Comprehensive judgment method for highway toll vehicle model

Abstract

A comprehensive judgment method for highway toll vehicle types relates to the technical field of highway toll, and comprises the following steps: constructing a standard vehicle type database and collecting vehicle data when a vehicle enters an entrance of a toll station; collecting real-time environment monitoring data at the entrance of the toll station; constructing a multi-layer neural network classification model; inputting the comprehensive feature vector into a trained multilayer neural network classification model, obtaining the predicted confidence coefficient of each vehicle category, and selecting the category with the highest confidence coefficient after calibration as a preliminary classification result; determining a final classification result; and matching the final classification result with the vehicle type information in the standard vehicle type database. According to the scheme, the accuracy of vehicle recognition is remarkably improved, and the necessity of manual intervention is reduced especially under the condition that the vehicle speed is high or the shooting angle is complex.

Description

Technical Field

[0001] The invention relates to the technical field of highway toll collection, and in particular to a comprehensive judgment method for a highway toll collection vehicle type. Background Art

[0002] With the rapid development of the transportation industry, the number of vehicles passing through highways continues to increase, and traditional vehicle classification and charging methods face huge challenges. Existing vehicle model recognition technology mainly relies on manual intervention and simple image processing methods, resulting in low vehicle classification accuracy, insufficient recognition efficiency, inaccurate charging standards and other problems. In addition, traditional methods have weak adaptability to different vehicle models and cannot effectively cope with the actual situation of high speed and large traffic volume, which not only increases the workload of toll stations, but also leads to longer waiting time for drivers to pass, affecting the overall traffic flow.

[0003] Publication No. CN118470811A discloses a highway toll collection method, system, device and medium based on vehicle type recognition, including: constructing a vehicle type library system, the vehicle type library system includes license plate information, license plate color information, vehicle type and percentage information, axle and percentage information and vehicle type information of various types of vehicles; when a target vehicle is detected entering a toll station, an identification system identifies first information of the target vehicle, the first information includes license plate information, license plate color information and vehicle type information, and sends the first information to the vehicle type library system; the vehicle type library system determines whether the license plate information, license plate color information and vehicle type information in the vehicle type library match the first information of the target vehicle; if so, the identification system sends the first information of the target vehicle to the toll collection system, so that the toll collection system charges according to the first information of the target vehicle;

[0004] If not, the vehicle model library system determines whether the vehicle type information of the target vehicle is special vehicle type information; if so, the vehicle model library system feeds back the license plate information, license plate color information, vehicle model and percentage information, axle and percentage and vehicle type information of the target vehicle to the recognition system, so that the recognition system re-recognizes the second information of the target vehicle and sends the second information to the charging system, the second information includes the license plate information, license plate color information, vehicle model information, axle information and special vehicle type information of the target vehicle, so that the charging system charges according to the second information; if not, the vehicle model library system feeds back the license plate information, license plate color information, vehicle model information, axle information and vehicle type information of the target vehicle to the recognition system, so that the recognition system sends the license plate information, license plate color information, vehicle model information, axle information and vehicle type information of the target vehicle to the charging system for billing.

[0005] The deficiencies of the above-mentioned existing technologies are mainly reflected in the following aspects: First, the existing technologies rely on traditional vehicle model libraries and basic image recognition technology, and generally have the problem of low recognition accuracy. This is mainly due to factors such as poor image shooting conditions and lagging algorithm updates, which lead to frequent misrecognition and require manual intervention for correction. Second, the existing vehicle model library system lacks a flexible update mechanism and cannot quickly adapt to the emergence of new models or environmental changes, resulting in poor performance of the recognition system on new models. Third, when dealing with abnormal situations, the existing technology relies on manual feedback and manual recognition, resulting in slow processing speed and low efficiency, which may cause congestion at toll stations. Fourth, when facing a variety of vehicles and complex traffic conditions, the existing technology lacks a reliable exception handling mechanism, which can easily lead to toll errors and affect the stability of the system. Summary of the invention

[0006] The purpose of the present invention is to provide a learning evaluation system and method based on big data to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A comprehensive judgment method for highway toll vehicle types includes the following steps:

[0009] Step 1: Build a standard vehicle model database. The standard vehicle model database stores the vehicle data and vehicle categories of all vehicle models that have passed through the toll station in the past. The vehicle data of vehicles entering the toll station entrance is collected in real time and integrated into a comprehensive feature vector. The vehicle data includes the vehicle's weight, length, width, number of axles, and height.

[0010] Step 2: Collect real-time environmental monitoring data at the entrance of the toll station, and modify the real-time collected environmental monitoring data according to the historical environmental monitoring data. The environmental monitoring data includes environmental temperature, humidity and wind speed data;

[0011] Step 3: Construct a multi-layer neural network classification model, use the vehicle data of the vehicle types that have passed through the toll station in the past as a training set, use the vehicle category of the corresponding vehicle as a label, train the multi-layer neural network classification model, and obtain the trained multi-layer neural network classification model;

[0012] Step 4: Input the comprehensive feature vector into the trained multi-layer neural network classification model to obtain the confidence of each predicted vehicle category, and select the category with the highest confidence after calibration as the preliminary classification result;

[0013] Step 5: If the confidence of the preliminary classification result is lower than the set threshold, the adaptive classification process is started, the vehicle data is corrected according to the corrected environmental monitoring data, and the classification result is secondary classified using the support vector machine based on the corrected vehicle data. The final classification result is obtained by weighted fusion of the preliminary classification result of MLP and the secondary classification result of SVM. Otherwise, the preliminary classification result is directly used as the final classification result.

[0014] Step 6: Match the final classification result with the vehicle model information in the standard vehicle model database. If the match is successful, perform the corresponding charging operation based on the vehicle model category. If the match fails, record the abnormal event and transfer to manual judgment of the vehicle model.

[0015] Furthermore, for cars, SUVs, trucks and other models, the standard model database stores detailed information on all models that have passed through the toll booths in the past, including the physical parameters of the vehicle and the corresponding model category. Each model record contains a unique model number, model name and detailed parameter information. The vehicle's own data is fused into a comprehensive feature vector V, which represents the main physical characteristics and appearance information of the vehicle. The expression of the feature vector is:

[0016] V = [L, W, H, M, A]

[0017] Among them, L is the length of the vehicle, W is the width, H is the height, M is the weight, and A is the number of axles;

[0018] The fused feature vector V is stored in the local database, and these data are arranged in the order of time when the vehicles enter the toll station.

[0019] Furthermore, the temperature T of the surrounding environment is collected in real time. current 、Humidity current and wind speed W current Based on the environmental monitoring data and historical data, historical data refers to the data collected by the same sensor in the previous 24 hours under similar environmental conditions. The average of various data collected in the previous 24 hours is calculated as the reference data for correction. Based on the difference between historical and current environmental data, the temperature correction factor F is calculated. T , humidity correction factor F H and wind speed correction factor F W .

[0020] Furthermore, the generated feature vector V is input into the multi-layer neural network classification model, and the historical vehicle data of the marked passing vehicles is extracted from the database. The historical vehicle data includes: the historical vehicle length L i , historical vehicle width W i , historical vehicle height H i , historical vehicle weight Mi , historical vehicle axle number A i , the historical features of the same passing vehicles are combined into a feature vector V i , the formula is:

[0021] V i =[L i , W i , H i 、M i , A i ]

[0022] For the class label of known models, each feature vector V i Need to correspond to a label y i ,y i Indicates the vehicle model corresponding to the feature vector of the i-th historical data, label y i The formula is:

[0023] y i ∈{SUB,SUA,SUC,SUD}

[0024] Among them, SUB is a sedan model, SUV is an off-road vehicle model, SUC is a truck model, and SUD is other models;

[0025] Use the cross entropy loss function as the optimization target to calculate the gap between the model's predicted label and the true label. The formula is:

[0026]

[0027] Among them, N is the number of samples, C is the number of categories, and y ij is the jth true label corresponding to the i-th feature vector, Confidence of predictions for multi-layer neural network classification models;

[0028] Use historical data and labels to train the model, update the weights, and perform multiple iterations until the model reaches the preset convergence standard, that is, the change in the loss function is less than the set threshold, which is 0.8. After the model training is completed, use the training set for prediction, calculate the confidence of each category, and select the category with the highest confidence as the preliminary classification result. The specific operation of classification is achieved through the following formula:

[0029]

[0030] in, is the preliminary classification result, and f(V) is the confidence vector output by the neural network.

[0031] Furthermore, if the confidence of the preliminary classification result is higher than the set threshold value of 0.6, the preliminary classification result will be matched with the vehicle model information in the standard vehicle model database. After the match is successful, the vehicle model information will be fed back to the charging system, and the corresponding charging operation will be performed according to the vehicle model category. If the confidence of the preliminary classification result is lower than the set threshold value of 0.6, the adaptive classification process will be started.

[0032] Furthermore, for each collected physical parameter of the vehicle, a correction factor is used for correction, and the dynamic correction formula is:

[0033] S corrected =S raw ·(1+α T ·F T +α H ·F H +α W ·F W )

[0034] Among them, α T , α H and α W are the weight coefficients of the impact of temperature, humidity and wind speed on sensor data, and 0.3<<α T <<0.7,0.1<<α H <<0.4,0.1<<α W <<0.4,S raw is a physical parameter collected, S corrected is a physical parameter after correction;

[0035] The corrected physical parameters of each vehicle are used as the training set, and the vehicle category of the corresponding vehicle is used as the label. The SVM model is trained using the above training set, and the decision value de between the output corrected label and the historical label is calculated. The confidence between the corrected label and the historical label is calculated. The formula is:

[0036]

[0037] Where: P(Y corrected / Y historical ) represents the confidence between the revised label and the historical label, ranging from [0, 1];

[0038] The environmental monitoring data is repeatedly used to calibrate the collected vehicle data so that the confidence between the calibrated label and the historical label is greater than 0.6, and then the corrected label replaces the previous label.

[0039] Furthermore, the final classification result is obtained by weighted fusion of the preliminary classification result of MLP and the secondary classification result of SVM. The fusion formula is:

[0040] C final =λ·C MLP +(1―λ)·f SVM (V)

[0041] Among them, C final is the final classification result, C MLP is the preliminary classification result, f SVM (V) is the secondary classification result, λ is the weight coefficient, that is, 0.4≤λ≤0.8.

[0042] Furthermore, a hash table is constructed to store the feature vector information of the standard vehicle model database. Each vehicle model is used as the key of the hash table, and the value of the hash table contains the feature vector of the vehicle model. final Match the vehicle model information in the standard vehicle model database, calculate the matching degree between the vehicle feature vector and the parameters of each vehicle model in the standard vehicle model database, and use the hash function for fast query. The formula is:

[0043] h(V)=hash(V)

[0044] Where h(V) is the hash value of the feature vector V, and hash(V) is the hash function;

[0045] When the matching degree is higher than the threshold, the match is considered successful. The system will feed back the vehicle model information to the charging system and perform the corresponding charging operation according to the vehicle model category. If the match fails or the confidence of the classification result is lower than the threshold, the secondary recognition process is triggered. The secondary recognition is manually determined by combining the vehicle's own data to determine the vehicle's charging standard.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention introduces a multi-layer neural network for deep learning, combined with the classification ability of SVM, and this solution significantly improves the accuracy of vehicle identification, especially when the vehicle speed is fast or the shooting angle is complex, reducing the necessity of manual intervention. By integrating multi-source information, the system can correct and filter out noise in the data in real time, making the input data more accurate, and enhancing the classification accuracy and robustness. At the same time, this mechanism reduces the dependence on a single sensor, improves the response speed, and optimizes the data utilization rate, which helps to reduce hardware costs. In addition, dynamic correction and data fusion ensure that the system adapts to environmental changes, supports multi-scenario applications, improves the learning ability of the model, and promotes the improvement of the overall intelligence level. Through the continuous learning mechanism, the recognition system can be dynamically adjusted according to real-time data, and the content of the vehicle model library is updated in time, which enhances the adaptability of the system and ensures stable performance in complex environments. Through the intelligent feedback mechanism and multi-level verification of vehicle information, this solution significantly reduces the charging anomalies caused by recognition errors and improves the stability and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic diagram of the overall method flow of the present invention. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0050] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0051] Example:

[0052] See also Figure 1 , the present invention provides a technical solution:

[0053] A comprehensive judgment method for highway toll vehicle types includes the following steps:

[0054] Step 1: Build a standard vehicle model database. The standard vehicle model database stores detailed information of all known vehicle models, including the vehicle's weight M, length L, width W, number of axles A, height H and corresponding vehicle model categories; building a standard vehicle model database facilitates matching the final classification result with the vehicle model information in the standard vehicle model database, and collects the vehicle's own data when the vehicle enters the toll station entrance. The vehicle's own data includes the vehicle's weight M, length L, width W, number of axles A, height H and corresponding vehicle model categories. All devices collect data synchronously, collect multiple sets of data, remove the highest and lowest values ​​of the same type of data in the multiple sets of data collected, and average the other data to obtain the final data used. All devices collect data at the same time and average the multiple sets of data to improve the accuracy of the collected data and reduce errors. The vehicle's own data is fused into a comprehensive feature vector. Fusion into a comprehensive feature vector is to integrate multi-dimensional data from different sensors to create a unified and accurate vehicle feature representation. By eliminating the deviation and inconsistency between sensors, the comprehensive feature vector V can provide a more reliable information basis, support subsequent vehicle model classification and matching algorithms, and improve the accuracy and efficiency of overall data processing.

[0055] Step 2, collect the temperature, humidity, and wind speed of the sensor's surrounding environment in real time, based on environmental monitoring data and historical data. Historical data refers to data collected by the same sensor within the previous 24 hours under similar environmental conditions. These data will be used to evaluate and correct the working status of the sensor under different environmental conditions to ensure the accuracy of the measurement data. Correct the current sensor data; real-time environmental monitoring can help identify the impact of environmental conditions on sensor data, such as changes in temperature and humidity may cause laser measurement errors. By monitoring these factors, sensor data can be dynamically corrected to ensure data accuracy and consistency, providing a reliable basis for subsequent processing;

[0056] Step 3, build a multi-layer neural network classification model, use the historical features of the marked passing vehicles as the training set, the historical features include historical vehicle length, historical vehicle width, historical vehicle height, historical vehicle weight and historical vehicle axle number, use the corresponding vehicle model as a label, use the historical features and labels to train the model, and obtain the trained multi-layer neural network classification model. The multi-layer neural network is MLP. The multi-layer neural network classification model has a strong nonlinear mapping ability, can effectively capture complex feature relationships, has strong adaptability, and can handle high-dimensional and diverse data. MLP can extract and learn feature hierarchies through multiple hidden layers to achieve deeper feature abstraction, and provide higher classification accuracy than traditional linear classification models. In addition, MLP has self-learning ability, continuously optimizes model parameters through the back-propagation algorithm, can perform adaptive training on different data sets, improve the generalization ability and robustness of the model, and thus show better performance in practical applications.

[0057] Step 4: After the model training is completed, the generated feature vector is input into the multi-layer neural network classification model. The neural network calculates the confidence of each vehicle type category through forward propagation and selects the category with the highest confidence as the preliminary classification result.

[0058] Step 5 If the confidence of the preliminary classification result is lower than the set threshold, the adaptive classification process is started, the vehicle data is corrected according to the corrected environmental monitoring data, and the classification result is secondary classified using the support vector machine based on the corrected vehicle data. The final classification result is obtained by weighted fusion of the preliminary classification result of MLP and the secondary classification result of SVM, otherwise the preliminary classification result is directly used as the final classification result; by fusing the results of the MLP and SVM models, the advantages of the two models can be fully utilized to improve the accuracy of classification. MLP is good at processing high-dimensional data, while SVM performs well in small sample conditions. Through weighted fusion, the robustness of the classification results can be improved, especially in boundary conditions, which can effectively reduce the risk of misclassification;

[0059] Step 6, match the final classification result with the vehicle model information in the standard vehicle model database. If the match is successful, perform the corresponding charging operation based on the vehicle model category. If the match fails, record the abnormal event and transfer to manual judgment of the vehicle model.

[0060] Furthermore, weighbridges, laser scanners and high-definition cameras are installed at the entrances of toll booths. When installed, the weighbridge should be placed on the normal driving path of the vehicle to accurately measure the vehicle weight. The vehicle weights of different types of vehicles are generally different. The following are the results: Light cars: Empty weight: approximately between 1,000 kg and 1,500 kg (approximately 2,200 to 3,300 pounds). Gross weight (including passengers and luggage): usually between 1,500 kg and 2,000 kg (approximately 3,300 to 4,400 pounds). SUVs and crossovers: Empty weight: approximately between 1,500 kg and 2,500 kg (approximately 3,300 to 5,500 pounds). Gross weight: usually between 2,000 kg and 3,000 kg (approximately 4,400 to 6,600 pounds). Trucks: Light trucks (such as minivans): Unladen weight: Approximately between 1,500 kg and 3,500 kg (about 3,300 to 7,700 pounds). Gross weight: Usually between 3,500 kg and 5,000 kg (about 7,700 to 11,000 pounds). Medium trucks: Unladen weight: Approximately between 3,500 kg and 7,500 kg (about 7,700 to 16,500 pounds). Gross weight: Usually between 7,500 kg and 12,000 kg (about 16,500 to 26,500 pounds). Heavy trucks (such as 18-wheel trucks): Unladen weight: Approximately between 7,500 kg and 15,000 kg (about 16,500 to 33,000 pounds). Gross weight: Usually between 20,000 kg and 40,000 kg (approximately 44,000 to 88,000 pounds), so measuring the weight of the vehicle can not only determine the approximate model based on the weight of the vehicle, but also determine whether the vehicle is overloaded. The laser scanner should be located on the side and top of the vehicle to ensure that the full length, width and height of the vehicle are covered. The high-definition camera should be installed in the direction facing the vehicle to ensure clear capture of the vehicle image to facilitate the acquisition of the vehicle model category. When the vehicle enters the entrance of the toll station, the vehicle's weight M, length L, width W, number of axles A and height H are collected synchronously, and the corresponding vehicle model category is captured through the camera, and the laser scanner and weighbridge are calibrated using standard measuring objects. By repeatedly measuring calibration objects of known size and weight, adjusting the initial settings of the equipment to ensure that the measurement error is within the allowable range (±0.5%), encrypting the data and quickly transmitting it to the data processing unit through the internal LAN. During the data transmission process, it is encrypted by the AES-256 symmetric encryption algorithm to ensure data security. The data processing unit performs weighted moving average (WMA stands for weighted moving average) processing on the raw data collected by the sensor to remove noise. The reason for choosing the WMA algorithm is that it can smooth data fluctuations while retaining important trend information. The formula of WMA is as follows:

[0061]

[0062] in, Represents the data value after filtering, X t―i is the original data value at the ti moment, w i is the weight coefficient, N is the window size;

[0063] Calculate the data length of the filtered data value and ensure that the length is a multiple of 128 bits. If the condition is not met, add 0 to the end of the data to make it a multiple of 128 bits.

[0064] Split the filtered data value whose length has been confirmed into multiple 128-bit blocks;

[0065] Use a predefined 256-bit key as the key for the AES-256 algorithm;

[0066] Generate a random 128-bit initialization vector as the initial state of the AES-256 algorithm;

[0067] For each 128-bit data block, the data block is XORed with the ciphertext of the previous data block (IV is used initially), and then the XORed data block is encrypted using the AES-256 algorithm and a predefined key. The encryption operation will generate a new ciphertext block; the newly generated ciphertext block is then used as the encrypted data value.

[0068] The AES encryption algorithm has the advantages of high security, fast performance, flexibility, wide application and good compatibility. It is widely recognized as a safe and reliable encryption standard, supports multiple encryption modes, and is suitable for various platforms and devices. Whether it is network communication, database encryption or file encryption, AES can provide reliable protection. Its open standard makes it widely supported and implemented in various programming languages ​​and operating systems. Therefore, the AES encryption algorithm is one of the most commonly used symmetric encryption algorithms.

[0069] Furthermore, vehicle categories include sedans, SUAs, trucks and other models. The standard vehicle model database stores detailed information of all known vehicle models, including the physical parameters of the vehicle and the corresponding vehicle model category. Each vehicle model record contains a unique vehicle model number (ID), vehicle model name and detailed parameter information. The vehicle data itself is integrated into a comprehensive feature vector V, which represents the main physical characteristics and appearance information of the vehicle. The expression of the feature vector is:

[0070] V = [L, W, H, M, A]

[0071] Where L is the length of the vehicle, W is the width, H is the height, M is the weight, and A is the number of axles;

[0072] The fused feature vector V is stored in the local database, and the data is arranged in the order of the time when the vehicle enters the toll station to ensure the orderly management and efficient retrieval of vehicle data. This time-series storage method makes subsequent analysis and query more convenient, and can quickly track and process vehicle information within a specific time period. In addition, through time-series arrangement, the system can facilitate traffic monitoring, toll statistics and anomaly detection, thereby optimizing toll management and traffic smoothness.

[0073] Furthermore, the temperature T of the surrounding environment is collected in real time. current 、Humidity current and wind speed W current Based on the environmental monitoring data and historical data, historical data refers to the data collected by the same sensor in the previous 24 hours under similar environmental conditions. The average of various data collected in the previous 24 hours is calculated as the reference data for correction. Based on the difference between historical and current environmental data, the temperature correction factor F is calculated. T , humidity correction factor F H and wind speed correction factor F W .

[0074]

[0075] Among them, T historical , H historical and W historical It indicates the average values ​​of temperature T, humidity H and wind speed W collected in the previous 24 hours;

[0076] Real-time environmental monitoring can help identify the impact of environmental conditions on sensor data, such as temperature and humidity changes that may cause laser measurement errors. By monitoring these factors, sensor data can be dynamically corrected to ensure data accuracy and consistency, providing a reliable basis for subsequent processing.

[0077] Furthermore, the generated feature vector V is input into the multi-layer neural network classification model to extract the marked historical features of passing vehicles from the database. The historical features include: historical vehicle length L i , historical vehicle width W i , historical vehicle height H i , historical vehicle weight M i , historical vehicle axle number A i , the historical features of the same passing vehicles are combined into a feature vector V i , the formula is:

[0078] V i =[L i , W i , H i 、M i , A i ]

[0079] For the class label of known models, each feature vector V i Corresponding to a label y i , known models include: sedan (SUB), SUA, truck (SUC) and other models (SUD), y i Indicates the vehicle model corresponding to the feature vector of the i-th historical data, label y i The formula is:

[0080] y i ∈{SUB,SUA,SUC,SUD}

[0081] Among them, SUB is a sedan model, SUC is a truck model, and SUD is other models;

[0082] Select an appropriate multi-layer neural network structure, including input layer, hidden layer and output layer. The number of nodes in the input layer should be the same as the dimension of the feature vector, and the number of nodes in the output layer should be the same as the number of vehicle types.

[0083] Use the cross entropy loss function as the optimization target to calculate the gap between the model's predicted label and the true label. The formula is:

[0084]

[0085] Among them, N is the number of samples, C is the number of categories, and y ij is the jth true label corresponding to the i-th feature vector, Confidence of predictions for multi-layer neural network classification models;

[0086] The main advantage of the cross entropy loss function is that it can handle probability outputs, so that the model not only focuses on whether it is correct or not during training, but also measures the accuracy of the predicted probability. By minimizing the cross entropy loss, the model tends to give a more accurate probability distribution, thereby improving the reliability of the classification results. This means that even if the classification is correct, it will be optimized according to the confidence of the prediction, which helps to improve the performance of the model for uncertain scenarios, especially in multi-classification problems. It can significantly improve the prediction quality of the model. Using the cross entropy loss function as the optimization target can effectively calculate the gap between the model's predicted label and the true label. It not only penalizes the results of incorrect predictions, but also optimizes the results of correct predictions but low confidence, thereby pushing the model output closer to the true probability distribution. This method can accelerate the convergence process of the model, improve the robustness and generalization ability of the model, especially in classification problems, which can help the model better handle uncertainty and complex multi-category scenarios.

[0087] Use historical data and labels to train the model, update the weights, and perform multiple iterations until the model reaches the preset convergence standard, that is, the change in the loss function is less than the set threshold, which is 0.8. After the model training is completed, use the training set for prediction, calculate the confidence of each category, and select the category with the highest confidence as the preliminary classification result. The specific operation of classification is achieved through the following formula:

[0088]

[0089] in, is the preliminary classification result, and f(V) is the confidence vector output by the neural network.

[0090] The advantage of using historical data and labels to train the model and updating the weights through multiple iterations until the model reaches the preset convergence standard (i.e. the change in the loss function is less than the set threshold) is that it ensures that the model fully learns the patterns and regularities in the data, which not only improves the accuracy of the model but also effectively prevents overfitting. Gradually reducing the change in the loss function can help the model converge steadily and find the best parameter configuration, thereby improving the generalization ability of the model and making its performance on new data more stable and reliable.

[0091] Furthermore, the confidence vector output by the neural network is a vector representation provided by the network when classifying the input sample, which is usually used for multi-class classification tasks. Each element of the vector corresponds to the predicted probability of each category, reflecting the model's confidence that the input sample belongs to each category. By processing the original output of the last layer of the neural network through the activation function, each value in the generated confidence vector is between 0 and 1, and the sum of all values ​​is 1. The higher the confidence value, the more confident the model is in predicting the category. For example, if the output confidence vector is [0.1, 0.7, 0.2], it means that the model believes that the input sample has a 70% probability of belonging to the second category, a 10% probability of belonging to the first category, and a 20% probability of belonging to the third category. Such an output makes the decision-making process clearer and facilitates the selection of prediction categories.

[0092] Furthermore, if the confidence of the preliminary classification result is higher than the set threshold value of 0.6, the preliminary classification result will be matched with the vehicle model information in the standard vehicle model database. After the match is successful, the vehicle model information will be fed back to the charging system, and the corresponding charging operation will be performed according to the vehicle model category. If the confidence of the preliminary classification result is lower than the set threshold value of 0.6, the adaptive classification process will be started.

[0093] Furthermore, correction factors are used to correct the collected physical parameters of each vehicle. The purpose of the correction is to improve the accuracy and reliability of the data. Since different sensors may have measurement errors, environmental influences, or system deviations, the use of correction factors can correct these inconsistencies, thereby ensuring that the obtained physical parameters are closer to the true values. This correction process helps to eliminate noise and interference, improve the quality of subsequent data analysis and processing, ensure that the model training and decision-making process are based on accurate data, and thus improve the performance and effectiveness of the overall system. The dynamic correction formula is:

[0094] S corrected =S raw ·(1+α T ·F T +α H ·F H +α W ·F W )

[0095] Among them, α T , α H and α W are the weight coefficients of the impact of temperature, humidity and wind speed on sensor data, and 0.3<<α T <<0.7,0.1<<α H <<0.4,0.1<<α W <<0.4,S raw is a physical parameter collected, S corrected is a physical parameter after correction;

[0096] The corrected physical parameters of each vehicle are used as the training set, and the vehicle category of the corresponding vehicle is used as the label. The SVM model is trained using the above training set, and the decision value de between the output corrected label and the historical label is calculated. The confidence between the corrected label and the historical label is calculated. The formula is:

[0097]

[0098] Where: P(Y corrected / Y historical ) represents the confidence between the revised label and the historical label, ranging from [0, 1];

[0099] The environmental monitoring data is repeatedly used to calibrate the collected vehicle data so that the confidence between the calibrated label and the historical label is greater than 0.6, and then the corrected label replaces the previous label.

[0100] The purpose of using support vector machine (SVM) to perform secondary classification on the classification results is to further improve the classification accuracy and refine the decision boundary. The first classification may not be able to effectively distinguish the subtle differences between certain similar categories. Through secondary classification, SVM can use new features or adjust the decision boundary on the basis of preliminary classification to more accurately identify and distinguish these categories. In addition, secondary classification can also be applied to deal with complex problems, such as imbalanced or overlapping categories in multi-class classification tasks, thereby enhancing the robustness and generalization ability of the model and improving the overall classification effect. In this way, higher accuracy and better classification performance can be achieved in the end.

[0101] Furthermore, the final classification result is obtained by weighted fusion of the preliminary classification result of MLP and the secondary classification result of SVM. The fusion formula is:

[0102] C final =λ·C MLP +(1―λ)·f SVM (V)

[0103] Among them, C final is the final classification result, C MLP is the preliminary classification result, f SVM (V) is the secondary classification result, λ is the weight coefficient, and 0.4≤λ≤0.8 is dynamically adjusted according to the actual traffic conditions;

[0104] By fusing the results of the MLP and SVM models, we can take full advantage of the advantages of both models and improve classification accuracy. MLP is good at processing high-dimensional data, while SVM performs well in small sample conditions. Through weighted fusion, we can improve the robustness of the classification results, especially in boundary conditions, and effectively reduce the risk of misclassification.

[0105] Furthermore, a hash table is constructed to store the feature vector information of the standard vehicle model database. Each vehicle model is used as the key of the hash table, and the value of the hash table contains the feature vector of the vehicle model. An empty hash table is created to facilitate the subsequent insertion of vehicle model data. A hash function is designed for the unique identifier of the vehicle model. The function should have a low conflict rate to ensure the uniform distribution of vehicle models in the hash table. Insert vehicle model data: store the standard vehicle model parameter information in the hash table. For each vehicle model, calculate the hash value of its unique identifier. If the corresponding position in the hash table is empty, insert it directly. If a conflict occurs (that is, there is already data at that position), use the chain address method (that is, establish a linked list at that position) to store all conflicting data. The final classification result C final Match the vehicle model information with the standard vehicle model database, traverse each vehicle model in the hash table, calculate the matching degree between the vehicle feature vector and the parameters of each vehicle model in the standard vehicle model database, and use the hash function for fast query. The formula is:

[0106] h(V)=hash(V)

[0107] Where h(V) is the hash value of the feature vector V, and hash(V) is the hash function;

[0108] Furthermore, a hash function is an algorithm that maps input data of arbitrary size (such as strings, files) to an output value of fixed size (called a hash value or digest). Hash functions are widely used in fields such as data structures (such as hash tables), data verification, digital signatures, password storage, and blockchain. A good hash function has several key properties: first, the output results should be evenly distributed to reduce conflicts (i.e., situations where different inputs produce the same hash value); second, the hash function should be one-way, meaning that the original input cannot be easily deduced from the hash value; finally, any slight change in the input should result in a significant change in the hash value. This makes hash functions very useful in data integrity verification and security.

[0109] When the matching degree is higher than the threshold, which is set to 0.7, the match is considered successful, and the system will feed the vehicle model information back to the toll collection system, and perform the corresponding toll collection operation according to the vehicle model category. After completing the toll collection judgment, the result data (including vehicle model, toll amount, etc.) will be sent to the toll station display terminal and displayed on the screen in real time for the driver to view. At the same time, the system will synchronize this information to the cloud database for subsequent query and statistical analysis.

[0110] Furthermore, in the hash table, different vehicle model identifiers may calculate the same hash value (hash conflict). To handle this situation, the chain address method is used: when a conflict occurs, a linked list is established at the corresponding position of the hash table to link all vehicle model records with the same hash value together, thereby ensuring that the information of each vehicle model can be accurately stored and accessed. Introducing a hash table for vehicle model feature matching: The feature information of the vehicle model is managed by the hash table to achieve fast query and efficient matching, thereby improving the overall performance of the system. Structured vehicle model parameter storage: The vehicle model feature vector and related information are structured and stored to ensure that each vehicle model has a clear feature representation, which is convenient for subsequent calculation and analysis. Efficient matching algorithm: By calculating the distance between the vehicle feature vector and the vehicle model feature vector, the best matching vehicle model is quickly determined to improve classification accuracy.

[0111] If the match fails or the confidence of the classification result is lower than the threshold, the system will automatically record the abnormal event. The system will detect possible abnormalities (such as classification errors, duplicate charges, etc.) by analyzing the differences between historical data and real-time data. Once an abnormality is found, the system will automatically trigger an alarm and notify the back-end staff for further verification and processing. The relevant data will be sent to the manual review system. The information of abnormal vehicles will be fed back to the front-end sensor system. For abnormal situations, the back-end staff can enter the correction opinions into the system through the manual feedback system after verification to correct the classification or charging errors. The corrected data will be used to update the database and improve the accuracy of future classification as part of the continuous learning of the model. The secondary identification process is triggered. The secondary identification is manually combined with the vehicle's own data to determine the vehicle's charging standards, and then fed back to the charging system to ensure the accuracy of the charges. At this time, the system will mark the vehicle's characteristic information and update it in the model library to optimize the future identification process. All processed vehicle characteristic information and classification results are automatically stored in the vehicle model database for future classification optimization and model training. The system dynamically adjusts the model parameters according to the impact of the external environment on the data to maintain high recognition accuracy in different environments. The anomaly detection mechanism ensures that the system can detect and handle potential errors or abnormal situations in a timely manner, thereby reducing user complaints caused by incorrect classification or charging issues. This mechanism improves the reliability and robustness of the system and reduces potential revenue losses and user dissatisfaction.

[0112] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0113] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0114] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0115] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be covered by the protection scope of the present application.

Claims

1. A comprehensive judgment method for highway toll vehicle types, characterized in that: The steps include: Step 1: Build a standard vehicle model database, which stores the vehicle data and vehicle categories of all vehicle models that have passed through the toll station in the past. The vehicle data of vehicles entering the toll station entrance is collected in real time, and the vehicle data is merged into a comprehensive feature vector. The vehicle data includes the vehicle's weight, length, width, number of axles, and height. Step 2: Collect real-time environmental monitoring data at the entrance of the toll station, and modify the real-time collected environmental monitoring data according to the historical environmental monitoring data, wherein the environmental monitoring data includes environmental temperature, humidity and wind speed data; Step 3: Construct a multi-layer neural network classification model, use the vehicle data of the vehicle types that have passed through the toll station in the past as a training set, use the vehicle category of the corresponding vehicle as a label, train the multi-layer neural network classification model, and obtain the trained multi-layer neural network classification model; Step 4: Input the comprehensive feature vector into the trained multi-layer neural network classification model to obtain the confidence of each predicted vehicle category, and select the category with the highest confidence after calibration as the preliminary classification result; Step 5: If the confidence of the preliminary classification result is lower than the set threshold, the adaptive classification process is started, the vehicle data is corrected according to the corrected environmental monitoring data, and the classification result is secondary classified using the support vector machine based on the corrected vehicle data. The final classification result is obtained by weighted fusion of the preliminary classification result of MLP and the secondary classification result of SVM. Otherwise, the preliminary classification result is directly used as the final classification result. Step 6: Match the final classification result with the vehicle model information in the standard vehicle model database. If the match is successful, perform the corresponding charging operation based on the vehicle model category. If the match fails, record the abnormal event and transfer to manual judgment of the vehicle model.

2. A comprehensive judgment method for highway toll vehicle types according to claim 1, characterized in that: The vehicle categories include: sedans, off-road vehicles, trucks and other models. The standard vehicle model database stores detailed information on all vehicle models that have passed through the toll station in the past, including the physical parameters of the vehicle and the corresponding vehicle model category. Each vehicle model record contains a unique vehicle model number, vehicle model name and detailed parameter information. The vehicle data itself is integrated into a comprehensive feature vector V, which represents the main physical characteristics and appearance information of the vehicle. The expression of the feature vector is: V = [L, W, H, M, A] Where L is the length of the vehicle, W is the width, H is the height, M is the weight, and A is the number of axles; The fused feature vector V is stored in the local database, and these data are arranged in the order of time when the vehicles enter the toll station.

3. A comprehensive judgment method for highway toll vehicle types according to claim 2, characterized in that: Real-time collection of ambient temperature T current 、Humidity current and wind speed W current Based on the environmental monitoring data and historical data, the historical data refers to the data collected by the same sensor in the previous 24 hours under similar environmental conditions. The average of various data collected in the previous 24 hours is calculated as the reference data for correction. Based on the difference between historical and current environmental data, the temperature correction factor F is calculated. T , humidity correction factor F H and wind speed correction factor F W .

4. A comprehensive judgment method for highway toll vehicle types according to claim 2, characterized in that: The generated feature vector V is input into the multi-layer neural network classification model, and the historical vehicle data of the marked passing vehicles is extracted from the database. The historical vehicle data includes: the historical vehicle length L i , historical vehicle width W i , historical vehicle height H i , historical vehicle weight M i , historical vehicle axle number A i , the historical features of the same passing vehicles are combined into a feature vector V i , the formula is: V i =[L i 、W i 、H i 、M i 、A i ] For the class label of known models, each feature vector V i Need to correspond to a label y i ,y i Indicates the vehicle model corresponding to the feature vector of the i-th historical data, label y i The formula is: y i ∈{SUB,SUV,SUC,SUD} Among them, SUB is a sedan model, SUV is an off-road vehicle model, SUC is a truck model, and SUD is other models; Use the cross entropy loss function as the optimization target to calculate the gap between the model's predicted label and the true label. The formula is: Among them, N is the number of samples, C is the number of categories, and y ij is the jth true label corresponding to the i-th feature vector, Confidence of predictions for multi-layer neural network classification models; Use historical data and labels to train the model, update the weights, and perform multiple iterations until the model reaches the preset convergence standard, that is, the change in the loss function is less than the set threshold, which is 0.

8. After the model training is completed, use the training set for prediction, calculate the confidence of each category, and select the category with the highest confidence as the preliminary classification result. The specific operation of classification is achieved through the following formula: in, is the preliminary classification result, and f(V) is the confidence vector output by the neural network.

5. A comprehensive judgment method for highway toll vehicle types according to claim 4, characterized in that: If the confidence level of the preliminary classification result is higher than the set threshold of 0.6, the preliminary classification result will be matched with the vehicle model information in the standard vehicle model database. After the match is successful, the vehicle model information will be fed back to the toll collection system, and the corresponding toll collection operation will be performed according to the vehicle model category. If the confidence level of the preliminary classification result is lower than the set threshold of 0.6, the adaptive classification process will be started.

6. A comprehensive judgment method for highway toll vehicle types according to claim 5, characterized in that: For each vehicle's physical parameters collected, correction factors are used for correction. The dynamic correction formula is: S corrected =S raw ·(1+a T ·F T +a H ·F H +a W ·F W ) Among them, α T , α H and α W are the weight coefficients of the impact of temperature, humidity and wind speed on sensor data, and 0.3<<α T <<0.7,0.1<<α H <<0.4,0.1<<α W <<0.4,S raw is a physical parameter collected, S corrected is a physical parameter after correction; The corrected physical parameters of each vehicle are used as the training set, and the vehicle category of the corresponding vehicle is used as the label. The SVM model is trained using the above training set, and the decision value de between the output corrected label and the historical label is calculated. The confidence between the corrected label and the historical label is calculated. The formula is: Where: P(Y corrected / Y historical ) represents the confidence between the revised label and the historical label, ranging from [0, 1]; The environmental monitoring data is repeatedly used to calibrate the collected vehicle data so that the confidence between the calibrated label and the historical label is greater than 0.6, and then the corrected label replaces the previous label.

7. A comprehensive judgment method for highway toll vehicle types according to claim 6, characterized in that: The final classification result is obtained by weighted fusion of the preliminary classification result of MLP and the secondary classification result of SVM. The fusion formula is: C final =λ·C MLP +(1―λ)·f SVM (V) Among them, C final is the final classification result, C MLP is the preliminary classification result, f SVM (V) is the secondary classification result, λ is the weight coefficient, that is, 0.4≤λ≤0.

8.

8. A comprehensive judgment method for highway toll vehicle types according to claim 7, characterized in that: Construct a hash table to store the feature vector information of the standard vehicle model database. Each vehicle model is used as the key of the hash table, and the value of the hash table contains the feature vector of the vehicle model. final Match the vehicle model information in the standard vehicle model database, calculate the matching degree between the vehicle feature vector and the parameters of each vehicle model in the standard vehicle model database, and use the hash function for fast query. The formula is: h(V)=hash(V) Where h(V) is the hash value of the feature vector V, and hash(V) is the hash function; When the matching degree is higher than the threshold, the match is considered successful, and the system will feed back the vehicle model information to the charging system and perform the corresponding charging operation according to the vehicle model category. If the match fails or the confidence of the classification result is lower than the threshold, the secondary recognition process is triggered. The secondary recognition is manually determined by combining the vehicle's own data to determine the vehicle's charging standard.

Images