A Method and System for Real-time Traffic Flow Parameter Statistics and Intelligent Congestion Assessment Based on Unmanned Aerial Vehicle (UAV) Inspection

Abstract

This invention discloses a method and system for real-time traffic flow parameter statistics and intelligent congestion determination based on unmanned aerial vehicle (UAV) inspection. The method includes: constructing a collaborative perception network, extracting traffic feature parameters, and building a three-level state system; for basic road sections, using a macroscopic traffic flow model and spatial constraint clustering algorithm to output state identification results; for hub interchange sections, using sample equalization processing and a binary classification tree to output state classification results; for entrance and exit ramp sections, using a weighted fusion of queuing models and capacity reduction models to determine the congestion level; collecting the identification results from each section to construct a road network state matrix, and using a graph neural network and time-series prediction network to output the congestion diffusion path of the entire road network; generating a visual chart based on the three-level state system and diffusion path, and outputting tiered handling instructions. This achieves accurate perception, zone identification, situational prediction, and tiered control of the entire highway network, significantly improving the accuracy of congestion determination and the precision of control.

Description

Technical Field

[0001] This invention belongs to the field of intelligent traffic management and control technology, specifically relating to a method and system for real-time statistics of traffic flow parameters and intelligent congestion determination based on unmanned aerial vehicle (UAV) inspection. Background Technology

[0002] As a vital inter-regional transportation corridor, the operational efficiency of highways directly impacts socio-economic development and the public's travel experience. With the continuous increase in motor vehicle ownership and the implementation of toll-free policies during holidays, highway congestion has become an increasingly prominent issue and a key challenge in traffic management.

[0003] Currently, the perception and assessment of highway traffic congestion mainly rely on the following technical means: First, the collection of cross-sectional traffic flow parameters based on fixed detectors (such as loop detectors and microwave radar) to determine the congestion status by setting speed or flow thresholds; second, the inference of the degree of congestion by calculating the average travel time of the section based on vehicle passage records of the ETC gantry system; and third, the determination of the operating status by analyzing vehicle density and speed based on image recognition technology from video surveillance.

[0004] Existing technologies have the following shortcomings in congestion assessment:

[0005] First, the data source is singular, limiting the perception capability. A single data source cannot comprehensively depict the traffic operation status: fixed detectors are sparsely deployed and have incomplete coverage; ETC gantry data has time delays, making real-time judgment difficult; video surveillance is greatly affected by environmental factors such as lighting and weather, and single-point monitoring cannot reflect the overall status of a section. The lack of an effective fusion mechanism between multi-source data makes it difficult to balance the accuracy and real-time performance of congestion judgment.

[0006] Second, the methods lack specificity for different sections, resulting in limited accuracy. The traffic flow characteristics of the basic sections of highways, interchange sections, and entrance / exit ramp sections differ significantly, but existing methods mostly use uniform speed or flow thresholds for discrimination, failing to construct differentiated discrimination models for the different operating mechanisms of each section. This leads to low recognition accuracy for typical congestion scenarios such as weaving conflicts at interchange sections and queuing overflows at ramp sections.

[0007] Third, the ability to trace the source of congestion and predict its spread is lacking. Existing methods mainly focus on identifying congestion at the current moment, lacking quantitative analysis of the patterns of congestion propagation, making it difficult to trace the source of congestion and predict its future spread path, thus limiting the effectiveness of proactive traffic control.

[0008] Fourth, the control strategies are crude and lack a tiered coordination mechanism. Existing control measures are mostly responsive strategies based on fixed rules, failing to dynamically match differentiated handling plans according to the level, scope, and evolution trend of congestion. Furthermore, there is a lack of coordination and linkage between main roads, ramps, and toll stations, making it difficult to achieve precise control.

[0009] To address the aforementioned technical challenges, there is an urgent need to develop an integrated approach that combines multi-source heterogeneous data, differentiated identification based on regional distribution, congestion source tracing and simulation, and hierarchical coordinated management and control, in order to improve the accuracy of highway congestion assessment and the precision of management and control. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of existing technologies, such as single data source, insufficient regional targeting, lack of congestion source tracing capabilities, and extensive management strategies. It provides a method and system for real-time statistics of traffic flow parameters and intelligent congestion judgment based on drone inspection, so as to realize accurate perception, regional identification, situational prediction, and hierarchical management of the entire highway network.

[0011] In a first aspect, embodiments of the present invention provide a method for real-time statistics of traffic flow parameters and intelligent congestion determination based on unmanned aerial vehicle (UAV) inspection, the method comprising:

[0012] A collaborative sensing network is constructed to collect multi-source data and perform spatiotemporal calibration, extract traffic feature parameters, and construct a three-level state system.

[0013] For basic road sections, a macroscopic traffic flow model and a spatially constrained clustering algorithm are used to output state identification results based on the aforementioned feature parameters;

[0014] For hub interchange sections, sample equalization processing and binary classification trees are used to output status classification results based on the aforementioned feature parameters and basic road segment identification results;

[0015] For the entrance and exit ramp sections, a weighted fusion of the queuing model and the capacity reduction model is adopted, and the congestion level is determined based on the aforementioned feature parameters, the identification results of the basic road sections and hub interchange sections;

[0016] The identification results of basic road sections, hub interchange sections and ramp sections are collected to construct a road network state matrix, and graph neural networks and time series prediction networks are used to output the congestion diffusion path of the entire road network.

[0017] Based on the three-level state system and diffusion path, a visual chart is generated and a graded handling instruction is output.

[0018] Secondly, embodiments of the present invention provide a real-time traffic flow parameter statistics and intelligent congestion determination system based on unmanned aerial vehicle (UAV) inspection, applied to the real-time traffic flow parameter statistics and intelligent congestion determination method based on UAV inspection as described in the first aspect, the system comprising:

[0019] The data acquisition module is used to build a collaborative sensing network for multi-source data acquisition and spatiotemporal calibration, extract traffic feature parameters, and construct a three-level state system.

[0020] The basic road segment identification module is used to identify basic road segments by employing a macroscopic traffic flow model and a spatial constraint clustering algorithm, and outputting state identification results based on the aforementioned feature parameters.

[0021] The hub interchange section identification module is used to identify hub interchange sections by using sample equalization processing and binary classification tree, and outputting status classification results based on the feature parameters and basic road section identification results.

[0022] The ramp segment identification module is used to determine the congestion level for entrance and exit ramp segments by weighted fusion of queuing model and capacity reduction model, based on the characteristic parameters, basic road segments and hub interchange segment identification results.

[0023] The situation simulation module is used to collect the identification results of basic road sections, hub interchange sections and ramp sections, construct the road network state matrix, and use graph neural networks and time series prediction networks to output the congestion spread path of the entire road network.

[0024] The strategy output module is used to generate a visual chart and output graded handling instructions based on the three-level state system and diffusion path.

[0025] Thirdly, embodiments of the present invention provide an electronic device, comprising:

[0026] processor;

[0027] Memory used to store processor-executable instructions;

[0028] The processor is configured to implement the method for real-time statistics of traffic flow parameters and intelligent congestion determination based on UAV inspection as described in the first aspect when executing the instructions.

[0029] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a program that instructs a device to execute the method for real-time statistics of traffic flow parameters and intelligent congestion determination based on unmanned aerial vehicle (UAV) inspection as described in the first aspect.

[0030] Compared with existing technologies, this invention has the following advantages: By constructing an integrated air-ground collaborative perception network composed of drones, roadside ETC gantries, and toll booths, it achieves spatiotemporal calibration and feature fusion of multi-source data, solving the problems of incomplete coverage and significant environmental influence of single data sources, and significantly improving the accuracy and robustness of traffic state perception. Addressing the differences in traffic flow characteristics among basic road sections, hub interchange sections, and entrance / exit ramp sections, it constructs discrimination mechanisms based on physical models and spatial constraints clustering, sample balancing and binary classification trees, queuing theory, and capacity reduction, respectively, enabling targeted identification of different sections and significantly improving the accuracy of congestion identification. By embedding the macroscopic traffic flow model as a physical constraint into the clustering algorithm, replacing random initialization with calibration parameters, and introducing a deviation correction mechanism during the iteration process, it achieves a deep integration of traffic flow physical mechanisms and data-driven methods, ensuring both discrimination accuracy and enhancing model interpretability. The identification results of basic road sections and hub interchange sections are quantified into influencing factors, and the arrival rate and reduction coefficient of the ramp section discrimination model are adjusted in reverse. This enables the cascading transmission of state information between upstream and downstream sections, significantly improving the spatiotemporal continuity of congestion assessment. By extracting road network topology relationships and quantifying congestion diffusion intensity through graph neural networks, and combining this with a time-series prediction network to perform rolling predictions of future states, the propagation path is traced backward from the congestion node, achieving accurate location of congestion sources and advanced prediction of diffusion trends. Based on state levels and diffusion paths, core congestion areas, diffusion impact areas, and peripheral unobstructed areas are dynamically divided, and differentiated first to third-level handling instructions are matched and pushed to multiple terminals for collaborative execution, realizing a complete closed loop and precise coordinated control from perception to decision-making to execution. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of a method for real-time statistics of traffic flow parameters and intelligent congestion determination based on drone inspection, provided in an embodiment of the present invention.

[0032] Figure 2 This is an architecture diagram of a traffic flow parameter real-time statistics and congestion intelligent determination system based on drone inspection, provided as an embodiment of the present invention.

[0033] Figure 3 This is a schematic diagram of an electronic device provided according to an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0035] It should be noted that in the embodiments of the present invention, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0036] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0037] Example 1

[0038] Figure 1 This is a schematic diagram illustrating a method for real-time traffic flow parameter statistics and intelligent congestion determination based on unmanned aerial vehicle (UAV) inspection, as provided in an embodiment of the present invention. Figure 1 As shown, a method for real-time statistics of traffic flow parameters and intelligent congestion determination based on drone inspection includes:

[0039] S1. Construct a collaborative sensing network for multi-source data collection and spatiotemporal calibration, extract traffic characteristic parameters, and build a three-level state system. Through a collaborative network composed of drones, roadside equipment, and toll booths, data is collected and spatiotemporally aligned for different sections of the highway. Characteristic indicators that reflect traffic operation status are extracted from this data, and an evaluation system including three levels—smooth flow, transition, and congestion—is established based on this.

[0040] Specifically, in this embodiment, the construction of a collaborative sensing network for multi-source data acquisition and spatiotemporal calibration, extraction of traffic feature parameters, and construction of a three-level state system specifically includes:

[0041] S1.1 The cooperative sensing network consists of a drone platform, a roadside ETC gantry, and toll booths. The drone platform is equipped with a dual-mode visible light and infrared camera module, the roadside ETC gantry is equipped with microwave radar and an antenna array, and the toll booths are equipped with license plate recognition cameras and vehicle detectors. In this embodiment, the cooperative sensing network consists of three parts: the drone platform is responsible for aerial video acquisition, the roadside ETC gantry is installed along the highway to read vehicle electronic tag information, and the toll booths are set at entrances and exits to record vehicle entry and exit. ETC is the abbreviation for Electronic Toll Collection, which refers to a technology that allows vehicles to complete toll settlement without stopping when passing through toll stations.

[0042] S1.2 The multi-source data acquisition adopts a differentiated sampling strategy: For basic road sections, drone aerial video and ETC transaction data are collected at a first preset frequency; for hub interchange sections, drone aerial video and ETC transaction data are collected at a second preset frequency; and for entrance and exit ramp sections, drone aerial video and toll gate vehicle data are collected at a third preset frequency. In this embodiment, different collection frequencies are used for different sections: Traffic flow is relatively stable on basic road sections, so drones can collect data at a lower frequency; traffic flow at hub interchange sections is complex, requiring higher frequency collection to capture instantaneous changes; entrance and exit ramp sections involve acceleration, deceleration, and queuing processes, so higher frequency collection is also used. The ETC gantry automatically records one data point each time a vehicle passes, and the toll gate generates one vehicle passage record each time a vehicle enters or exits.

[0043] S1.3 The spatiotemporal calibration uses the Global Navigation Satellite System (GNSS) to add a unified timestamp to the collected data, and uses the geographical coordinates of the ETC gantry as a spatial reference to perform perspective transformation on the UAV aerial video to establish a mapping relationship between pixel coordinates and road surface coordinates. In this embodiment, firstly, a unified timestamp is added to each piece of collected data using the GNSS to ensure that data from different sources are aligned in time. Then, the latitude and longitude coordinates of the ETC gantry are used as a spatial reference to perform perspective transformation on the video images captured by the UAV. Perspective transformation refers to converting an image viewed from a tilted angle into a vertically tilted planar image, thereby establishing a correspondence between the video pixel positions and the actual road surface positions.

[0044] S1.4 The traffic characteristic parameters include at least traffic volume, average vehicle speed, and traffic density. The three-level state system includes free-flow, transitional, and congested flow states. Each state corresponds to a preset traffic volume threshold range, vehicle speed threshold range, and density threshold range. In this embodiment, three core indicators are extracted from the processed data: traffic volume refers to the number of vehicles passing through a certain cross-section per unit time; average vehicle speed refers to the average speed of vehicles passing through a certain road segment; and traffic density refers to the number of vehicles per unit length of road segment. Based on the values ​​of these three indicators, the traffic state is divided into three levels: free-flow indicates sparse traffic and smooth travel; transitional flow indicates that traffic is close to saturation and speed begins to decrease; and congested flow indicates dense traffic, slow travel, or stagnation. Each level corresponds to a preset value range; for example, when traffic volume is below a certain value and vehicle speed is above a certain value, it is determined to be a free-flow state.

[0045] S2. For basic road sections, a macroscopic traffic flow model and spatially constrained clustering algorithm are used to output state identification results based on the aforementioned feature parameters. For straight highway mainline sections, a traffic flow model describing the relationship between speed and density is used, combined with a clustering algorithm that considers the distance between road sections. The collected data is compared with preset category centers to determine whether the current state is smooth, transitional, or congested.

[0046] Specifically, in this embodiment, the step of using a macroscopic traffic flow model and spatial constraint clustering algorithm to output state identification results for basic road segments includes:

[0047] S2.1 The macroscopic traffic flow model adopts a nonlinear relationship model between speed and density. Model parameters are determined through historical data fitting, and three key parameters—free-flow speed, congestion density, and saturation flow—are calibrated. The calibration results are used as the initial cluster centers for the clustering algorithm. In this embodiment, historical traffic flow data for basic road segments is first collected, including vehicle speeds and corresponding traffic densities at different times. These data points are plotted on a coordinate system and fitted using a nonlinear curve. Fitting refers to finding a curve that best represents the data distribution trend. Based on the fitted curve, three key parameters are calibrated: free-flow speed refers to the highest speed vehicles can reach when traffic density is close to zero; congestion density refers to the limit vehicle density of the road segment when vehicle speed drops to near zero; and saturation flow refers to the maximum number of vehicles that can pass through the road segment in a steady state. The values ​​corresponding to these three parameters are used as the initial cluster centers, replacing the traditional random initialization method.

[0048] The nonlinear curve is constructed based on a three-segment macroscopic traffic flow model, and its mathematical expression is:

[0049] ,

[0050] in, This is a traffic flow function, which is the number of vehicles passing through a road segment section per unit time, expressed in vehicles per hour. Traffic density is the number of vehicles per unit length of road segment, expressed in vehicles per kilometer. Free-flow velocity, which is the highest speed a vehicle can reach when the density approaches zero, is expressed in kilometers per hour. Saturation flow is the maximum number of vehicles that can pass through a road segment under steady-state conditions, expressed in vehicles per hour. Congestion wave velocity, which is the speed at which congestion propagates upstream, is measured in kilometers per hour. Congestion density refers to the maximum vehicle density on a road segment when vehicle speed drops to near zero, measured in vehicles per kilometer.

[0051] The system collects the density of each road segment in real time. As input, calculate the values ​​of the three sub-expressions respectively, where the first sub-expression... This represents the linear relationship between flow rate and density under free-flow conditions, and is suitable for scenarios with low density and where vehicles can move freely; the second sub-equation This represents the nonlinear relationship between flow rate and saturation flow rate asymptotically with increasing density under transitional flow conditions, applicable to scenarios where density approaches a critical value; the third sub-equation. This represents the linear decrease in traffic flow with increasing density under congested flow conditions, suitable for scenarios with high density and slow vehicle movement. The three sub-equations characterize the three stages of traffic flow evolution from free flow to congested flow. The minimum value is chosen because actual traffic flow is limited by the minimum possible capacity under the current density conditions, a recognized physical constraint in traffic flow theory. The system collects the density ρ of each road segment in real time as input, calculates the values ​​of the three sub-equations, and takes the minimum value as the current theoretical traffic flow. By comparing the residuals with the measured flow rate, the least squares method is used for dynamic adjustment. and This allows the model output to approximate the actual observations. The calibrated parameters are used as the initial centers for the clustering algorithm, replacing the traditional random initialization method.

[0052] S2.2 The spatially constrained clustering algorithm adopts the fuzzy mean clustering framework, introducing a decay weight based on the spatial distance between road segments into the objective function. This decay weight decreases exponentially with the increase of the spatial distance between road segments. In this embodiment, the fuzzy mean clustering algorithm is used to classify traffic conditions. Fuzzy mean clustering is a clustering method that allows a sample point to belong to multiple categories simultaneously. Each sample point has a membership degree between 0 and 1 for each category. To consider the spatial relationship between road segments, a decay weight is introduced into the clustering objective function. This weight decreases exponentially with the increase of the spatial distance between two road segments. The exponential decrease means that the weight decreases faster the distance, making spatially adjacent road segments have a greater mutual influence during clustering, while road segments that are farther apart have a smaller influence, thus ensuring the spatial continuity of the clustering results.

[0053] The attenuation weight is expressed in the form of a Gaussian radial basis function, and its calculation formula is as follows:

[0054] ,

[0055] in, For road section and road section The spatial constraint weights between them have a range of (0,1]; For road section and road section The Euclidean spatial distance between the centers of mass, in meters; Spatial influence scale parameter, controlling the rate at which the weight decays with distance, in meters; : using natural constant The system uses an exponential function with base 0.5. Based on road network geographic information, it establishes a centroid coordinate table for road segments and calculates the Euclidean distance between any two road segments. .

[0056] Spatial influence scale parameters The decay rate used to control spatial weights needs to be determined based on the geometric characteristics of the road network. Specifically, the mean centroid distance between all adjacent road segments in the road network is first calculated. ,Right now , where E is the set of adjacent road segment pairs, and |E| is the total number of adjacent road segment pairs. Then... Values Between 1.5 and 2 times, and preferably taken in this embodiment. =1.8* For scenarios lacking historical road network data, Alternatively, cross-validation can be used to determine the dataset: historical traffic data is divided into training and validation sets based on spatial distance, and the set that maximizes the silhouette coefficient of the clustering result is selected. Value. When < hour, A value approaching 1 indicates strong constraints between adjacent road segments; when Much larger hour, A value close to 0 indicates weak constraints between long-distance road segments. This weight matrix remains unchanged during the clustering iteration process, serving as a multiplicative correction term for the objective function.

[0057] S2.3. The membership degree of each sample point to each cluster center is calculated iteratively, and the sample point is assigned to the category with the highest membership degree. The results of identifying the three states of the basic road segment—smooth, transitional, or congested—are output. In this embodiment, the membership degree of each sampling point relative to each cluster center is calculated iteratively. Iterative optimization refers to repeatedly executing the calculation process, adjusting parameters each time based on the current result, until the result no longer changes significantly. For each sampling point, its membership degree to each cluster center is compared, and it is assigned to the category with the highest membership degree. The three categories correspond to the three states—smooth, transitional, and congested—and the final output is the state identification result of the basic road segment in each time period.

[0058] Furthermore, the coupling method between the macroscopic traffic flow model and the spatially constrained clustering algorithm is as follows:

[0059] An initial state class center matrix is ​​constructed using the key parameters obtained from the macroscopic traffic flow model calibration, including free-flow class centers, transitional flow class centers, and congested flow class centers. In this embodiment, the three key parameters obtained from the macroscopic traffic flow model calibration correspond to the initial centers of the three state categories, respectively. Free-flow class centers consist of free-flow velocity and corresponding low-density and medium-flow rates; transitional flow class centers consist of critical velocity and corresponding optimal density and saturation flow rates; congested flow class centers consist of near-zero velocity and corresponding congestion density and low flow rates. This set of initial centers is arranged in matrix form as the starting point for subsequent clustering operations. A matrix refers to an array structure arranged in rows and columns, where each row represents a state category and each column represents a traffic characteristic parameter.

[0060] The initial state cluster center matrix is ​​used as the starting point for the clustering algorithm iteration, replacing the traditional random initialization method. In this embodiment, the traditional clustering algorithm randomly selects the initial cluster centers, which may lead to different results each time and may deviate from actual traffic patterns. This invention uses the parameters obtained from the calibration of the macroscopic traffic flow model as the input to the clustering algorithm with fixed initial centers, so that the clustering iteration starts from a position that conforms to the physical laws of traffic flow, avoiding the uncertainty caused by random initialization, and accelerating the convergence speed of the algorithm. Convergence refers to the state where the position of the cluster centers no longer changes significantly after multiple iterations.

[0061] During the clustering iteration process, the speed-density relationship determined by the macroscopic traffic flow model is used as the physical constraint boundary. When the cluster center deviates from the speed-density relationship curve by more than a preset deviation threshold during the iteration process, a correction mechanism is triggered to return the cluster center to the nearest legal position on the curve.

[0062] In this embodiment, the ideal speed-density relationship curve determined by the macroscopic traffic flow model is used as the physical constraint boundary. During the clustering iteration process, when the cluster center deviates from this curve by more than a preset allowable range after an update, the system automatically triggers a correction mechanism. The specific correction method is as follows: calculate the distance from the cluster center point to each point on the curve, find the closest point, and revert the cluster center to this legal position. The preset deviation threshold is set according to the actual road conditions and accuracy requirements; for example, it can be set to a speed deviation of no more than five kilometers per hour or a density deviation of no more than five vehicles per kilometer. This mechanism ensures that the clustering results always conform to the basic physical laws of traffic flow.

[0063] After iterative convergence, the final cluster centers, corrected for physical constraints, and their corresponding state identification results are output. In this embodiment, the process of clustering iteration, bias judgment, and correction rollback is repeated until the cluster centers converge. The final cluster centers obtained at this point reflect both the distribution characteristics of the data itself and satisfy the speed and density constraints of the macroscopic traffic flow model. Based on the final cluster centers, the membership degree of each sampling point is calculated and assigned to the corresponding state category, outputting the identification results of the basic road segment's smooth, transitional, or congested state in each time period.

[0064] S3. For hub interchange sections, sample balancing and binary classification trees are used to output status classification results based on the aforementioned feature parameters and basic road segment identification results. For hub areas where traffic flows from different directions converge, the few categories in the sample data are first expanded to balance the sample distribution. Then, a binary tree classification model is used to output the smooth flow, warning, or congestion judgment results for the area based on features such as the ratio of upstream and downstream traffic flows.

[0065] Specifically, in this embodiment, the process of using sample balancing and binary classification tree to output state classification results for the hub interchange section includes:

[0066] S3.1 Construct a feature vector for the hub interchange section. The feature vector includes at least three dimensions: upstream and downstream traffic flow feature ratio, lane change rate, and weaving section density. The upstream and downstream traffic flow feature ratio is calculated based on the traffic volume in the feature parameters.

[0067] In this embodiment, a feature vector with three dimensions is constructed to address the weaving characteristics of traffic flow in the hub interchange section. The upstream-downstream traffic flow characteristic ratio refers to the ratio of traffic flow in the upstream segment to that in the downstream segment. This ratio is calculated based on collected traffic volume data; when the upstream flow is significantly greater than the downstream flow, congestion may occur. The lane change rate refers to the frequency with which vehicles switch lanes between the mainline and ramps, reflecting the lane-changing intensity in the weaving area. The weaving segment density refers to the number of vehicles simultaneously present within a unit length of weaving area; higher density indicates more frequent conflicts. These three features collectively characterize the operational status of the hub interchange section.

[0068] S3.2. The congestion category samples are augmented using synthetic oversampling technology to generate new samples with a distribution similar to the original samples. The neighbor sample cleaning algorithm is used to remove boundary noise points and outliers in the augmented sample set, so that the number of samples in different state categories reaches a preset balance ratio.

[0069] A synthetic oversampling technique is used to augment the congestion category samples, generating new samples with a distribution similar to the original samples. Specific parameters are set as follows: for each sample in the congestion category sample set... Randomly select a sample from its K nearest neighbor congestion category samples. ,exist and A new synthetic sample is randomly generated on the connection line. ,in The value is a uniform random number between 0 and 1. K is set to 5, and the target oversampling factor is to make the number of congested category samples reach 80% of the number of uncongested category samples. After expansion, a neighbor sample cleaning algorithm is used to remove boundary noise points and outliers: for each sample, the categories of its 5 nearest neighbor samples are checked. If more than 3 of them belong to other categories, the sample is identified as a boundary noise point and removed from the sample set.

[0070] In this embodiment, because congestion occurs infrequently in actual highway operation, the number of congestion categories in the collected samples is far less than the number of smooth-flowing categories. Synthetic oversampling technology involves selecting samples from the minority class and randomly generating new synthetic samples along the lines connecting them to neighboring samples, thereby increasing the number of minority class samples. The neighbor sample cleaning algorithm involves checking the category of each sample's neighboring samples within a certain range; if most neighbors belong to other categories, the sample is determined to be a boundary noise point or outlier and is removed. After augmentation and cleaning, the number of samples in different state categories reaches a pre-set equilibrium ratio.

[0071] S3.3 Using the Gini coefficient as the evaluation criterion for node splitting, a binary classification tree is recursively constructed. Each internal node of the binary classification tree corresponds to a feature and its splitting threshold, and each leaf node corresponds to a state category.

[0072] In this embodiment, a binary classification tree is used to classify the status of the hub interconnection segment. A binary classification tree is a tree-structured decision model where each internal node represents a judgment on a feature, and each leaf node represents the final classification result. The Gini coefficient is an indicator of the purity of the node's data; a smaller Gini coefficient indicates that the samples in that node are more concentrated in the same category. The algorithm starts from the root node, traverses all possible splitting thresholds for each feature, selects the feature and threshold that causes the Gini coefficient of the child node to decrease the most, and then repeats the above process for each child node, recursively constructing the entire tree.

[0073] S3.4. During the construction process, a cost complexity pruning strategy is adopted. The subtree with the lowest misclassification rate in the subtree is selected as the final classifier through cross-validation. The feature parameters and basic road segment identification results are input into the final classifier, and the three status classification results of the hub interchange section are output: smooth, warning, or blocked.

[0074] In this embodiment, a fully grown decision tree may suffer from overfitting, meaning it classifies training data accurately but performs poorly on new data. The cost-complexity pruning strategy comprehensively considers the classification error and complexity of the tree, calculating a comprehensive cost index for each subtree and selecting the subtree with the lowest cost as the final classifier. Cross-validation involves dividing the data into multiple parts, using one part for training and another for validation in turn, to evaluate the model's generalization ability. The feature parameters and the identification results of the basic road segments are input into the final classifier. Based on the judgment results of the samples at each node, the classification is passed down the tree to the leaf nodes, outputting the classification results of the hub interchange section as either unobstructed, under warning, or congested.

[0075] S4. For entrance and exit ramp sections, a weighted fusion of queuing model and capacity reduction model is used. The congestion level is determined based on the aforementioned feature parameters, basic road sections, and hub interchange identification results. For toll station and ramp areas, on the one hand, vehicle waiting length is calculated using queuing theory model; on the other hand, capacity is reduced based on the turning angle and length of the ramp. The two are then weighted and combined, and the levels are classified as smooth, warning, or congested according to the score.

[0076] Specifically, in this embodiment, the weighted fusion of the queuing model and the capacity reduction model to determine the congestion level for the entrance and exit ramp sections includes:

[0077] S4.1 The queuing model adopts a multi-service station queuing framework. Based on the traffic volume as the arrival rate, the number of service channels at the toll gate as the number of service stations, and the historical average service time as the service rate, the theoretical queue length and average waiting time are calculated.

[0078] In this embodiment, a multi-service-desk queuing framework is used to simulate the vehicle queuing process in a toll station area. A multi-service-desk queuing framework refers to a queuing system with multiple parallel service windows. Vehicles are assigned to available windows on a first-come, first-served basis upon arrival. Traffic volume (the number of vehicles arriving at the toll station per unit time) is used as the arrival rate; the number of open lanes at the toll booths is used as the number of service counters (the number of service windows that can process vehicles simultaneously); and the historical average service time is used as the service rate (the average time required for each vehicle to complete service). Based on these parameters, the theoretical queue length (the number of vehicles waiting for service at the toll station) and the average waiting time (the average waiting time for each vehicle from arrival to the start of service) are calculated.

[0079] This invention employs an M / M / c multi-server queuing theory model, assuming that vehicle arrivals follow a Poisson distribution and service times follow a negative exponential distribution, with a theoretical average queue length. and average waiting time The calculation formula is:

[0080] ,

[0081] ,

[0082] Among them, the system idle probability It is given by the following formula:

[0083] ,

[0084] System busy rate Defined as:

[0085] ,

[0086] in, The average vehicle arrival rate, i.e., the number of vehicles arriving at the toll station per unit time, is calculated based on the traffic volume Q extracted from S1.4, using the formula λ = Q / 3600, where Q is in vehicles / hour and λ is in vehicles / second. μ is the average service rate per lane, i.e., the number of vehicles that each toll lane can serve per unit time, derived from the historical average service duration. The reciprocal is determined, i.e., μ = 1 / . The method for obtaining μ is as follows: extract the passage records of all vehicles from the historical data of the toll booths, calculate the time interval from the arrival of each vehicle at the toll window to the completion of service and departure from the window, and take the arithmetic mean after removing the top 5% extreme values. If historical data is lacking, μ can be set with empirical values ​​according to the toll station type: μ = 0.5 vehicles / second (i.e., average service time of 2 seconds) for ETC dedicated lanes, and μ = 0.2 vehicles / second (i.e., average service time of 5 seconds) for MTC manual lanes. This represents the number of service lanes currently open at the toll booth, and is an integer. System busy rate represents the proportion of time the toll collection system is in a busy state; This represents the system idle probability, which is the probability that no vehicle is receiving service in any of the service channels. The theoretical average queue length is the number of vehicles waiting for service at the toll station, expressed in vehicles. The theoretical average waiting time is the average time required for a vehicle to start receiving service from its arrival, expressed in seconds.

[0087] The system reads the number of open lanes at the toll booths in real time. Arrival rate calculated from traffic volume data Service rate obtained by combining historical statistics First, calculate the busy rate. ,like ≥1 indicates that the system is oversaturated, and the queue length will grow indefinitely, directly classified as severe congestion; if If the value is less than 1, then calculate sequentially. , and Output queue score.

[0088] S4.2 The capacity reduction model reduces the baseline capacity based on the turning angle and length of the ramp. The larger the turning angle or the shorter the ramp length, the larger the reduction coefficient. The reduced capacity is equal to the baseline capacity multiplied by the turning reduction factor and then by the length reduction factor.

[0089] In this embodiment, the capacity of a ramp is affected by its geometry and needs to be reduced. The steering angle refers to the turning radius of a vehicle entering the ramp from the main line or merging into the main line from the ramp. The larger the steering angle, the more significant the vehicle deceleration, and the greater the decrease in capacity. The ramp length refers to the distance from the ramp's starting point to its ending point. The shorter the ramp length, the less space there is for vehicle acceleration or deceleration, and the lower the capacity. The reduced capacity equals the base capacity multiplied by the steering reduction factor and then multiplied by the length reduction factor. The steering reduction factor is a coefficient less than 1, and its value decreases as the steering angle increases; similarly, the length reduction factor is also a coefficient less than 1, and its value decreases as the ramp length decreases.

[0090] To enable the reduction factor to quantify the impact of geometric parameters, the following capacity reduction model is constructed:

[0091] ,

[0092] in, The effective traffic capacity after reduction is expressed in vehicles per hour. The baseline capacity is the maximum capacity under ideal geometric conditions, expressed in vehicles per hour. Ramp turning angle, which is the turning radius of a vehicle when it enters a ramp from the main line or merges into the main line from a ramp, is measured in degrees. The maximum steering angle is usually taken as 90 degrees as the normalization reference. The length of the ramp is the distance from the start to the end of the ramp, measured in meters. The average length of the ramp is taken as the benchmark for the length attenuation characteristic scale, and the unit is meters. This is the steering reduction factor, with a value range of [0,1], and a default value of 0.3; The length reduction factor has a value range of [0,1], with a default value of 0.5.

[0093] The system obtains the steering angle based on the design parameters of the ramp. and ramp length Substitute into the above formula to calculate the effective throughput. Compare the current actual traffic with The ratio is used as a reduction score, and the closer the ratio is to 1, the more strained the traffic capacity.

[0094] S4.3. The queuing score output by the queuing model and the capacity reduction score output by the capacity reduction model are weighted and fused together, wherein the weight coefficients of the queuing score and the capacity reduction score are dynamically adjusted according to the real-time traffic conditions.

[0095] In this embodiment, the queuing score output by the queuing model and the capacity reduction score output by the capacity reduction model are weighted and fused. Weighted fusion means multiplying each score by its respective weight coefficient and then adding them together to obtain a comprehensive score. The weight coefficient represents the importance of each score in the fusion result, and the sum of the two weight coefficients equals 1. The weight coefficients of the queuing score and the capacity reduction score are dynamically adjusted according to real-time traffic conditions: when a continuous increase in the queue length before the toll station is detected, the weight of the queuing score is appropriately increased; when a significant shortage of ramp capacity is detected, the weight of the capacity reduction score is appropriately increased. Dynamic adjustment means that the weight coefficients are not fixed but are updated in real time according to changes in real-time monitoring data.

[0096] The weighted fusion comprehensive score Calculated using the following formula:

[0097] ,

[0098] The weighting coefficients satisfy the following: , The score is a comprehensive score, with a range of [0,1]. A higher score indicates more severe congestion. The dynamic weight for queuing scoring is initially set to 0.5; To reduce the dynamic weight of the score, the initial value is 0.5; This represents the current theoretical queue length, in vehicles. This is the preset maximum allowable queue length, in units of vehicles; This represents the current actual traffic flow, expressed in vehicles per hour. The effective traffic capacity after reduction is expressed in vehicles per hour.

[0099] The system first calculates the queuing score. and reduction score The weights are dynamically adjusted based on real-time monitoring data. The queue length growth rate is calculated by sampling the queue length every 30 seconds. Calculate the difference Δ between two consecutive samples. = (t)- (t-30), growth rate r=Δ / 30, in vehicles / second. When the growth rate r exceeds the preset threshold of 0.5 vehicles / second and persists for more than two sampling periods, it is determined to be a queuing surge state, and weight adjustment is performed: =min( +0.1, 0.8), =1- When the capacity saturation = / If the value exceeds 0.9 and persists for more than three sampling periods, it is determined to be a state of insufficient traffic capacity, and weight adjustment is performed: =min( +0.1, 0.8), =1- The weight adjustment has an upper limit of 0.8 and a lower limit of 0.2 to prevent any single weight from becoming excessively dominant and causing fusion failure. When both queue surge and insufficient throughput are simultaneously met, priority is given to executing the weight adjustment. The adjustments have been made.

[0100] The weights are dynamically adjusted based on real-time monitoring data: when the queue length growth rate exceeds a threshold... Increase by 0.1 The corresponding reduction is 0.1; when the traffic capacity saturation exceeds 0.9, Increase by 0.1 Decrease accordingly by 0.1. Calculate the overall score. Then, the congestion level is determined by comparing it with a preset threshold.

[0101] S4.4. Based on the weighted and fused comprehensive score, compare it with the preset congestion level threshold. When the comprehensive score exceeds the first threshold, it is determined to be a congestion level. When it is between the second threshold and the first threshold, it is determined to be a warning level. When it is below the second threshold, it is determined to be a smooth level.

[0102] In this embodiment, the congestion level of the ramp segment is determined by comparing the weighted and fused comprehensive score with a preset congestion level threshold. The preset threshold includes a first threshold and a second threshold, where the first threshold is higher than the second threshold. The first threshold corresponds to the lowest comprehensive score for the congestion level, and the second threshold corresponds to the lowest comprehensive score for the warning level. When the comprehensive score exceeds the first threshold, it is determined to be at the congestion level, indicating that the ramp segment is severely congested and vehicles need to wait for a long time; when the comprehensive score is between the second threshold and the first threshold, it is determined to be at the warning level, indicating that queuing has begun to appear on the ramp segment but has not yet caused severe congestion; when the comprehensive score is lower than the second threshold, it is determined to be at the smooth flow level, indicating that vehicles can pass smoothly.

[0103] Furthermore, the weighted fusion also integrates the identification results of basic road sections and hub interchange sections, specifically including:

[0104] S4.3.1 Obtain the state identification results of the basic road segments and the state classification results of the hub interchange segments, and convert them into corresponding quantitative influence factors. In this embodiment, the state identification results of the basic road segments and the state classification results of the hub interchange segments are converted into quantitative influence factors that can be used for numerical calculation.

[0105] The quantitative impact factor refers to mapping discrete state categories to continuous numerical values. The mapping rules are determined through statistical analysis of historical data. Specifically, the method involves collecting data on the corresponding state categories of each road segment and the actual queue lengths of downstream ramps within historical time periods, and calculating the average downstream queue lengths under different state categories. In a smooth state As a benchmark value Define the mapping values ​​of other state categories as / Based on actual measurements and statistics, the following mapping rules are adopted in this embodiment: Smooth flow is mapped to 1.0; transitional state (basic road section) or warning state (hub interchange section) is mapped to between 1.6 and 2.0 (specifically dynamically calibrated based on historical congestion probability; 1.8 is used in this embodiment); and congested state is mapped to between 2.5 and 3.5 (3.0 is used in this embodiment). This quantitative impact factor is used as a multiplicative coefficient, directly multiplied by the arrival rate λ of the queuing model or the reduction coefficient of the capacity reduction model.

[0106] The specific mapping relationship can be determined according to the actual road conditions. This transformation allows the qualitative state description to participate in subsequent mathematical operations and model parameter adjustments.

[0107] S4.3.2 Adjust the arrival rate parameter of the queuing model based on the state identification results of the basic road segment. When the basic road segment is in a congested state, increase the arrival rate to reflect the cumulative effect of upstream vehicles. In this embodiment, the arrival rate parameter of the queuing model is dynamically adjusted based on the state identification results of the basic road segment. The arrival rate refers to the number of vehicles arriving at the toll station area per unit time and is a key input parameter of the queuing model. When the basic road segment is in a congested state, it indicates that congestion has occurred upstream, and a large number of vehicles are queuing and moving slowly downstream. At this time, it is necessary to increase the arrival rate parameter of the queuing model to reflect the cumulative effect of upstream vehicles on the ramp section and avoid underestimating the number of arriving vehicles, which would lead to an underestimation of the queue length prediction.

[0108] S4.3.3 Adjust the reduction coefficient of the capacity reduction model based on the status classification results of the interchange section. When the interchange section is in a warning or congestion state, increase the reduction coefficient to reflect the blocking effect of congestion propagating upstream. In this embodiment, the reduction coefficient of the capacity reduction model is dynamically adjusted based on the status classification results of the interchange section. The reduction coefficient reflects the degree to which various adverse factors weaken the ramp capacity. When the interchange section is in a warning or congestion state, it indicates that congestion has occurred in the traffic convergence area. This congestion will propagate upstream, blocking the merging of traffic on connected ramps. At this time, it is necessary to increase the reduction coefficient of the capacity reduction model, that is, further reduce the estimated capacity of the ramp, to reflect the restrictive effect of downstream congestion on upstream incoming traffic.

[0109] S4.3.4. The adjusted queuing score and the reduced score are weighted and fused together to recalculate the comprehensive score. The congestion level of the ramp segment is determined again based on the updated comprehensive score, realizing the cascading transmission and fusion judgment of the status information of upstream and downstream segments. In this embodiment, the adjusted queuing score and the reduced score are weighted and fused together again to recalculate the comprehensive score. Cascading transmission refers to the transmission of the status information of upstream or adjacent segments to downstream segments level by level along the traffic flow direction, affecting the judgment result of that segment. The congestion level of the ramp segment is determined again based on the updated comprehensive score, realizing the fusion judgment of the status information of upstream and downstream segments. Through this mechanism, the congestion status of basic road segments and the congestion status of hub interchanges can be transmitted to the judgment model of ramp segments, so that the congestion level judgment of ramp segments not only depends on local queuing and traffic capacity data, but also reflects the influence of traffic status from upstream and downstream adjacent segments, thereby improving the spatiotemporal continuity and accuracy of the judgment.

[0110] S5. The identification results of basic road sections, hub interchange sections, and ramp sections are collected to construct a road network state matrix. A graph neural network and a time-series prediction network are used to output the congestion spread path of the entire road network. The judgment results of the aforementioned sections are summarized to construct a matrix reflecting the changes in the state of the entire road network. A graph neural network is used to analyze the degree of mutual influence between road sections, and time-series prediction technology is combined to predict the state change trend over a future period, ultimately depicting how congestion spreads from its source to the surrounding areas.

[0111] Specifically, in this embodiment, the process of collecting the identification results of basic road sections, hub interchange sections, and ramp sections to construct a road network state matrix, and using graph neural networks and time-series prediction networks to output the congestion diffusion path of the entire road network, specifically includes:

[0112] S5.1. Using each road segment in the highway network as a node and the topological connection relationship between adjacent road segments as an edge, construct the road network topology graph structure.

[0113] In this embodiment, the highway network is abstracted as a graph structure. The graph structure consists of nodes and edges. Nodes represent individual road segment units within the network, and edges represent the topological connections between adjacent road segments. A topological connection means that two road segments are spatially connected end-to-end, allowing vehicles to directly enter from one road segment to another. By constructing the road network topology graph, the physical road network is transformed into a data structure that can be processed by a computer.

[0114] S5.2 Map the status identification results of basic road segments, the status classification results of hub interchange segments, and the congestion levels of ramp segments to the corresponding graph nodes. Use the current status level of each node as the node attribute value to construct a road network status matrix, where the row index of the road network status matrix corresponds to the road segment number and the column index corresponds to the time series.

[0115] In this embodiment, the identification results of basic road segments, hub interchange segments, and ramp segments are mapped to corresponding graph nodes, with the current state level of each node serving as its attribute value. The road network state matrix is ​​a two-dimensional table, where rows correspond to different road segment numbers, columns correspond to different time sampling points, and each cell records the state level of the road segment at the corresponding time. By constructing the state matrix, the discrete identification results scattered across various segments are integrated into a unified time-series data structure, providing a foundation for subsequent situational analysis.

[0116] S5.3. A graph neural network is used to extract node features and edge relationships in the road network topology graph. The state information of adjacent nodes is aggregated through graph convolution operations. The state difference value between each node and its neighboring nodes is calculated. The diffusion intensity of congestion between different road segments is quantified based on the state difference value.

[0117] In this embodiment, a graph neural network is used to extract node features and edge relationships from the road network topology graph. A graph neural network is a deep learning model specifically designed for graph-structured data, capable of aggregating the state information of neighboring nodes to the current node through graph convolution operations. Graph convolution operation involves collecting the state information of each node's neighbors, performing a weighted summation according to certain weights, and then updating the node's feature representation. Based on the aggregated node features, the state difference value between each node and its neighbors is calculated. The larger the difference value, the greater the likelihood of congestion propagating from that node to its neighbors, thereby quantifying the intensity of congestion spread between different road segments.

[0118] The diffusion intensity is calculated using a graph convolutional network, with the specific formula as follows:

[0119] ,

[0120] The node feature update formula is as follows:

[0121] ,

[0122] in, For the node To the node The intensity of congestion diffusion indicates the spread of congestion from road segment Spread to the road section The probability of; For nodes The feature vector contains information such as the road segment's status level, traffic flow, and density; For nodes eigenvectors; For nodes The one-hot encoded vector is used to index the target node; This is a trainable diffusion strength weight matrix; For the first Nodes after layer graph convolution eigenvectors; For nodes The set of neighboring nodes; For nodes The degree, that is, the number of nodes adjacent to this node; For the first The trainable weight matrix of the layer; The ReLU function is used as the non-linear activation function. For linear rectifier functions, defined as follows: .

[0123] The system inputs the road network state matrix into the trained graph convolutional network model. The graph convolutional network consists of two graph convolutional layers and one output layer. Node feature vectors The graph has 5 dimensions, specifically including: current state level (unobstructed is coded as 1, transitional as 2, and congested as 3), the ratio of current flow to saturated flow (saturation), the ratio of current density to congestion density (density), the average state level of upstream neighbor nodes, and the average state level of downstream neighbor nodes. The first layer of graph convolution maps the 5-dimensional features to 16-dimensional latent features, and the second layer of graph convolution maps the 16-dimensional latent features to 8-dimensional high-level features. The output layer uses the Softmax activation function to convert the high-level features into the propagation probability of each node to other nodes, forming an N×N diffusion intensity matrix, where N is the total number of nodes in the road network. Node pairs with a propagation probability greater than 0.6 are selected as effective diffusion paths.

[0124] During the forward propagation process, each layer updates the feature representation of the current node by aggregating the features of its neighboring nodes. The final output layer uses the Softmax function to normalize the diffusion intensity into a probability distribution, selects edges with a probability greater than 0.6 as effective diffusion paths, and outputs them to the temporal prediction network.

[0125] S5.4 Input the road network state matrix and the calculated diffusion intensity within the current time window into the time series prediction network. The time series prediction network takes the historical time series as input, captures the evolution of the state over time through a cyclic gating mechanism, and outputs the predicted state sequence of each node within several future time windows.

[0126] In this embodiment, the road network state matrix within the current time window and the calculated diffusion intensity are input into the time-series prediction network.

[0127] The temporal prediction network employs a gated recurrent unit (GRU) architecture, which requires fewer parameters and trains faster than Long Short-Term Memory (LSTM) networks, while achieving comparable performance in this application scenario. The input time window is set to 12 time steps, each corresponding to 1 minute, meaning it uses the historical state sequence of the past 12 minutes to predict future states. The output prediction step size is set to 6 time steps, predicting the state sequence of the next 6 minutes. The network structure consists of two GRU layers, each containing 64 hidden units, followed by a fully connected layer that maps the hidden states to state level outputs. During training, the cross-entropy loss function is used, employing the Adam optimizer with an initial learning rate of 0.001. The training run consists of 200 epochs, and an early stopping mechanism is implemented where training stops after 20 consecutive epochs if the loss does not decrease.

[0128] A cyclic gating mechanism refers to a network with memory units that selectively retain historical information, forget irrelevant information, and output current information, thereby capturing the patterns of traffic state evolution over time. The network takes a sequence of states from multiple historical moments as input and outputs a predicted sequence of states for each node at several future moments, enabling proactive prediction of future congestion trends.

[0129] S5.5 Based on the predicted state sequence, starting from the node currently in a blocked state, trace the propagation path of state deterioration along the direction of the greatest diffusion intensity until the predicted state is restored to smooth flow, and generate a congestion diffusion path and evolution trend map of the entire road network.

[0130] In this embodiment, starting from the currently congested node, the propagation path of the worsening situation is traced segment by segment along the direction of maximum diffusion intensity. The direction of maximum diffusion intensity refers to the direction with the largest state difference value among the neighboring nodes starting from the current node. The tracing process continues until the predicted state returns to unobstructed. The traced node sequence is connected to form a complete congestion propagation path. At the same time, the state change process of each node is plotted as an evolution trend graph to visually demonstrate how congestion spreads from the source to surrounding road segments, when it reaches its peak, and when it gradually dissipates.

[0131] S6. Based on the three-level state system and the diffusion path, generate a visual chart and output graded handling instructions. Overlay the current state level and the predicted diffusion path on the map to generate a visual chart. According to the severity of congestion and the scope of impact, divide the road network into core area, affected area and peripheral area. Match corresponding guidance, traffic restriction or information release instructions for different areas and send them to roadside equipment or navigation system for execution.

[0132] Specifically, in this embodiment, the step of generating a visual chart based on the three-level state system and diffusion path and outputting graded handling instructions includes:

[0133] S6.1 Overlay the current status level of each road segment in the entire road network and the predicted congestion spread path onto the geographic information system map, use different colors to indicate the three status levels of free flow, transitional flow and congested flow, and use arrows to indicate the propagation direction of the congestion spread path, and generate a congestion situation visualization chart that includes a status distribution layer and a spread path layer.

[0134] In this embodiment, the current status level of each road segment in the entire road network and the predicted congestion propagation path are overlaid and displayed on a geographic information system (GIS) map. A GIS map is an electronic map containing spatial information such as road location, direction, and connectivity. The three status levels are indicated by different colors: green for free flow, yellow for transitional flow, and red for congested flow. The propagation direction of the congestion propagation path is indicated by arrows, with the arrow pointing in the direction the congestion spreads from the current area to downstream areas. The final generated visualization contains two layers: a status distribution layer displaying the current color status of each road segment, and a propagation path layer displaying the arrow paths of congestion propagation.

[0135] S6.2 Based on the three-level state system and diffusion path, the highway network is divided into three control areas: the core congestion area, the diffusion impact area, and the peripheral unobstructed area. The core congestion area is the continuous road segment that is currently in a congested flow state and has the greatest diffusion intensity. The diffusion impact area is the road segment that is predicted to be affected by congestion within a set time window in the future. The peripheral unobstructed area is the remaining road segment.

[0136] In this embodiment, based on the three-level state system and diffusion path, the highway network is divided into three control zones. The core congestion zone is the continuous road segment currently in a congested flow state with the highest diffusion intensity. Diffusion intensity Defined as node i to all downstream neighbor nodes The sum of the diffusion intensities, i.e. Downstream ( ) is a node The set of downstream neighbor nodes. Among all nodes in a blocked flow state, select... The largest node is used as the core source node, and then the flow expands from this node along the upstream and downstream directions, connecting adjacent nodes that are in a blocked flow state. Larger than the core source node A contiguous section of road with a value of 50% is classified as the core congestion zone. The highest diffusion intensity means that the congestion in this area is rapidly spreading to the surrounding areas, and is the source of the entire congestion event.

[0137] The diffusion impact zone refers to road sections that, based on forecasts, will be affected by congestion within a predetermined time window. This time window can be set as needed, for example, five or ten minutes. The outer unobstructed zone refers to other road sections that do not fall into either of the above two categories; these road sections are currently unobstructed and will not be affected by congestion in the near future.

[0138] S6.3 Match corresponding graded handling instructions for different control areas: generate first-level handling instructions including forced diversion, ramp closure and variable speed limit for core congestion areas; generate second-level handling instructions including dynamic guidance, progressive speed limit and toll station adjustment for diffusion impact areas; and generate third-level handling instructions including information release and route recommendation for peripheral unobstructed areas.

[0139] In this embodiment, corresponding levels of handling instructions are matched for different control areas. The first-level handling instructions target the core congestion area, and measures include forced diversion (guiding vehicles to leave the highway early and take alternative routes), ramp closure (temporarily closing entrance ramps to prevent more vehicles from entering), and variable speed limits (dynamically reducing speed limits to smooth traffic flow). The second-level handling instructions target the spreading impact area, and measures include dynamic guidance (using navigation to suggest alternative routes to indicate upcoming congestion), progressive speed limits (gradually reducing speed limits in stages), and toll station adjustments (adjusting toll lane configurations and toll collection strategies). The third-level handling instructions target the outer unobstructed areas, and measures include information dissemination (informing about upcoming road conditions and estimated travel time) and route recommendations (suggesting faster alternative routes).

[0140] S6.4. The graded handling instructions are encapsulated into a standardized data structure containing the target road segment identifier, instruction type, execution parameters and effective time period, and pushed to the roadside variable information signs, navigation applications and toll station control systems for execution.

[0141] In this embodiment, the tiered handling instructions are encapsulated into a standardized data structure. The standardized data structure refers to an information packet organized according to a unified format, containing the following fields: target road segment identifier (the specific road segment number for which the instruction is executed), instruction type (whether it belongs to Level 1, Level 2, or Level 3), execution parameters (specific values ​​such as speed limit and diversion ratio), and effective time period (the start and end times of the instruction's effectiveness). After encapsulation, the instruction is pushed to three types of terminal devices for execution: roadside variable message signs (electronic displays installed above or on the side of the road), navigation applications (navigation software installed on mobile phones or in-vehicle devices), and toll station control systems (the toll gates and control computers of the toll lanes).

[0142] Example 2

[0143] like Figure 2 As shown in the diagram, this invention provides a system architecture diagram for real-time traffic flow parameter statistics and intelligent congestion determination based on drone inspection. This system is applied to the real-time traffic flow parameter statistics and intelligent congestion determination method based on drone inspection as described in Embodiment 1, and includes:

[0144] The data acquisition module 210 is used to construct a collaborative sensing network for multi-source data acquisition and spatiotemporal calibration, extract traffic feature parameters, and construct a three-level state system.

[0145] The basic road segment identification module 220 is used to output the status identification result based on the feature parameters of the basic road segment by adopting a macro traffic flow model and a spatial constraint clustering algorithm.

[0146] The hub interchange section identification module 230 is used to output the status classification result based on the feature parameters and the basic road section identification result for the hub interchange section by using sample equalization processing and binary classification tree.

[0147] The ramp segment identification module 240 is used to determine the congestion level for entrance and exit ramp segments by weighted fusion of queuing model and capacity reduction model, based on the characteristic parameters, basic road segments and hub interchange segment identification results.

[0148] The situation simulation module 250 is used to collect the identification results of basic road sections, hub interchange sections and ramp sections, construct the road network state matrix, and use graph neural networks and time series prediction networks to output the congestion diffusion path of the entire road network.

[0149] The strategy output module 260 is used to generate a visual chart and output hierarchical handling instructions based on the three-level state system and diffusion path.

[0150] Figure 3 This is an electronic device provided in one embodiment of the present invention. For example... Figure 3As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0151] In this embodiment of the invention, the memory 300 is used to store executable instructions of the processor 301, which is configured to implement the method as described in the first aspect when executing the instructions.

[0152] In an embodiment of the present invention, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0153] In one embodiment of the present invention, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0154] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0155] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0156] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0157] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0158] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for real-time statistics of traffic flow parameters and intelligent congestion determination based on unmanned aerial vehicle (UAV) inspection, characterized in that, Includes the following steps: A collaborative sensing network is constructed to collect multi-source data and perform spatiotemporal calibration, extract traffic feature parameters, and construct a three-level state system. For basic road sections, a macroscopic traffic flow model and a spatially constrained clustering algorithm are used to output state identification results based on the aforementioned feature parameters; For hub interchange sections, sample equalization processing and binary classification trees are used to output status classification results based on the aforementioned feature parameters and basic road segment identification results; For the entrance and exit ramp sections, a weighted fusion of the queuing model and the capacity reduction model is adopted, and the congestion level is determined based on the aforementioned feature parameters, the identification results of the basic road sections and hub interchange sections; The identification results of basic road sections, hub interchange sections and ramp sections are collected to construct a road network state matrix, and graph neural networks and time series prediction networks are used to output the congestion diffusion path of the entire road network. Based on the three-level state system and diffusion path, a visual chart is generated and a graded handling instruction is output.

2. The method according to claim 1, characterized in that, The construction of a collaborative sensing network for multi-source data acquisition and spatiotemporal calibration, extraction of traffic feature parameters, and construction of a three-level state system specifically includes: The collaborative sensing network consists of a drone platform, a roadside ETC gantry, and a toll booth. The drone platform is equipped with a visible light and infrared dual-mode camera module, the roadside ETC gantry is equipped with microwave radar and an antenna array, and the toll booth is equipped with a license plate recognition camera and a vehicle detector. The multi-source data acquisition adopts a differentiated sampling strategy: drone aerial videos and ETC transaction data are collected at a first preset frequency for basic road sections, drone aerial videos and ETC transaction data are collected at a second preset frequency for hub interchange sections, and drone aerial videos and toll gate vehicle passage data are collected at a third preset frequency for entrance and exit ramp sections. The spatiotemporal calibration uses the Global Navigation Satellite System to add a unified timestamp to the collected data, and uses the geographical coordinates of the ETC gantry as a spatial reference to perform perspective transformation on the drone aerial video to establish a mapping relationship between pixel coordinates and road surface coordinates. The traffic characteristic parameters include at least traffic volume, average vehicle speed, and traffic density. The three-level state system includes free flow, transitional flow, and congested flow, and each state corresponds to a preset traffic volume threshold range, vehicle speed threshold range, and density threshold range.

3. The method according to claim 1, characterized in that, The state identification results for basic road sections, generated using a macroscopic traffic flow model and spatial constraint clustering algorithm, specifically include: The macroscopic traffic flow model adopts a nonlinear relationship model between speed and density. The model parameters are determined by fitting historical data, and three key parameters, namely free flow speed, congestion density and saturation flow, are calibrated. The calibration results are used as the initial cluster centers of the clustering algorithm. The spatially constrained clustering algorithm adopts a fuzzy mean clustering framework and introduces a decay weight based on the spatial distance between road segments into the objective function. The decay weight decreases exponentially as the spatial distance between road segments increases. By iteratively optimizing the calculation of the membership degree of each sample point to each cluster center, the sample points are assigned to the category with the highest membership degree, and the three status identification results of the basic road segment—smooth, transitional, or congested—are output.

4. The method according to claim 3, characterized in that, The coupling method between the macro-traffic flow model and the spatially constrained clustering algorithm is as follows: An initial state class center matrix is ​​constructed using the key parameters obtained from the macroscopic traffic flow model calibration, including free flow class centers, transition flow class centers, and congested flow class centers. The initial state class center matrix is ​​used as the starting point for the clustering algorithm iteration, replacing the traditional random initialization method; During the clustering iteration process, the speed-density relationship determined by the macro traffic flow model is used as the physical constraint boundary. When the cluster center deviates from the speed-density relationship curve by more than a preset deviation threshold during the iteration process, a correction mechanism is triggered to revert the cluster center to the nearest legal position on the curve. After iterative convergence, the final cluster centers and corresponding state recognition results after physical constraint correction are output.

5. The method according to claim 1, characterized in that, The specific steps for processing sample balancing and outputting state classification results using a binary classification tree for the hub interchange section include: Construct a feature vector for the hub interchange section. The feature vector includes at least three dimensions: upstream and downstream traffic flow feature ratio, lane change rate, and weaving section density. The upstream and downstream traffic flow feature ratio is calculated based on the traffic volume in the feature parameters. Synthetic oversampling technology is used to augment the samples of the congestion category, generating new samples with a distribution similar to the original samples. The neighbor sample cleaning algorithm is used to remove boundary noise points and outliers in the augmented sample set, so that the number of samples of different state categories reaches a preset balance ratio. Using the Gini coefficient as the evaluation criterion for node splitting, a binary classification tree is recursively constructed. Each internal node of the binary classification tree corresponds to a feature and its splitting threshold, and each leaf node corresponds to a state category. During the construction process, a cost complexity pruning strategy is adopted. The subtree with the lowest misclassification rate is selected as the final classifier through cross-validation. The feature parameters and basic road segment identification results are input into the final classifier, and the output results are classified into three states: smooth, warning, or congested for the hub interchange.

6. The method according to claim 1, characterized in that, The method for determining congestion levels for on- and off-ramp sections using a weighted fusion of queuing and capacity reduction models specifically includes: The queuing model adopts a multi-service station queuing framework. Based on the traffic volume as the arrival rate, the number of service lanes at toll booths as the number of service stations, and the historical average service time as the service rate, the theoretical queue length and average waiting time are calculated. The capacity reduction model reduces the baseline capacity based on the turning angle and length of the ramp. The larger the turning angle or the shorter the ramp length, the larger the reduction factor. The reduced capacity is equal to the baseline capacity multiplied by the turning reduction factor and then multiplied by the length reduction factor. The queuing score output by the queuing model and the capacity reduction score output by the capacity reduction model are weighted and fused together, and the weight coefficients of the queuing score and the capacity reduction score are dynamically adjusted according to the real-time traffic conditions. The weighted and integrated comprehensive score is compared with the preset congestion level threshold. When the comprehensive score exceeds the first threshold, it is determined to be a congestion level; when it is between the second threshold and the first threshold, it is determined to be a warning level; and when it is below the second threshold, it is determined to be a smooth level.

7. The method according to claim 6, characterized in that, The weighted fusion also integrates the identification results of basic road sections and hub interchange sections, specifically including: Obtain the status identification results of basic road sections and the status classification results of hub interchange sections, and convert them into corresponding quantitative influence factors; The arrival rate parameter of the queuing model is adjusted based on the status identification results of the basic road segment. When the basic road segment is in a congested state, the arrival rate is increased to reflect the cumulative effect of upstream vehicles. The reduction coefficient of the capacity reduction model is adjusted according to the status classification results of the hub interchange section. When the hub interchange section is in a warning or congestion state, the reduction coefficient is increased to reflect the blocking effect of congestion propagating upstream. The adjusted queuing score and the reduced score are weighted and merged to recalculate the comprehensive score. The congestion level of the ramp section is determined again based on the updated comprehensive score, so as to realize the cascading transmission and fusion judgment of the status information of upstream and downstream sections.

8. The method according to claim 1, characterized in that, The identification results of basic road sections, hub interchange sections, and ramp sections are collected to construct a road network state matrix. A graph neural network and a time-series prediction network are then used to output the congestion propagation path of the entire road network, specifically including: Using each segment of the highway network as a node and the topological connection relationship between adjacent segments as an edge, a road network topology graph structure is constructed. The status identification results of basic road segments, the status classification results of hub interchange segments, and the congestion levels of ramp segments are mapped to the corresponding graph nodes. The current status level of each node is used as the node attribute value to construct a road network status matrix, where the row index of the road network status matrix corresponds to the road segment number and the column index corresponds to the time series. A graph neural network is used to extract node features and edge relationships in the road network topology graph. The state information of adjacent nodes is aggregated through graph convolution operation. The state difference value between each node and its neighboring nodes is calculated. The diffusion intensity of congestion between different road segments is quantified based on the state difference value. The road network state matrix and the calculated diffusion intensity within the current time window are input into the time series prediction network. The time series prediction network takes the historical time series as input, captures the evolution of the state over time through a cyclic gating mechanism, and outputs the predicted state sequence of each node in several future time windows. Based on the predicted state sequence, starting from the node currently in a congested state, the propagation path of state deterioration is traced along the direction of the greatest diffusion intensity until the predicted state returns to smooth flow, generating a congestion diffusion path and evolution trend map of the entire road network.

9. The method according to claim 1, characterized in that, The process of generating visual charts and outputting tiered handling instructions based on a three-level state system and diffusion path specifically includes: The current status level of each road segment in the entire road network and the predicted congestion spread path are overlaid on the geographic information system map. The three status levels of free flow, transitional flow and congested flow are marked with different colors, and the direction of congestion spread path is marked with arrows. A congestion situation visualization chart containing status distribution layer and spread path layer is generated. Based on the aforementioned three-level state system and diffusion path, the highway network is divided into three control areas: the core congestion area, the diffusion impact area, and the peripheral unobstructed area. The core congestion area is the continuous road segment that is currently in a congested flow state and has the greatest diffusion intensity. The diffusion impact area is the road segment that is predicted to be affected by congestion within a set time window in the future. The peripheral unobstructed area is the remaining road segment. The system assigns corresponding tiered response instructions to different control areas: Level 1 instructions, including forced diversion, ramp closure, and variable speed limits, are generated for core congestion areas; Level 2 instructions, including dynamic guidance, progressive speed limits, and toll station adjustments, are generated for areas with diffuse impact; and Level 3 instructions, including information dissemination and route recommendations, are generated for peripheral unobstructed areas. The tiered handling instructions are encapsulated into a standardized data structure containing the target road segment identifier, instruction type, execution parameters, and effective time period, and then pushed to roadside variable message signs, navigation applications, and toll station control systems for execution.

10. A real-time road traffic operation status assessment and congestion intelligent identification system based on unmanned aerial vehicle (UAV) patrol, applied to the method described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition module is used to build a collaborative sensing network for multi-source data acquisition and spatiotemporal calibration, extract traffic feature parameters, and construct a three-level state system. The basic road segment identification module is used to identify basic road segments by employing a macroscopic traffic flow model and a spatial constraint clustering algorithm, and outputting state identification results based on the aforementioned feature parameters. The hub interchange section identification module is used to identify hub interchange sections by using sample equalization processing and binary classification tree, and outputting status classification results based on the feature parameters and basic road section identification results. The ramp segment identification module is used to determine the congestion level for entrance and exit ramp segments by weighted fusion of queuing model and capacity reduction model, based on the characteristic parameters, basic road segments and hub interchange segment identification results. The situation simulation module is used to collect the identification results of basic road sections, hub interchange sections and ramp sections, construct the road network state matrix, and use graph neural networks and time series prediction networks to output the congestion spread path of the entire road network. The strategy output module is used to generate a visual chart and output graded handling instructions based on the three-level state system and diffusion path.

Images