the edge weights can be determined jointly by a spatial distance attenuation term and a traffic flow interaction enhancement term:

[0141] Where A_ij(t) represents the edge weight between node v_i and node v_j at time t; d_ij represents the spatial distance between the center points of two grid cells; d_0 represents the spatial attenuation scale; and λ_c represents the traffic flow interaction enhancement coefficient. Therefore, when lane changes between adjacent lanes are frequent and the speed difference between the lane-changing vehicle and the target lane vehicle is large, C_ij(t) increases, corresponding to an increase in the graph edge weight A_ij(t), indicating a stronger traffic flow coupling relationship between the adjacent lanes; when the number of lane changes is less or the relative speed is lower, C_ij(t) decreases, corresponding to a decrease in the graph edge weight.

[0142] For example, within a road grid segment p of length 100m, with a design calculation window T_c = 10s, and a total of 3 valid lane-changing events occur between lane 2 and lane 3, then:

[0143]

[0144] If the relative speeds corresponding to the three lane-changing events are 9 m / s, 7 m / s, and 8 m / s respectively, then:

[0145] Let F_0 = 0.3, V_0 = 15 m / s, then:

[0146]

[0147]

[0148] If the interaction intensity of the previous fusion cycle was C_23, p(t-Δt)=0.40, and the smoothing coefficient was α_c=0.7, then the current traffic flow interaction intensity is:

[0149] This value is then assigned to the graph edges between the corresponding adjacent lane grid nodes, used for the adjacency matrix construction and perturbation field prediction of the spatiotemporal graph convolutional neural network.

[0150] Step 3: Generate a low-disturbance flight path. Based on the disturbance field prediction map and combined with the flight performance constraints of the compound-wing UAV, with the objective function of minimizing the flight disturbance index, an improved fast expanding random tree algorithm is used to plan a dynamic flight path that avoids high-disturbance areas, maintains a safe altitude, and meets the inspection coverage requirements.

[0151] Specifically, in step 3, the planned path for the compound-wing UAV is set as follows:

[0152] in:

[0153] Indicates the start time of the currently planned path. Indicates the termination time of the current planned path; Indicates the drone at a certain time The longitudinal coordinates of the road; Indicates the drone at a certain time The lateral coordinates of the road; Indicates the drone at a certain time Flight altitude relative to road surface.

[0154] The flight disturbance index is defined as:

[0155] in, Representing a path The flight disturbance index; This indicates the perturbation intensity of the three-dimensional perturbation field output in step 2 at the corresponding spatial location and time. This represents the high sensitivity coefficient.

[0156] The high sensitivity coefficient is defined as:

[0157] in, Indicates the minimum safe flight altitude. This represents the high attenuation constant. Because... The closer , The larger the value, the more sensitive the drone is to vehicle wakes and near-ground airflow disturbances when flying at low altitudes.

[0158] In discrete implementation, let:

[0159] in, , If the time interval between the walk and the time interval is long, then the flight disturbance index is approximately:

[0160] Improved RRT The search state of the algorithm is defined as follows:

[0161] in, This indicates the desired cruising speed.

[0162] During the random sampling phase, sampling points The sampling probability is determined by the predicted value of the disturbance intensity:

[0163] in, Indicates the minimum sampling probability. Indicates the sampling point at the expected arrival time. The predicted value of the disturbance intensity, This represents the scale parameter of the disturbance intensity. As can be seen from this formula, high disturbance regions correspond to lower sampling probabilities, while low disturbance regions correspond to higher sampling probabilities.

[0164] Candidate Path The total cost function is defined as:

[0165] in, Indicates the total cost of the candidate paths; Indicates the flight disturbance index; Indicates the path length; Indicates the first Flight altitude of each waypoint; Indicates the first Segment path curvature; , , , These represent the weighting coefficients for the perturbation index, path length, height smoothing, and curvature smoothing, respectively.

[0166] Path planning must meet the following constraints:

[0167]

[0168]

[0169]

[0170] in, Indicates the minimum turning radius of the path. Minimum allowable turning radius; Indicates the rate of increase or decrease. Indicates the maximum climb rate; Indicates the heading angle. This represents the maximum rate of change of heading angle.

[0171] Improved RRT The role of the algorithm in this scheme is to transform the three-dimensional disturbance field prediction map obtained in step 2 into a low-disturbance three-dimensional trajectory that can be executed by the compound wing UAV, so that the path planning can simultaneously meet the requirements of low disturbance, low energy consumption, flight safety and inspection coverage.

[0172] Step 4: Perform adaptive inspection flight, control the compound wing UAV to fly along the dynamic flight path, receive updated traffic flow data in real time, and when a significant change in the disturbance field is detected, trigger the local path replanning mechanism and adjust the flight attitude and image acquisition parameters to maintain stable imaging.

[0173] Specifically, in step 4, the compound-wing UAV is equipped with a vertical takeoff and landing rotor and a fixed-wing propulsion system, with a cruising speed in the range of 80–120 km / h, a mission endurance of ≥90 minutes, and a three-axis gimbal-stabilized camera with an image resolution of at least 4096×3072 pixels, supporting visible light and infrared dual-mode imaging. The flight control system reads data from the inertial measurement unit (IMU) and global navigation satellite system (GNSS) at a frequency of 200Hz, and performs multi-source fusion positioning combined with visual odometry, achieving a position accuracy better than 0.5 meters. During flight, edge computing nodes send updated disturbance field prediction maps to the UAV every 5 seconds. The system sets a disturbance field change detection threshold: if the change in disturbance intensity of any grid cell within 500 meters ahead of the current flight path exceeds 2 Pa, it is judged as a "significant change," immediately triggering a local path replanning mechanism. This mechanism is limited to a range of 1000 meters before and after the current flight segment, employing a rolling window optimization strategy, that is, only the sub-path within the affected window is replanned, while the rest remains unchanged. After the new path is generated, it is continuously spliced ​​with the original path at the connection point using a fourth-order Bézier curve (C2) to ensure curvature continuity and avoid abrupt changes. The rate of change of heading angle is strictly limited to no more than 15 degrees / second to avoid secondary disturbances caused by violent maneuvers. Simultaneously, the flight control system dynamically adjusts the flight attitude based on the estimated disturbance intensity of the new path: in the medium disturbance region (2–4 Pa), the pitch damping gain is increased by 10%; in the high disturbance region (>4 Pa), the gimbal active compensation mode is activated, increasing the image acquisition frame rate from 30fps to 60fps and enabling electronic image stabilization to maintain image clarity. All parameter adjustment commands are issued via the MAVLink protocol, with a transmission latency of less than 20 milliseconds.

[0174] Step 5: Assess the level of traffic interference. During flight, monitor the frequency of lane changes, lateral acceleration fluctuations, frequency of emergency braking events, and the percentage of time the pilot's head is raised (optionally anonymized) through the airborne vision system to calculate the traffic interference index of the UAV on the traffic flow. If the index exceeds the preset threshold, automatically increase the flight altitude or shift the flight path centerline away from the traffic flow until the interference index falls back to a safe range.

[0175] Specifically, in step 5, the airborne vision system utilizes forward-looking and downward-looking dual cameras working in tandem. The forward-looking camera (FOV 90°) captures the driving trajectory of vehicles within 500 meters ahead at a frame rate of 15fps, detects vehicle bounding boxes in real time using the YOLOv5s model, and generates vehicle trajectory sequences in conjunction with the SORT tracking algorithm, thereby calculating the number of lane-crossing events per unit time, i.e., lane-changing frequency (times / minute / kilometer).

[0176] The downward-facing camera focuses on the road section directly below and acquires vehicle motion images at a preset frame rate. The longitudinal velocity, lateral velocity, longitudinal acceleration, and lateral acceleration of the vehicle in the road coordinate system are estimated using the OpenCV optical flow method. Among them, the lateral acceleration of the vehicle is used to calculate the lateral acceleration variance to characterize abnormal lateral maneuvering and lane change disturbances of the vehicle. The longitudinal acceleration of the vehicle is used to identify emergency braking events. When the longitudinal acceleration of the vehicle is less than a preset deceleration threshold and the duration exceeds a preset duration, it is determined to be an emergency braking event.

[0177] After the forward-facing camera captures images, the YOLOv5s model is used to detect vehicle bounding boxes.

[0178] in, , Indicates the center pixel coordinates of the target bounding box; , Indicates the width and height of the target bounding box; Indicates the confidence level of the detection; Indicates the vehicle category.

[0179] The center point of the bottom edge of the target bounding box is used as the pixel coordinates of the vehicle's grounding point and mapped to the road coordinate system using a homography matrix:

[0180] in, This represents the homography matrix from the camera coordinate system to the road coordinate system.

[0181] The vehicle state vector in the SORT tracker is defined as:

[0182] The SORT tracker uses Kalman filtering for state prediction and a Hungarian algorithm to match the detection results with existing trajectories, generating a vehicle trajectory sequence.

[0183] in, Indicates vehicle In the The lane number to which the frame belongs.

[0184] When the vehicle trajectory satisfies:

[0185]

[0186] and:

[0187] If a vehicle maintains its position in the original lane and the target lane for a continuous period of time before and after changing lanes for a period of time not less than a preset threshold, it is considered a valid lane change event.

[0188] The frequency of lane changes within the statistical window is defined as follows:

[0189] in, This indicates the frequency of lane changes, measured in times per minute per kilometer. This indicates the number of valid lane change events within the statistics window; Indicates the length of the statistical window; Indicates the length of the observed road segment.

[0190] The driver's head-up behavior is estimated using a non-identity recognition method. The airborne vision system only estimates the attitude angle or line-of-sight angle of the driver's head area to determine whether the driver has a short-term head-up behavior towards the drone; it does not extract, compare, or store driver identity information or facial feature templates. For vehicles... The head posture estimation module outputs the driver's head or line of sight elevation angle. This angle is only used as an anonymized statistic for calculating the level of traffic interference.

[0191] The theoretical formula for calculating the gaze elevation angle is:

[0192] in, Indicates vehicle The driver at all times The theoretical elevation angle corresponding to when observing the drone; This indicates the head or gaze direction elevation angle output by the head pose estimation module; Indicates the drone's flight altitude; Indicates vehicle Driver's eye level; Indicates the ground projection point of the drone and the vehicle The horizontal distance between them; This represents a minimal constant to prevent the denominator from being zero.

[0193] When the following conditions are met:

[0194] And the duration satisfies:

[0195] Then determine the vehicle The driver exhibited head-up posture towards the drone during that time period.

[0196] in and For preset parameters (e.g.) The system accumulates the percentage of time each observed vehicle's head posture is raised within every 10 seconds. Traffic Disturbance Index. The traffic disturbance index is calculated using a normalized weighted method, determined by the variance of vehicle lateral acceleration, the frequency of emergency braking events, and the percentage of time the driver's head is raised (optionally anonymized). Since these three indicators have different dimensions, they are normalized using their respective reference values ​​before calculation, making the traffic disturbance index a dimensionless indicator.

[0197] Among them, the vehicle lateral acceleration variance is used to characterize the degree of abnormal lateral vehicle handling, the frequency of emergency braking events is used to characterize sudden longitudinal braking behavior, and the optional anonymized driver head-up duration percentage is used to characterize the degree of visual interference to the driver from the UAV. When the traffic interference index exceeds a preset safety threshold, the flight control system activates an interference suppression strategy, first increasing the flight altitude; if the index does not decrease after a preset time, the flight path centerline is further shifted away from the traffic flow, with the maximum shift not exceeding a preset maximum shift.

[0198] The traffic disruption index D is calculated using the following normalized weighting method:

[0199] in, Let Variance be the lateral acceleration of the vehicle. This is a reference value for the variance of lateral acceleration; For the frequency of emergency braking events, This serves as a reference value for the frequency of emergency braking events. For optional anonymized driver head-up duration percentage, The reference value is the percentage of head-tilt duration; α, β, and γ are weighting coefficients, and satisfy the following:

[0200] In this embodiment, the following can be adopted:

[0201] Safety threshold Set to 1.2. When the following conditions are met:

[0202] At this time, the flight control system activates the interference suppression strategy: first, the flight altitude is increased by 10 meters; if D does not fall back below the safe threshold after 5 seconds, the flight path centerline is further shifted away from the traffic flow by 5 meters, with a maximum shift not exceeding 30 meters. This process is regulated by a PI controller to ensure smooth flight altitude adjustment and flight path shift.

[0203] To further enhance system robustness, the method also includes establishing a disturbance field-flight stability mapping database. This database is constructed during system initialization through flight calibration: under different traffic densities (0.1–0.8 vehicles / 100m / lane) and speed combinations (40–120km / h), the UAV's attitude angle fluctuations (standard deviations of roll, pitch, and yaw), image blur (measured by Laplace variance), and positioning drift (relative to an RTK-GNSS benchmark) are recorded. The data is stored in four-dimensional tensor form, with dimensions of [traffic density, average vehicle speed, disturbance intensity, stability index]. During online operation, the system compares the currently observed stability indexes with the expected values ​​under the corresponding conditions in the database. If the deviation exceeds 15%, online correction of the disturbance field prediction model is triggered: a lightweight LSTM network is used to learn the deviation pattern, and a compensation term is applied to the model output, thereby improving the robustness of path planning.

[0204] In a multi-drone cooperative mode, the cooperative optimization problem is used to solve for the target flight altitude of each compound-wing UAV and the longitudinal distance between any two UAVs. Let the number of UAVs participating in the cooperative inspection be... , No. The drone's flight altitude is , No. The drone and the first The longitudinal spacing between the drones is:

[0205] in, and They represent the first drones and the first The position coordinates of the drone along the longitudinal direction of the road.

[0206] The collaborative optimization problem is represented as:

[0207] The constraints are:

[0208]

[0209]

[0210] in, Represents the optimized first... The target flight altitude of the drone; Represents the optimized first... The drone and the first The longitudinal distance between targets between drones; Indicates the first A drone at high altitude The flight disturbance index corresponding to the flight path; This represents the peak disturbance intensity after the superposition of multiple machine disturbance fields; Indicates the peak disturbance intensity when a single machine is operating independently; This represents the threshold value for the disturbance superposition factor, which is set to:

[0211] This represents the penalty coefficient for superimposed disturbances; Indicates the penalty coefficient for deviation of longitudinal spacing; Indicates the desired vertical spacing; and These represent the minimum and maximum safe longitudinal spacing, respectively.

[0212] In this embodiment:

[0213]

[0214] This formula is not an ordinary equation, but an optimization problem expression used to solve for the target flight altitude and longitudinal spacing of multiple drones, so that multi-drone inspection can improve coverage efficiency while avoiding excessive superposition of disturbance fields.

[0215] The above constitutes a relatively complete embodiment of the present invention. To illustrate the application process of the present invention in a high-density traffic scenario on a highway, an exemplary test scenario is constructed: the target road segment is a two-way six-lane highway segment with a length of 10 to 15 kilometers, and a one-way traffic flow of 3,000 to 5,000 vehicles per hour during peak hours. Roadside sensing units are continuously deployed along the target road segment, with the spacing between adjacent roadside sensing units not exceeding 500 meters, and together with the vehicle-mounted communication terminal and the traffic management center interface, they form a multi-source fusion traffic status dataset. In this test scenario, a compound-wing UAV takes off from the roadside take-off and landing point, with an initial flight altitude of 75 to 85 meters, and the centerline of the flight path maintains a preset lateral offset distance relative to the ground projection of the centerline of the highway median. When the system detects that vehicles are slowing down, lane occupancy is increasing, or the proportion of large vehicles is increasing ahead of the target road segment, the traffic flow disturbance field prediction model outputs a three-dimensional disturbance field prediction map within the future time window. If the predicted disturbance intensity of a local grid cell exceeds a preset disturbance threshold, the path planning module regenerates a low-disturbance dynamic flight path based on the flight disturbance index, path length, altitude change smoothing term, and curvature smoothing term, and controls the compound-wing UAV to increase its flight altitude or lateral offset distance. During the inspection process, the traffic interference assessment module calculates the traffic interference index based on the lateral acceleration variance of ground vehicles, the frequency of emergency braking events, and the percentage of anonymized pilot head-up time. When the traffic interference index exceeds a preset safety threshold, the flight control module prioritizes increasing the flight altitude; if the traffic interference index does not decrease within a preset time, it further controls the flight path centerline to shift away from the traffic flow. Through the above processing, the compound-wing UAV can maintain a relatively stable flight state under high-density traffic flow and local sudden traffic fluctuations, and reduce visual and behavioral interference to ground traffic flow.

[0216] In another implementation, the traffic flow disturbance field prediction model employs an alternative architecture: a Transformer encoder-decoder structure replaces the spatiotemporal graph convolutional network. The input sequence consists of time-series features of each lane, including traffic volume, speed, and occupancy, which are then encoded at location and input to the Transformer encoder. The decoder generates a future disturbance field sequence using an autoregressive approach, outputting a three-dimensional disturbance field prediction map at each step, composed of disturbance distributions at different flight altitudes. This model is suitable for scenarios with long prediction time windows, but it has high computational overhead and can be deployed in a regional cloud computing center, transmitting prediction results back via fiber optic cable. The path planning algorithm is correspondingly adjusted to a model-based predictive control framework, with rolling optimization in 1-second control cycles. The state variables include the UAV's position, speed, and disturbance field estimation, with constraints identical to those in Implementation Example 1. This scheme is suitable for mountainous highway inspection scenarios with long prediction time requirements.

[0217] In another implementation, interference level assessment uses pure behavioral indicators, omitting the estimation of non-identity-based driver head-up behavior. The traffic interference index is composed solely of the vehicle's lateral acceleration variance and emergency braking frequency, with weights of 0.6 and 0.4, respectively. This simplified scheme is suitable for nighttime or low-visibility scenarios where visual recognition reliability decreases. The system analyzes vehicle acceleration distribution using millimeter-wave radar echo signals, avoiding reliance on optical imaging. The flight altitude adjustment strategy is correspondingly relaxed, with a safety threshold set at 1.5 to balance interference suppression and inspection efficiency. This scheme is suitable for scenarios where visual recognition reliability decreases, such as nighttime, foggy weather, or strong backlighting, and helps improve the system's adaptability in complex environments.

[0218] Example 2: A low-disturbance highway inspection system based on traffic flow prediction using a hybrid wing UAV includes a roadside sensing unit, an onboard communication receiving unit, a traffic management center interface, an edge computing node, a hybrid wing UAV, a flight control module, a path planning module, and a traffic interference assessment module. The roadside sensing unit, vehicle-mounted communication receiving unit, and traffic management center interface are used to acquire real-time traffic flow data of the highway and form a multi-source fusion traffic status dataset. The edge computing node is used to build a traffic flow disturbance field prediction model based on the traffic state dataset and output a three-dimensional disturbance field prediction map within a future time window. The path planning module is used to generate a low-disturbance dynamic flight path based on the three-dimensional disturbance field prediction map and the flight performance constraints of the compound wing UAV. The flight control module is used to control the compound wing UAV to perform inspection flight along the low-disturbance dynamic flight path, and to perform local path replanning when the disturbance field changes significantly. The traffic interference assessment module is used to calculate the traffic interference index based on the frequency of lane changes of ground vehicles, lateral acceleration fluctuations, frequency of emergency braking events, and the optional percentage of anonymized driver head-up time, and to send altitude adjustment or route deviation commands to the flight control module based on the traffic interference index.

[0219] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A low-disturbance highway inspection method using a compound-wing UAV based on traffic flow prediction, characterized in that, Includes the following steps: Step 1: Acquire real-time traffic flow data on highways. Collect traffic volume, average speed, lane occupancy rate, vehicle type distribution, vehicle trajectory and emergency information of the target road segment through roadside sensing units, vehicle communication terminals and traffic management center interfaces to form a multi-source fusion traffic status dataset. Step 2: Construct a traffic flow disturbance field prediction model. Discretize the target road segment into three-dimensional grid cells along the longitudinal direction, transverse direction, and flight altitude direction of the road. Use the three-dimensional grid cells as graph nodes and the spatial adjacency relationship and traffic flow interaction relationship between adjacent lanes as graph edges. The traffic flow interaction relationship is obtained by normalizing the vehicle lane-changing frequency and relative speed within the sliding time window. Based on the traffic state dataset, predict the disturbance intensity of each three-dimensional grid cell within the future time window to generate a three-dimensional disturbance field prediction map. Step 3: Generate a low-disturbance flight path. Adjust the random sampling probability of the fast expanding random tree algorithm according to the three-dimensional disturbance field prediction map, and construct the path cost function with the flight disturbance index, path length, altitude change smoothing term and curvature smoothing term. Under the conditions of satisfying the minimum turning radius, maximum climb rate, minimum safe altitude and inspection coverage constraints of the compound wing UAV, generate a dynamic flight path that avoids the high-disturbance area. Step 4: Execute adaptive inspection flight, control the compound wing UAV to fly along the dynamic flight path, and receive updated traffic flow data and disturbance field prediction map in real time. When the change in disturbance intensity within the preset range ahead of the current flight path exceeds the preset change threshold, trigger local path replanning, and continuously splice the replanned path with the original path before switching to execution. Step 5: Assess the level of traffic interference. During flight, monitor the frequency of lane changes, lateral acceleration fluctuations, frequency of emergency braking events, and the percentage of time the pilot's head is raised (optionally anonymous) using the airborne vision system. Calculate the traffic interference index of the UAV on the traffic flow. When the traffic interference index exceeds a preset safety threshold, control the compound-wing UAV to increase its flight altitude or shift away from the traffic flow centerline until the traffic interference index falls back to a safe range.

2. The low-disturbance highway inspection method based on traffic flow prediction using a compound-wing UAV according to claim 1, characterized in that: The traffic flow disturbance field prediction model adopts any one of the following: spatiotemporal graph convolutional neural network, temporal convolutional network, or Transformer temporal prediction network; Among them, the node features of each graph node include at least the traffic flow, vehicle speed standard deviation, proportion of large vehicles, historical acceleration change rate mean, lane occupancy rate, emergency event signs, height layer coding, and lateral distance from the lane centerline within the road area mapped by the corresponding three-dimensional grid cell. The edge weight of any graph edge is determined by the spatial distance attenuation term and the traffic flow interaction intensity. The traffic flow interaction intensity is obtained by normalizing the effective lane-changing frequency between adjacent lanes and the relative speed between lane-changing vehicles and vehicles in the target lane within the sliding time window. The traffic flow disturbance field prediction model outputs the disturbance intensity prediction value of each three-dimensional grid cell within the future time window, and the disturbance intensity prediction value is used for subsequent low-disturbance flight path planning.

3. The low-disturbance highway inspection method based on traffic flow prediction using a compound-wing UAV according to claim 2, characterized in that: The instantaneous interaction intensity between adjacent lanes is defined as: in, This refers to the normalized lane-changing frequency. The relative velocity after normalization; To avoid abrupt changes in graph edge weights caused by a single abnormal lane change, an exponential moving average method is used to obtain the final traffic flow interaction intensity. in, This represents the traffic flow interaction intensity used to construct the graph edge weights at the current moment; Δt represents the traffic flow interaction intensity of the previous fusion cycle; Δt represents the interaction intensity update cycle. This represents the smoothing coefficient, which can be between 0.6 and 0.

8.

4. The low-disturbance highway inspection method based on traffic flow prediction using a compound-wing UAV according to claim 1, characterized in that, The flight disturbance index in step three is defined as follows: in, Representing a path The flight disturbance index; This indicates the perturbation intensity of the three-dimensional perturbation field output in step two at the corresponding spatial location and time. Indicates the high sensitivity coefficient; The high sensitivity coefficient is defined as: in, Indicates the minimum safe flight altitude. Indicates the high attenuation constant; due to The closer , The larger the value, the more sensitive the drone is to vehicle wakes and near-ground airflow disturbances when flying at low altitudes.

5. The low-disturbance highway inspection method based on traffic flow prediction using a compound-wing UAV according to claim 1, characterized in that: The fast expanding random tree algorithm described in step three suppresses the sampling probability of high-disturbance areas and increases the sampling probability of low-disturbance areas during the random sampling phase. The path cost function includes a disturbance exponent, a path length penalty term, an altitude change smoothing term, and a curvature smoothing term. The projection line of the center line of the highway median onto the ground is used as the lateral offset reference. The dynamic flight path is controlled to maintain a preset offset distance in the lateral direction of the road. The initial offset distance is 15 meters. When the density of large vehicles in the adjacent lane exceeds 0.4 vehicles / 100 meters, the preset offset distance is increased to 25 meters.

6. The low-disturbance highway inspection method based on traffic flow prediction using a compound-wing UAV according to claim 1, characterized in that: The local path replanning described in step four is limited to a range of 1000 meters before and after the current flight segment. A rolling window optimization strategy is adopted. The new path is continuously spliced ​​with the original path at the connection point by a fourth-order Bezier curve using C2. The heading angle change rate does not exceed 15 degrees / second. The flight control system dynamically adjusts the flight attitude and image acquisition parameters according to the estimated disturbance intensity of the new path, including increasing the pitch damping gain and activating the gimbal active compensation mode.

7. The low-disturbance highway inspection method based on traffic flow prediction using a compound-wing UAV according to claim 1, characterized in that: The traffic interference index is composed of the vehicle lateral acceleration variance, the frequency of emergency braking events, and the proportion of optional anonymized driver head tilt duration, all normalized and weighted. The proportion of optional anonymized driver head tilt duration is obtained by estimating the deviation between the driver's head tilt angle and the theoretical gaze tilt angle of the UAV relative to the vehicle. The airborne vision system does not extract, compare, or store driver identity information or facial feature templates. When the traffic interference index exceeds a safety threshold, the flight control system first increases the flight altitude. If the altitude does not drop after a preset time, the flight path centerline is further shifted away from the traffic flow.

8. The low-disturbance highway inspection method based on traffic flow prediction using a compound-wing UAV according to claim 1, characterized in that: It also includes establishing a disturbance field-flight stability mapping database to record the attitude angle fluctuations, image blurring, and positioning drift of UAVs under different combinations of traffic flow density and vehicle speed. During the online operation phase, the actual observed stability indicators are compared with the expected values ​​of the corresponding working conditions in the database. If the deviation exceeds 15%, the online correction of the disturbance field prediction model is triggered. A lightweight LSTM network is used to learn the deviation pattern and apply a compensation term to the model output.

9. The low-disturbance highway inspection method based on traffic flow prediction using a compound-wing UAV according to claim 1, characterized in that: When the coverage of a single drone is insufficient, at least two hybrid-wing drones are dispatched to take off from the take-off and landing platform. Each hybrid-wing drone coordinates its flight altitude and longitudinal spacing through a distributed consensus algorithm. Each hybrid-wing drone periodically broadcasts its own position, speed, disturbance field perception results, and remaining battery power. After receiving information from other hybrid-wing drones, it solves a collaborative optimization problem with the flight disturbance index of each drone, the peak value of the multi-drone disturbance field superposition, and the longitudinal spacing deviation as the objective terms. This updates the flight altitude and longitudinal spacing, ensuring that the peak value of the multi-drone disturbance field superposition does not exceed a preset multiple of the peak value of a single drone's independent operation, and that the longitudinal spacing between adjacent drones remains within a preset safe distance range.

10. A low-disturbance highway inspection system based on traffic flow prediction using a compound-wing unmanned aerial vehicle (UAV), characterized in that: It includes a roadside sensing unit, a vehicle-mounted communication receiving unit, a traffic management center interface, an edge computing node, a compound-wing UAV, a flight control module, a path planning module, and a traffic interference assessment module; The roadside sensing unit, vehicle-mounted communication receiving unit, and traffic management center interface are used to acquire real-time traffic flow data of the highway and form a multi-source fusion traffic status dataset. The edge computing node is used to build a traffic flow disturbance field prediction model based on the traffic state dataset and output a three-dimensional disturbance field prediction map within a future time window. The path planning module is used to generate a low-disturbance dynamic flight path based on the three-dimensional disturbance field prediction map and the flight performance constraints of the compound wing UAV. The flight control module is used to control the compound wing UAV to perform inspection flight along the low-disturbance dynamic flight path, and to perform local path replanning when the disturbance field changes significantly. The traffic interference assessment module is used to calculate the traffic interference index based on the frequency of lane changes of ground vehicles, lateral acceleration fluctuations, frequency of emergency braking events, and the optional percentage of anonymized driver head-up time, and to send altitude adjustment or route deviation commands to the flight control module based on the traffic interference index.

Images