An ETC vehicle identity consistency deep verification method, system and device

Abstract

The present application relates to the field of intelligent transportation, and particularly relates to an ETC vehicle identity consistency deep verification method, system and device. The present application solves the problem that a single point static comparison cannot identify a label change and fee evasion in the prior art by constructing a vehicle trusted identity file and establishing a "file-real vehicle" comparison mechanism, placing a single pass event in the historical trajectory dimension for verification. And through multi-dimensional feature vectorization expression and similarity calculation, identity consistency can still be determined according to other stable features in the case of single feature failure caused by license plate pollution, poor lighting and the like, greatly improving environmental adaptability and verification accuracy. Finally, through dynamic comparison of the consistency comprehensive score and the set threshold, second-level perception and automatic early warning of the change of the OBU and vehicle binding relationship are realized, and the defect that the prior art has no perception of "vehicle-OBU" unbinding is solved, providing a precise and efficient technical means for highway fee evasion auditing.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation, and in particular to a method, system and device for deep verification of ETC vehicle identity consistency. Background Technology

[0002] Electronic Toll Collection (ETC) is an advanced road and bridge toll collection method widely used globally. In China, the system primarily follows national recommended standards and employs Dedicated Short Range Communication (DSRC) technology in the 5.8GHz band to enable information exchange between the Roadside Unit (RSU) installed in the ETC lane at toll stations and the On-Board Unit (OBU) installed inside the vehicle's windshield, thereby automating toll calculation and deduction. The vehicle's OBU is issued by provincial operating companies in each province (municipality) and contains key information unique to each vehicle, including license plate number, license plate color, and vehicle type (classified by passenger car, freight car, and specific number of axles and seats). When a vehicle passes through the mainline ETC gantry or toll station lane, the RSU reads this information from the OBU and compares it with vehicle images captured by high-definition license plate recognition equipment, serving as the primary basis for toll calculation and deduction.

[0003] Currently, ETC (Electronic Toll Collection) on-board units are being widely promoted and installed nationwide. Due to the tight construction schedule and massive issuance volume, inconsistencies have arisen between the vehicle information (license plate, vehicle type, etc.) recorded in some OBUs and the actual vehicle information during the issuance process. Furthermore, there are illegal activities exploiting loopholes in the ETC system rules to evade tolls. One typical and particularly harmful method is "large vehicle, small label," where criminals illegally install OBUs for small vehicles (such as Class I passenger cars) on large vehicles (such as Class VI freight trucks). This causes the system to charge large vehicles according to the toll rate recorded in the OBU when they pass through ETC gantries, thus allowing them to evade large amounts of tolls.

[0004] To address the above problems, several solutions already exist in the existing technology, mainly including:

[0005] (1) A method and system for identifying abnormal behavior in ETC services based on OCR recognition and information comparison (publication number: CN120086755A): This method captures vehicle images, extracts information such as license plate and vehicle type, and compares them with the registration information in the OBU to determine whether it is abnormal. However, it relies too much on the accuracy of OCR recognition and is a single-point, static comparison, which is easily affected by environmental factors and has a high misjudgment rate.

[0006] (2) A time-series-based method for identifying cross-provincial toll evasion vehicles on highways (publication number: CN115599836A): This method identifies toll evasion behavior by analyzing the time-series characteristics of gantry transaction data. However, it is highly dependent on the accuracy of the transaction data. Incorrect license plate recognition or OBU replacement can lead to incorrect record association. Furthermore, it mainly focuses on the time dimension and lacks the ability to identify anomalies in vehicle type, toll rate, etc.

[0007] (3) A dynamic path identification method, device and ETC system based on multi-sensor fusion (publication number: CN120333481A): This method integrates multi-source data such as millimeter-wave radar and visual monitoring to identify paths. However, it relies on the additional deployment of multiple sensors, resulting in high hardware costs. Furthermore, it is isolated from the existing toll collection system network, making it difficult to associate OBU information with vehicle characteristics, thus limiting its versatility.

[0008] (4) A toll evasion audit system based on the integration of Beidou positioning and ETC technology (publication number: CN223078716U): This method detects toll evasion by comparing Beidou positioning trajectories with ETC gantry transaction sequences. However, it requires vehicles to be equipped with Beidou positioning devices, which is not applicable to the vast majority of vehicles that have not installed such devices, resulting in poor universality.

[0009] In summary, existing ETC vehicle identity verification methods mainly adopt single-point, static comparison methods, overly rely on OCR recognition accuracy, and fail to perform spatiotemporal trajectory analysis formed by continuous vehicle passage in the road network. This makes it difficult to detect toll evasion behaviors such as "large vehicle with small label" from a continuity perspective. At the same time, existing technologies usually use fixed weights or simple rules when fusing information, and cannot dynamically adjust the weight of each feature in the decision-making process according to changes in observation quality caused by license plate damage, bad weather, etc., resulting in poor environmental adaptability and high false judgment rate. In addition, existing methods lack a dual verification mechanism across time and space, and cannot sensitively detect whether the binding relationship between OBU and vehicle physical identity has been broken or changed.

[0010] Therefore, there is a need for a more continuous, environmentally adaptable, and accurate method, system, and equipment for in-depth verification of the consistency of ETC vehicle identity. Summary of the Invention

[0011] The purpose of this invention is to overcome the problems of over-reliance on OCR recognition accuracy, poor environmental adaptability, and high false judgment rate in the existing technology, and to provide a method, system and device for deep verification of ETC vehicle identity consistency.

[0012] To achieve the above-mentioned objectives, the present invention provides the following technical solution: A method for deep verification of ETC vehicle identity consistency includes the following steps: S1: Obtain real-time data of the vehicle to be verified; the real-time data includes OBU information and image data; S2: Preprocess the real-time data and extract features to output the observed feature vector; S3: Search and match the observed feature vector in the pre-constructed trusted identity profile database; if the match is successful, retrieve the baseline feature vector of the corresponding trusted identity profile or the profile to be verified; if the match fails, store the observed feature vector as the profile to be verified. S4: Calculate the similarity between the observed feature vector and the benchmark feature vector, and generate a consistency score for the observed feature vector; S5: Compare the overall consistency score with a set threshold, and output the identity consistency depth verification result of the vehicle to be verified; the identity consistency depth verification result includes: If the overall consistency score is greater than or equal to the set threshold, it is determined that the binding relationship between the vehicle to be verified and the OBU has not changed. If the overall consistency score is less than the set threshold, it is determined that the binding relationship between the vehicle to be verified and the OBU is abnormal, and the set alarm mechanism is triggered.

[0013] As a preferred embodiment of the present invention, the real-time data in S1 includes a timestamp and a location identifier, and the image data includes multi-angle captured images of the vehicle to be verified.

[0014] As a preferred embodiment of the present invention, step S2 includes the following steps: S21: Preprocess the real-time data to generate a single passage event record; the preprocessing includes data cleaning and spatiotemporal alignment. S22: Extract the static and dynamic information of the OBU from the OBU information, and the static and dynamic visual features from the image data; and calculate the quality assessment score for each feature component based on the confidence level of the license plate OCR and the sharpness of the image data; the quality assessment score includes the image quality pre-score and the feature confidence level; S23: Integrate the various feature components extracted in S22 and output them as the observed feature vector.

[0015] As a preferred embodiment of the present invention, the image quality pre-scoring includes the following steps: Receive pre-processed captured vehicle images I raw ; For captured vehicle images I raw Perform multi-indicator analysis and calculate sub-scores for each indicator; Normalization and thresholding: The total score S is obtained by weighted summation of all sub-scores. image Its expression is: S image = Σ(w i * s i ) * k weather , Among them, w i Assign weights to indicator i, s i Assign weights to index i, where i∈{sharp,expo,noise}, k weather As a weather impact penalty factor; when S image When the value is less than 0.4, the current image is marked as a "low-quality image," and the current image is either deleted or used only for calculating low-weight features.

[0016] As a preferred embodiment of the present invention, the multi-index analysis includes: Sharpness index: Calculated using the variance of the Laplacian operator, its expression is: S sharp = min(σ Laplace / T sharp , 1); Among them, S sharp For the sub-score of the sharpness index, σ Laplace Let T be the variance of the captured vehicle image to be scored after applying the Laplacian operator. sharp This is the preset resolution threshold; Exposure index: Calculated by measuring the ratio of dark to bright areas in the histogram, its expression is: S expo = max(0, 1 - |μ brightness - μ target | / δ); Among them, S expo For the sub-score of the exposure metric, μ brightness The average value of the brightness of all pixels in the captured image of the vehicle to be scored, μ target The set average brightness is the desired target, and δ is the set maximum brightness fluctuation. Noise level index: calculated based on a three-dimensional block matching algorithm, its expression is: S noise = 1 - (noise estimate / N max ); Among them, S noise For the sub-rating of the noise level index, noise estimate To estimate the noise level, N max The set upper limit for noise tolerance; Weather impact penalty factor: A lightweight classifier is used to determine whether the current image is subject to a set severe weather condition; if it is, let k... weather If it equals the set value, otherwise k weather It equals 1.

[0017] As a preferred embodiment of the present invention, the calculation of the feature confidence level includes the following steps: Calculate the feature confidence level of the license plate information; the formula for calculating the feature confidence level of the license plate information is: char_confidence i = max(p i ), plate_confidence = prod(char_confidence i ) 1 / N , , , Among them, char_confidence i p represents the single-character confidence score of the i-th character. i is the softmax probability vector of the i-th character, i∈[1,N], where N is the total number of characters in the license plate; plate_confidence is the overall OCR confidence of the license plate, and prod() is the multiplication operation; Location confidence level for license plate information. The coordinates of the predicted license plate bounding box. The coordinates of the reference box are given, and Area() is the function to calculate the area of ​​the rectangle. The feature confidence scores for the license plate information are w1=0.5, w2=0.3, and w3=0.2. Calculate the feature confidence level of vehicle attribute information; the formula for calculating the feature confidence level of vehicle attribute information is: , , , , , in, The confidence level for color features. This is the dominant color vector extracted from the current vehicle image. For the pre-established color cluster center vector, The standard deviation of the color cluster; Confidence level of brand characteristics The highest softmax probability value in the brand category is output by the classification network. The confidence level of the model characteristics. The classification network outputs the highest softmax probability value in the vehicle category. The confidence level of the label features. The original line type output of the label sticker before Sigmoid activation; The confidence level of the scratch feature. The original linear output before Sigmoid activation due to scratch damage; Calculate the dynamic weight of the feature credibility of each piece of information; its expression is: , in, Let be the dynamic weight of the feature credibility of the i-th piece of information. The basic weight for the feature credibility of the i-th piece of information is set to an initial value. Let be the feature credibility of the i-th piece of information, where i, j ∈ [1, n], and n is the number of pieces of information; Based on the feature credibility of each piece of information and its corresponding dynamic weight, a weighted sum is performed to obtain the feature credibility of the current feature.

[0018] As a preferred embodiment of the present invention, the retrieval and matching in S3 is performed using the feature component with the highest quality assessment score as the primary key. The trusted identity profile database consists of several trusted identity profiles, and the trusted identity profiles include a baseline feature vector, feature stability score, and associated OBU history. The baseline feature vector is a weighted average of several historical observation feature vectors; The feature stability score is the stability score of each feature component in the baseline feature vector; The associated OBU history is used to record all OBU serial numbers that have been bound to the corresponding vehicle and their start and end times.

[0019] As a preferred embodiment of the present invention, after the file to be verified has been matched a set number of times, the file to be verified is converted into a trusted identity file.

[0020] As a preferred embodiment of the present invention, step S4 includes the following steps: S41: Calculate the similarity between each corresponding feature component in the observed feature vector and the reference feature vector; S42: Calculate the dynamic weights of the similarity of each feature component; S43: Multiply the similarity of all feature components by their corresponding dynamic weights and sum them to generate a consistency score for the observed feature vector.

[0021] As a preferred embodiment of the present invention, in step S41, the corresponding similarity calculation method is selected based on the feature type of the current feature component; When the feature component is a numerical vector or an embedded vector, cosine similarity is used as the similarity. When the feature component is a categorical feature, exact matching or fuzzy matching based on category distance is used as the similarity. When the feature component is a license plate number feature, edit distance or Jaro-Winkler distance is used to measure string similarity.

[0022] In a preferred embodiment of the present invention, in step S42, the dynamic weight = static importance weight ⊙ observation quality dynamic weight ⊙ spatiotemporal correlation dynamic weight; where ⊙ is the Hadamard product; the static importance weight is the base weight, and its initial value is a set value; the observation quality dynamic weight is generated based on the quality assessment score of the current feature component; the spatiotemporal correlation dynamic weight is used to evaluate the contribution of the benchmark feature vector to the current judgment.

[0023] As a preferred embodiment of the present invention, the spatiotemporal correlation dynamic weight includes a time decay weight and a spatial logic weight; Time decay weight: ; Where Δt is the time difference between historical observation records and the current time. It is the attenuation coefficient of feature i; Spatial logical weights:

[0024]

[0025]

[0026] in, For the k-th consecutive gantry, γ is the spatial logic enhancement coefficient; It is the combined similarity score between two observations, where G(Δt) is a Gaussian function. The time for detection at the k-th detection position. The average time window for a vehicle to pass through two consecutive detection locations. and The minimum and maximum values ​​for a reasonable time window. The standard deviation is denoted as .

[0027] As a preferred embodiment of the present invention, the expression for the consistency comprehensive score in S43 is:

[0028] in, For consistency score, Let N be the final weight of the i-th feature component, and N be the number of feature components. Let be the feature similarity of the i-th feature component.

[0029] As a preferred embodiment of the present invention, the threshold setting in S5 is adaptively adjusted based on vehicle type, road segment type, and vehicle historical behavior, including the following steps: Generate a unique identifier for each vehicle according to the established business rules; A dynamic threshold is established based on the historical data of each unique identifier using the sliding window statistical method. The system acquires the current consistency score in real time and compares it with the threshold. When alarms are triggered continuously or not for a long time, the system automatically adjusts the sensitivity parameters of the dynamic threshold and the detection window size according to the set rules.

[0030] As a preferred embodiment of the present invention, it further includes S6: when the consistency comprehensive score is greater than or equal to a set threshold, a smooth averaging or Kalman filtering method is used to add the currently observed feature vector to the trusted identity file of the corresponding vehicle for dynamic updating.

[0031] An ETC vehicle identity consistency deep verification system is used to execute any of the above-described ETC vehicle identity consistency deep verification methods; the system includes a data acquisition layer and a data processing and analysis layer that are interconnected. The data acquisition layer is deployed at every ETC gantry and ETC lane at toll stations along the highway to acquire real-time data of vehicles to be verified. The data processing and analysis layer is used to perform in-depth identity consistency verification on the vehicle to be verified based on the real-time data.

[0032] An ETC vehicle identity consistency deep verification device includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform any of the above-described ETC vehicle identity consistency deep verification methods.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a trusted vehicle identity profile and establishes a "profile-vehicle" comparison mechanism, placing a single passage event within the historical trajectory dimension for verification. This solves the problem of existing technologies' difficulty in identifying toll evasion through single-point static comparison. Furthermore, through multi-dimensional feature vectorization and similarity calculation, it can still determine identity consistency based on other stable features (such as vehicle model, annual inspection sticker, and vehicle scratches) even when single features fail due to license plate damage or poor lighting, significantly improving environmental adaptability and verification accuracy. Finally, by dynamically comparing the overall consistency score with a set threshold, it achieves second-level perception and automatic warning of changes in the OBU-vehicle binding relationship, overcoming the deficiency of existing technologies in detecting "vehicle-OBU" unbinding. This provides a precise and efficient technical means for highway toll evasion auditing. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the ETC vehicle identity consistency deep verification method according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the structure of an ETC vehicle identity consistency deep verification system according to Embodiment 3 of the present invention; Figure 3 This is a schematic diagram of the structure of an ETC vehicle identity consistency deep verification device as described in Embodiment 4 of the present invention. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0036] Example 1 like Figure 1 As shown, a method for deep verification of ETC vehicle identity consistency includes the following steps: S1: Obtain real-time data of the vehicle to be verified; the real-time data includes OBU information and image data.

[0037] S2: Preprocess the real-time data and extract features to output the observed feature vector.

[0038] S3: Search and match the observed feature vector in the pre-constructed trusted identity profile database; if the match is successful, retrieve the baseline feature vector of the corresponding trusted identity profile or the profile to be verified; if the match fails, store the observed feature vector as the profile to be verified.

[0039] S4: Calculate the similarity between the observed feature vector and the benchmark feature vector, and generate a consistency comprehensive score for the observed feature vector.

[0040] S5: Compare the overall consistency score with a set threshold, and output the identity consistency depth verification result of the vehicle to be verified; the identity consistency depth verification result includes: If the overall consistency score is greater than or equal to the set threshold, it is determined that the binding relationship between the vehicle to be verified and the OBU has not changed. If the overall consistency score is less than the set threshold, it is determined that the binding relationship between the vehicle to be verified and the OBU is abnormal, and the set alarm mechanism is triggered.

[0041] Example 2 This embodiment is a specific implementation of the deep verification method for ETC vehicle identity consistency described in Embodiment 1, including the following steps: A method for deep verification of ETC vehicle identity consistency includes the following steps: S1: Obtain real-time data of the vehicle to be verified; the real-time data includes OBU information and image data.

[0042] The real-time data includes timestamps and location identifiers, and the image data includes multi-angle captured images of the vehicle to be verified.

[0043] Specifically, when a vehicle equipped with an OBU passes through a checkpoint, the checkpoint's RSU (Remote Unit) completes a transaction with the OBU, capturing OBU information and transaction details. Simultaneously, images of the vehicle are captured from multiple angles. This heterogeneous data is packaged and given precise timestamps and location identifiers, outputting real-time data of the vehicle to be verified at the current checkpoint.

[0044] S2: Preprocess the real-time data and extract features to output the observed feature vector.

[0045] S21: Preprocess the real-time data to generate a single passage event record; the preprocessing includes data cleaning and spatiotemporal alignment.

[0046] The data cleaning is used to remove invalid and duplicate data.

[0047] The spatiotemporal alignment operation uses timestamps and gantry / lane IDs to precisely link OBU transaction data and multiple captured images from different angles generated by the same vehicle at the same time and location, forming a complete "Single Pass Record (SPR)".

[0048] S22: Perform feature extraction on the preprocessed real-time data. Calculate the quality assessment score for each feature component based on the confidence level of the license plate OCR and the sharpness of the image data.

[0049] Specifically, assuming the vehicle has N dimensions of features extracted, These characteristics can be divided into four categories: (a) Static information of OBU (F_OBU_S): such as license plate number, color, vehicle type, etc., which are fixed in the OBU.

[0050] (ii) OBU dynamic information (F_OBU_D): such as transaction time, location, status, etc.

[0051] (iii) Visual static features (F_VIS_S): These are relatively stable features extracted from an image, such as brand, model, color, annual inspection mark, sticker, scratches, etc.

[0052] (iv) Visual dynamic features (F_VIS_D): These are volatile features, such as interior decorations of the vehicle and the driver (this feature is extracted using only fuzzy features, such as clothing color and whether the driver is wearing glasses, while protecting privacy).

[0053] In this embodiment, taking the detection point at time t and position l as an example, the observation feature set O(t, l) extracted from the real-time data is converted into a d-dimensional feature vector V(t, l) through feature encoding:

[0054] Among them, v i It is the vectorized representation of the i-th feature, denoted as feature component v. i , i∈[1,N]. For example, license plate numbers can be represented by embeddings of OCR results, vehicle models can be one-hot vectors, colors can be vectors of RGB values, and complex visual features such as brand, model, and scratches can be represented by feature embeddings extracted by deep learning models (such as ResNet, ViT).

[0055] At the same time, for each feature component v i A quality assessment score q is generated along with this. i (t, l)∈[0, 1], representing the confidence level of this feature in this observation. For example, q plate This can be the confidence level of OCR recognition. The quality assessment score includes an image quality pre-score and feature confidence; the specific calculation includes the following steps: Image quality pre-scoring: (1) Receive the preprocessed captured vehicle image I raw ; (2) For captured vehicle images I raw Perform multi-indicator analysis; obtain a sub-score s through the analysis of each indicator. iThen, the total score S is obtained by weighted summation of all sub-scores. image The multi-index analysis includes: ① Sharpness: Calculated using the variance of the Laplacian operator, its expression is: S sharp = min(σ Laplace / T sharp , 1.0); Among them, S sharp For the sub-score of the sharpness index, σ Laplace T represents the variance of the captured vehicle image after applying the Laplacian operator, reflecting the richness of edge information in the image. A larger variance indicates more high-frequency variations (i.e., sharp edges), while a smaller variance may indicate blurriness or motion blur. sharp This is a preset sharpness threshold. When the sharpness exceeds the threshold, the score saturates to 1.0 to prevent outlier interference.

[0056] ② Exposure: Calculated by measuring the ratio of dark areas (L_dark) to bright areas (L_bright) in the histogram. Its expression is: S expo = max(0, 1 - |μ brightness - μ target | / δ); Among them, S expo For the sub-score of the exposure metric, μ brightness The average of the brightness values ​​of all pixels in the captured vehicle image to be scored, typically within the range of [0, 255]; μ target δ is the set target average brightness, which is the brightness level we expect the image to reach under ideal conditions; δ is the set maximum brightness fluctuation value. When the actual brightness deviates from the target brightness by more than δ, the exposure score will linearly decrease to 0.

[0057] ③ Noise Level: The denoising residual is evaluated based on the 3D Block Matching 3D (BM3D) algorithm, and its expression is as follows: S noise = 1 - (noise estimate / N max ); Among them, S noise For the sub-rating of the noise level index, noise estimate To estimate the noise level, it can be quantified by analyzing smooth regions of the image or using residuals from algorithms such as BM3D; N maxThis is the set upper limit for noise tolerance; exceeding this value is considered extremely noisy. This percentage represents the current noise level relative to the maximum tolerable noise level; the higher the noise level, the more points are deducted.

[0058] ④ Weather impact penalty factor: Use a lightweight classifier (MobileNetV2) to determine whether it is a set severe weather (such as rainy or foggy weather).

[0059] If the weather is set to be severe, let k weather equals the set value; otherwise k weather It equals 1.

[0060] If the weather is heavy rain or fog, let k weather =0.7; when the weather is light rain, let k weather =0.9; adjust the settings according to your needs.

[0061] (3) Normalization and thresholding: The total score S is obtained by weighted summation of all sub-scores. image Its expression is: S image = Σ(w i * s i ) * k weather , Among them, w i Assign weights to indicator i, s i Assign weights to index i, where i∈{sharp,expo,noise}.

[0062] If S image If the value is less than 0.4, mark the current image as a "low-quality image", delete the current image, or use it only to calculate low-weight features.

[0063] Feature credibility: The calculation of the feature confidence level includes the following steps: (1) Calculate the feature credibility of the license plate information; the calculation of the feature credibility of the license plate information includes the following steps: ①OCR confidence level: Generate a softmax probability vector for each character using an optical character recognition model (such as CRNN or Transformer-based OCR model); The expression for calculating the single-character confidence score for each character is as follows: char_confidence i = max(p i ); Among them, char_confidence i p represents the single-character confidence score of the i-th character.i Let be the softmax probability vector of the i-th character, i∈[1,N], where N is the total number of characters in the license plate; The overall license plate confidence score is calculated as the OCR confidence score (geometric mean), and its expression is as follows: plate_confidence = prod(char_confidence i ) 1 / N ; Where plate_confidence is the overall license plate confidence level, and prod() is a multiplication operation used to calculate the geometric mean of the confidence levels of all characters; ②Location confidence: The location score is calculated by using the IoU of the license plate information detection bounding box and the GT (predicted by the Anchor model trained on the artificial dataset).

[0064] ; in, Location confidence level for license plate information. The coordinates of the predicted license plate bounding box. The coordinates of the reference box (or historical average box) are given, and Area() is the function to calculate the area of ​​the rectangle. ③ Calculate the overall quality score As a feature of the license plate information: ; Where w1=0.5, w2=0.3, w3=0.2; (2) Calculate the feature confidence level of the vehicle attribute information; the feature confidence level of the vehicle attribute information includes: ① Color: K-Means clustering is performed based on the HSV color space. The closer the cluster centers are, the higher the confidence level.

[0065] , in, The confidence level for color features. This is the dominant color vector extracted from the current vehicle image. For the pre-established color cluster center vector, Let be the standard deviation of the color cluster.

[0066] ② Brand and Model: The maximum probability output by the Softmax classification network.

[0067] , , in, Confidence level of brand characteristics The highest softmax probability value in the brand category is output by the classification network. The confidence level of the model characteristics. The classification network outputs the highest softmax probability value in the vehicle category.

[0068] ③ Micro-features: The sigmoid probability output by the binary classification network is regarded as the quality score.

[0069] , , in, The confidence level of the label features. The original line type output of the label sticker before Sigmoid activation; The confidence level of the scratch feature. This is the original linear output before Sigmoid activation for scratch damage.

[0070] (3) Calculate the dynamic weight of the feature credibility of each piece of information; its expression is: , in, Let be the dynamic weight of the feature credibility of the i-th piece of information, and =1, The basic weight for the feature credibility of the i-th piece of information is set to a predetermined value (e.g., license plate is set to 0.35, vehicle type is set to 0.25, etc.). Let be the feature credibility of the i-th piece of information, where i and j ∈ [1, n], and n is the number of pieces of information.

[0071] (4) Based on the feature credibility of each piece of information and the corresponding dynamic weight, perform weighted summation to obtain the feature credibility of the current feature.

[0072] S23: Integrate the various feature components extracted in S22 and output them as the observed feature vector.

[0073] S3: Search and match the observed feature vector in the pre-constructed trusted identity profile database; if the match is successful, retrieve the baseline feature vector of the corresponding trusted identity profile or the profile to be verified; if the match fails, store the observed feature vector as the profile to be verified.

[0074] In S3, the retrieval and matching are performed using the feature component with the highest quality assessment score as the primary key. The trusted identity profile database consists of several trusted identity profiles, and the trusted identity profiles include the baseline feature vector, feature stability score, and associated OBU history. The reference feature vector V profile This is a weighted average of several historical observation feature vectors; it represents the most stable and reliable representation of the vehicle's various features. Furthermore, this feature can also be a weighted average of high-quality historical observation data selected based on set filtering criteria.

[0075] The feature stability score A stability score is assigned to each feature component in the baseline feature vector. Where s i The historical stability of the i-th feature is represented by calculating the variance or information entropy of its historical observations. A more stable feature s... i The higher the value.

[0076] The associated OBU history H OBU This is used to record all OBU serial numbers that have been bound to the corresponding vehicle and their start and end times.

[0077] Furthermore, after the file to be verified has been matched a set number of times, the file to be verified is converted into a trusted identity file.

[0078] Furthermore, the archive updates employ an exponential moving average (EMA) strategy to smoothly incorporate new observational data, using the baseline feature vector V. profile For example: V profile—new =α * V(t, l) + (1 - α) * V profile_old Among them, V profile—new V is the updated baseline feature vector. profile_old The baseline feature vector before the update is α∈(0, 1), which is the learning rate. It can be dynamically adjusted according to the observation quality q(t, l). High-quality observations can have a larger α.

[0079] S4: Calculate the similarity between the observed feature vector and the benchmark feature vector, and generate a consistency comprehensive score for the observed feature vector.

[0080] S41: Calculate the similarity between each corresponding feature component in the observed feature vector and the reference feature vector.

[0081] When there is a new observed feature vector V current At that time, it is necessary to calculate its relationship with file V. profile The similarity is calculated by first calculating the basic similarity Sim for each feature component. i : Sim i = Similarity(vi_current , v i_profile ); Among them, v i_current Let v be the observed feature vector of detection point i. i_profile Let i be the baseline feature vector of the detection point i.

[0082] Furthermore, in step S41, the corresponding similarity calculation method is selected based on the feature type of the current feature component; When the feature components are numerical vectors or embedded vectors (such as brand, model, scratch features), cosine similarity is used as the similarity. Sim i = (v i_current * v i_profile ) / (||v i_current || * ||v i_profile ||); When the feature component is a categorical feature (such as vehicle model or color), exact matching (i.e., similarity of 1 or 0) or fuzzy matching based on category distance is used as the similarity. When the feature component is a license plate number feature (such as OBU license plate and visual license plate), edit distance or Jaro-Winkler distance is used to measure string similarity in order to deal with the situation where individual characters are misrecognized by OCR.

[0083] S42: Calculate the dynamic weights of the similarity of each feature component.

[0084] Furthermore, the dynamic weights =Static importance weight ⊙Dynamic weight of observation quality Spatiotemporal correlation dynamic weights ; where ⊙ represents the Hadamarda accumulation.

[0085] (1) Static importance weights : = [w_1_static, ..., w_N_static]; These are basic weights set based on prior knowledge, reflecting the inherent importance of different features in distinguishing vehicle identities. For example, their initial values ​​are set according to general requirements, such as: w_OBU_info_static = 0.3 (Overall weight of information within the OBU) w_plate_static = 0.3 (Visual license plate) w_type_static = 0.15 (Visual vehicle model) w_brand_static = 0.1 (Visual Brand) w_finegrained_static=0.15 (sum of weights of other micro-features). Micro-features are very unique, and once a match is found, the weights should be increased appropriately.

[0086] These weights are set during system initialization and can be updated over a set period. The weights are then optimized through model training based on the detection results within that period.

[0087] (2) Dynamic weight of observation quality : = [w_1_quality, ..., w_N_quality]; The observation quality dynamic weight is based on the current feature component's quality assessment score Q. current = [q1, ...,q N Generate; used to evaluate the contribution of the benchmark feature vector to the current judgment.

[0088] This example uses the quality score directly (in actual use, coefficients and other influencing factors can also be added): w_i_quality = q i Furthermore, to maintain the relative stability of the total weight, this embodiment introduces a weight normalization redistribution mechanism. When the quality q of a certain important feature (such as a license plate)... plate When a feature decreases, its weight decreases, while the weights of other features increase proportionally.

[0089] This embodiment uses a license plate as an example for illustration. Let W_base = W_static ⊙ W_quality, that is, w_i_base = w_i_static * q i .

[0090] When the license plate is damaged, q plate As the value decreases from 1 to 0.2, the dynamic weight of the license plate, w_plate_base, decreases accordingly. The resulting weight loss Δw = w_plate_static * (1 - q) plate It needs to be assigned to other features, and proportionally based on the static weights of those other features: w_j_final_quality = w_j_base + Δw * (w_j_static / Σ_{k≠plate} w_k_static) (for j ≠ plate) w_plate_final_quality = w_plate_base.

[0091] Wherein, W_base is the base weight, W_static is the static weight, W_quality is the observation quality weight, w_i_base is the base ratio of the i-th feature component, w_i_static is the static correction factor of the i-th feature component, j is the feature dimension index (e.g., j=1 represents the license plate OCR score, j=2 represents the color confidence, etc.), k is the time / space context label (e.g., k=Highway-01 represents a specific road segment), and w_j_final_quality is the final weight of the j-th feature component after considering quality.

[0092] In this way, the system intelligently "reduces its reliance on unreliable evidence and instead places more trust in other reliable evidence."

[0093] (3) Spatiotemporal correlation dynamic weights : The spatiotemporal correlation dynamic weight includes a time decay weight and a spatial logical weight. This weight is mainly used to update trusted identity profiles and, during identity verification, to evaluate the contribution of different observation points in historical trajectories to the current judgment.

[0094] Time decay weight: ; Where Δt is the time difference between historical observation records and the current time. λ is the decay coefficient of feature i; core stable features (such as brand and model) have smaller λ and decay slowly; volatile features (such as car interior decorations) have larger λ and decay quickly, reflecting that the more recent the observation record, the greater its influence in judging the current identity.

[0095] Spatial logical weights: This weight is used to enhance confidence in continuous, reasonable trajectories. Assume vehicles continuously pass through a gantry sequence. ,exist Verification will be conducted.

[0096] If the vehicle is Observation results and Highly consistent, and arrive travel time Within a reasonable time window calculated based on road network topology and speed limits [ Inside, then An improvement has been achieved. Current spatial logical weights. It is an enhancing factor. ≥ 1. Its expression is:

[0097]

[0098]

[0099] in, For the k-th consecutive gantry, γ is the spatial logic enhancement coefficient; It is the combined similarity score between two observations; the higher the score, the closer M is to 1; G(Δt) is a Gaussian function, when Δt falls within [ The value is largest (close to 1) in the central region of the symbol, and decreases rapidly when it deviates from the central region. The time for detection at the k-th detection position. The average time window for a vehicle to pass through two consecutive detection locations. and The minimum and maximum values ​​for a reasonable time window. The standard deviation is denoted as .

[0100] If multiple consecutive gantry frames satisfy this condition, w spatial This accumulation of information allows the system to have a very high degree of confidence in the identity of vehicles along this trajectory chain. Conversely, if a vehicle's appearance is highly consistent with the earlier part of the trajectory, but its travel time or OBU information changes abruptly at a certain point, the system will immediately detect the anomaly.

[0101] S43: Multiply the similarity of all feature components by their corresponding dynamic weights and sum them to generate a consistency score for the observed feature vector.

[0102] Final consistency score It is a weighted sum of all feature similarities and their final dynamic weights. The expression for the consistency comprehensive score is:

[0103] in, For consistency score, Let N be the final weight of the i-th feature component, and N be the number of feature components. Let be the feature similarity of the i-th feature component.

[0104] S5: Compare the overall consistency score with the set threshold, and output the identity consistency depth verification result of the vehicle to be verified.

[0105] Will Compared to dynamic thresholds, this threshold is not fixed but can be adaptively adjusted based on factors such as vehicle type, route, and historical behavior. For example, for vehicles suspected of toll evasion, the threshold can be appropriately increased to subject them to stricter scrutiny.

[0106] Furthermore, the identity consistency deep verification results include: If the overall consistency score is greater than or equal to the set threshold, it is determined that the binding relationship between the vehicle to be verified and the OBU has not changed. If the overall consistency score is less than the set threshold, it is determined that the binding relationship between the vehicle to be verified and the OBU is abnormal, and the set alarm mechanism is triggered.

[0107] Furthermore, the threshold is set adaptively based on vehicle type, road segment type, and vehicle historical behavior, including the following steps: (1) Data normalization and clustering: Generate a unique identifier for each vehicle according to business rules.

[0108] group_id = f"{vehicle_type}_{road_segment}" (Example: "SUV_108 Road"); Where vehicle_type is the vehicle type identifier, and road_segment is the road segment identifier; This embodiment is grouped by "vehicle type" and "road segment". The difficulty and speed distribution of OCR for different types of license plates are significantly different.

[0109] (2) Constructing a historical baseline model: Establish dynamic thresholds based on the historical data of each group_id using the sliding window statistical method.

[0110] Initialize the 'adaptive threshold', receive the historical data sequence of group_id, and specify the sliding window size (default is 200) and the standard deviation factor k (default is 2.5), and take the most recent N records as the reference window; Extract the data points at the end of the historical data sequence that are set to the size of a window, and use them as a reference window.

[0111] Calculate the mean and standard deviation of the scores within this window.

[0112] Let the dynamic threshold = mean + k * standard deviation.

[0113] Among them, the parameter k determines the sensitivity: a large k results in fewer false alarms but more false negatives; a small k results in the opposite.

[0114] When historical scores are stable, the threshold is low, making it easier to detect minor anomalies; when historical scores fluctuate drastically, the threshold increases accordingly, reducing false alarms caused by normal fluctuations.

[0115] (3) Real-time anomaly detection: The online system obtains the current consistency score Score_current and compares it with the threshold.

[0116] If the threshold is exceeded, an alarm will be triggered.

[0117] (4) Threshold adaptive feedback mechanism: When alarms are triggered continuously or not triggered for a long time, the size of k or window is automatically adjusted to adapt to the new environment.

[0118] For example: When the alarm over-rate exceeds 30%, set the updated k = 1.1 * the original k to reduce the false alarm rate; Alternatively, if there is no alarm for a long time, set the updated k to 0.9 * the original k to enhance sensitivity.

[0119] Furthermore, this embodiment also includes S6: when the consistency comprehensive score is greater than or equal to a set threshold, a smooth averaging or Kalman filtering method is used to add the currently observed feature vector to the trusted identity file of the corresponding vehicle for dynamic updating.

[0120] For passage records deemed "consistent," the system will utilize their high-quality observation data (V). current Q current This is used to update the vehicle's Trusted Identity Profile (TIP). The update process employs a smoothing average or Kalman filter approach, enabling the profile to adapt to minor, normal changes in the vehicle (such as new small scratches, seasonal tire replacements, etc.) while maintaining the stability of core features.

[0121] Meanwhile, this invention periodically utilizes massive amounts of confirmed "consistent" and "abnormal" sample data to retrain and fine-tune the neural network model for feature extraction and the model for weight calculation, thereby achieving continuous iterative optimization of algorithm performance.

[0122] In summary, the solution proposed in this invention has the following advantages: (i) The accuracy of detection has been greatly improved, achieving a leapfrog audit from "point" to "line": This invention, by introducing spatiotemporal trajectory continuity analysis, no longer relies on single-point, single-time "snapshot" comparisons. It strings together multiple vehicle passage records into a trajectory chain, making judgments by analyzing the continuity and stability of the vehicle's multidimensional "biological features" along the entire chain. This "linear" tracking approach effectively filters out misjudgments caused by accidental factors such as poor image quality at a single point or OCR recognition errors. In tests simulating real ETC gantry data, for typical toll evasion behaviors such as "large vehicle, small label" mid-journey label replacement, the detection accuracy of this invention can be improved from approximately 90% of traditional single-point OCR comparison methods to over 98.5%. This is because even if criminals replace the OBU at a certain gantry, the vehicle's appearance characteristics will still highly match the characteristics on the historical trajectory, while the OBU information undergoes a sudden change. This strong contradictory signal is accurately captured by the algorithm model of this invention.

[0123] (ii) The false alarm rate has been significantly reduced, improving the efficiency of audit work: The dynamic weighted fusion of this invention can adaptively adjust the weight of features in decision-making based on the real-time observation quality of the features. For example, when the OCR confidence of a license plate is low due to dirt, rain, snow, or strong light, the system will automatically reduce the weight of the license plate information and instead rely more on other stable features such as vehicle model, brand, color, and annual inspection sticker. This effectively avoids false alarms caused by the unreliability of a single information source. Compared to traditional methods that rely on fixed rules and thresholds, this invention can reduce the false alarm rate caused by environmental factors and technical limitations from over 5% to below 1.5%, greatly reducing the workload of manual auditing in the background and allowing auditors to focus on truly high-suspicious events.

[0124] (III) It has achieved accurate perception and real-time early warning of changes in the "vehicle-OBU" binding relationship: The core of this invention lies in establishing a dynamic "trusted identity feature file" for each physical vehicle and continuously verifying the consistency between the vehicle characteristics and OBU information in the current passage event and this file. This mechanism fundamentally solves the logical flaw of existing technologies that use OBU or license plate identification as vehicle identifiers. Once an OBU is illegally moved to another vehicle, the system will immediately detect that the physical characteristics of the current vehicle are seriously inconsistent with the vehicle characteristics in the OBU's historical file, thus determining that the "vehicle-OBU" binding relationship has changed. The system can complete the calculation and generate an early warning within seconds, with a response time far lower than the hours or days of traditional post-event audits, providing the possibility for on-site coordinated handling or near real-time interception.

[0125] (iv) Enhanced system robustness and adaptability, and reduced dependence on hardware: This invention, through in-depth analysis of existing ETC gantry RSUs and high-definition images, and innovative algorithm models, eliminates the need for large-scale deployment of additional hardware such as millimeter-wave radar and infrared sensors, resulting in high economic efficiency and scalability. The redundant design of multi-dimensional features and the dynamic weighting mechanism enable the system to make relatively reliable judgments based on other features even when some features (such as license plates) are missing or of poor quality, demonstrating strong robustness.

[0126] (v) It resolved historical issues related to issuance errors and improved the quality of ETC services: In addition to combating proactive illegal activities such as "large vehicles with small license plates," this invention can also effectively identify and report historical issues stemming from operational errors during the early large-scale issuance of ETC, such as discrepancies between OBU information and actual vehicle details. Through long-term operation and data accumulation, the system can establish a very high level of confidence in the file for each vehicle, thereby automatically detecting long-term mismatches between file information and OBU registration information. This provides issuers with precise clues to clean up and correct erroneous data, improving the service quality and data accuracy of the entire ETC system.

[0127] Example 3 An ETC vehicle identity consistency deep verification system is provided, which is used for the ETC vehicle identity consistency deep verification method described in the above embodiments; the system includes a data acquisition layer and a data processing and analysis layer that are interconnected.

[0128] The data acquisition layer is used to acquire real-time data of the vehicle to be verified. This layer is deployed at every ETC gantry or ETC lane at toll stations along the highway and is responsible for capturing raw data. It mainly consists of existing or slightly modified equipment, requiring no large-scale hardware investment.

[0129] (i) Roadside Unit (RSU): Compliant with the national standard GB / T 20851 series, it communicates with the on-board unit (OBU) via the 5.8GHz DSRC protocol, and reads and uploads fixed static vehicle information (license plate number, license plate color, vehicle type, issuer information, OBU serial number, etc.) and dynamic transaction information (transaction timestamp, gantry / lane number, transaction status, billing amount, etc.) in real time. This is the main source of the vehicle's "electronic identity".

[0130] (ii) High-definition license plate recognition cameras: These are typically installed on the front or back of the ETC gantry or in the toll lane to capture high-definition images of vehicles. These images form the basis for OCR license plate recognition to assist in verifying OBU license plate information. They are also the basic data source for extracting macroscopic features of the vehicle's front (such as brand logo and grille style) and microscopic features of the windshield (such as annual inspection stickers, environmental protection stickers, various stickers, and interior items such as pendants and mobile phone holders).

[0131] (III) Panoramic / Side High-Definition Cameras (Optional): It is recommended to install three-lens cameras at different angles on the side of the gantry or in the toll lane where conditions permit, to capture images of the vehicle's side and overall outline from different angles. This is an important data source for extracting key "vehicle biometrics" such as body color, specific vehicle model (different from the general vehicle classification in OBU), body waistline, window proportions, wheel style, vehicle advertising stickers, special paintwork, scratches and damage.

[0132] (iv) Edge Computing Unit (Optional): In gantries with high traffic volume or limited network conditions, high-performance industrial control computers or embedded AI devices can be deployed. This unit can perform preliminary processing on the massive image data collected at the front end, such as image quality judgment and screening, target detection and cropping, preliminary feature extraction and encoding, etc., so that only structured feature data and high-quality images are transmitted to the center, greatly reducing the computing pressure on the central server and the occupation of network bandwidth.

[0133] The communication connection mainly relies on the dedicated fiber optic communication network already built on the highway / 5G / 4G and other wireless communication technologies to transmit various types of data acquired by the acquisition layer to the data processing and analysis layer securely, reliably and with low latency, ensuring low latency and high bandwidth of data transmission; at the same time, all data must be encrypted before transmission to ensure data security.

[0134] The data processing and analysis layer is used to perform deep identity verification on the vehicle to be verified based on the real-time data. This is the core computing hub of the invention, typically deployed in a provincial or road segment data center, used to implement the core algorithm logic of the invention. Specifically, it includes the following modules: (i) Data Preprocessing Module: This module is responsible for receiving raw data streams from all collection points across the entire road network. It performs data cleaning (removing invalid and duplicate data), format standardization conversion, and most importantly, executes spatiotemporal alignment. This involves using timestamps and gantry / lane IDs to precisely link OBU transaction data generated by the same vehicle at the same time and location, as well as multiple captured images from different angles, forming a complete "Single Pass Record (SPR)".

[0135] (ii) Multi-feature extraction and encoding module: responsible for extracting multi-dimensional vehicle features from the preprocessed data using various algorithms and converting them into standardized feature vectors.

[0136] (1) Macro feature extraction: Use mature CNN models (such as YOLO series, ResNet, etc.) to perform target detection and classification on vehicle images, and extract license plate number (including confidence level), license plate color, vehicle type, main / secondary body color, vehicle brand and sub-brand / model, etc.

[0137] (2) Micro-feature (vehicle biometrics) extraction: Employ more refined models, such as Re-ID (Re-identification) models, feature point detection models, and image segmentation models, to extract more unique features for each individual vehicle. ① Annual inspection sticker / insurance sticker: Locate the upper right corner of the windshield and extract and analyze its color, shape, year number and other features.

[0138] ② Vehicle attachments: Check for roof racks, tow hooks at the rear, and special stickers or decals on the vehicle body.

[0139] ③ Damage and scratches: Using anomaly detection or texture analysis algorithms, identify permanent or semi-permanent marks such as dents and scratches in specific areas of the vehicle body.

[0140] ④ Interior items: Analyze the inside of the windshield to check for hanging ornaments, decorations on the center console, cards on the sun visor, etc.

[0141] (3) All the extracted discrete or continuous features are uniformly transformed into a high-dimensional, standardized feature vector V through embedding technology. Each sub-vector represents a type of feature. This vectorized representation facilitates efficient similarity calculation and mathematical modeling.

[0142] (III) Spatio-temporal correlation and trajectory construction module: This module is responsible for linking multiple "single passage event records" from different gantries / toll lanes belonging to the same physical vehicle in chronological order. This module uses the most stable features of the vehicle (such as a high-quality license plate number + vehicle type) as the initial index, and combines travel time and road network topology logic (for example, the travel time from gantry A to gantry B should be within a reasonable range) to construct the "spatio-temporal trajectory chain (STC)" for each vehicle in the road network.

[0143] (iv) Multi-feature dynamic weighted matching and consistency verification module: This is the core of the algorithm of this invention, responsible for calculating the consistency score between the current observed features of the vehicle and the historical reliable archives.

[0144] (v) Trusted Identity Profile Generation and Update Module: This module is responsible for maintaining a dynamic Trusted Identity Profile (TIP) for each unique physical vehicle in the system (identified by its feature vector). This profile includes not only the vehicle's baseline feature vector but also the stability score for each feature, historical change records, and associated OBU historical information. The profile evolves dynamically, continuously optimizing and strengthening itself by incorporating new, confirmed, and consistent observation data.

[0145] Furthermore, such as Figure 2 As shown, this embodiment also includes an application and decision-making layer for storing a trusted identity profile database and reporting to staff, specifically including the following structure: (i) Audit and Alarm Subsystem: When the consistency verification module determines that the "vehicle-OBU" binding relationship has changed or is highly suspicious, the system will automatically generate an alarm work order. The work order records in detail the current passage information, historical trajectory, feature comparison details and the basis for the anomaly judgment of the suspected vehicle, and pushes it to the work platform of the auditors for manual review and processing with high priority.

[0146] (ii) Trusted Identity Profile Database: Stores and manages the feature profiles (TIPs) of all vehicles, providing rich query, statistical and analysis functions to provide data support for traffic management, big data analysis and other fields.

[0147] (III) Data Visualization and Reporting Subsystem: Using intuitive methods such as maps, charts, and dashboards, the system dynamically displays the driving trajectory and characteristic changes of suspected vehicles, as well as the spatiotemporal distribution heat map and statistical reports of abnormal behaviors such as "large vehicles with small labels" across the entire road network, providing managers with macro-level decision-making support.

[0148] Example 3 like Figure 3 As shown, an ETC vehicle identity consistency deep verification device includes at least one processor, a memory communicatively connected to the at least one processor, and at least one input / output interface communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the ETC vehicle identity consistency deep verification method described in the foregoing embodiments. The input / output interface may include a display, keyboard, mouse, and USB interface for inputting and outputting data.

[0149] Furthermore, the ETC vehicle identity consistency deep verification device can be a desktop computer, mobile phone, tablet computer, wearable ETC vehicle identity consistency deep verification device, or any other ETC vehicle identity consistency deep verification device capable of deep information recognition.

[0150] Furthermore, the processor may include one or more processing cores. The processor connects to various parts within the ETC vehicle identity consistency deep verification device using various interfaces and lines. It executes various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory, and by calling data stored in memory. Optionally, the processor may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also be implemented separately as a communication chip, without being integrated into the processor.

[0151] The memory may include random access memory (RAM) or read-only memory (ROM). The memory can be used to store instructions, programs, code, code sets, or instruction sets, such as instructions or code sets used to implement the ETC vehicle identity consistency deep verification method provided in this application embodiment. The memory may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for implementing at least one function, instructions for implementing the various method embodiments described above, etc. The data storage area may also store data created during the use of the ETC vehicle identity consistency deep verification device (such as a modulation sequence-depth mapping table, image data, spectrogram data, etc.).

[0152] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory, magnetic disks, or optical disks.

[0153] When the integrated units of the present invention are implemented as software functional units and sold or used as independent products, they can also be stored in a computer-readable storage medium. The computer-readable storage medium stores program code, which can be called by a processor to execute the methods described in the above method embodiments. Based on this understanding, the technical solution of the embodiments of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes electronic memories such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Optionally, the computer-readable storage medium includes a non-transitory computer-readable storage medium. The computer-readable storage medium has storage space for program code that executes any of the method steps described above. This program code can be read from or written to one or more computer program products. The program code can be compressed, for example, in an appropriate form.

[0154] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for deep verification of ETC vehicle identity consistency, characterized in that, Includes the following steps: S1: Obtain real-time data of the vehicle to be verified; the real-time data includes OBU information and image data; S2: Preprocess the real-time data and extract features to output the observed feature vector; S3: Search and match the observed feature vector in the pre-constructed trusted identity profile database; if the match is successful, retrieve the baseline feature vector of the corresponding trusted identity profile or the profile to be verified; if the match fails, store the observed feature vector as the profile to be verified. S4: Calculate the similarity between the observed feature vector and the benchmark feature vector, and generate a consistency score for the observed feature vector; S5: Compare the overall consistency score with a set threshold, and output the identity consistency depth verification result of the vehicle to be verified; the identity consistency depth verification result includes: If the overall consistency score is greater than or equal to the set threshold, it is determined that the binding relationship between the vehicle to be verified and the OBU has not changed. If the overall consistency score is less than the set threshold, it is determined that the binding relationship between the vehicle to be verified and the OBU is abnormal, and the set alarm mechanism is triggered.

2. The method for deep verification of ETC vehicle identity consistency according to claim 1, characterized in that, The real-time data in S1 includes timestamps and location identifiers, and the image data includes multi-angle captured images of the vehicle to be verified.

3. The method for deep verification of ETC vehicle identity consistency according to claim 2, characterized in that, S2 includes the following steps: S21: Preprocess the real-time data to generate a single passage event record; the preprocessing includes data cleaning and spatiotemporal alignment. S22: Extract the static and dynamic information of the OBU from the OBU information, and the static and dynamic visual features from the image data; and calculate the quality assessment score for each feature component based on the confidence level of the license plate OCR and the sharpness of the image data; the quality assessment score includes the image quality pre-score and the feature confidence level; S23: Integrate the various feature components extracted in S22 and output them as the observed feature vector.

4. The method for deep verification of ETC vehicle identity consistency according to claim 3, characterized in that, The image quality pre-scoring includes the following steps: Receive pre-processed captured vehicle images I raw ; For captured vehicle images I raw Perform multi-indicator analysis and calculate sub-scores for each indicator; Normalization and thresholding: The total score S is obtained by weighted summation of all sub-scores. image Its expression is: S image = Σ(w i *s i ) * k weather , Among them, w i Assign weights to indicator i, s i Assign weights to index i, where i∈{sharp,expo,noise}, k weather As a weather impact penalty factor; when S image When the value is less than 0.4, the current image is marked as a "low-quality image", and the current image is deleted or used only for calculating low-weight features.

5. The method for deep verification of ETC vehicle identity consistency according to claim 4, characterized in that, The multi-indicator analysis includes: Sharpness index: Calculated using the variance of the Laplacian operator, its expression is: S sharp = min(σ Laplace / T sharp , 1); Among them, S sharp For the sub-score of the sharpness index, σ Laplace Let T be the variance of the captured vehicle image to be scored after applying the Laplacian operator. sharp This is the preset resolution threshold; Exposure index: Calculated by measuring the ratio of dark to bright areas in the histogram, its expression is: S expo = max(0, 1 - |μ brightness - m target | / d); Among them, S expo For the sub-score of the exposure metric, μ brightness The average value of the brightness of all pixels in the captured image of the vehicle to be scored, μ target The set average brightness is the desired target, and δ is the set maximum brightness fluctuation. Noise level index: calculated based on a three-dimensional block matching algorithm, its expression is: S noise = 1 - (noise estimate / N max ); Among them, S noise For the sub-rating of the noise level index, noise estimate To estimate the noise level, N max The set upper limit for noise tolerance; Weather impact penalty factor: A lightweight classifier is used to determine whether the current image is subject to a set severe weather condition; if it is, let k... weather If it equals the set value, otherwise k weather It equals 1.

6. The method for deep verification of ETC vehicle identity consistency according to claim 4, characterized in that, The calculation of the feature confidence level includes the following steps: Calculate the feature confidence level of the license plate information; the formula for calculating the feature confidence level of the license plate information is: char_confidence i = max(p i ), plate_confidence = prod(char_confidence i ) 1 / N , ， ， Among them, char_confidence i p represents the single-character confidence score of the i-th character. i is the softmax probability vector of the i-th character, i∈[1,N], where N is the total number of characters in the license plate; plate_confidence is the overall OCR confidence of the license plate, and prod() is the multiplication operation; Location confidence level for license plate information. The coordinates of the predicted license plate bounding box. The coordinates of the reference box are given, and Area() is the function to calculate the area of ​​the rectangle. The feature confidence scores for the license plate information are w1=0.5, w2=0.3, and w3=0.

2. Calculate the feature confidence level of vehicle attribute information; the formula for calculating the feature confidence level of vehicle attribute information is: ， ， ， ， ， in, The confidence level for color features. This is the dominant color vector extracted from the current vehicle image. For the pre-established color cluster center vector, The standard deviation of the color cluster; Confidence level of brand characteristics The highest softmax probability value in the brand category is output by the classification network. The confidence level of the model characteristics. The classification network outputs the highest softmax probability value in the vehicle category. The confidence level of the label features. The original line type output of the label sticker before Sigmoid activation; The confidence level of the scratch feature. The original linear output before Sigmoid activation due to scratch damage; Calculate the dynamic weight of the feature credibility of each piece of information; its expression is: ， in, Let be the dynamic weight of the feature credibility of the i-th piece of information. The basic weight for the feature credibility of the i-th piece of information is set to an initial value. Let be the feature credibility of the i-th piece of information, where i, j ∈ [1, n], and n is the number of pieces of information; Based on the feature credibility of each piece of information and its corresponding dynamic weight, a weighted sum is performed to obtain the feature credibility of the current feature.

7. The method for deep verification of ETC vehicle identity consistency according to claim 3, characterized in that, In S3, the retrieval and matching are performed using the feature component with the highest quality assessment score as the primary key. The trusted identity profile database consists of several trusted identity profiles, and the trusted identity profiles include the baseline feature vector, feature stability score, and associated OBU history. The baseline feature vector is a weighted average of several historical observation feature vectors; The feature stability score is the stability score of each feature component in the baseline feature vector; The associated OBU history is used to record all OBU serial numbers that have been bound to the corresponding vehicle and their start and end times.

8. The method for deep verification of ETC vehicle identity consistency according to claim 7, characterized in that, After the file to be verified has been matched a set number of times, the file to be verified will be converted into a trusted identity file.

9. The method for deep verification of ETC vehicle identity consistency according to claim 7, characterized in that, S4 includes the following steps: S41: Calculate the similarity between each corresponding feature component in the observed feature vector and the reference feature vector; S42: Calculate the dynamic weights of the similarity of each feature component; S43: Multiply the similarity of all feature components by their corresponding dynamic weights and sum them to generate a consistency score for the observed feature vector.

10. The method for deep verification of ETC vehicle identity consistency according to claim 9, characterized in that, In step S41, the corresponding similarity calculation method is selected based on the feature type of the current feature component. When the feature component is a numerical vector or an embedded vector, cosine similarity is used as the similarity. When the feature component is a categorical feature, exact matching or fuzzy matching based on category distance is used as the similarity. When the feature component is a license plate number feature, edit distance or Jaro-Winkler distance is used to measure string similarity.

11. The method for deep verification of ETC vehicle identity consistency according to claim 10, characterized in that, In step S42, the dynamic weight = static importance weight ⊙ observation quality dynamic weight ⊙ spatiotemporal correlation dynamic weight; where ⊙ is the Hadamard product; the static importance weight is the base weight, and its initial value is a set value; the observation quality dynamic weight is generated based on the quality assessment score of the current feature component; the spatiotemporal correlation dynamic weight is used to evaluate the contribution of the benchmark feature vector to the current judgment.

12. The method for deep verification of ETC vehicle identity consistency according to claim 11, characterized in that, The spatiotemporal correlation dynamic weights include time decay weights and spatial logical weights; Time decay weight: ; Where Δt is the time difference between historical observation records and the current time. It is the attenuation coefficient of feature i; Spatial logical weights: in, For the k-th consecutive gantry, γ is the spatial logic enhancement coefficient; It is the combined similarity score between two observations, where G(Δt) is a Gaussian function. The time for detection at the k-th detection position. The average time window for a vehicle to pass through two consecutive detection locations. and The minimum and maximum values ​​for a reasonable time window. The standard deviation is denoted as .

13. The method for deep verification of ETC vehicle identity consistency according to claim 12, characterized in that, The expression for the consistency comprehensive score in S43 is: in, For consistency score, Let N be the final weight of the i-th feature component, and N be the number of feature components. Let be the feature similarity of the i-th feature component.

14. The method for deep verification of ETC vehicle identity consistency according to claim 12, characterized in that, The threshold set in S5 is adaptively adjusted based on vehicle type, road segment type, and vehicle historical behavior, including the following steps: Generate a unique identifier for each vehicle according to the established business rules; A dynamic threshold is established based on the historical data of each unique identifier using the sliding window statistical method. The system acquires the current consistency score in real time and compares it with the threshold. When alarms are triggered continuously or not for a long time, the system automatically adjusts the sensitivity parameters of the dynamic threshold and the detection window size according to the set rules.

15. The method for deep verification of ETC vehicle identity consistency according to claim 14, characterized in that, It also includes S6: When the overall consistency score is greater than or equal to a set threshold, a smooth averaging or Kalman filtering method is used to add the currently observed feature vector to the trusted identity file of the corresponding vehicle for dynamic updating.

16. A deep verification system for ETC vehicle identity consistency, characterized in that, A method for performing deep verification of ETC vehicle identity consistency according to any one of claims 1 to 15; the system includes a data acquisition layer and a data processing and analysis layer that are interconnected. The data acquisition layer is deployed at every ETC gantry and ETC lane at toll stations along the highway to acquire real-time data of vehicles to be verified. The data processing and analysis layer is used to perform in-depth identity consistency verification on the vehicle to be verified based on the real-time data.

17. An ETC vehicle identity consistency deep verification device, characterized in that, It includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform an ETC vehicle identity consistency deep verification method according to any one of claims 1 to 15.

Images