Urban traffic management intelligent evaluation system and method based on large language model

Abstract

The invention provides an urban traffic management intelligent evaluation system and method based on a large language model, and belongs to the technical field of intelligent traffic management. The system comprises a data processing module, an MECA traffic flow analysis module, a feature engineering module, a driving behavior evaluation module, a clustering analysis module, an LLM intelligent decision module, a visualization generation module and a visualization module. The system analyzes the road traffic data collected by the unmanned aerial vehicle, adopts the MECA technology to automatically identify the road type and the traffic environment, intelligently judges the congestion level, evaluates the driving behavior, and combines a big language model to generate a targeted traffic management optimization suggestion. According to the method, the adaptive context sensing technology is innovatively combined with multi-criterion learning, adaptive congestion judgment of different road types is realized, different traffic characteristics of urban expressways, common urban roads and expressways can be accurately recognized, and differentiated management strategies are provided accordingly. The system supports real-time processing of large-scale traffic data, and provides scientific and accurate decision support for urban traffic management departments.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent traffic management, and particularly relates to a city traffic management intelligent evaluation system and method based on multi-source data fusion. The system realizes comprehensive evaluation of city traffic conditions and generation of optimization suggestions by combining vehicle trajectory data collected by unmanned aerial vehicles, MECA (Multidimensional Environment Contextualized Assessment) technology, multidimensional driving behavior evaluation, clustering analysis, and large language models. BACKGROUND

[0002] With the acceleration of urbanization and the rapid growth of motor vehicle ownership, urban traffic congestion problems are becoming increasingly serious, and traditional traffic management methods have been difficult to meet the needs of modern city traffic management. The existing traffic management system mainly has the following technical problems:

[0003] First, the data collection method is single and the coverage is limited. Traditional traffic data collection mainly relies on fixed position monitoring equipment and ground coils, which is difficult to obtain dynamic traffic information of the whole road, especially the continuous trajectory data of vehicles and driving behavior characteristics.

[0004] Second, the traffic state evaluation dimension is single. Most of the existing systems only focus on traffic flow, average speed and other macro indicators, lack comprehensive evaluation of driving behavior safety, efficiency, comfort and other multi-dimensional, and cannot fully reflect the traffic operation quality.

[0005] Third, the road type recognition relies on manual annotation. The traffic characteristics and management needs of different types of roads (highway, urban road, rural road) are quite different, but the existing system lacks the ability to automatically identify road types and requires a lot of manual annotation work.

[0006] Fourth, the congestion determination standard lacks adaptability. The existing system usually uses fixed speed threshold to determine congestion status, without considering the differentiated standards under different road types and traffic environments, resulting in low accuracy of congestion identification.

[0007] Fifth, the driving behavior evaluation lacks individualization and context awareness. Traditional methods use uniform evaluation standards without considering the rationality differences of driving behavior under different road environments and traffic densities, and the evaluation results lack pertinence.

[0008] Sixth, the traffic management suggestion generation capability is insufficient. Most of the existing systems only provide data statistics and state display, lack intelligent optimization suggestion generation capability, and cannot provide executable improvement schemes for traffic management departments.

[0009] Therefore, there is an urgent need for a city traffic management evaluation system that can comprehensively utilize multi-source data, achieve multi-dimensional evaluation, and have intelligent analysis and suggestion generation capabilities.

[0010] After searching, the application publication number CN120297546A, an automatic driving evaluation method and device based on a large language model, relates to the field of data processing. The method includes: obtaining driving scene data to be evaluated, preprocessing the driving scene data to be evaluated, and obtaining driving decisions to be evaluated; inputting the driving decisions to be evaluated into a pre-trained driving evaluation model to obtain an automatic driving evaluation result; wherein the driving evaluation model is constructed based on a large language model integrated driving evaluation Agent; the driving evaluation Agent obtains the automatic driving evaluation result through a pre-constructed automatic driving evaluation system; and the automatic driving evaluation system is constructed based on a multi-modal driving evaluation database.

[0011] Essential differences in evaluation methodology: First, the fundamental difference between the two systems in methodology. The CN120297546A patent uses a "language understanding driven" evaluation method, which attempts to have a large language model understand the driving scene and give an evaluation. This method essentially converts the driving evaluation problem into a natural language processing problem. It's like asking a literary critic to judge the merits of an engineering design - while the critic may give a seemingly reasonable evaluation, it lacks the depth of understanding of engineering expertise. In contrast, the current system uses a "numerical calculation driven" evaluation method. The system quantifies driving behavior into specific physical indicators based on traffic engineering theory, such as headway, acceleration change rate, and lane holding time. Each indicator has a clear physical meaning and calculation formula. For example, the safety score is based on the time to collision (TTC), and when TTC is less than a certain threshold, the system will reduce the safety score. This evaluation method based on physical laws ensures the objectivity and repeatability of the results.

[0012] Fundamental differences in data processing architecture: One of the key features of the CN120297546A patent is the reliance on multi-modal data processing, which requires converting various formats of data (images, text, sensor signals) into a form that the language model can understand. This process carries a significant risk of information loss. The current system uses a direct numerical processing architecture. The trajectory data received by the system remains in its original numerical form and features are extracted directly through mathematical operations. For example, the system assesses driving smoothness by calculating the standard deviation of the speed sequence and analyzes lane ID sequence changes to count lane changes. This processing method avoids information loss caused by data format conversion, ensuring the accuracy of the evaluation.

[0013] Adaptability of evaluation criteria: One of the core challenges faced by the patent method is the rigidity of evaluation criteria. The evaluation criteria of large language models come from training data, and once the training is completed, these criteria are basically fixed. When facing new road environments or changes in traffic rules, the model has difficulty in quickly adapting. The system solves this problem through a parameterized evaluation system. The system defines different evaluation parameter sets for different road types. This parameterized design allows the system to flexibly adjust the evaluation criteria according to the actual application scenario without the need to retrain the entire model.

[0014] Trade-off between computational efficiency and real-time performance: The method of CN120297546A patent needs to call a large language model for each evaluation task, which brings huge computational overhead. According to the current performance of large language models, processing a complex scenario may take several seconds or even longer. When hundreds of scenarios need to be processed simultaneously, the response delay of the system will become unacceptable. The system realizes efficient processing through an optimized computing process. The system adopts a three-stage processing strategy: the first stage, batch extraction of all vehicle basic features; the second stage, parallel calculation of each dimension score; the third stage, aggregation to generate the final result. This pipeline processing method allows the system to complete the evaluation of hundreds of vehicles within seconds, meeting the needs of real-time traffic management. SUMMARY

[0015] The present invention aims to solve the problems of the prior art. A method is proposed. The technical solution of the present invention is as follows:

[0016] An intelligent evaluation system for urban traffic management based on a large language model, comprising:

[0017] a data processing module, a MECA traffic flow analysis module, a feature engineering module, a driving behavior evaluation module, a clustering analysis module, an LLM intelligent decision-making module, a visualization generation module, and a visualization module; wherein,

[0018] The data processing module is used to receive and preprocess the vehicle trajectory data collected by the unmanned aerial vehicle, and output the processed data to the MECA traffic flow analysis module and the feature engineering module respectively;

[0019] The MECA traffic flow analysis module is used to automatically identify road types through adaptive context perception technology, analyze traffic density, calculate speed distribution, determine congestion levels based on different road types' congestion standards, analyze lane usage and lane changing behavior, and evaluate traffic efficiency and safety risks;

[0020] The feature engineering module is used to extract driving behavior features from the original trajectory data, including speed features, acceleration features, lane change features, following features, and surrounding environment features;

[0021] The driving behavior evaluation module integrates the MECA evaluator to adaptively adjust the scoring criteria according to the road context, evaluate the driving behavior of each driver in multiple dimensions, and generate driving scores and levels.

[0022] The LLM intelligent decision-making module generates overall traffic condition evaluations, identifies main problems, provides optimization suggestions, formulates implementation priorities, and predicts improvement effects based on traffic analysis results and driving behavior evaluation results through a large language model.

[0023] The visualization module generates comprehensive traffic management reports, including HTML-formatted visual reports, detailed data files, and text summaries.

[0024] Further, the MECA traffic flow analysis module includes a road context recognition unit, a congestion detection unit, a lane analysis unit, and a safety evaluation unit.

[0025] The road context recognition unit automatically identifies road types, including highways, urban expressways, and ordinary urban roads, by analyzing traffic flow characteristics and calculates recognition confidence.

[0026] The congestion detection unit selects corresponding congestion determination criteria based on the identified road type, sets differentiated free-flow thresholds according to different road types, and calculates congestion indices based on the ratio of average speed to free-flow speed.

[0027] The lane analysis unit analyzes lane usage distribution, detects lane-changing behavior for each vehicle, calculates lane-changing frequency and amplitude, identifies driving modes such as aggressive weaving, stable cruising, and left lane occupancy, and analyzes the spatiotemporal distribution pattern of lane changes.

[0028] The safety evaluation unit analyzes collision time (TTC), sudden braking events, emergency braking events, and speed differences to evaluate road safety risk levels.

[0029] Further, the lane change detection method of the lane analysis unit includes:

[0030] Sort the trajectory data of each vehicle by time frame;

[0031] Calculate the lane ID difference between adjacent frames and count the number of lane changes;

[0032] Calculate the lane-changing amplitude, i.e., the number of lanes crossed in one lane change;

[0033] Identify dangerous lane-changing behavior, including crossing multiple lanes and rapid consecutive lane changes;

[0034] Based on lane-changing frequency, lane-changing interval, and lane-keeping time, classify drivers into different driving modes.

[0035] Further, the driving behavior evaluation module adopts an MECA evaluator, including: a feature extraction unit for processing vehicle trajectory data and extracting multi-dimensional driving features; an adaptive scoring unit for automatically adjusting scoring weights according to road types and traffic densities;

[0036] a grade division unit for dividing drivers into five grades of A, B, C, D and E based on comprehensive scores.

[0037] Further, the LLM intelligent decision-making module includes:

[0038] a data summary generation unit for arranging traffic indicators and evaluation results into a structured summary;

[0039] an intelligent analysis unit for analyzing traffic data and generating professional recommendations based on traffic management expert roles by calling a large language model;

[0040] a report analysis unit for identifying chapter markers, segmenting content, cleaning and verifying, and intelligently completing missing content using a multi-stage analysis strategy;

[0041] a backup solution unit for generating a preset traffic management recommendation template when calling fails.

[0042] Further, the driving behavior evaluation module also provides driving style recognition results based on driver cluster analysis, and the driver cluster analysis process includes five main steps, starting from data preparation, through determining the optimal cluster number, K-means++ initial clustering and SVM optimized clustering, finally completing cluster naming and feature analysis. Driver cluster analysis aims to identify different driving styles and behavior patterns; combining K-means++ initial clustering and support vector machine (SVM) boundary optimization, drivers are accurately classified into different driving style types; K-means++ first provides preliminary clustering based on natural data distribution, while SVM uses nonlinear boundary learning ability to further optimize the clustering boundary.

[0043] A traffic management evaluation method based on any of the systems, comprising the following steps:

[0044] Step S1: loading traffic data collected by a drone through a data processing module, performing data cleaning and trajectory extraction, and screening effective vehicle data meeting the minimum frame number requirement;

[0045] Step S2: performing comprehensive traffic analysis through an MECA traffic flow analysis module, including identifying road types, calculating traffic density, analyzing speed distribution, detecting congestion conditions, analyzing lane usage, calculating traffic efficiency and evaluating safety risks;

[0046] Step S3: Process the filtered vehicle data using the MECA evaluator, extract driving features, train the evaluation model, conduct behavior evaluation for each driver, generate driving score and rating;

[0047] Step S4: Generate traffic management recommendations through the LLM intelligent decision module, including overall evaluation of traffic conditions, identification of main problems, optimization suggestions, implementation priority and expected improvement effect;

[0048] Step S5: Generate visual charts, including speed distribution chart, traffic density timeline, lane usage distribution chart, congestion state timeline, driving score distribution chart;

[0049] Step S6: Generate comprehensive traffic management report through the visualization module, integrate all analysis results and recommendations, output HTML report, data file and text summary.

[0050] Further, the congestion detection method in step S2 includes:

[0051] Obtain vehicle average speed data, intelligently identify speed column and perform unit conversion;

[0052] Calculate the average speed of each vehicle, and then calculate the mean of the average speeds of all vehicles as the average speed of the road section;

[0053] According to the road type identified by MECA, select the corresponding congestion determination standard;

[0054] Based on the comparison of average speed and threshold, determine the congestion level as smooth, slow, congested or severely congested;

[0055] Calculate the congestion index = (free flow speed - average speed) / free flow speed x 100%;

[0056] Statistical proportion of vehicles in different congestion levels.

[0057] Further, the driving behavior evaluation in step S3 includes:

[0058] Initialize the MECA evaluator and set the road context parameters;

[0059] Extract multi-dimensional features for each vehicle, including speed stability, acceleration smoothness, lane discipline, safety distance maintenance, etc.

[0060] Adaptively adjust the weight of each dimension according to the road type and traffic density;

[0061] Calculate the comprehensive driving score using weighted summation;

[0062] Determine the rating threshold value based on the score distribution;

[0063] An evaluation result including a vehicle ID, dimension scores, a comprehensive score, and a rating is generated.

[0064] Further, the LLM intelligent suggestion generation in step S4 includes:

[0065] Traffic management expert system prompt words are constructed, and an analysis framework and output requirements are defined.

[0066] A data summary including road information, traffic flow conditions, congestion analysis, lane usage, safety risks, and driving behavior is prepared.

[0067] A large language model is called, and the system prompt words and data summary are passed in.

[0068] A multi-stage parsing strategy is used to process the response content, including identifying chapter markers, dividing the content by markers, cleaning up format markers, and verifying content integrity.

[0069] Intelligent completion is performed on missing content to ensure the completeness of each part of the report.

[0070] A backup template is used to generate a basic suggestion when the large language model fails to call.

[0071] The advantages and beneficial effects of the present application are as follows:

[0072] The present application first applies a large language model to traffic management decision-making, innovatively uses the understanding and generation capabilities of LLM to achieve intelligent conversion from data to suggestions, and provides professional management decision support. The MECA adaptive evaluation method is proposed, which automatically adjusts the evaluation standard according to the road environment, realizes the dynamic and accurate evaluation, and improves the scientificity of the evaluation result. A four-dimensional comprehensive evaluation system is constructed, covering safety, efficiency, comfort, and discipline, and a scientific weight determination method is used to realize comprehensive evaluation of driving behavior. Multi-level traffic state analysis is realized, from macro traffic flow to micro driving behavior, from overall evaluation to individual analysis, from present situation diagnosis to improvement suggestion. An intelligent visual analysis system is developed, which automatically generates professional visual charts, provides multi-format comprehensive report output and intuitive traffic situation presentation.

[0073] The present application has strong data processing capability, supports multiple data formats and naming conventions, uses intelligent data cleaning strategies, has efficient trajectory extraction algorithms and perfect outlier processing mechanisms. The analysis dimensions are comprehensive, including a four-dimensional driving behavior evaluation system, multi-level traffic flow analysis, refined lane usage evaluation, and comprehensive safety risk identification. The evaluation method is scientific, using CRITIC-entropy weight method combined weight, AHP dimension weight, adaptive evaluation standard, and data-driven clustering analysis.

[0074] The system has high intelligence degree, has MECA automatic road recognition, LLM intelligent suggestion generation, self-adaptive congestion determination and intelligent clustering naming functions. The system has strong practicability, provides professional visual presentation, multi-format report output, specific operable suggestions and complete implementation priority. The system has good expansibility, adopts a modular system architecture, a standardized data interface, configurable evaluation parameters and flexible output formats.

[0075] The core technical innovation point of the present application

[0076] Self-adaptive context perception evaluation technology based on MECA

[0077] The present application proposes an innovative MECA self-adaptive context perception multi-criteria learning system, which breaks through the technical limitation of the traditional driving evaluation system using fixed evaluation standards. In the traditional technical solution, the same evaluation standard and weight system is used regardless of whether the vehicle is driving on a highway, urban road or rural road. This "one-size-fits-all" evaluation method cannot truly reflect the rationality of driving behavior under different road environments.

[0078] The MECA technology of the present application can automatically determine the current road type and traffic state based on the speed distribution characteristics, lane change frequency, vehicle spacing distribution, parking ratio and other multi-dimensional characteristics in the vehicle trajectory data by constructing a road environment automatic recognition module. On this basis, the system dynamically adjusts the weight distribution and scoring standard of each evaluation dimension. For example, when the system identifies a city congestion road environment, it will automatically reduce the punishment weight for frequent parking and low-speed driving, and increase the evaluation weight for maintaining a safe vehicle distance. In the highway environment, the system will strengthen the evaluation weight for lane stability and speed consistency.

[0079] Traffic management decision generation technology driven by large language model

[0080] The present application creatively applies large language model technology to the field of urban traffic management decision, realizing the intelligent conversion from traffic data analysis to management suggestion generation. Unlike the traditional system which can only output data statistical report, the present application enables the large language model to play the role of a senior traffic management expert through a carefully designed prompt engineering framework, analyzes the traffic conditions in depth and generates executable management suggestions.

[0081] The innovation of the technology is reflected in three aspects: first, the system constructs a professional knowledge prompt template in the field of traffic management, including traffic flow theory, road engineering principles, traffic management strategies and other professional knowledge, so that the large language model can analyze based on a professional perspective; second, the system realizes intelligent mapping of multi-dimensional data to structured suggestions, which can generate specific and actionable management measures such as "suggest implementing variable lane management on this road section" and "optimize signal timing scheme" according to indicators such as congestion index, safety risk level, and lane utilization rate; finally, the system can also prioritize and evaluate the expected effect of the suggestions, providing decision support for traffic management departments.

[0082] Multi-dimensional weight automatic determination mechanism based on CRITIC-entropy weight method

[0083] The present application innovatively combines the CRITIC method with the entropy weight method to solve the technical problems of strong subjectivity and lack of scientific basis in determining the weight in traditional evaluation systems. The CRITIC method determines the weight by analyzing the comparison intensity and conflict between evaluation indexes, which can identify the correlation and information overlap between indexes, and avoid highly correlated indexes from obtaining too high total weight. The entropy weight method determines the weight by calculating the information entropy of each index, and the greater the variation of the index, the more information it contains, and the greater the corresponding weight.

[0084] The present application innovatively combines the two methods, considering both the mutual relationship between indexes and the information content of the indexes, and obtaining the final weight through weighted average. This combined weighting method overcomes the limitations of single weighting method, making the weight determination of the evaluation system more objective and scientific. At the same time, the system also uses the analytic hierarchy process to determine the weight relationship between the four dimensions of safety, efficiency, comfort and lane discipline, and constructs a multi-level and multi-dimensional comprehensive evaluation system. BRIEF DESCRIPTION OF DRAWINGS

[0085] Figure 1 is the overall architecture schematic diagram of the preferred embodiment system provided by the present application;

[0086] Figure 2 is the internal structure diagram of the MECA traffic flow analysis module;

[0087] Figure 3 is the driving behavior evaluation flowchart;

[0088] Figure 4 is the workflow diagram of the LLM intelligent decision module;

[0089] Figure 5 is the lane change detection algorithm schematic diagram;

[0090] Figure 6is an example of a system-generated speed distribution analysis chart;

[0091] Figure 7 is an example of a traffic density time variation chart;

[0092] Figure 8 is an example of a driving score level distribution chart;

[0093] Figure 9 is an example of LLM recommendation feedback.

[0094] Figure 10 is the workflow diagram of the intelligent evaluation system for urban traffic management based on large language models. DETAILED DESCRIPTION

[0095] The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments of the present application.

[0096] The technical solution of the present application to solve the above technical problems is:

[0097] The core innovation of the present application is:

[0098] Adaptive road recognition technology: automatically recognize road types through MECA technology, without manual annotation, to distinguish between expressways, urban expressways and ordinary urban roads, and select differentiated evaluation standards accordingly.

[0099] Multi-dimensional comprehensive evaluation: not only analyze traffic flow conditions, but also deeply evaluate individual driving behavior, identify aggressive driving, frequent lane changing and other problem behaviors, and provide basis for precise management.

[0100] Intelligent decision support: integrate large language models, automatically generate professional traffic management recommendations based on data analysis results, including signal optimization, lane management, driving behavior guidance and other specific measures.

[0101] Full-process automation: from data processing to report generation, greatly improve analysis efficiency and support real-time decision-making.

[0102] The present application is realized by the following technical solutions:

[0103] An intelligent evaluation system for urban traffic management based on MECA, comprising:

[0104] Data processing module: responsible for loading and preprocessing vehicle trajectory data collected by drones. This module supports CSV and Excel formats, can automatically identify data structure, clean abnormal data, and extract valid vehicle trajectories. By setting a minimum frame threshold (default 200 frames), it filters out vehicles with too short observation time, ensuring the reliability of the analysis data.

[0105] MECA Traffic Flow Analysis Module: This is the core module of the system, which integrates the road state recognition and lane detection technology of MECA with an improved traffic analyzer. This module can:

[0106] • Automatically identify road types (highway / urban / rural) by analyzing traffic flow characteristics

[0107] • Apply differentiated congestion criteria according to different road types

[0108] • In-depth analysis of lane usage patterns and lane changing behavior

[0109] • Evaluate traffic safety risks

[0110] • Provide comprehensive traffic state assessment

[0111] Driving context classification is one of the key innovations of the MECA method, which can identify the current driving environment and provide context information for subsequent feature extraction and scoring. The method mainly classifies the driving context from two dimensions: road type and traffic density.

[0112] (1) Road type classification: The method classifies the road environment into three categories: highway, urban road, and rural road based on multiple features such as speed distribution, lane number, trajectory geometric characteristics, etc. The classification process uses feature scoring method, as shown in the formula:

[0113]

[0114] where f i is the i-th feature, i is the corresponding weight, and n is the total number of features. The method calculates the score of each road type and selects the type with the highest score as the final classification result.

[0115] (2) Traffic density classification: The method classifies traffic density into three levels: high, medium, and low based on vehicle spacing, presence of preceding vehicle, number of surrounding vehicles, etc. Consider the following features: Proportion of preceding vehicle: the proportion of frames with preceding vehicles in the total number of frames, DHW, THW, Proportion of surrounding vehicles: the proportion of vehicles in six directions (left front, left side, left rear, right front, right side, right rear)

[0116] The calculation of surrounding vehicle density is shown in the formula:

[0117]

[0118] where D is the set of six directions, P(V dir ≠0) is the probability of the presence of vehicles in direction dir

[0119] The MECA method adopts Principal Component Analysis (PCA) as the core feature representation learning method, achieving effective dimension reduction and representation learning of driving features without relying on deep learning frameworks. As a classic unsupervised dimension reduction technique, PCA can capture the main variation patterns in the data and generate compact low-dimensional feature representations, providing a foundation for subsequent driving style clustering and scoring.

[0120] The goal of feature representation learning is to convert high-dimensional driving features into low-dimensional embedding vectors while preserving key information in the original features. The method achieves this goal through three key steps: feature preprocessing, common feature identification, and principal component analysis, ensuring that the generated feature representations are both informationally complete and computationally efficient.

[0121] The encoding process includes three steps: feature vector preparation, standardization, and PCA projection:

[0122]

[0123] x new : the newly input driving feature vector, containing various driving behavior features such as speed, acceleration, and car distance data, μ: the feature mean vector calculated during training, σ: the feature standard deviation vector calculated during training, V k : the PCA projection matrix composed of the eigenvectors corresponding to the first k largest eigenvalues, which maps the standardized high-dimensional feature space to the low-dimensional principal component space.

[0124] To dynamically adjust feature weights based on driving context, this paper integrates a context-aware mechanism into the driving evaluation method, which can dynamically adjust the importance weights of different features based on context information such as road type and traffic density. The threshold for measuring the risk level of driving behavior will also change under different road conditions and traffic density, such as the safe speed and safe following distance on highways being much higher than on urban roads. The method presets a dimension weight mapping table for each driving context (combination of road type and traffic density).

[0125] The method looks up the corresponding weights from the mapping table based on the current driving context and calculates the maximum weight for each dimension, as shown in the formula:

[0126]

[0127] where w d is the weight of dimension d, w d,c is the weight of d under context c, and C is the set of contexts.

[0128] The core idea of the context-aware mechanism is to identify feature categories and dynamically adjust weights according to driving contexts. First, the method classifies all features into four categories: safety features, efficiency features, comfort features, and lane discipline features.

[0129] Dimension weight determination: Determine the weights of the four dimensions according to the context information. According to a pre-set context weight mapping table, define the dimension weights for different road types and traffic densities.

[0130] The method assigns weights to each specific feature by matching the feature name to determine its dimension. For feature f, the weight calculation formula is:

[0131] W f =W dim(f)

[0132] where dim(f) represents the dimension to which feature f belongs.

[0133] Weight normalization: Ensure that the sum of all feature weights is 1.

[0134]

[0135] where W f ′ is the normalized feature weight, and n is the total number of features.

[0136] Through the above steps, the method realizes the context-aware dynamic adjustment of feature weights. For example, in the case of high traffic density, the speed decreases, and the weight of safety features will decrease slightly, making the scoring result more in line with the requirements of the current driving environment.

[0137] Feature engineering module: Extract rich driving behavior features from raw trajectory data, including speed features (average speed, speed variance), acceleration features (frequent acceleration / braking frequency), lane features (number of lane changes, lane holding time), following features (headway, safety distance), etc.

[0138] The goal of the feature weight optimization module is to dynamically calculate feature weights according to driving environments, so that the evaluation method focuses on different driving features in different environments. For example, in highway environments, safety distance and lane stability should be paid more attention to; while in urban congestion environments, brake stability and low-speed control ability are more important.

[0139] The feature weight optimization module combines the CRITIC method and the entropy weight method, including four parts: feature variability analysis, feature conflict calculation, feature information calculation, and weight generation. The CRITIC method objectively evaluates the importance of features by considering the standard deviation of features and the correlation between features; the entropy weight method measures the information amount of features by information entropy, and assigns higher weights to features with low entropy (large information amount).

[0140] The feature weight calculation process based on the CRITIC method is as follows: first, calculate the standard deviation of each feature, representing the contrast intensity of the feature; then, calculate the correlation coefficient matrix between features and convert it to a conflict matrix; next, calculate the information content of each feature, which is the product of the standard deviation and the conflict; finally, normalize the information content to obtain the feature weight.

[0141] The CRITIC method is mathematically expressed as shown in the formula:

[0142]

[0143] where σ j is the standard deviation of feature j, x ij is the value of feature j representing sample i, is the mean value of feature j, r jk is the correlation coefficient of feature j and feature k, c j represents the information content of feature j, and m is the number of features. represents the weight of feature j based on the CRITIC method.

[0144] The entropy weight method is based on the information entropy principle, and the feature weight calculation process is as follows: first, normalize the feature values to the [0, 1] interval; then, calculate the information entropy of the feature; next, calculate the difference coefficient (1-entropy) of the feature; finally, normalize the difference coefficient to obtain the feature weight.

[0145] The mathematical expression is shown in the formula:

[0146]

[0147] where p ij is the normalized feature value, and n is the number of samples. e j represents the information entropy of feature j, represents the weight of feature j based on the entropy weight method.

[0148] The method combines the two methods to calculate the final feature weight, as shown in the formula:

[0149]

[0150] where w j represents the final weight of feature j, and α is the combination parameter, controlling the relative importance of the two methods.

[0151] Driving behavior evaluation module: integrates the MECA evaluator, which can adaptively adjust the scoring standard according to the road environment. For example, more emphasis on speed stability and lane discipline in highway environment, and more attention to flexible response and safety distance in urban road environment.

[0152] The standard scoring method combines feature representation and feature weight to comprehensively evaluate driving behavior from four dimensions: safety, efficiency, comfort, and lane discipline. The multi-standard scoring method is designed based on the following principles: (1) different dimensions of features should have different calculation methods; (2) the scoring standard should be dynamically adjusted according to the driving context; (3) the scoring result should be interpretable.

[0153] The scoring process includes two steps: dimension scoring and overall scoring.

[0154] The score of each dimension is calculated as shown in the formula:

[0155]

[0156] where Score d is the score of dimension d, w j is the weight of feature j, is the normalized feature value, sign j is the direction of feature j (1 means the larger the value the better, -1 means the smaller the value the better), m d is the number of features included in dimension d.

[0157] The overall score is calculated as shown in the formula:

[0158]

[0159] where OverallScore is the overall score, w d is the dimension d weight, and D is the total number of dimensions (4 in this method).

[0160] The method classifies drivers into five levels: A, B, C, D, and E, corresponding to score segments of 90-100, 75-90, 60-75, 40-60, and 0-40. At the same time, the method also provides detailed scoring of each dimension and driving style recognition results based on driver cluster analysis, providing a comprehensive evaluation of driving behavior for drivers.

[0161] Driver cluster analysis aims to identify different driving styles and behavior patterns. Combining K-means++ initial clustering and support vector machine (SVM) boundary optimization, drivers are accurately classified into different driving style types. In this hybrid framework, K-means++ first provides preliminary clustering based on the natural distribution of data, while SVM further optimizes the clustering boundary using its powerful non-linear boundary learning ability. This "unsupervised + supervised" hybrid learning strategy makes the clustering result more accurate and meaningful.

[0162] The driver clustering process contains five main steps, starting from data preparation, through determining the optimal number of clusters, K-means++ initial clustering, and SVM optimized clustering, finally completing cluster naming and feature analysis. The specific steps are as follows:

[0163] Standardization

[0164] To ensure that each dimension contributes equally in the clustering process, the method standardizes all features using the Z-score standardization method:

[0165]

[0166] where X s is the standardized feature, μ is the mean vector of each dimension, and σ is the standard deviation vector of each dimension. This standardization ensures that different dimensionless scoring indicators have equal importance when calculating distances.

[0167] K-means++ initial clustering

[0168] K-means++ selects initial cluster centers through a probabilistic weighting method. First, a random data point is selected as the first cluster center c1. For each subsequent cluster center c i (i=2,3,...,k), calculate the distance D(x) from each data point x to the nearest selected cluster center, and select a new cluster center with probability P(x), where:

[0169]

[0170] This selection method ensures that the initial cluster centers are far apart, reducing the likelihood of the algorithm falling into a local optimum. In driver clustering, this means that the initial clusters are more likely to capture significantly different driving styles.

[0171] After determining the initial centers, the algorithm performs standard K-means iterations. Since it may not be able to capture complex nonlinear relationships between driving behaviors, SVM is introduced for boundary optimization. SVM can map data to high-dimensional space through kernel functions, thereby identifying complex nonlinear boundaries.

[0172] The basic idea of SVM is to find an optimal hyperplane that maximally separates data points of different classes. For non-linearly separable data, SVM uses kernel tricks to map data to high-dimensional feature space:

[0173] K(x i ,x j )=φ(x i ) T φ(x j )

[0174] The Radial Basis Function (RBF) kernel used in this paper is defined as:

[0175] K(x i ,x j )=exp(-γ||x i -x j || 2 )

[0176] where γ is a parameter that controls the flexibility of the decision boundary. The γ parameter is set using the default'scale' value:

[0177]

[0178] The clustering results of K-means++ are used as initial labels to train the SVM classifier. The SVM training process solves the following optimization problem:

[0179]

[0180] Subject to the constraints:

[0181] y i ×(w T φ(x i )+b)≥1-ξ i ,ξ i ≥0

[0182] where: w is the normal vector of the hyperplane, b is the bias term of the hyperplane, ξ i is the slack variable that allows some data points to violate the margin constraint, C is the penalty parameter with a default value of 1.0, φ is the function that maps the data to a high-dimensional space, y i is the label provided by K-means++.

[0183] The SVM model can provide classification probability estimates, providing more rich information for subsequent analysis. After training, the decision function of SVM is:

[0184]

[0185] where α i is the Lagrange multiplier obtained by solving the dual problem. Due to the use of the RBF kernel, SVM can generate a non-linear decision boundary in the original feature space, more accurately dividing the driver data with complex distribution characteristics.

[0186] The LLM Intelligent Decision-Making Module is the core innovation of the system. Leveraging the understanding and generation capabilities of large language models, it transforms data analysis results into professional management recommendations. The system assigns a professional traffic management expert role to the LLM module to ensure the professionalism of the generated content. This role is defined as a senior urban traffic management expert, whose tasks include analyzing current road traffic conditions, identifying major traffic problems, assessing safety risks, and providing specific and feasible traffic management optimization suggestions.

[0187] The system constructs a detailed analytical context, including basic road information such as type, length, and number of lanes; traffic flow conditions such as density, speed, and congestion level; safety risk assessment such as hazardous event statistics; and driving behavior distribution such as rating statistics. LLM generates recommendations based on multiple dimensions, including traffic flow optimization (measures to improve traffic flow), congestion mitigation (targeted traffic management solutions), safety enhancement (methods to reduce accident risk), lane management (optimized lane usage strategies), signal control (traffic signal optimization suggestions), and driving behavior guidance (measures to improve overall driving behavior).

[0188] The system employs a multi-stage parsing strategy to ensure accurate extraction of content from each chapter. The first stage identifies explicit chapter titles, the second stage segments content based on chapter tags, the third stage cleans and verifies the content, and the fourth stage intelligently completes any missing content. When an LLM call fails, the system provides preset professional suggestion templates to ensure service continuity, including standardized suggestions based on congestion levels, differentiated solutions considering road types, and comprehensive suggestions including implementation priorities.

[0189] Visualization Generation Module: This module generates intuitive data visualization charts to help managers quickly understand traffic conditions. The speed distribution histogram displays the distribution of vehicle speeds, including speed frequency distribution, average speed markings, congestion threshold markings, and kernel density estimation curves. The histogram provides a clear understanding of the main driving speed ranges, the dispersion of speed distribution, and its relationship with congestion thresholds.

[0190] The traffic density timeline displays how traffic density changes over time. The X-axis represents time frames, the Y-axis represents the number of vehicles, and the filled area shows the density trend. This chart helps identify peak traffic periods, density fluctuation patterns, and abnormal traffic flow. The lane usage pie chart displays the utilization rate of each lane in pie chart format. The size of each sector represents the proportion of lane usage, different colors distinguish different lanes, and the precise percentage of utilization is marked. This chart can be used to assess the balance of lane utilization, overload of specific lanes, and the rationality of lane resource allocation.

[0191] The congestion state timeline displays the time evolution of congestion states, with scatter points colored to represent different congestion levels: green for free flow, yellow for slow, orange for congestion, and red for severe congestion, while showing threshold lines for each congestion level. This chart can track congestion evolution, identify congestion formation and dissipation patterns, and assess traffic state stability.

[0192] The rating distribution chart includes various charts that display the evaluation results of driving behavior. The score distribution histogram displays the overall score distribution, marks the average score position, and shows the rating intervals. The dimension score boxplot compares the score distribution of each dimension, displays the median and quartiles, and identifies outliers. The rating distribution pie chart displays the proportion of drivers at levels A to E, providing an intuitive understanding of overall driving level. The system uses professional visualization design, including scientific color mapping color scheme, automatically configured Chinese font support, grid settings with primary and secondary grid lines to improve readability, and complete legends and annotations.

[0193] The report output module generates comprehensive traffic management reports in multiple formats. The HTML visual report uses responsive design, with the report header containing the title and generation time, core indicator overview cards, and congestion level visualization labels. The traffic condition analysis section includes an overall evaluation text box, road base data cards, traffic flow analysis tables, and congestion distribution grids. The safety risk assessment section includes risk event statistics, dangerous indicator highlights, and safety improvement suggestions. The driving behavior analysis section includes rating distribution statistics, score comparison charts, and behavior pattern analysis. The management suggestions section includes major problem identification, optimization suggestion list, implementation priority, and expected improvement effect.

[0194] The system supports multiple data formats for export. CSV formats include traffic_metrics.csv traffic metrics data, driver_evaluation_results.csv driving evaluation results, and MECA_features.csv feature data. JSON formats include llm_recommendations.json LLM-generated recommendations, containing debugging information and content length statistics. TXT formats include traffic_summary.txt simplified text reports, suitable for quick browsing and email sending. The report uses professional style design, including gradient color summary boxes, responsive grid layout, status indicator color system, and print-friendly styles. Although it is a static report, it includes anchor navigation support, table sorting reservations, chart hover tips, and responsive zooming interactive function designs.

[0195] The system, device, module or unit illustrated in the above embodiments can be implemented by a computer chip or entity, or by a product with certain functions.

[0196] It is also to be noted that the terms "comprising", "including", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0197] The above examples are to be understood only as illustrative of the application and not a restriction on the scope of protection of the application. After reading the specification, the skilled person can make various changes or modifications to the application, and these equivalent changes and modifications also fall within the scope defined by the claims of the application.

Claims

1. An intelligent evaluation system for urban traffic management based on a large language model, characterized in that, include: The module includes a data processing module, a MECA traffic flow analysis module, a feature engineering module, a driving behavior assessment module, a cluster analysis module, an LLM intelligent decision-making module, a visualization generation module, and a visualization module; among them, The data processing module is used to receive and preprocess vehicle trajectory data collected by the UAV, and output the processed data to the MECA traffic flow analysis module and the feature engineering module respectively. The MECA (Multidimensional Environmental Context Assessment) traffic flow analysis module is used to automatically identify road types, analyze traffic density, calculate speed distribution, determine congestion levels based on congestion standards for different road types, analyze lane usage and lane-changing behavior, and assess traffic efficiency and safety risks through adaptive context awareness technology. The feature engineering module is used to extract driving behavior features from raw trajectory data, including speed features, acceleration features, lane change features, following features, and surrounding environment features. The driving behavior assessment module integrates the MECA evaluator, which is used to adaptively adjust the scoring criteria according to the road context, to evaluate the driving behavior of each driver in multiple dimensions, and to generate driving scores and grades. The LLM (Large Language Model) intelligent decision-making module is used to generate an overall assessment of traffic conditions, identify major problems, provide optimization suggestions, formulate implementation priorities, and predict improvement effects based on traffic analysis results and driving behavior assessment results. The visualization module is used to generate comprehensive traffic management reports, including HTML-formatted visualization reports, detailed data files, and text summaries.

2. The system according to claim 1, characterized in that, The MECA traffic flow analysis module includes a road context recognition unit, a congestion detection unit, a lane analysis unit, and a safety assessment unit; among which, The road context recognition unit is used to automatically identify road types, including highways, urban expressways, and ordinary urban roads, by analyzing traffic flow characteristics, and to calculate the recognition confidence level. The congestion detection unit is used to select the corresponding congestion judgment criteria according to the identified road type. Differentiated smooth flow thresholds are set according to different road types, and the congestion index is calculated based on the ratio of average speed to free flow speed. The lane analysis unit is used to statistically analyze lane usage distribution, detect lane-changing behavior of each vehicle, calculate lane-changing frequency and magnitude, identify driving modes including aggressive weaving, stable cruise, and left lane occupation, and analyze the spatiotemporal distribution patterns of lane changes. The safety assessment unit is used to analyze time-to-collision (TTC), sudden braking events, emergency braking events, and speed differences to assess the level of road safety risk.

3. The system according to claim 2, characterized in that, The lane change detection method of the lane analysis unit includes: The trajectory data for each vehicle is sorted by time frame; Calculate the lane ID difference between adjacent frames and count the number of lane changes; Calculate the lane change range, that is, the number of lanes crossed in one lane change; Identify dangerous lane-changing behaviors, including crossing multiple lanes and rapid, consecutive lane changes; Drivers are categorized into different driving modes based on lane change frequency, lane change interval, and lane keeping time.

4. The system according to claim 1, characterized in that, The driving behavior assessment module uses the MECA evaluator and includes: a feature extraction unit for processing vehicle trajectory data and extracting multi-dimensional driving features; and an adaptive scoring unit for automatically adjusting the scoring weights based on road type and traffic density. The rating system categorizes drivers into five levels: A, B, C, D, and E, based on a comprehensive score.

5. The system according to claim 1, characterized in that, The LLM intelligent decision-making module includes: The data summary generation unit is used to organize traffic indicators and evaluation results into structured summaries; The intelligent analysis unit, by calling upon a large language model and based on the role settings of traffic management experts, analyzes traffic data and generates professional suggestions; The report parsing unit employs a multi-stage parsing strategy to identify chapter markers, segment content, clean and verify information, and intelligently complete missing content. The backup solution unit generates a preset traffic management suggestion template when the call fails.

6. The system according to claim 1, characterized in that, The driving behavior assessment module also provides driving style identification results based on driver clustering analysis. The driver clustering analysis process includes five main steps, starting with data preparation, determining the optimal number of clusters, initial K-means++ clustering, and SVM optimization of clustering, and finally completing cluster naming and feature analysis. Driver clustering analysis aims to identify different driving styles and behavior patterns. Combining K-means++ initial clustering and support vector machine (SVM) boundary optimization, drivers are accurately classified into different driving style types. K-means++ first provides preliminary clustering based on the natural distribution of data, while SVM utilizes nonlinear boundary learning capabilities to further optimize the cluster boundaries.

7. A traffic management evaluation method based on the system described in any one of claims 1-6, characterized in that, Includes the following steps: Step S1: Load the traffic data collected by the drone through the data processing module, perform data cleaning and trajectory extraction, and filter out valid vehicle data that meets the minimum frame count requirement; Step S2: Perform comprehensive traffic analysis using the MECA traffic flow analysis module, including identifying road types, calculating traffic density, analyzing speed distribution, detecting congestion, analyzing lane usage, calculating traffic efficiency, and assessing safety risks. Step S3: Use the MECA evaluator to process the filtered vehicle data, extract driving features, train the evaluation model, evaluate the behavior of each driver, and generate driving scores and grades; Step S4: Generate traffic management recommendations through the LLM intelligent decision-making module, including an overall assessment of traffic conditions, identification of major problems, optimization suggestions, implementation priorities, and expected improvement effects; Step S5: Generate visualization charts, including speed distribution map, traffic density timeline, lane usage distribution map, congestion status timeline, and driving score distribution map; Step S6: Generate a comprehensive traffic management report through the visualization module, integrate all analysis results and recommendations, and output HTML report, data file and text summary.

8. The method according to claim 7, characterized in that, The congestion detection method in step S2 includes: Acquire average vehicle speed data, intelligently identify speed columns, and perform unit conversion; Calculate the average speed of each vehicle, and then calculate the average of the average speeds of all vehicles as the average speed of the road segment. Select the corresponding congestion assessment criteria based on the road type identified by MECA; Based on the comparison of average speed and threshold, the congestion level is determined as smooth, slow, congested, or severe congestion. Congestion index calculation = (free flow velocity - average velocity) / free flow velocity × 100%; The proportion of vehicles at different levels of congestion was statistically analyzed.

9. The method according to claim 7, characterized in that, The driving behavior assessment in step S3 includes: Initialize the MECA evaluator and set the road context parameters; For each vehicle, multi-dimensional features are extracted, including speed stability, acceleration smoothness, lane discipline, and safe distance maintenance. The weights of each dimension are adaptively adjusted based on road type and traffic density. The overall driving score is calculated using a weighted summation method. The threshold for classifying levels is automatically determined based on the score distribution. Generate an evaluation result that includes vehicle ID, scores for each dimension, overall score, and rating.

10. The method according to claim 7, characterized in that, The LLM intelligent suggestion generation in step S4 includes: Construct prompts for the traffic management expert system, define the analysis framework and output requirements; Prepare a data summary that includes road information, traffic flow conditions, congestion analysis, lane usage, safety risks, and driving behavior; Call the large language model, passing in the system prompt words and data summary; A multi-stage parsing strategy is used to process the response content, including identifying chapter markers, splitting content by markers, cleaning up format markers, and verifying content integrity. Intelligent completion of missing content ensures the integrity of all parts of the report; Enable alternative template generation for basic suggestions when the large language model call fails.

Images