Second-hand car market mixed flow control method and system based on multi-target recognition

Abstract

The invention relates to the technical field of traffic control, in particular to a second-hand car market mixed flow traffic control method and system based on multi-target recognition. The method comprises the following steps: in response to an activation signal of a barrier gate triggering device, collecting multi-dimensional identification original data of vehicles passing through a barrier gate of a second-hand vehicle market through a multi-source identification device; performing cross-device data calibration processing on the multi-dimensional recognition original data, performing feature matching, and generating a vehicle identity recognition result and an identity confidence coefficient parameter; according to the vehicle identity recognition result and the identity confidence coefficient parameter, a dynamic passing strategy engine is called to generate a vehicle passing permission evaluation result and a strategy adaptation coefficient; and generating vehicle passing state record data according to the passing authority evaluation result and the strategy adaptation coefficient, performing classification analysis, generating an intervention processing work order, and feeding back the intervention processing work order to the system optimization module for iteratively updating the identification model and the passing strategy. The barrier gate passing efficiency of the second-hand car market can be greatly improved.

Description

A method and system for controlling mixed traffic in the used car market based on multi-target recognition Technical Field

[0001] This invention relates to the field of traffic control technology, and in particular to a method and system for controlling mixed traffic in a used car market based on multi-target recognition. Background Technology

[0002] With the continuous expansion of the used car market, vehicle liquidity and transaction frequency have significantly increased. Market entrances and exits, as the core hubs for the centralized management of new cars and the frequent entry and exit of other vehicles, directly impact market operation order and management effectiveness through their efficiency and security. Used car markets generally exhibit the operational characteristic of "mixed flow of new cars and other vehicles." New cars need to complete entry and exit verification according to the inventory management system, while other vehicles encompass various identity types, including visitors, customers driving themselves, and industry-related personnel. This places higher demands on the recognition accuracy, identity differentiation capabilities, and strategy adaptability of the access control system.

[0003] In addition, a similar patent, CN114038074A, discloses a vehicle passage control method and system to reduce the occurrence of other vehicles cutting in line and using ETC (Electronic Toll Collection) to pass, thereby reducing the passage time for vehicles waiting to pass. This application includes: acquiring the first license plate information bound to the ETC card of the vehicle waiting to pass; acquiring a first image of the vehicle waiting to pass in the lane using a first image acquisition device; when the second license plate information identified in the first image matches the first license plate information, controlling the barrier to be in a waiting-to-leave state; acquiring a second image of the lane using a second image acquisition device within a preset time; determining whether the third license plate information identified in the second image matches the second license plate information; if they do not match, determining that the vehicle corresponding to the second image is a vehicle cutting in line and using ETC, and controlling the barrier to change from the waiting-to-leave state to an interception state. This invention improves the accuracy of ETC lane traffic control, but the solution focuses on preventing queue-jumping in ETC lanes on highways or in industrial parks. It is not designed for the core needs of the used car market where "new cars and social vehicles are mixed"—it lacks the ability to integrate and identify multi-dimensional identity markers such as RFID tags and QR codes for new cars, and cannot adapt to the differentiated needs of new car inventory verification and social vehicle classification management; it has not built a dynamic traffic strategy engine that links with the used car market inventory management system, and the flexibility of traffic rule configuration is insufficient. Summary of the Invention

[0004] In order to solve the above-mentioned technical problems existing in the existing vehicle traffic control process, the present invention provides a method that can simultaneously collect license plates, RFID tags, QR codes and environmental parameters through multiple source devices, adapt to the mixed traffic scenario of commercial vehicles and social vehicles, and make up for the limitations of a single identification mode. After cross-device calibration and standardization, data consistency and reliability are ensured, laying the foundation for identification. Multimodal fusion recognition, combined with a market-specific identity information database, enables accurate vehicle identity verification and classification. A dynamic access strategy engine, linked with the inventory management system, generates differentiated access schemes to meet the needs of rapid passage for new cars and classified management of social vehicles. An anomaly intervention mechanism and closed-loop optimization system ensure traffic order while continuously improving system adaptability. The overall solution significantly improves the efficiency and control accuracy of the used car market gate access, thereby achieving the dual goals of new car inventory verification and differentiated access management. The mixed-traffic control method for the used car market based on multi-target recognition includes the following steps: In response to the activation signal of the gate trigger device, the license plate reader, RFID reader, QR code scanner, and environmental perception sensor are activated through the multi-source recognition device integration layer to collect multi-dimensional identification raw data of vehicles passing through the used car market gate. The multi-dimensional identification raw data includes vehicle license plate image data, RFID tag storage information, QR code encoding data, and traffic scene environmental parameters; cross-device data calibration processing is performed on the multi-dimensional identification raw data. The system aligns the acquisition timestamps of different recognition devices using a device timing synchronization algorithm, employs a noise filtering model to remove environmental interference data, and performs format standardization conversion on unstructured image data and encoded data to obtain a standardized recognition dataset with a unified dimension. Based on this standardized recognition dataset, a multimodal identity fusion recognition module extracts vehicle identity feature vectors, performs feature matching in conjunction with a used car market vehicle identity information database, and generates vehicle identity recognition results and identity confidence parameters. According to the vehicle identity recognition results and identity confidence parameters, a dynamic access control engine is invoked, combining the current gate operation status parameters with the market access management rule base to generate vehicle access permission evaluation results and strategy adaptation coefficients. Based on the access permission evaluation results and strategy adaptation coefficients, the gate control decision layer outputs gate action control commands, and simultaneously generates vehicle access status record data. The vehicle access status record data is classified and analyzed, and a corresponding management intervention notification mechanism is triggered for vehicles with abnormal access types, generating an intervention processing work order and pushing it to the market management terminal. At the same time, the entire process data is fed back to the system optimization module for iterative updates to the recognition model and access strategy.

[0005] This invention integrates a multi-source identification device layer to simultaneously collect vehicle license plates, RFID tags, QR codes, and environmental parameters, comprehensively covering both the exclusive identification of new cars and the identity information of other vehicles. It perfectly adapts to the core scenario of the used car market where "new cars and other vehicles are mixed," overcoming the shortcomings of similar patents' single identification modes in meeting inventory verification and classification management requirements. Through cross-device time synchronization, noise filtering, and format standardization, data consistency and reliability are ensured, laying a solid foundation for accurate identification. The multimodal identity fusion identification module, combined with a market-specific identity information database, achieves accurate vehicle identity verification and type differentiation, generating confident identification results. The dynamic access strategy engine, linked with the inventory management system and market rule base, generates differentiated access schemes for different types of vehicles, such as new cars and other vehicles, significantly improving access flexibility and management accuracy. The gate control decision-making level quickly outputs action commands and records data throughout the entire process. The abnormal intervention mechanism promptly handles vehicles that violate regulations. The data feedback throughout the entire process enables iterative optimization of the identification model and passage strategy. The overall solution not only ensures the efficiency of rapid turnover of new cars, but also strengthens the management and order of social vehicles, significantly improving the efficiency of gate passage and the level of refined management in the used car market, and fully meeting the core needs of dynamic market management.

[0006] Preferably, the step of activating the license plate recognition device, RFID reader, QR code scanner, and environmental perception sensor through the multi-source identification device integration layer to collect multi-dimensional identification raw data of vehicles passing through the used car market gate includes the following steps: calling the device status monitoring interface to detect the operating status parameters of the license plate recognition device, RFID reader, QR code scanner, and environmental perception sensor in real time, and generating a device availability assessment result; based on the device availability assessment result, assigning data acquisition priorities and acquisition timing windows to each identification device, and generating a device acquisition scheduling plan; according to the device acquisition scheduling plan, activating the acquisition function of available identification devices, capturing high-definition images of vehicle license plates and image acquisition angle parameters through the license plate recognition device, and transmitting the data through R... The RFID reader reads the identification information and tag signal strength data from the vehicle's RFID tag. A QR code scanner obtains the encoding information and scanning distance parameters of the vehicle's associated QR code. Environmental sensors collect data on light intensity, obstruction status, and weather conditions at the gate to generate vehicle license plate image data, RFID tag storage information, QR code encoding data, and environmental parameters for the passage scenario. Device identifiers, collection timestamps, and environmental parameter tags are added to the raw data collected by each device to generate a multi-source raw data set with metadata tags. Based on the data source type of the multi-source raw data set, data integrity verification rules are established, missing data is marked, and the reasons for the missing data are recorded, generating multi-dimensional identification raw data.

[0007] This invention completely solves the core pain points of "poor equipment coordination, incomplete data, and weak scenario adaptability" in multi-source data collection for used car market gates by using equipment status monitoring and refined data collection scheduling. It calls the equipment status monitoring interface to assess the availability of each identification device, and allocates collection priorities and timing windows based on the assessment results, ensuring efficient coordination of available devices and avoiding data collection failures caused by single device malfunctions. Simultaneous collection from multiple devices is initiated, covering vehicle license plate images, RFID tag information, QR code encoding, and environmental parameters. This captures both the RFID and QR code identification unique to new cars and the license plate information of other vehicles, perfectly adapting to the scenario of "mixed traffic of new and other vehicles." Device identifiers, timestamps, and environmental tags are added to the collected data, and integrity verification rules are established to mark missing data, ensuring data traceability and integrity. This step breaks through the limitations of the single identification mode of similar patents, achieving multi-dimensional, traceable, and high-quality raw data collection, providing comprehensive and reliable data support for subsequent accurate identification and differentiated management, and completely changing the status quo of traditional solutions with single data sources and poor adaptability.

[0008] Preferably, the step of extracting vehicle identity feature vectors based on the standardized recognition dataset using a multimodal identity fusion recognition module and performing feature matching in conjunction with a used car market vehicle identity information database includes the following steps: separating standardized vehicle license plate image data, standardized RFID identity data, and standardized QR code encoding data from the standardized recognition dataset to generate a single-modal recognition data subset; performing feature extraction on each of the single-modal recognition data subsets; extracting character structure features, license plate color features, and texture features from the standardized license plate image data using a preset license plate feature extraction model to generate a license plate feature vector; and extracting the vehicle's unique identifier, ownership information, and tag binding time features from the standardized RFID identity data using a preset RFID data parsing model to generate an RFID identity feature vector. The process involves: extracting vehicle-related business codes and access authorization features from standardized QR code encoding data to generate a QR code feature vector; invoking a multimodal feature fusion algorithm to weight and fuse the license plate feature vector, RFID identity feature vector, and QR code feature vector, calculating the confidence weight of each modality feature, and generating a fused vehicle identity feature vector; calculating the similarity between the vehicle identity feature vector and the registered identity feature data in the used car market vehicle identity information database to generate a feature matching similarity matrix; determining the vehicle identity matching result based on the feature matching similarity matrix and a preset identity recognition threshold parameter, and simultaneously calculating the confidence parameter of the identity recognition result based on the confidence weight and similarity value of each modality feature, thus generating the vehicle identity recognition result and identity confidence parameter.

[0009] This invention addresses the shortcomings of traditional solutions, such as single-modal feature extraction and multimodal fusion matching, by precisely overcoming the limitations of single-modal and inaccurate identity recognition. It separates single-modal data subsets from a standardized dataset and extracts character, color, and texture features from license plates, unique identifiers and ownership characteristics from RFID tags, and business codes and authorization identifiers from QR codes, enabling in-depth mining of identity information across various modalities. A multimodal feature fusion algorithm is then used to generate a unified vehicle identity feature vector by combining the confidence weights of each modality, avoiding the limitations of single-modal recognition. Feature matching is performed with a dedicated vehicle identity information database for the used car market, which includes both new car inventory registration information and social vehicle registration data, accurately distinguishing vehicle types and verifying new car inventory. The generated identity recognition results and confidence parameters provide a quantitative basis for subsequent access control assessment, fundamentally changing the current situation where similar patents rely solely on license plate and ETC matching, failing to meet the multi-identity recognition needs of the used car market, and significantly improving the accuracy and reliability of vehicle identity recognition.

[0010] Preferably, the step of invoking a multimodal feature fusion algorithm to weight and fuse the license plate feature vector, RFID identity feature vector, and QR code feature vector, and calculating the credibility weight of each modality feature, includes the following steps: normalizing the dimensions of the license plate feature vector, RFID identity feature vector, and QR code feature vector to unify the dimensional space of the feature vectors and generate a standardized single-modal feature vector set; based on the standardized single-modal feature vector set, extracting the information entropy and feature discrimination parameters of each modality feature to generate a modality feature quality assessment dataset; and invoking a modality weight allocation model, combining the modality feature quality assessment dataset with the equipment acquisition environment parameters, to analyze the various modal features. The reliability of modal features in the current scenario is assessed by calculating the initial credibility weights of each modal feature. A feature correlation analysis algorithm is used to calculate the linear correlation coefficients and semantic association degrees between different modal feature vectors, generating an intermodal association matrix. Based on this intermodal association matrix, the initial credibility weights are iteratively corrected, reducing weight redundancy in highly correlated modalities and increasing the weight proportion of complementary modalities, generating an optimized set of modal credibility weights. A weighted summation fusion strategy is employed to perform matrix operations on the standardized single-modal feature vector set and the corresponding optimized modal credibility weight set, generating a fused vehicle identity feature vector. Simultaneously, a weight allocation log is recorded during the fusion process.

[0011] This invention comprehensively solves the problems of fixed weights, feature redundancy, and poor adaptability in traditional fusion algorithms by using dynamic weight allocation and multimodal feature fusion. The dimensionality of each modality's feature vector is normalized to ensure feature space consistency, laying the foundation for fusion. Information entropy and discriminative parameters of each modality are extracted, and feature quality is evaluated in conjunction with acquisition environment parameters to calculate initial reliability weights, ensuring the weight allocation aligns with the actual acquisition scenario. By analyzing the linear correlation coefficient and semantic association between modalities, the initial weights are iteratively corrected, weakening redundant feature weights and strengthening the proportion of complementary features, avoiding recognition bias caused by redundant feature calculations. A weighted summation strategy is used to generate the fused feature vector, integrating multi-dimensional identity information from license plates, RFID, and QR codes, while highlighting the contribution of high-reliability modalities through dynamic weight allocation, thus improving the representational capability of the fused features. This step achieves intelligent fusion of multimodal features, effectively addressing performance fluctuations of various devices under different acquisition environments, ensuring accurate extraction of vehicle identity features even in complex scenarios, and providing core technical support for subsequent accurate identification and differentiated access control.

[0012] Preferably, the iterative correction of the initial confidence weights based on the inter-modal correlation matrix, weakening the weight redundancy of highly correlated modes and strengthening the weight proportion of complementary modes, includes the following steps: performing eigenvalue decomposition on the inter-modal correlation matrix, extracting eigenvalues ​​and eigenvectors of the inter-modal correlation matrix, and determining the principal component direction and correlation strength threshold of the modal correlation; based on the eigenvalues ​​and eigenvectors, identifying highly correlated mode combinations and complementary mode combinations, and generating modal correlation classification results; for highly correlated mode combinations, calculating the redundancy coefficient of each mode within the mode combination, and reducing the redundancy coefficient proportionally. The initial confidence weights of each modality are reduced to generate intermediate weights after redundancy correction. For complementary modal combinations, the information gain value of each modality within the complementary modal combination is calculated, and the intermediate weights of each modality are increased proportionally based on the information gain value to generate a complementary enhanced weight set. The weight normalization function is called to map the complementary enhanced weight set to a preset weight interval to ensure that the sum of the confidence weights of each modality is a preset value, generating a preliminary optimized weight set. Based on historical fusion effect feedback data, a weight correction model is constructed to calibrate the error of the preliminary optimized weight set, generating an optimized modal confidence weight set.

[0013] This invention addresses the core pain points of traditional multimodal fusion, namely fixed weights, feature redundancy, and poor adaptability, through modal correlation analysis and iterative weight correction. It performs feature decomposition on the intermodal correlation matrix to accurately identify highly correlated and complementary modal combinations, providing a scientific basis for weight optimization. For highly correlated modalities, redundancy coefficients are calculated and weights are reduced to minimize recognition bias caused by redundant feature calculations. For complementary modalities, information gain values ​​are calculated and weights are increased to strengthen the complementary effect of different dimensions of identity information. Weight normalization ensures the rationality of weight allocation, and error calibration is performed based on historical fusion effect feedback to generate an optimized weight set adapted to the current scenario. This step achieves dynamic and intelligent allocation of credibility weights, avoiding the drawbacks of fixed weights being unable to handle complex scenarios. It also enhances the representational ability of fused features by weakening redundancy and strengthening complementarity, ensuring that vehicle identity feature vectors can comprehensively and accurately integrate multi-dimensional identity information. This provides core technical support for subsequent accurate identification and differentiated management, fundamentally changing the poor adaptability of traditional fusion algorithms.

[0014] Preferably, the step of calling the dynamic access control engine based on the vehicle identification results and identity confidence parameters, and combining the current gate operation status parameters with the market access control rule base to generate vehicle access permission evaluation results and strategy adaptation coefficients includes the following steps: parsing the vehicle identification results, extracting vehicle type identifier, entity type, business association status, and historical access record information to generate a vehicle identity attribute dataset; based on the vehicle identity attribute dataset, associating it with the classification management rules in the used car market access control rule base to determine the applicable basic access control strategy type and generate a strategy matching candidate set; collecting the current gate operation status parameters, including the current gate opening / closing status and traffic flow statistics. Based on equipment load parameters and fault warning information, a barrier gate operation status dataset is generated. Combining the vehicle identity attribute dataset, barrier gate operation status dataset, and identity confidence parameters, the adaptability of the basic access strategy in the current scenario is analyzed, and the expected effect parameters of strategy execution are calculated. Based on the expected effect parameters of strategy execution, the basic access strategies in the strategy matching candidate set are prioritized, and the ranking results are dynamically adjusted based on real-time market management demand parameters to generate the optimal access strategy scheme. The core decision rules and execution conditions of the optimal access strategy scheme are extracted, and vehicle access permission evaluation results are generated. Simultaneously, based on the strategy adaptability analysis results and the satisfaction of execution conditions, the strategy adaptability coefficient is calculated.

[0015] This invention, through a dynamic strategy engine and multi-dimensional scenario adaptation, precisely addresses the problem of rigid traffic rules in traditional solutions, which fail to meet the differentiated management needs of the used car market. It analyzes vehicle identification results, extracts attributes such as vehicle type, owner, and business association status, and generates a basic strategy candidate set by associating it with market classification management rules, perfectly adapting to the core scenario of "mixed traffic of new and private vehicles." It collects gate operation status parameters (open / close status, traffic flow, equipment load, etc.), combines them with vehicle identity attributes and identity confidence parameters, analyzes the scenario adaptability of the basic strategy, and calculates the expected effect, ensuring that strategy execution meets both management requirements and adapts to the current operating status. Based on the expected effect ranking and dynamic adjustments combined with real-time management needs, it generates the optimal traffic strategy, achieving differentiated management on a "one vehicle, one policy" basis—new vehicles can be quickly released, private vehicles can pass according to rules, and vehicles violating regulations are restricted from passing. This step breaks through the limitations of fixed rules in similar patents, constructing a dynamic strategy system linked to used car market inventory management and traffic management, significantly improving the flexibility and accuracy of traffic control.

[0016] Preferably, the step of combining the vehicle identity attribute dataset, the barrier gate operation status dataset, and identity confidence parameters to analyze the adaptability of the basic access strategy in the current scenario and calculate the expected effect parameters of strategy execution includes the following steps: constructing a strategy adaptation evaluation index system, which includes access efficiency indicators, security control indicators, management compliance indicators, and user experience indicators, and assigning scenario adaptation weights to each indicator; based on the vehicle identity attribute dataset, extracting key attribute factors affecting access strategy adaptation, including vehicle access priority, business urgency, and historical violation records, and generating an attribute influence factor set; and combining the barrier gate operation status dataset to analyze the constraints of the current access scenario. The system takes into account conditions, including traffic capacity, equipment operating limits, and channel occupancy status, to generate a set of scenario constraint parameters. It then calls a multi-factor comprehensive evaluation algorithm, substituting the attribute influence factor set, scenario constraint parameter set, and identity confidence parameter into the strategy adaptation evaluation index system to calculate the score of each basic access strategy under each evaluation index. Based on the scores of each evaluation index and the scenario adaptation weight, it uses a weighted summation method to calculate the comprehensive adaptation score of each basic access strategy. Finally, it calls a preset effect prediction model, combining the comprehensive adaptation score with historical strategy execution effect data, to predict the traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate after the implementation of each basic access strategy, generating expected strategy execution effect parameters.

[0017] This invention comprehensively addresses the shortcomings of traditional solutions in terms of one-sided and unscientific strategy adaptability assessments through a multi-dimensional evaluation system and scenario-based effect prediction. It constructs a multi-dimensional evaluation index system covering traffic efficiency, security control, management compliance, and user experience, providing a comprehensive framework for strategy evaluation. Key vehicle identity attribute factors (traffic priority, business urgency, etc.) are extracted, and the constraints of the barrier gate operation scenario (flow capacity, equipment limits, etc.) are analyzed to ensure the evaluation aligns with actual application scenarios. A multi-factor comprehensive evaluation algorithm is invoked, combining identity confidence parameters to calculate the comprehensive adaptability score of each strategy, avoiding the limitations of single-indicator evaluation. Based on the comprehensive score and historical data, the strategy execution effect (traffic efficiency improvement rate, violation risk reduction rate, etc.) is predicted, providing a quantitative basis for strategy selection. This step achieves scientific and scenario-based evaluation of traffic strategies, ensuring that the selected strategy meets the differentiated management needs of the used car market while balancing traffic efficiency and management security. It completely changes the traditional approach of relying on experience to formulate strategies with poor adaptability, providing reliable decision support for dynamic traffic strategy engines.

[0018] Preferably, the step of calling a preset effect prediction model, combining the comprehensive adaptation score and historical strategy execution effect data, to predict the traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate after the implementation of each basic traffic strategy, and generating expected strategy execution effect parameters includes the following steps: collecting historical traffic strategy execution data, including execution records of each basic traffic strategy in different scenarios, corresponding comprehensive adaptation scores, and actual execution effect indicators, to generate a historical strategy effect dataset; performing data preprocessing on the historical strategy effect dataset, removing abnormal data and invalid records, and unifying the indicator dimensions through data normalization to generate a standardized historical dataset; and constructing training samples for the effect prediction model based on the standardized historical dataset. In this episode, the comprehensive adaptation score, vehicle identity attribute features, and scenario constraint parameters are used as input features, and the traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate are used as output labels. A machine learning algorithm is used to train the model on the training sample set, and the model parameters are optimized through cross-validation to generate an effect prediction model. The comprehensive adaptation score of the current basic traffic strategy, the corresponding set of vehicle identity attribute influencing factors, and the set of scenario constraint parameters are input into the effect prediction model, which outputs the predicted values ​​of traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate for each basic traffic strategy. The credibility of these predicted values ​​is verified, and corrections are made based on the model prediction error rate to generate the expected effect parameters of strategy execution.

[0019] This invention addresses the core pain points of traditional traffic strategy evaluation—namely, the lack of quantitative basis and insufficient predictive accuracy—by constructing a historical data-driven effect prediction model. Historical strategy execution data from different scenarios is collected and preprocessed to generate a standardized dataset, ensuring the reliability and consistency of the training data. Using comprehensive adaptation scores, vehicle identity attributes, and scenario constraint parameters as input features, and traffic efficiency, violation risk, and management cost-related indicators as output labels, a training sample set accurately matching the mixed-traffic scenario of the used car market is constructed. Machine learning algorithms are used to train the effect prediction model, and cross-validation is combined to optimize parameters and improve the model's prediction accuracy. Current scenario parameters are input into the model to generate predicted values, which are then verified for credibility and corrected for errors, resulting in quantified parameters for the expected effect of strategy execution. This step achieves scientific prediction and quantitative evaluation of traffic strategy effects, fundamentally changing the traditional approach of relying on experience to formulate strategies and ensuring adaptability. It provides data support for the dynamic traffic strategy engine to select the optimal solution, ensuring that the selected strategy meets the differentiated needs of rapid passage for new cars and the classified management of social vehicles, while balancing traffic efficiency, safety control, and management costs, significantly improving the precision of traffic management in the used car market.

[0020] Preferably, the method of using machine learning algorithms to train the model on the training sample set and optimizing the model parameters through cross-validation includes the following steps: dividing the training sample set into a training subset, a validation subset, and a test subset, and setting the data proportion and data distribution characteristics of each subset; initializing the network structure and hyperparameters of the performance prediction model, including the number of hidden layer neurons, the initial learning rate, and the iteration threshold; iteratively training the model using the gradient descent optimization algorithm based on the training subset, calculating the loss function value for each iteration, and adjusting the model parameters through the backpropagation algorithm; during each iteration, inputting the validation subset into the model at the current training stage, calculating the validation set prediction error, and when the validation set prediction error increases consecutively... Upon startup, an early stopping mechanism is triggered to halt model training. A cross-validation algorithm is invoked to divide the training sample set into multiple mutually exclusive subsets for multiple rounds of model training and validation, recording the optimal model parameters and corresponding prediction errors for each round. Based on the results of multiple rounds of cross-validation, the optimal distribution of each model parameter is statistically analyzed, and a parameter fusion algorithm is used to determine the final model parameters. These final model parameters are then substituted into the model structure, and the model's performance is evaluated using a test subset. The model's prediction accuracy, recall, and F1 score are calculated, generating a model performance evaluation report. When the model performance evaluation report meets preset standards, the performance prediction model is output. If it does not meet the standards, the model network structure is adjusted and retrained until a satisfactory performance prediction model is generated.

[0021] This invention precisely addresses the technical shortcomings of traditional machine learning models, such as overfitting and weak generalization ability, through a scientific model training and parameter optimization process. The training sample set is divided into training, validation, and testing subsets to ensure a reasonable data distribution and provide a reliable foundation for model training and evaluation. The model network structure and hyperparameters are initialized, and gradient descent is used for iterative training, combined with an early stopping mechanism to avoid overfitting, ensuring the effectiveness of model training. Multiple rounds of training and validation are conducted using a cross-validation algorithm to statistically determine the optimal parameter distribution and fuse it to determine the final parameters, improving the model's generalization ability and stability. Model performance is evaluated based on the test subset, generating an evaluation report including prediction accuracy, recall, and F1 score to ensure the model meets preset standards. This comprehensive training system, encompassing "data partitioning - iterative training - early stopping to prevent overfitting - cross-validation - performance evaluation," generates a predictive model with high accuracy and strong generalization ability. It can accurately predict the execution effect of different traffic strategies in the mixed-traffic scenario of the used car market, providing core technical support for the scientific decision-making of dynamic traffic strategies and completely changing the current situation of non-standardized model training processes and poor prediction results.

[0022] As a preferred embodiment, the second technical solution of the present invention is: a used car market mixed traffic control system based on multi-target recognition, used to execute the used car market mixed traffic control method based on multi-target recognition as described above. This used car market mixed traffic control system based on multi-target recognition includes a multi-source recognition device integration layer, a data processing layer, an identity recognition layer, a strategy decision layer, a control execution layer, and a data feedback layer. Each layer interacts through a standardized data interface. The multi-source recognition device integration layer includes a license plate reader, an RFID reader, a QR code scanner, and an environmental perception sensor, used to collect multi-dimensional vehicle identification raw data and perform preliminary preprocessing. The data processing layer is used to process multi-target recognition data... The system performs calibration, noise reduction, and standardization on the raw data to generate a standardized recognition dataset. The identity recognition layer extracts vehicle identity feature vectors using a multimodal feature fusion algorithm, combines them with an identity information database to complete vehicle identity recognition, and generates identity recognition results and confidence parameters. The strategy decision layer calls the dynamic access control engine, combines the gate's operating status and management rule base, and generates access permission evaluation results and strategy adaptation coefficients. The control execution layer outputs gate control commands, generates access status record data, and triggers an abnormal vehicle management intervention mechanism. The data feedback layer collects data from the entire process to provide data support for recognition model optimization and access control strategy iteration.

[0023] This invention comprehensively solves the problems of poor scenario adaptability, fragmented functions, and inability to meet the needs of mixed traffic management in traditional traffic control systems through a layered architecture design and full-process collaborative linkage. The multi-source identification device integration layer realizes the comprehensive collection of multi-dimensional vehicle identity data and environmental parameters, adapting to mixed traffic scenarios of commercial vehicles and social vehicles; the data processing layer completes data calibration, noise reduction, and standardization to ensure data quality; the identity recognition layer accurately identifies vehicle identities through multi-modal feature fusion algorithms, meeting the needs of commercial vehicle inventory verification and vehicle classification; the strategy decision layer calls the dynamic traffic strategy engine, combining the gate operation status and management rule base to generate differentiated traffic schemes, breaking through the limitations of traditional fixed rules; the control execution layer outputs precise control commands and triggers abnormal intervention to ensure traffic order; and the data feedback layer collects full-process data to support model and strategy iterative optimization. Each layer interacts efficiently through standardized interfaces, constructing a complete closed-loop system of "collection-processing-identification-decision-execution-optimization". This not only solves the core needs of multi-dimensional identity recognition and differentiated management in the mixed-traffic scenario of the used car market, but also enables the continuous evolution of the system, completely changing the status quo of traditional control systems with single functions and poor adaptability, and providing the used car market with an efficient, accurate and flexible access control solution.

[0024] The following benefits are achieved: (1) By integrating and collecting multi-dimensional traffic data through multi-source identification devices, the core pain point of "single identity identification and incomplete data coverage" in the passage of the second-hand car market gate is completely solved. After responding to the gate activation signal, the license plate recognition device, RFID card reader, QR code scanner and environmental perception sensor are activated simultaneously to collect vehicle license plate images, RFID tag information, QR code data and environmental parameters. It covers the identity identification of the exclusive RFID tag and inventory QR code of the commercial vehicle, as well as the license plate information of social vehicles, which is perfectly adapted to the scenario characteristics of "mixed flow of commercial vehicles and social vehicles" in the second-hand car market. Compared with the single identification mode of similar patents that only rely on ETC and license plate matching, this step realizes the synchronous collection of multi-dimensional identity data, which provides rich material for subsequent accurate identity identification. At the same time, the passage scenario parameters collected by the environmental perception sensor provide support for subsequent data calibration and anti-interference processing, which completely changes the status quo of the traditional solution with single data source and difficulty in meeting the differentiated needs of commercial vehicle inventory verification and social vehicle classification management, and lays a solid data foundation for accurate management of the whole process.

[0025] (2) By cross-device calibration and standardization, the technical defects of traditional solutions, such as data asynchrony, messy formats, and weak anti-interference ability, are precisely solved. The device time synchronization algorithm is used to align the collection timestamps of each recognition device to ensure the time consistency of multi-source data and avoid identity matching errors caused by collection time difference; the noise filtering model is used to remove interference data such as ambient light and occlusion, thereby improving data reliability; unstructured image data and coded data are converted into a standardized dataset with a unified format, which solves the problem of data format incompatibility of multiple devices. This step realizes the triple optimization of multi-source data of "synchronization-purification-standardization", ensuring the consistency, accuracy and usability of data, and providing high-quality input for subsequent multimodal identity fusion recognition. Compared with the limitations of similar patents that lack data calibration mechanism and are susceptible to environmental interference, this step greatly improves data quality, provides technical guarantee for accurately distinguishing commercial vehicles from social vehicles and meeting differentiated management and control needs, and avoids traffic control errors caused by data problems.

[0026] (3) By using multimodal fusion recognition technology, the core problems of single identity recognition and insufficient accuracy in traditional solutions are completely solved. The multimodal identity fusion recognition module extracts multi-dimensional identity feature vectors of vehicles from standardized datasets, integrates feature information from multiple identity identifiers such as license plates, RFID tags, and QR codes, and avoids the limitations of single identifier recognition; it combines feature matching with a vehicle identity information database specific to the used car market, which includes both new car inventory information and social vehicle registration data, and can accurately distinguish between new cars and social vehicles, while also completing new car inventory verification. The generated vehicle identity recognition results and identity confidence parameters provide a quantitative basis for subsequent access permission assessment, completely changing the current situation where similar patents only rely on license plate and ETC matching and cannot adapt to the multi-identity recognition needs of the used car market. This step greatly improves the accuracy and reliability of vehicle identity recognition, and provides core support for new car inventory management and social vehicle classification access.

[0027] (4) The dynamic access strategy engine comprehensively solves the problems of rigid access rules and inability to meet the needs of differentiated management in traditional solutions. Based on the identity recognition results and confidence parameters, combined with the current gate operation status (such as traffic flow and equipment load) and the market access management rule library (including differentiated rules such as fast passage for commercial vehicles, registration and passage for social vehicles, and prohibition of passage), the access permission assessment results and strategy adaptation coefficients are generated, realizing dynamic management and control of "one policy for each vehicle". Compared with the shortcomings of fixed access rules and insufficient flexibility of similar patents, the dynamic strategy engine built in this step can be linked with the used car market inventory management system. Commercial vehicles can be quickly released with valid identity tags, social vehicles need to register and pass according to the rules, and vehicles that violate the rules are restricted from passing, perfectly adapting to the core needs of commercial vehicle inventory verification and social vehicle classification management and control. This step improves the flexibility and control accuracy of gate access, ensuring the efficiency of commercial vehicle circulation and strengthening market access order.

[0028] (5) By precisely controlling decision-making and recording all status data, the problems of delayed control response and weak data traceability in traditional solutions are precisely solved. Based on the access permission assessment results and strategy adaptation coefficient, the gate control decision layer quickly outputs gate action control commands (release, intercept, delayed release), ensuring the timeliness and accuracy of control response and improving passage efficiency. At the same time, passage status record data containing information such as vehicle identity, passage time, control commands, and equipment status is generated, realizing data traceability of the entire passage process. Compared with similar patents that only focus on preventing queue jumping and lack full-process data recording, the status record data in this step provides a basis for subsequent anomaly handling and accumulates materials for system optimization, completely changing the status of traditional solutions that "only control, do not record" or record incompletely. This step ensures the accuracy of gate control and the integrity of passage data, providing dual support for market management and system iteration.

[0029] (6) Through anomaly intervention mechanisms and closed-loop iterative optimization, the problems of lack of proactive control and continuous decline in system adaptability in traditional solutions are comprehensively solved. Traffic status record data is classified and analyzed. For vehicles violating regulations such as unauthorized passage or identity mismatch, a management intervention notification mechanism is triggered, generating work orders that are pushed to the market management terminal, enabling rapid handling of abnormal situations and strengthening market traffic control. Simultaneously, all process data (identification data, control data, and intervention data) is fed back to the system optimization module for iterative updates to the identification model and traffic strategy. This allows the identification model to continuously adapt to new identity types and environmental interference scenarios, and the traffic strategy to be continuously optimized to meet the needs of market management rule adjustments. Compared to similar patents that lack closed-loop optimization mechanisms and suffer from decreased accuracy after long-term use, the "data-analysis-intervention-optimization" closed-loop system constructed in this step ensures continuous system evolution, improving the effectiveness of current traffic control and guaranteeing the system's long-term stable adaptability, fully meeting the core needs of dynamic management in the used car market. Attached Figure Description

[0030] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 is a flowchart illustrating the steps of the multi-target recognition-based mixed-traffic control method for used car markets according to the present invention; Figure 2 is a block diagram illustrating the modules of the multi-target recognition-based mixed-traffic control system for used car markets according to the present invention; Figure 3 is a flowchart illustrating the vehicle entry process of the present invention; Figure 4 is a flowchart illustrating the vehicle exit process of the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0032] To achieve the above objectives, please refer to Figure 1. Embodiment 1 of the present invention provides a method for controlling mixed traffic in a used car market based on multi-target recognition, including the following steps: S01: In response to the activation signal of the gate triggering device, the license plate recognition device, RFID card reader, QR code scanner and environmental perception sensor are activated through the multi-source recognition device integration layer to collect multi-dimensional recognition raw data of vehicles passing through the used car market gate. The multi-dimensional recognition raw data includes vehicle license plate image data, RFID tag storage information, QR code encoding data and traffic scene environmental parameters. In this embodiment of the present invention, the ground induction coil triggering device at the used car market gate detects the entry of a vehicle (trigger signal strength threshold ≥800mV), generates an activation signal and transmits it to the multi-source recognition device integration layer. The integrated layer responds to the activation signal, simultaneously activating the license plate recognition device, RFID reader, QR code scanner, and environmental perception sensor: The license plate recognition device captures high-definition images of the vehicle license plate at a resolution of 1920×1080 and a sampling frequency of 30 frames / second, recording the image acquisition angle (0° horizontally, 15° vertically); The RFID reader reads the RFID tag on the vehicle's windshield via the 13.56MHz frequency band, obtaining 15-digit identification information (such as "CSC20240601008") and the tag signal strength (-62dBm); The QR code scanner recognizes the QR code affixed to the vehicle's side window at a scanning frequency of 500ms / scan, parsing out 12-digit encoded information (such as "QR2024CS000256") and the scanning distance (28cm); The environmental perception sensor collects the light intensity (650lux), obstruction status (unobstructed), weather conditions (sunny day), and temperature (25℃) at the gate. When collecting data, each device automatically adds a device identifier (license plate recognition device ID001, RFID reader ID002, etc.) and a collection timestamp (accurate to milliseconds, such as "2024-06-01 14:25:36.789"), generating a multi-dimensional recognition raw data set containing vehicle license plate image data (10 consecutive frames), RFID tag storage information, QR code encoding data, and traffic scene environmental parameters.

[0033] S02: Perform cross-device data calibration processing on the multi-dimensional recognition raw data. Align the acquisition timestamps of different recognition devices using a device timing synchronization algorithm, remove environmental interference data using a noise filtering model, and standardize the format of unstructured image data and coded data to obtain a standardized recognition dataset with unified dimensions. In this embodiment of the invention, cross-device data calibration processing is performed on the multi-dimensional recognition raw data: using a device timing synchronization algorithm, with the acquisition timestamp of the license plate recognition device as the benchmark, and aligning the acquisition times of all devices through timestamp deviation compensation (RFID reader deviation +120ms, QR code scanner deviation +80ms, environmental sensor deviation +50ms), ensuring that the data time difference is ≤10ms; calling noise... An acoustic filtering model (based on a Gaussian filtering algorithm with a 3×3 filter kernel) is used to remove interference data with an intensity < -80dBm from RFID signals, invalid encoded data with ambiguity > 0.3 from QR code scanning, and abnormal illumination values ​​with fluctuations > 20% from environmental sensors. Unstructured data undergoes format standardization: license plate image data is uniformly compressed to 800×400 resolution JPG format, and color interference is removed through grayscale processing (grayscale value range 0-255); RFID identification and QR code encoding are converted to string format, with added field descriptions (e.g., "Identification - Type: Vehicle"); environmental parameters are numerically normalized (illuminance 650 lux converted to 0.65, temperature 25℃ converted to 0.5). Finally, a standardized identification dataset with unified dimensions (consistent data format, timestamp, and numerical range) is generated, containing one image file, three types of structured data, and four environmental parameters.

[0034] S03: Based on the standardized recognition dataset, the vehicle identity feature vector is extracted through the multimodal identity fusion recognition module, and feature matching is performed in conjunction with the used car market vehicle identity information database to generate vehicle identity recognition results and identity confidence parameters; In this embodiment of the invention, the multimodal identity fusion recognition module receives the standardized recognition dataset and extracts features in modules: the license plate feature extraction submodule uses a convolutional neural network to extract character structure features (64-dimensional vector, including character aspect ratio and stroke spacing) and license plate color features (8-dimensional vector, with blue background and white characters corresponding to a feature value of 0) from the standardized license plate image. The .92 and texture features (32-dimensional vector, license plate surface reflectivity) are concatenated to form a 104-dimensional license plate feature vector. The RFID feature extraction submodule parses the identity identifier, extracting the vehicle's owner code (8-dimensional vector), tag binding time (4-dimensional vector, normalized to 0-1), and signal stability parameters (4-dimensional vector), generating a 16-dimensional RFID identity feature vector. The QR code feature extraction submodule parses the encoding information, extracting the business type identifier (8-dimensional vector), access authorization level (4-dimensional vector), and encoding generation time (4-dimensional vector), generating a 16-dimensional QR code feature vector. A weighted fusion algorithm is called to sum the three types of feature vectors based on the confidence weights of each modality feature (license plate 0.45, RFID 0.32, QR code 0.23), generating a 128-dimensional vehicle identity feature vector. The vector was compared with the vehicle identity information database of the used car market (containing 128-dimensional feature vectors of 6,000 registered vehicles) by cosine similarity calculation. The registered vehicle ID "BA20240515012" was matched with a similarity value of 0.93. The vehicle identity recognition result (matched registered vehicle, type is commercial vehicle, and the merchant is "Chengxin Car Dealership") and identity confidence parameter (93%) were generated.

[0035] S04: Based on the vehicle identification results and identity confidence parameters, the dynamic access strategy engine is invoked. Combining the current gate operation status parameters with the market access management rule base, a vehicle access permission evaluation result and strategy adaptation coefficient are generated. In this embodiment, the dynamic access strategy engine receives the vehicle identification results and identity confidence parameters (93% ≥ 85%). First, it parses the identity results to extract vehicle attributes: vehicle type (commercial vehicle), owner (merchant), business status (normally for sale, no mortgage), and historical access records (15 entries and 13 exits in the last 30 days, no violations), generating a vehicle identity attribute dataset. Then, it collects the current gate operation status parameters: gate open / closed status (closed), traffic flow (4 vehicles entered and 3 vehicles exited in the last 5 minutes, low traffic), equipment load (CPU utilization 38%, memory utilization 45%), and fault warning (no faults), generating a gate operation status dataset. The engine is linked to the used car market access management rule library, matching the basic access strategies (priority access, inspection-free access) for new cars and participating merchants. Combining gate status and confidence parameters, the system analyzes strategy suitability: the inspection-free access strategy is expected to improve efficiency by 40% and reduce safety risk by 1%, while the priority access strategy is expected to improve efficiency by 30% and reduce safety risk by 0%. Through multi-factor comprehensive evaluation (efficiency weight 0.35, safety weight 0.3, compliance weight 0.2, experience weight 0.15), the comprehensive score for the inspection-free access strategy is calculated to be 92 points, and for the priority access strategy, it is 88 points, determining the inspection-free access strategy as the optimal strategy. Vehicle access permission evaluation results are generated (allowed passage, inspection-free release). Simultaneously, based on the strategy suitability score (92 points) and the execution condition satisfaction rate (100%), the strategy suitability coefficient is calculated (0.92×0.8+1.0×0.2=0.936).

[0036] S05: Based on the access permission assessment result and strategy adaptation coefficient, the gate control decision layer outputs gate action control commands and simultaneously generates vehicle passage status record data. In this embodiment of the invention, the gate control decision layer receives the access permission assessment result (allowed passage, inspection-free release) and the strategy adaptation coefficient (0.936), and calls the gate control command generation algorithm: based on the gate mechanical parameters (maximum lifting height 1.8 meters, lifting speed 0.3 meters / second), calculates the lifting time of 6 seconds and the holding time of 15 seconds (to meet vehicle passage requirements), and generates gate action control commands (lifting height 1.8 meters, lifting duration 6 seconds, holding for 15 seconds, then automatic lowering). The commands are transmitted to the gate actuator via the RS485 interface. After receiving the commands, the actuator starts the motor and completes the lifting, holding, and lowering actions according to preset parameters. At the same time, the control decision layer automatically generates vehicle passage status record data, including vehicle ID (“BA20240515012”), passage time (“2024-06-01 14:26:12”), strategy type (inspection-free passage), gate action parameters, equipment operating status and identity confidence level. The data format is uniformly JSON, and the field integrity is ≥98%.

[0037] S06: Classify and analyze the vehicle passage status record data, trigger the corresponding management intervention notification mechanism for vehicles with abnormal passage types, generate an intervention processing work order and push it to the market management terminal, and at the same time feed the full process data back to the system optimization module for iterative updates to the identification model and passage strategy.

[0038] In this embodiment of the invention, vehicle passage status records are analyzed by categorizing passage types: normal passage (compliant with policy, no abnormalities) and abnormal passage (identity questionable, violation records, unauthorized). If an abnormally passing vehicle is identified (e.g., identity confidence level 68% < 85%, or a vehicle locked by financial institutions), a corresponding management intervention notification mechanism is triggered: the system automatically generates an intervention processing work order, including basic vehicle information (license plate, identity identifier), abnormality type (identity questionable / locked status), passage time, gate location, and processing suggestions (interception and verification / contacting the merchant). The work order is pushed to the market management terminal (computer or mobile APP) in real time via the WebSocket protocol. Upon receiving the work order, the terminal triggers an audio-visual reminder (volume 60dB, light flashing frequency 2 times / second). Simultaneously, the system aggregates all process data (equipment-collected data, standardized datasets, identity recognition results, strategy decision-making process, access records, and intervention work orders) and transmits them to the system optimization module via the data bus. The optimization module adds abnormal data to the recognition model training set after labeling it for incremental model training (learning rate 0.05, 5 training rounds). It also feeds back strategy execution performance data (such as a 38% increase in actual efficiency of inspection-free access) to the strategy engine to adjust strategy weight parameters (efficiency weight is fine-tuned from 0.35 to 0.37), thereby achieving iterative updates of the recognition model and access strategy.

[0039] The above method can be applied to both vehicle entry and exit processes. For vehicle entry, when a vehicle enters the gate entrance and triggers the inductive loop or camera, the gate's identification equipment starts. The system collects and integrates data from multiple identification devices, including license plate readers, RFID readers, and QR code scanners. Based on the integrated identification data, the system identifies the vehicle and, according to the current gate's access policy, determines whether to allow the vehicle to pass, then controls the gate equipment to take appropriate action. For vehicles returning without paying, the system will notify the market management in real time, who will then handle and intervene according to the management rules (as shown in Figure 3). The vehicle exit process can be as follows: when a vehicle enters the gate exit and triggers the ground loop or camera, the gate's identification equipment at the exit starts. The system collects and integrates data from multiple identification devices such as license plate recognition devices, RFID card readers, and QR code scanners. Based on the integrated identification data, the system completes vehicle identification and determines whether to allow the vehicle to pass according to the current gate's access policy, and then controls the gate equipment to take corresponding actions. For financial vehicles and vehicles in a locked state, the system will notify the market management in real time. For social vehicles that need to pay, the system will remind the user to make the payment, and the vehicle can only be released after the payment is successful (as shown in Figure 4).

[0040] Furthermore, the step of activating the license plate recognition device, RFID reader, QR code scanner, and environmental perception sensor through the multi-source identification device integration layer to collect multi-dimensional identification raw data of vehicles passing through the used car market gate includes the following steps: calling the device status monitoring interface to detect the operating status parameters of the license plate recognition device, RFID reader, QR code scanner, and environmental perception sensor in real time, and generating a device availability assessment result; based on the device availability assessment result, assigning data acquisition priorities and acquisition timing windows to each identification device, and generating a device acquisition scheduling plan; according to the device acquisition scheduling plan, activating the acquisition function of available identification devices, capturing high-definition images of vehicle license plates and image acquisition angle parameters through the license plate recognition device, and transmitting the data through the R... The RFID reader reads the identification information and tag signal strength data from the vehicle's RFID tag. A QR code scanner obtains the encoding information and scanning distance parameters of the vehicle's associated QR code. Environmental sensors collect data on light intensity, obstruction status, and weather conditions at the gate to generate vehicle license plate image data, RFID tag storage information, QR code encoding data, and environmental parameters for the passage scenario. Device identifiers, collection timestamps, and environmental parameter tags are added to the raw data collected by each device to generate a multi-source raw data set with metadata tags. Based on the data source type of the multi-source raw data set, data integrity verification rules are established, missing data is marked, and the reasons for the missing data are recorded, generating multi-dimensional identification raw data.

[0041] In this embodiment of the invention, by calling the device status monitoring interface, the operating status parameters of the license plate recognition device, RFID reader, QR code scanner, and environmental perception sensor at the used car market gate are detected in real time, including the device power supply voltage (standard range 220V±10%), communication link latency (normal threshold ≤50ms), and hardware module response speed (startup time after triggering ≤100ms). The device availability assessment results are generated as follows: license plate recognition device (power supply 225V, latency 30ms, response 80ms, available), RFID reader (power supply 218V, latency 45ms, response 90ms, available), QR code scanner (power supply 222V, latency 35ms, response 75ms, available), and environmental perception sensor (power supply 220V, latency 25ms, response 60ms, available). Based on the evaluation results, data acquisition priorities and acquisition time windows are assigned to each device: license plate recognition device (priority 1, acquisition time window 0-300ms), RFID card reader (priority 2, acquisition time window 100-400ms), QR code scanner (priority 3, acquisition time window 200-500ms), and environmental sensing sensor (priority 4, continuous acquisition time window 0-500ms). A device acquisition scheduling scheme is generated to ensure that data acquisition is conflict-free and sequential. According to the scheduling plan, the data acquisition function of available identification equipment is activated: the license plate recognition device captures high-definition images of vehicle license plates (resolution 1920×1080, JPG format) and image acquisition angle parameters (horizontal angle 0°, vertical angle 15°); the RFID reader reads the identification information (such as "CSC20240501001") and tag signal strength data (-65dBm) in the vehicle's RFID tag through the 13.56MHz frequency band; the QR code scanner obtains the encoding information (such as "QR2024CS000123") and scanning distance parameters (30cm) of the QR code associated with the vehicle's windshield; the environmental perception sensor continuously collects the light intensity (500lux), obstruction status (no obstruction), and meteorological condition data (sunny, no precipitation) at the gate, generating vehicle license plate image data, RFID tag storage information, QR code encoding data, and traffic scene environmental parameters. Add device identifiers (license plate reader ID001, RFID reader ID002, QR code scanner ID003, environmental sensor ID004), collection timestamps (such as "2024-05-01 09:30:22.156") and collection environment parameter tags (light intensity 500 lux - unobstructed - sunny) to the raw data collected by each device, and generate a multi-source raw data set with metadata tags.Data integrity verification rules are established based on data source type: license plate images must contain complete character areas, RFID data must contain 15-digit identification, QR code data must contain 12-digit codes, and environmental parameters must contain three types of core data. Missing data is marked and the reason for the absence is recorded (e.g., vehicles without QR codes are marked "no QR code pasted"), and multi-dimensional identification raw data is generated.

[0042] Furthermore, the step of extracting vehicle identity feature vectors based on the standardized recognition dataset using a multimodal identity fusion recognition module and performing feature matching in conjunction with a used car market vehicle identity information database includes the following steps: separating standardized vehicle license plate image data, standardized RFID identity data, and standardized QR code encoding data from the standardized recognition dataset to generate a single-modal recognition data subset; performing feature extraction on the single-modal recognition data subset respectively; extracting character structure features, license plate color features, and texture features from the standardized license plate image data using a preset license plate feature extraction model to generate a license plate feature vector; and extracting the vehicle's unique identifier, ownership information, and tag binding time features from the standardized RFID identity data using a preset RFID data parsing model to generate an RFID identity feature vector. The process involves: extracting vehicle-related business codes and access authorization features from standardized QR code encoding data to generate a QR code feature vector; invoking a multimodal feature fusion algorithm to weight and fuse the license plate feature vector, RFID identity feature vector, and QR code feature vector, calculating the confidence weight of each modality feature, and generating a fused vehicle identity feature vector; calculating the similarity between the vehicle identity feature vector and the registered identity feature data in the used car market vehicle identity information database to generate a feature matching similarity matrix; determining the vehicle identity matching result based on the feature matching similarity matrix and a preset identity recognition threshold parameter, and simultaneously calculating the confidence parameter of the identity recognition result based on the confidence weight and similarity value of each modality feature, thus generating the vehicle identity recognition result and identity confidence parameter.

[0043] In this embodiment of the invention, three single-modal recognition data subsets are generated by separating standardized vehicle license plate image data (resolution compressed to 800×400, grayscale processing), standardized RFID identity data (identity identifier format standardized, signal strength normalized to the 0-1 range), and standardized QR code encoding data (encoding format parsed into strings and redundant characters removed) from a standardized recognition dataset. Feature extraction is performed on each single-modal data subset: using a preset license plate feature extraction model (based on a convolutional neural network architecture, trained on 100,000 license plate images), character structure features (character aspect ratio, stroke spacing, contour features, generating a 64-dimensional vector), license plate color features (blue background with white characters corresponding to a feature value of 0.9, yellow background with black characters corresponding to 0.7, generating an 8-dimensional vector), and texture features (license plate surface texture density, reflectivity, generating a 32-dimensional vector) from the standardized license plate image data, concatenating them into a 104-dimensional license plate feature vector; using a preset RFID data parsing model, from the standardized RFID... The vehicle's unique identifier (15-bit code converted to numerical features, generating a 64-dimensional vector), the owner's information (corresponding to the market merchant code, generating a 16-dimensional vector), and the tag binding time feature (binding time from the current time converted to a normalized value, generating an 8-dimensional vector) are extracted from the identity data and concatenated into an 88-dimensional RFID identity feature vector. From the standardized QR code encoding data, the vehicle's associated business code (the first 8 bits of the code are parsed into a business type feature, generating a 32-dimensional vector) and the access authorization identifier feature (the last 4 bits of the code are parsed into an authorization level, generating a 16-dimensional vector) are extracted and concatenated into a 48-dimensional QR code feature vector. A multimodal feature fusion algorithm is then used to weight and fuse the 104-dimensional license plate feature vector, the 88-dimensional RFID identity feature vector, and the 48-dimensional QR code feature vector: based on the environmental parameters of each device's acquisition (sufficient lighting, no obstruction) and data integrity (no missing data), the credibility weights of each modality feature are calculated (license plate 0.4, RFID 0.35, QR code 0.25), generating a 128-dimensional fused vehicle identity feature vector. Based on this fused feature vector, cosine similarity is calculated with the registered identity feature data (including 128-dimensional feature vectors of 5,000 registered vehicles) in the used car market vehicle identity information database to generate a feature matching similarity matrix (the matrix element values ​​range from 0 to 1, representing the degree of matching between the current vehicle and the registered vehicle).Based on the similarity matrix and the preset identity recognition threshold parameter (0.85), the vehicle identity matching result is determined (e.g., the similarity with the registered vehicle ID "BA20240420005" is 0.92, which is higher than the threshold, so the match is considered successful). Based on the confidence weight of each modality and the similarity value, the confidence parameter of the identity recognition result is calculated (0.92×0.4+0.95×0.35+0.88×0.25=0.917), and the vehicle identity recognition result (matching the registered vehicle ID "BA20240420005") and the identity confidence parameter (91.7%) are generated.

[0044] Furthermore, the step of invoking the multimodal feature fusion algorithm to weight and fuse the license plate feature vector, RFID identity feature vector, and QR code feature vector, and calculating the credibility weight of each modality feature, includes the following steps: Dimensional normalization processing is performed on the license plate feature vector, RFID identity feature vector, and QR code feature vector to unify the dimensional space of the feature vectors and generate a standardized single-modal feature vector set; based on the standardized single-modal feature vector set, the information entropy and feature discrimination parameters of each modality feature are extracted to generate a modality feature quality assessment dataset; a modality weight allocation model is invoked, and the modality feature quality assessment dataset is combined with the device acquisition environment parameters to analyze each… The reliability of modal features in the current scenario is assessed by calculating the initial credibility weights of each modal feature. A feature correlation analysis algorithm is used to calculate the linear correlation coefficients and semantic association degrees between different modal feature vectors, generating an intermodal association matrix. Based on this intermodal association matrix, the initial credibility weights are iteratively corrected, reducing weight redundancy in highly correlated modalities and increasing the weight proportion of complementary modalities, generating an optimized set of modal credibility weights. A weighted summation fusion strategy is employed to perform matrix operations on the standardized single-modal feature vector set and the corresponding optimized modal credibility weight set, generating a fused vehicle identity feature vector. Simultaneously, a weight allocation log is recorded during the fusion process.

[0045] In this embodiment of the invention, the 104-dimensional license plate feature vector, the 88-dimensional RFID identity feature vector, and the 48-dimensional QR code feature vector are normalized. Linear interpolation is then used to extend the RFID identity feature vector to 104 dimensions and the QR code feature vector to 104 dimensions, unifying the dimensional space of the feature vectors (104 dimensions) and generating a standardized single-modal feature vector set (104-dimensional license plate, 104-dimensional RFID, and 104-dimensional QR code). Based on this standardized vector set, the information entropy (measuring feature uncertainty, target range 0.5-1.5) and feature discrimination parameters (measuring the contribution of features to identity recognition, target range 0.6-1.0) of each modality feature are extracted: license plate feature information entropy 0.8, discrimination 0.92; RFID feature information entropy 0.9, discrimination 0.88; QR code feature information entropy 0.7, discrimination 0.85, generating a modal feature quality assessment dataset. The modal weight allocation model (trained based on the gradient boosting tree algorithm) is invoked, and combined with the quality assessment dataset and equipment acquisition environment parameters (500 lux illumination, no obstruction, sunny day) to analyze the reliability of each modal feature in the current scene: license plate features are less affected by illumination and have high discriminative power, with an initial confidence weight of 0.42; RFID features have stable signals and moderate information entropy, with an initial confidence weight of 0.36; QR code features have slightly lower discriminative power, with an initial confidence weight of 0.22, generating an initial confidence weight set (0.42, 0.36, 0.22). The linear correlation coefficient and semantic association degree between different modal feature vectors are calculated using the Pearson correlation coefficient algorithm: the correlation coefficient between license plate and RFID features is 0.3 (weak correlation), the correlation coefficient between license plate and QR code features is 0.25 (weak correlation), and the correlation coefficient between RFID and QR code features is 0.4 (moderate correlation), generating a 3×3 intermodal association matrix. Based on this correlation matrix, the initial confidence weights are iteratively adjusted: RFID and QR codes have a moderate correlation, so the RFID weight is reduced by 0.02 and the QR code weight by 0.01. The released weights are allocated to the highly complementary license plate features (increased by 0.03), strengthening the weight ratio of complementary modalities, and generating an optimized modal confidence weight set (0.45, 0.34, 0.21). A weighted summation fusion strategy is adopted, performing matrix operations on the standardized single-modal feature vector set and the optimized weight set (each modal vector is multiplied by its corresponding weight and then summed) to generate a 104-dimensional fused vehicle identity feature vector. Simultaneously, a weight allocation log is recorded during the fusion process (including initial weights, correction basis, final weights, and the contribution ratio of each modality) to ensure the traceability of the fusion process.

[0046] Furthermore, the iterative correction of the initial confidence weights based on the inter-modal correlation matrix, weakening the weight redundancy of highly correlated modes and strengthening the weight proportion of complementary modes, includes the following steps: performing eigenvalue decomposition on the inter-modal correlation matrix, extracting eigenvalues ​​and eigenvectors of the inter-modal correlation matrix, and determining the principal component direction and correlation strength threshold of the modal correlation; based on the eigenvalues ​​and eigenvectors, identifying highly correlated mode combinations and complementary mode combinations, and generating modal correlation classification results; for highly correlated mode combinations, calculating the redundancy coefficient of each mode within the mode combination, and reducing the redundancy coefficient proportionally. The initial confidence weights of each modality are reduced to generate intermediate weights after redundancy correction. For complementary modal combinations, the information gain value of each modality within the complementary modal combination is calculated, and the intermediate weights of each modality are increased proportionally based on the information gain value to generate a complementary enhanced weight set. The weight normalization function is called to map the complementary enhanced weight set to a preset weight interval to ensure that the sum of the confidence weights of each modality is a preset value, generating a preliminary optimized weight set. Based on historical fusion effect feedback data, a weight correction model is constructed to calibrate the error of the preliminary optimized weight set, generating an optimized modal confidence weight set.

[0047] In this embodiment of the invention, the 3×3 intermodal correlation matrix (license plate-RFID 0.3, license plate-QR code 0.25, RFID-QR code 0.4) is decomposed using eigenvalue decomposition. The singular value decomposition algorithm is used to extract eigenvalues ​​(λ1=0.85, λ2=0.32, λ3=0.18) and corresponding eigenvectors. A correlation strength threshold of 0.35 is set. Principal component directions with eigenvalues ​​≥ 0.35 correspond to the RFID-QR code correlation dimension. Correlation strengths ≥ 0.35 are determined to be highly correlated, and < 0.35 are considered complementary. Based on the eigenvalues ​​and eigenvectors, highly correlated modal combinations (RFID-QR code, correlation strength ≥ 0.35) and complementary modal combinations (license plate-RFID, license plate-QR code, both correlation strengths < 0.35) are identified, generating modal correlation classification results. For the highly correlated combination of RFID and QR code, the redundancy coefficient of each mode in the mode combination is calculated: Redundancy coefficient = correlation strength × 0.5, RFID redundancy coefficient 0.2, QR code redundancy coefficient 0.2. Based on the redundancy coefficient, the initial confidence weight of each mode is reduced proportionally (RFID initial 0.36×(1-0.2)=0.288, QR code initial 0.22×(1-0.2)=0.176), and the intermediate weight set after redundancy correction is generated (license plate 0.42, RFID 0.288, QR code 0.176). For complementary mode combinations, the information gain value of each mode is calculated: the information gain of license plate to RFID is 0.15, and the information gain of license plate to QR code is 0.18. Based on the information gain value, the intermediate weights are increased proportionally (license plate increases by 0.15 × 0.1 = 0.042, RFID increases by 0.15 × 0.05 = 0.0144, and QR code increases by 0.18 × 0.05 = 0.0088), generating a complementary enhanced weight set (0.462, 0.3024, 0.1848). The SoftMax weight normalization function is called to map the weight set to a preset weight range of 0-1, ensuring that the sum of the weights of each mode is 1, and the preliminary optimized weight set (0.47, 0.31, 0.22) is calculated. Based on historical feedback data from 5000 fusion results (including matching accuracy and confidence bias), a weight correction model based on random forest was constructed to calibrate the initial optimized weight set: the license plate weight was reduced by 0.02, the RFID weight was increased by 0.01, and the QR code weight was increased by 0.01, generating an optimized modal confidence weight set (0.45, 0.32, 0.23), ensuring that the weight allocation matches the actual recognition effect.

[0048] Furthermore, the step of generating vehicle access permission assessment results and strategy adaptation coefficients based on the vehicle identification results and identity confidence parameters, by calling the dynamic access strategy engine and combining the current gate operation status parameters with the market access management rule base, includes the following steps: parsing the vehicle identification results, extracting vehicle type identifier, entity type, business association status, and historical access record information to generate a vehicle identity attribute dataset; based on the vehicle identity attribute dataset, associating it with the classification management rules in the used car market access management rule base to determine the applicable basic access strategy type and generate a strategy matching candidate set; and collecting the current gate operation status parameters, including the current gate opening / closing status and traffic flow statistics. Based on equipment load parameters and fault warning information, a barrier gate operation status dataset is generated. Combining the vehicle identity attribute dataset, barrier gate operation status dataset, and identity confidence parameters, the adaptability of the basic access strategy in the current scenario is analyzed, and the expected effect parameters of strategy execution are calculated. Based on the expected effect parameters of strategy execution, the basic access strategies in the strategy matching candidate set are prioritized, and the ranking results are dynamically adjusted based on real-time market management demand parameters to generate the optimal access strategy scheme. The core decision rules and execution conditions of the optimal access strategy scheme are extracted, and vehicle access permission evaluation results are generated. Simultaneously, based on the strategy adaptability analysis results and the satisfaction of execution conditions, the strategy adaptability coefficient is calculated.

[0049] In this embodiment of the invention, the vehicle identity recognition result (matching the registered vehicle ID "BA20240420005") is parsed to extract the vehicle type identifier (commercial vehicle), the type of the entity to which it belongs (market merchant), the business association status (normally for sale, no mortgage lock-up), and historical passage record information (12 entries and 10 exits in the past 30 days, no violation records), generating a vehicle identity attribute dataset. Based on this dataset, the classification management rules in the used car market passage management rule base are associated: commercial vehicle - merchant affiliation - normal status corresponds to the basic passage strategy of "priority passage" and "no payment required", generating a strategy matching candidate set (priority passage strategy, regular passage strategy, inspection-free passage strategy). The current gate operation status parameters are collected: the current gate opening and closing status (closed), passage flow statistics (3 vehicles entered and 2 vehicles exited in the past 5 minutes, flow is at a low peak), equipment load parameters (CPU utilization rate 35%, memory utilization rate 42%), and fault warning information (no fault), generating a gate operation status dataset. Combining vehicle identity attribute datasets (high priority, no violations), gate operation status datasets (low flow, normal equipment), and identity confidence parameters (91.7%), the adaptability of basic passage strategies is analyzed: the priority passage strategy is expected to have no congestion and low safety risk, the regular passage strategy is inefficient, and the inspection-free passage strategy meets the condition of no violation records. Expected effect parameters for strategy execution are calculated (priority passage efficiency improvement of 30%, safety risk of 0%; inspection-free passage efficiency improvement of 40%, safety risk of 1%). Based on the expected effect parameters, the candidate strategy set is ranked (inspection-free passage > priority passage > regular passage). Combined with real-time market management demand parameters (currently, improved passage efficiency is required, allowing low-risk inspection exemptions), the ranking results are dynamically adjusted (maintaining the original ranking) to generate the optimal passage strategy scheme (inspection-free passage strategy). Extract the core decision rules of the scheme (commercial vehicles + merchants + no violations + low traffic → inspection-free passage) and execution conditions (identity confidence ≥ 85%, equipment no faults), generate vehicle passage permission assessment results (allow passage, inspection-free release), and calculate the strategy adaptation coefficient (0.95 × 0.8 + 1.0 × 0.2 = 0.96) based on the strategy adaptability analysis results (adaptability 95%) and the execution condition satisfaction (100%).

[0050] Furthermore, the step of combining the vehicle identity attribute dataset, the barrier gate operation status dataset, and identity confidence parameters to analyze the adaptability of the basic access strategy in the current scenario and calculate the expected effect parameters of strategy execution includes the following steps: constructing a strategy adaptation evaluation index system, which includes access efficiency indicators, security control indicators, management compliance indicators, and user experience indicators, and assigning scenario adaptation weights to each indicator; based on the vehicle identity attribute dataset, extracting key attribute factors affecting the adaptation of the access strategy, including vehicle access priority, business urgency, and historical violation records, and generating an attribute influence factor set; and combining the barrier gate operation status dataset to analyze the constraints of the current access scenario. The system takes into account conditions, including traffic capacity, equipment operating limits, and channel occupancy status, to generate a set of scenario constraint parameters. It then calls a multi-factor comprehensive evaluation algorithm, substituting the attribute influence factor set, scenario constraint parameter set, and identity confidence parameter into the strategy adaptation evaluation index system to calculate the score of each basic access strategy under each evaluation index. Based on the scores of each evaluation index and the scenario adaptation weight, it uses a weighted summation method to calculate the comprehensive adaptation score of each basic access strategy. Finally, it calls a preset effect prediction model, combining the comprehensive adaptation score with historical strategy execution effect data, to predict the traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate after the implementation of each basic access strategy, generating expected strategy execution effect parameters.

[0051] In this embodiment of the invention, a strategy adaptation evaluation index system is constructed, including traffic efficiency indicators (traffic time, pass rate), security control indicators (identity misjudgment risk, probability of unauthorized release), management compliance indicators (degree of compliance with rules, record completeness), and user experience indicators (waiting time, operation complexity). Scenario adaptation weights are assigned to each indicator (traffic efficiency 0.35, security control 0.3, management compliance 0.2, user experience 0.15). Based on the vehicle identity attribute dataset, key attribute factors are extracted: vehicle traffic priority (high priority corresponding to the merchant's vehicle, quantified value 0.9), business urgency (no emergency transport sign, quantified value 0.5), and historical violation records (no violations, quantified value 1.0), generating an attribute influence factor set. Based on the gate operation status dataset, the constraints of the current passage scenario are analyzed: flow capacity (10 vehicles / minute during off-peak hours, current flow is 3 vehicles / minute, lenient constraint, quantification value 0.9), equipment operating limits (CPU maximum load 80%, current 35%, lenient constraint, quantification value 0.95), and channel occupancy status (no occupancy, quantification value 1.0), generating a scenario constraint parameter set. The multi-factor comprehensive evaluation algorithm of the analytic hierarchy process is called, and the attribute influence factor set, scenario constraint parameter set, and identity confidence parameter (0.917) are substituted into the evaluation index system to calculate the scores of each basic passage strategy: priority passage strategy (passage efficiency 85 points, security control 90 points, management compliance 95 points, user experience 88 points); inspection-free passage strategy (passage efficiency 92 points, security control 85 points, management compliance 90 points, user experience 95 points); regular passage strategy (passage efficiency 70 points, security control 92 points, management compliance 95 points, user experience 75 points). Based on the indicator scores and scenario adaptation weights, the comprehensive adaptation score is calculated using a weighted summation method: Priority passage: 85×0.35+90×0.3+95×0.2+88×0.15=88.45 points; Inspection-free passage: 92×0.35+85×0.3+90×0.2+95×0.15=90.05 points; Regular passage: 70×0.35+92×0.3+95×0.2+75×0.15=81.35 points. By calling a linear regression-based performance prediction model and combining comprehensive fit scores with historical strategy execution performance data, the system predicts the traffic efficiency improvement rate (priority passage 30%, inspection-free passage 40%, regular passage 15%), violation risk reduction rate (priority passage 20%, inspection-free passage 10%, regular passage 30%), and management cost control rate (priority passage 15%, inspection-free passage 25%, regular passage 10%) after the implementation of each strategy. This generates expected performance parameters for strategy execution, providing data support for strategy selection.

[0052] Furthermore, the step of calling the preset effect prediction model, combining the comprehensive adaptation score and historical strategy execution effect data, to predict the traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate after the implementation of each basic traffic strategy, and generating expected strategy execution effect parameters includes the following steps: collecting historical traffic strategy execution data, including execution records of each basic traffic strategy in different scenarios, corresponding comprehensive adaptation scores, and actual execution effect indicators, to generate a historical strategy effect dataset; performing data preprocessing on the historical strategy effect dataset, removing abnormal data and invalid records, and unifying the indicator dimensions through data normalization to generate a standardized historical dataset; and constructing training samples for the effect prediction model based on the standardized historical dataset. In this episode, the comprehensive adaptation score, vehicle identity attribute features, and scenario constraint parameters are used as input features, and the traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate are used as output labels. A machine learning algorithm is used to train the model on the training sample set, and the model parameters are optimized through cross-validation to generate an effect prediction model. The comprehensive adaptation score of the current basic traffic strategy, the corresponding set of vehicle identity attribute influencing factors, and the set of scenario constraint parameters are input into the effect prediction model, which outputs the predicted values ​​of traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate for each basic traffic strategy. The credibility of these predicted values ​​is verified, and corrections are made based on the model prediction error rate to generate the expected effect parameters of strategy execution.

[0053] In this embodiment of the invention, historical traffic strategy execution data from the used car market over the past year is collected, covering 12 scenarios including off-peak / peak hours, sunny / rainy days, and new / private vehicles. This includes execution records (8000 records in total) for three basic strategies: priority passage, regular passage, and inspection-free passage. Corresponding comprehensive adaptation scores (60-95 points) and actual execution effect indicators (traffic efficiency improvement rate 5%-40%, violation risk reduction rate 0%-35%, management cost control rate 5%-30%) are used to generate a historical strategy effect dataset. This dataset is preprocessed: outlier data (e.g., 32 extreme records with traffic efficiency improvement rate > 50%) and invalid records (e.g., 156 records missing comprehensive adaptation scores) are removed using the interquartile range method; all indicators are uniformly converted to the 0-1 range using an extreme value normalization algorithm (e.g., a 30% traffic efficiency improvement rate is converted to 0.75), generating a standardized historical dataset (7812 valid records). A training sample set was constructed based on this dataset. Input features included a comprehensive fit score (normalized 0.6-0.95), vehicle identity attributes (vehicle type, owner, and quantified violation records), and scenario constraint parameters (traffic capacity and equipment operating status). Output labels were normalized values ​​of traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate. The training set (5468 records), validation set (1562 records), and test set (782 records) were divided in a 7:2:1 ratio. A gradient boosting tree machine learning algorithm was used to train the model on the training sample set, with 100 decision trees and a tree depth of 6. The model parameters were optimized using 5-fold cross-validation (learning rate 0.08, minimum sample split number 20) to generate a performance prediction model. The model is input into the comprehensive adaptation scores of the current basic traffic control strategy (exemption from inspection 0.9005, priority access 0.8845, regular access 0.8135), the corresponding vehicle identity attribute influencing factor set (commercial vehicle 0.9, no violations 1.0, merchant affiliation 0.95), and the scenario constraint parameter set (traffic constraint 0.9, device status 0.95). The output predicted values ​​are: exemption from inspection (efficiency improvement rate 0.4, violation risk reduction rate 0.1, cost control rate 0.25), priority access (efficiency 0.3, risk 0.2, cost 0.15), and regular access (efficiency 0.15, risk 0.3, cost 0.1). The reliability of the predicted values ​​is verified; the model's historical prediction error rate is 3%. After correction, the expected effect parameters for strategy execution are generated as follows: exemption from inspection (efficiency 40%, risk 10%, cost 25%), priority access (efficiency 30%, risk 20%, cost 15%), and regular access (efficiency 15%, risk 30%, cost 10%).

[0054] Furthermore, the step of using machine learning algorithms to train the model on the training sample set and optimizing the model parameters through cross-validation includes the following steps: dividing the training sample set into a training subset, a validation subset, and a test subset, and setting the data proportion and data distribution characteristics of each subset; initializing the network structure and hyperparameters of the performance prediction model, including the number of hidden layer neurons, the initial learning rate, and the iteration threshold; iteratively training the model using the gradient descent optimization algorithm based on the training subset, calculating the loss function value for each iteration, and adjusting the model parameters through the backpropagation algorithm; during each iteration, inputting the validation subset into the model at the current training stage, calculating the validation set prediction error, and when the validation set prediction error increases consecutively multiple times... Upon startup, an early stopping mechanism is triggered to halt model training. A cross-validation algorithm is invoked to divide the training sample set into multiple mutually exclusive subsets for multiple rounds of model training and validation, recording the optimal model parameters and corresponding prediction errors for each round. Based on the results of multiple rounds of cross-validation, the optimal distribution of each model parameter is statistically analyzed, and a parameter fusion algorithm is used to determine the final model parameters. These final model parameters are then substituted into the model structure, and the model's performance is evaluated using a test subset. The model's prediction accuracy, recall, and F1 score are calculated, generating a model performance evaluation report. When the model performance evaluation report meets preset standards, the performance prediction model is output. If it does not meet the standards, the model network structure is adjusted and retrained until a satisfactory performance prediction model is generated.

[0055] In this embodiment of the invention, the training sample set is divided into a training subset (5468 records), a validation subset (1562 records), and a test subset (782 records) in a 7:2:1 ratio to ensure that the data distribution characteristics of each subset are consistent (the proportion of vehicle type and scene type deviates from the original dataset by ≤2%). The network structure (gradient boosting tree, 100 decision trees, each tree depth 6) and hyperparameters (initial learning rate 0.1, iteration threshold 200, loss function is mean squared error) of the effect prediction model are initialized. Based on the training subset, the model is iteratively trained using the stochastic gradient descent optimization algorithm. The loss function value is calculated in each iteration (initial loss 0.12), and the splitting nodes and weight parameters of the decision trees are adjusted through the backpropagation algorithm. During the iteration process, the loss function value gradually decreases. After each iteration, the validation subset is input into the current model, and the validation set prediction error is calculated (initial error 0.11). When the iteration reaches 85 rounds, the validation set error increases for three consecutive rounds (0.032 in round 83, 0.033 in round 84, and 0.035 in round 85), triggering the early stopping mechanism and stopping model training. The 5-fold cross-validation algorithm is then used to divide the training sample set into five mutually exclusive subsets (each subset containing 1093-1094 data points). Five rounds of model training and validation are performed, recording the optimal model parameters (learning rate 0.07-0.09, tree depth 5-7) and the corresponding prediction error (0.03-0.035) for each round. Based on the results of the five rounds, the optimal distribution of each parameter is statistically analyzed: learning rate 0.08 (occurring 3 times), tree depth 6 (occurring 4 times), and number of decision trees 100 (occurring 5 times). A weighted average parameter fusion algorithm is used to determine the final model parameters (learning rate 0.08, tree depth 6, number of decision trees 100, minimum number of sample splits 20). Substitute the final parameters into the model structure and evaluate performance using a test subset: prediction accuracy (efficiency improvement 92%, violation risk reduction 89%, cost control 90%), recall (88%, 85%, 87%), and F1 score (0.90, 0.87, 0.88). Generate a model performance evaluation report. If the evaluation report meets the preset standards (accuracy ≥ 85%, F1 score ≥ 0.85), output the predicted model; if not, adjust the number of decision trees to 120 and the tree depth to 7, then retrain until the performance meets the standards.

[0056] Furthermore, Embodiment 2 of the present invention also provides a used car market mixed traffic control system based on multi-target recognition, used to execute the used car market mixed traffic control method based on multi-target recognition as described above. This used car market mixed traffic control system based on multi-target recognition includes a multi-source recognition device integration layer, a data processing layer, an identity recognition layer, a strategy decision layer, a control execution layer, and a data feedback layer. Each layer interacts through a standardized data interface. The multi-source recognition device integration layer includes a license plate reader, an RFID reader, a QR code scanner, and an environmental perception sensor, used to collect multi-dimensional vehicle identification raw data and perform preliminary preprocessing. The data processing layer is used to process multi-target recognition data... The system performs calibration, noise reduction, and standardization on the raw data to generate a standardized recognition dataset. The identity recognition layer extracts vehicle identity feature vectors using a multimodal feature fusion algorithm, combines them with an identity information database to complete vehicle identity recognition, and generates identity recognition results and confidence parameters. The strategy decision layer calls the dynamic access control engine, combines the gate's operating status and management rule base, and generates access permission evaluation results and strategy adaptation coefficients. The control execution layer outputs gate control commands, generates access status record data, and triggers an abnormal vehicle management intervention mechanism. The data feedback layer collects data from the entire process to provide data support for recognition model optimization and access control strategy iteration.

[0057] In this embodiment of the invention, the multi-target recognition-based mixed-traffic control system for the used car market includes a multi-source recognition device integration layer, a data processing layer, an identity recognition layer, a strategy decision-making layer, a control execution layer, and a data feedback layer. Each layer interacts through a RESTful standardized data interface (transmission latency ≤20ms). The multi-source recognition device integration layer includes a license plate reader, an RFID reader, a QR code scanner, and environmental sensing sensors, installed at the entrance and exit of the barrier gate. It collects high-definition images of vehicle license plates, RFID identification tags, QR code encoding, and multi-dimensional raw data such as light intensity and weather conditions. The device's built-in preprocessing module performs image noise reduction and data format conversion to generate pre-processed raw data. The data processing layer receives the raw data, uses an image calibration algorithm to correct license plate image distortion, filters and reduces RFID signal interference, and standardizes all data (unifying the encoding format and normalizing the numerical range) to generate a standardized recognition dataset, which is then pushed to the identity recognition layer. The identity recognition layer invokes a multimodal feature fusion algorithm to extract license plate feature vectors, RFID identity feature vectors, and QR code feature vectors from a standardized dataset. Based on modal confidence weights (0.45, 0.32, 0.23), a weighted fusion is performed to generate a vehicle identity feature vector. This vector is then matched with the used car market vehicle identity information database (containing feature data from 5000 registered vehicles) using cosine similarity to generate an identity recognition result (e.g., matching the registered vehicle ID "BA20240420005") and a confidence parameter (91.7%). The strategy decision layer invokes a dynamic access control engine, linking it to the used car market access control rule base (including classification rules for new cars, private cars, and financially locked cars). Combining the gate operation status data (traffic flow, equipment load) with the identity recognition result, a multi-factor comprehensive evaluation is used to generate an access control assessment result (allowed passage, inspection-free release) and a strategy adaptation coefficient (0.96). The control execution layer receives the strategy decision results, outputs the gate lifting control command (lifting height 1.8 meters, lifting time 2 seconds), and records the passage status data (passage time, vehicle ID, strategy type). If an abnormal vehicle is identified (such as a vehicle locked by a financial institution or a vehicle evading tolls), the management intervention mechanism is triggered (a warning message is pushed to the market management terminal, and the gate remains closed). The data feedback layer collects data from the entire process (equipment-collected data, recognition results, strategy execution effects, and manual correction records), summarizes and generates data reports daily, and sends them back to the data processing layer and the identity recognition layer through interfaces. This provides data support for incremental training of the recognition model and optimization of passage strategy parameters, realizing closed-loop iteration of the system.

[0058] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0059] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for controlling mixed traffic in a used car market based on multi-target recognition, characterized in that, Includes the following steps: In response to the activation signal from the barrier gate triggering device, the license plate reader, RFID reader, QR code scanner, and environmental sensing sensor are activated through the multi-source identification device integration layer to collect multi-dimensional identification raw data of vehicles passing through the used car market barrier gate. This multi-dimensional identification raw data includes vehicle license plate image data, RFID tag storage information, QR code encoded data, and environmental parameters of the passage scenario. The multi-dimensional identification raw data undergoes cross-device data calibration processing. A device timing synchronization algorithm aligns the collection timestamps of different identification devices, a noise filtering model removes environmental interference data, and unstructured image data and encoded data undergo format standardization conversion to obtain a standardized identification dataset with unified dimensions. Based on this standardized identification dataset, a multi-modal identity fusion identification module extracts vehicle identity features. The system uses a feature vector, combined with a used car market vehicle identity information database, to perform feature matching and generate vehicle identity recognition results and identity confidence parameters. Based on these results, a dynamic access control engine is invoked, and combined with the current gate operation status parameters and the market access management rule base, to generate vehicle access permission assessment results and strategy adaptation coefficients. According to these results, the gate control decision layer outputs gate action control commands and simultaneously generates vehicle access status record data. This data is then categorized and analyzed; for vehicles with abnormal access types, a corresponding management intervention notification mechanism is triggered, generating an intervention processing work order and pushing it to the market management terminal. Simultaneously, the entire process data is fed back to the system optimization module for iterative updates to the recognition model and access control strategy.

2. The method for controlling mixed traffic in the used car market based on multi-target recognition according to claim 1, characterized in that, The process of activating the license plate recognition device, RFID reader, QR code scanner, and environmental sensing sensor through the multi-source identification device integration layer to collect multi-dimensional identification raw data of vehicles passing through the used car market gate includes the following steps: Calling the device status monitoring interface to monitor the operating status parameters of the license plate recognition device, RFID reader, QR code scanner, and environmental sensing sensor in real time, and generating a device availability assessment result; based on the device availability assessment result, assigning data acquisition priorities and acquisition timing windows to each identification device, and generating a device acquisition scheduling plan; according to the device acquisition scheduling plan, activating the acquisition function of available identification devices, capturing high-definition images of vehicle license plates and image acquisition angle parameters through the license plate recognition device, and transmitting the images via RFI. The D-reader reads the identification information and tag signal strength data from the vehicle's RFID tag, obtains the encoding information and scanning distance parameters of the vehicle's associated QR code through the QR code scanner, and collects data on light intensity, obstruction status, and weather conditions at the gate through environmental perception sensors to generate vehicle license plate image data, RFID tag storage information, QR code encoding data, and traffic scene environmental parameters. Device identifiers, collection timestamps, and collection environment parameter tags are added to the raw data collected by each device to generate a multi-source raw data set with metadata tags. Based on the data source type of the multi-source raw data set, data integrity verification rules are established, missing data is marked and the reasons for the missing data are recorded, generating multi-dimensional identification raw data.

3. The method for controlling mixed traffic in the used car market based on multi-target recognition according to claim 1, characterized in that, The process of extracting vehicle identity feature vectors based on the standardized recognition dataset using a multimodal identity fusion recognition module and performing feature matching in conjunction with a used car market vehicle identity information database includes the following steps: separating standardized vehicle license plate image data, standardized RFID identity data, and standardized QR code encoding data from the standardized recognition dataset to generate a single-modal recognition data subset; performing feature extraction on each of the single-modal recognition data subsets; extracting character structure features, license plate color features, and texture features from the standardized license plate image data using a preset license plate feature extraction model to generate a license plate feature vector; and extracting the vehicle's unique identifier, ownership information, and tag binding time features from the standardized RFID identity data using a preset RFID data parsing model to generate an RFID identity feature vector; from... Vehicle-related business codes and access authorization identifiers are extracted from standardized QR code encoding data to generate a QR code feature vector. A multimodal feature fusion algorithm is then invoked to weight and fuse the license plate feature vector, RFID identity feature vector, and QR code feature vector, calculating the confidence weight of each modality feature to generate a fused vehicle identity feature vector. Based on this vehicle identity feature vector, similarity calculation is performed with the registered identity feature data in the used car market vehicle identity information database to generate a feature matching similarity matrix. According to the feature matching similarity matrix and a preset identity recognition threshold parameter, the vehicle identity matching result is determined. Simultaneously, based on the confidence weight and similarity value of each modality feature, the confidence parameter of the identity recognition result is calculated, generating the vehicle identity recognition result and identity confidence parameter.

4. The method for controlling mixed traffic in the used car market based on multi-target recognition according to claim 3, characterized in that, The process of invoking a multimodal feature fusion algorithm to weight and fuse the license plate feature vector, RFID identity feature vector, and QR code feature vector, and calculating the credibility weight of each modality feature, includes the following steps: Normalizing the dimensions of the license plate feature vector, RFID identity feature vector, and QR code feature vector to unify the dimensional space of the feature vectors and generate a standardized single-modal feature vector set; based on the standardized single-modal feature vector set, extracting the information entropy and feature discrimination parameters of each modality feature to generate a modality feature quality assessment dataset; and invoking a modality weight allocation model, combining the modality feature quality assessment dataset with the equipment acquisition environment parameters, to analyze the credibility weight of each modality feature. To assess the reliability of features in the current scenario, the initial credibility weights of each modal feature are calculated. A feature correlation analysis algorithm is used to calculate the linear correlation coefficient and semantic association degree between different modal feature vectors, generating an inter-modal association matrix. Based on this inter-modal association matrix, the initial credibility weights are iteratively corrected, reducing weight redundancy in highly correlated modalities and increasing the weight proportion of complementary modalities, generating an optimized set of modal credibility weights. A weighted summation fusion strategy is employed to perform matrix operations on the standardized single-modal feature vector set and the corresponding optimized modal credibility weight set, generating a fused vehicle identity feature vector, while simultaneously recording the weight allocation log during the fusion process.

5. The method for controlling mixed traffic in the used car market based on multi-target recognition according to claim 4, characterized in that, The iterative correction of the initial confidence weights based on the inter-modal correlation matrix, weakening the weight redundancy of highly correlated modes and strengthening the weight proportion of complementary modes, includes the following steps: performing eigenvalue decomposition on the inter-modal correlation matrix, extracting eigenvalues ​​and eigenvectors of the inter-modal correlation matrix, and determining the principal component direction and correlation strength threshold of the modal correlation; based on the eigenvalues ​​and eigenvectors, identifying highly correlated mode combinations and complementary mode combinations, and generating modal correlation classification results; for highly correlated mode combinations, calculating the redundancy coefficient of each mode within the mode combination, and reducing the weight proportion of each mode proportionally based on the redundancy coefficient. The initial confidence weights of the modalities are generated, and intermediate weights are generated after redundancy correction. For complementary modal combinations, the information gain value of each modality within the complementary modal combination is calculated, and the intermediate weights of each modality are increased proportionally based on the information gain value to generate a complementary enhanced weight set. The weight normalization function is called to map the complementary enhanced weight set to a preset weight interval to ensure that the sum of the confidence weights of each modality is a preset value, generating a preliminary optimized weight set. Based on historical fusion effect feedback data, a weight correction model is constructed to calibrate the error of the preliminary optimized weight set, generating an optimized modal confidence weight set.

6. The method for controlling mixed traffic in the used car market based on multi-target recognition according to claim 1, characterized in that, The process of generating vehicle access permission assessment results and strategy adaptation coefficients based on the vehicle identification results and identity confidence parameters, by calling the dynamic access strategy engine and combining the current gate operation status parameters with the market access management rule base, includes the following steps: parsing the vehicle identification results to extract vehicle type identifier, entity type, business association status, and historical access record information to generate a vehicle identity attribute dataset; based on the vehicle identity attribute dataset, associating it with the classification management rules in the used car market access management rule base to determine the applicable basic access strategy type and generate a strategy matching candidate set; and collecting the current gate operation status parameters, including the current gate opening / closing status, traffic flow statistics, etc. Equipment load parameters and fault warning information are used to generate a barrier gate operation status dataset. Combining the vehicle identity attribute dataset, barrier gate operation status dataset, and identity confidence parameters, the adaptability of the basic access strategy in the current scenario is analyzed, and the expected effect parameters of strategy execution are calculated. Based on the expected effect parameters of strategy execution, the basic access strategies in the strategy matching candidate set are prioritized, and the ranking results are dynamically adjusted based on real-time market management demand parameters to generate the optimal access strategy scheme. The core decision rules and execution conditions of the optimal access strategy scheme are extracted to generate vehicle access permission evaluation results. Simultaneously, based on the strategy adaptability analysis results and the satisfaction of execution conditions, the strategy adaptability coefficient is calculated.

7. The method for controlling mixed traffic in the used car market based on multi-target recognition according to claim 6, characterized in that, The process of analyzing the adaptability of the basic access control strategy in the current scenario and calculating the expected performance parameters of the strategy, by combining the vehicle identity attribute dataset, the barrier gate operation status dataset, and identity confidence parameters, includes the following steps: Constructing a strategy adaptation evaluation index system, which includes access efficiency indicators, security control indicators, management compliance indicators, and user experience indicators, and assigning scenario adaptation weights to each indicator; Based on the vehicle identity attribute dataset, extracting key attribute factors affecting the adaptability of the access control strategy, including vehicle access priority, business urgency, and historical violation records, and generating an attribute influence factor set; and Combining the barrier gate operation status dataset, analyzing the constraints of the current access scenario. The process involves several steps: first, generating a set of scenario constraint parameters, including traffic capacity, equipment operating limits, and channel occupancy status; second, invoking a multi-factor comprehensive evaluation algorithm to substitute the attribute influence factor set, scenario constraint parameter set, and identity confidence parameter into the strategy adaptation evaluation index system, calculating the score of each basic access strategy under each evaluation index; third, calculating the comprehensive adaptation score of each basic access strategy using a weighted summation method based on the scores of each evaluation index and the scenario adaptation weight; and fourth, invoking a preset effect prediction model and combining the comprehensive adaptation score with historical strategy execution effect data to predict the traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate after the implementation of each basic access strategy, generating expected strategy execution effect parameters.

8. The method for controlling mixed traffic in the used car market based on multi-target recognition according to claim 7, characterized in that, The process of calling a preset effect prediction model, combining the comprehensive adaptation score with historical strategy execution effect data, to predict the improvement rate of passage efficiency, the reduction rate of violation risk, and the control rate of management cost after the implementation of each basic passage strategy, and generating expected effect parameters for strategy execution includes the following steps: collecting historical passage strategy execution data, including execution records of each basic passage strategy in different scenarios, corresponding comprehensive adaptation scores, and actual execution effect indicators, to generate a historical strategy effect dataset; performing data preprocessing on the historical strategy effect dataset, removing abnormal data and invalid records, and unifying indicator dimensions through data normalization to generate a standardized historical dataset; and constructing a training sample set for the effect prediction model based on the standardized historical dataset. The model uses comprehensive adaptation score, vehicle identity attribute features, and scenario constraint parameters as input features, and traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate as output labels. A machine learning algorithm is used to train the model on the training sample set, and the model parameters are optimized through cross-validation to generate an effect prediction model. The comprehensive adaptation score of the current basic traffic strategy, the corresponding set of vehicle identity attribute influencing factors, and the set of scenario constraint parameters are input into the effect prediction model, which outputs predicted values ​​for traffic efficiency improvement rate, violation risk reduction rate, and management cost control rate for each basic traffic strategy. The reliability of these predicted values ​​is verified, and corrections are made based on the model prediction error rate to generate expected effect parameters for strategy execution.

9. The method for controlling mixed traffic in the used car market based on multi-target recognition according to claim 8, characterized in that, The process of training the model on the training sample set using machine learning algorithms and optimizing the model parameters through cross-validation includes the following steps: dividing the training sample set into a training subset, a validation subset, and a test subset, and setting the data proportion and data distribution characteristics of each subset; initializing the network structure and hyperparameters of the performance prediction model, including the number of hidden layer neurons, the initial learning rate, and the iteration threshold; iteratively training the model using the gradient descent optimization algorithm based on the training subset, calculating the loss function value for each iteration, and adjusting the model parameters using the backpropagation algorithm; during each iteration, inputting the validation subset into the model for the current training stage, calculating the validation set prediction error, and determining when the validation set prediction error increases consecutively multiple times. The system triggers an early stopping mechanism to halt model training; it then invokes a cross-validation algorithm to divide the training sample set into multiple mutually exclusive subsets for multiple rounds of model training and validation, recording the optimal model parameters and corresponding prediction errors for each round; based on the results of multiple rounds of cross-validation, it statistically analyzes the optimal value distribution of each model parameter and uses a parameter fusion algorithm to determine the final model parameters; it substitutes the final model parameters into the model structure and evaluates the model's performance using a test subset, calculating the model's prediction accuracy, recall, and F1 score, and generating a model performance evaluation report; when the model performance evaluation report meets preset standards, it outputs the performance prediction model; otherwise, it adjusts the model network structure and retrains until a satisfactory performance prediction model is generated.

10. A mixed-traffic control system for a used car market based on multi-target recognition, characterized in that, The system is used to implement the multi-target recognition-based mixed-traffic control method for the used car market as described in claim 1. The multi-target recognition-based mixed-traffic control system includes a multi-source recognition device integration layer, a data processing layer, an identity recognition layer, a strategy decision-making layer, a control execution layer, and a data feedback layer. Each layer interacts through a standardized data interface. The multi-source recognition device integration layer includes a license plate reader, an RFID reader, a QR code scanner, and environmental perception sensors, used to collect multi-dimensional vehicle identification raw data and perform preliminary preprocessing. The data processing layer is used to calibrate, reduce noise, and standardize the multi-dimensional identification raw data. The system performs a transformation to generate a standardized recognition dataset. The identity recognition layer extracts vehicle identity feature vectors using a multimodal feature fusion algorithm, combines them with an identity information database to complete vehicle identity recognition, and generates identity recognition results and confidence parameters. The strategy decision layer calls a dynamic access control engine, combines the gate's operating status and management rule base, and generates access permission evaluation results and strategy adaptation coefficients. The control execution layer outputs gate control commands, generates access status record data, and triggers an abnormal vehicle management intervention mechanism. The data feedback layer collects data from the entire process, providing data support for recognition model optimization and access control strategy iteration.

Images