Multi-source data fusion and traffic mechanism driven bayonet equipment closed-loop evaluation method

Abstract

The invention discloses a multi-source data fusion and traffic mechanism driven checkpoint equipment closed-loop evaluation method, relates to the technical field of checkpoint equipment evaluation, and solves the technical problem of poor evaluation effect of existing checkpoint equipment. The method comprises the following steps: constructing an upstream equipment mapping table of the bayonet equipment based on road network information and position information of the bayonet equipment; verifying the historical traffic mode of the gate based on the historical traffic data, calculating the real-time accuracy of the gate device based on the floating car trajectory data, and finally comparing the traffic conditions of the to-be-evaluated gate device and the upstream gate device through the upstream device mapping table to judge whether the traffic is conserved or not; determining a mistakenly shot license plate number according to a license plate verification rule; the real license plate number of the mistakenly shot license plate number, the association algorithm and the hardware information are determined by using similar vehicles in the upstream and downstream vehicles, factors which can influence the equipment effect are analyzed, and improvement measures are adopted in a targeted manner; according to the method, environment interference, algorithm output characteristics and hardware state parameters are associated to form a closed-loop feedback mechanism.

Description

Technical Field

[0001] This invention relates to the field of checkpoint equipment evaluation technology, specifically to a closed-loop evaluation method for checkpoint equipment driven by multi-source data fusion and traffic mechanism. Background Technology

[0002] Existing checkpoint equipment evaluation technologies have significant technical bottlenecks in terms of data dimensions, traffic feature mining, and fault diagnosis accuracy, specifically manifested in the following limitations:

[0003] I. A single data source leads to a one-sided evaluation dimension.

[0004] Existing methods mostly rely on a single type of data to build evaluation models, which makes it difficult to cover the multi-dimensional verification needs of device performance.

[0005] 1. Limitations of historical statistics and upstream / downstream correlation: For example, the Chinese patent "A method for judging checkpoint equipment failure based on vehicle passage data quality" (patent application number: CN202311668966.8, publication number: CN117746667A) mainly relies on the historical mean and variance of vehicle passages at a single checkpoint and the correlation of traffic flow between upstream and downstream checkpoints for failure judgment. However, in real-world scenarios, sudden traffic events can cause historical data distribution to become invalid, easily misjudging normal traffic fluctuations as equipment failure.

[0006] 2. The inherent defects in the coverage of positioning data: For example, the Chinese patent "Data Quality Inspection Method of Checkpoint Equipment Based on Positioning Data" (patent application number: CN201710115608.2, publication number: CN106940931B) introduces floating car positioning data to evaluate the accuracy of checkpoints, but it relies on the coverage of positioning equipment of specific vehicles such as taxis and buses (i.e. "floating car penetration rate").

[0007] II. Missing latent faults due to incomplete traffic mechanism modeling

[0008] Current technologies lack in-depth analysis of vehicle travel patterns and the spatiotemporal correlations of road networks, resulting in an inability to effectively identify hidden device anomalies in complex scenarios.

[0009] 1. Lack of a travel chain integrity verification mechanism: Parking lot entry and exit data are not integrated;

[0010] 2. Lack of individual travel pattern analysis: The lack of modeling for daily vehicle travel patterns makes it impossible to identify abnormal snapshots that deviate from the norm, and existing technologies cannot indicate the possibility of false snapshots based on commuting patterns.

[0011] Third, the detection of incorrect snapshots remains at the level of rule verification and lacks guidance for algorithm optimization.

[0012] Existing methods for detecting incorrect images only filter surface features, making it difficult to pinpoint the root cause of technical defects and thus unable to achieve a closed-loop business process.

[0013] 1. Superficial verification of national standard rules: Most solutions only remove records whose license plate numbers do not conform to the GA-36-2018 rule, but lack targeted analysis for high-frequency mis-capture scenarios such as "misidentification of similar characters" and "missed detection of partially obscured license plates", and cannot provide effective feedback for feature extraction and classifier optimization of license plate recognition algorithms.

[0014] 2. Ambiguity in attributing incorrect image types: Only the proportion of incorrect images to the total number of images is calculated, without distinguishing specific causes such as equipment hardware failure, algorithm defects, or environmental interference. Current technology cannot correlate environmental variables with failure modes, making it difficult for maintenance personnel to accurately formulate debugging strategies. Summary of the Invention

[0015] To address the problems existing in the prior art, this invention provides a closed-loop evaluation method for checkpoint equipment based on multi-source data fusion and traffic mechanism-driven approaches. This method solves the technical problem that the prior art does not comprehensively consider the data dimensions of the data source, traffic feature mining, and the impact of fault diagnosis accuracy on the evaluation effect when evaluating checkpoint equipment.

[0016] A closed-loop evaluation method for checkpoint equipment driven by multi-source data fusion and traffic mechanism includes the following steps:

[0017] Step 1: Construct an upstream device mapping table for the checkpoint devices based on road network information and checkpoint device location information;

[0018] Step 2: Verify the historical traffic pattern of the checkpoint based on historical traffic data, calculate the real-time accuracy of the checkpoint equipment based on floating car trajectory data, and finally compare the traffic situation of the checkpoint equipment to be evaluated with that of the upstream checkpoint equipment through the upstream equipment mapping table to determine whether the traffic is conserved; if all the above verifications pass, skip to step 3; if they fail, directly determine that the equipment is abnormal and skip to step 6.

[0019] Step 3: Verify the format of the license plate numbers collected by the checkpoint equipment according to the license plate verification rules to obtain normal license plate numbers and abnormal license plate numbers;

[0020] Step 4: Combine spatiotemporal rules and travel patterns to further determine the wrongly photographed license plate number by judging normal and abnormal license plate numbers;

[0021] Step 5: Determine the real license plate number of the wrongly photographed license plate number by using similar vehicles in the upstream and downstream vehicles, and associate the algorithm and hardware information;

[0022] Step 6: Analyze the real-time evaluation results from Step 2, and then analyze the OCR error types obtained in Step 5; analyze the degree of influence of external factors on the device's recognition effect, and take targeted improvement measures.

[0023] Further, step 1 includes:

[0024] Step 1.1: Construct a road network map and equipment mapping: Construct a directed road network map using road data. Nodes in the directed road network map are the starting and ending points of roads, and edges represent road segments. The attributes of the edges include the segment number and the road length.

[0025] Step 1.2: Load device data, establish a mapping from checkpoint device number to road segment number, and construct a reverse mapping for subsequent lookup;

[0026] Step 1.3: Construct an upstream device mapping table: By traversing the road network map in reverse, search for the upstream road segments that each checkpoint device can reach, calculate the path distance between the current road segment and the upstream road segment, filter out the upstream devices whose distance is within the maximum search distance, construct an upstream device mapping table, and record the upstream device and distance corresponding to each device.

[0027] Further, step 2 includes:

[0028] Step 2.1, Verification based on historical traffic patterns at checkpoints: First, aggregate historical traffic data by hourly granularity to construct a time series and calculate the fluctuation threshold of historical data; monitor the current traffic in real time and compare it with the distribution of historical values ​​for the same period. If the deviation exceeds the threshold for multiple consecutive periods, the checkpoint equipment may be abnormal.

[0029] Step 2.2, Accuracy assessment based on floating car trajectory data: Statistically analyze the data hourly, compare the license plate numbers of floating cars passing through the checkpoint equipment with the license plate numbers identified by the checkpoint equipment, calculate the real-time accuracy rate of the checkpoint equipment, and judge the corresponding checkpoint equipment to be abnormal if the real-time accuracy rate is lower than the minimum standard.

[0030] Step 2.3, Traffic Conservation Verification: Count the traffic flow of each checkpoint device on an hourly basis. For each checkpoint device, obtain the traffic information of its upstream device according to the upstream device mapping table, calculate the expected traffic considering traffic propagation, and judge whether the traffic is conserved by comparing the actual traffic of the current device with the expected traffic. If the difference exceeds the allowable fluctuation range, the checkpoint device is considered to be abnormal.

[0031] Further, step 3 includes:

[0032] Step 3.1, License Plate Format Verification: Use the GA-36-2018 industry standard to verify the format of the license plate number. The verification content includes the length of the license plate number, the abbreviation of the province, whether it contains prohibited characters, and whether it conforms to the rules of ordinary license plates or new energy vehicle license plates, which are divided into normal license plate numbers and abnormal license plate numbers.

[0033] Step 3.2, Statistics on Abnormal License Plate Numbers: Statistically record the number of times each abnormal license plate number was captured by different checkpoints for the license plate numbers that were verified as abnormal. For abnormal license plate numbers that appeared only once, they were directly determined as miscaptured license plate numbers. For abnormal license plate numbers that appeared multiple times, due to the possible influence of historical national standards and other factors, it is necessary to further combine spatiotemporal rules and travel patterns for judgment.

[0034] Step 3.3, Processing of Normal License Plate Numbers: Statistically record the number of times each normal license plate number is captured by different checkpoints for license plate numbers that are verified as normal. For normal license plate numbers that appear only once, judge them based on travel patterns. If the travel patterns are not met, they are considered as mistakenly captured license plate numbers. For normal license plate numbers that appear multiple times, judge them based on spatiotemporal rules and travel patterns. If the conditions are not met, they are considered as mistakenly captured license plate numbers.

[0035] Further, step 4 includes:

[0036] Step 4.1, Consistency judgment of travel time and traffic conditions: Introduce real-time congestion index data, establish a travel time distribution model, combine the time of the vehicle between the two checkpoints, and use the 3σ principle to judge whether there is an anomaly. If there is an anomaly, it is considered that the travel time is inconsistent with the traffic conditions, and it is regarded as a wrong license plate number.

[0037] Step 4.2, Route Reversal Judgment: Check the nodes passed by the vehicle during the journey to determine if there is a route reversal. If the nodes passed by the vehicle during the journey are repeated, it is considered that there is an unreasonable route reversal.

[0038] Step 4.3, Determining Travel Patterns: Combine parking lot data and vehicle travel patterns to make judgments, analyze information such as vehicle travel time, route, and stopping locations, and suggest spatiotemporal probability distributions; determine whether the probability of the vehicle currently appearing is greater than 0.05, if not, consider the license plate number to be a mistake.

[0039] Further, step 5 includes:

[0040] Step 5.1, Data Preprocessing: Merge the vehicle record data considered as mis-captured with the device-road segment mapping data, and convert the time field to datetime type for subsequent processing;

[0041] Step 5.2: Find candidate license plates: For each wrongly captured license plate record, find the upstream and downstream related road segments according to the upstream device mapping table, and set a time window based on the road segment length and time buffer. Select vehicle records that pass through the upstream and downstream related road segments within the time window and are not captured by the current device as candidate license plates.

[0042] Step 5.3: Calculate similarity: For each wrongly photographed license plate, calculate its similarity score with the candidate license plates. The similarity score takes into account the overall similarity of the license plates and the matching of the province abbreviation. Select the candidate license plates with similarity scores greater than or equal to the threshold, sort them in descending order of score, and take the top 5 as possible real license plate numbers.

[0043] Step 5.4: Information Association: Record information such as wrongly captured license plates, candidate license plates, similarity scores, and device numbers, and associate them with information such as environmental interference, algorithm and hardware conditions to facilitate targeted improvements in the future.

[0044] Further, step 6 includes:

[0045] 6.1 Analysis of the real-time evaluation results in step 2: The time series data of the real-time evaluation results obtained in step 2 are preprocessed and standardized, and then clustered into periodic and random devices using a clustering algorithm. The device accuracy is correlated with time and external environmental factors, and the relationship is fitted by a random forest model. Finally, the SHAPA method is used to interpret and determine the degree of influence of external factors on the device accuracy, and targeted improvements are made.

[0046] 6.2 OCR error type analysis obtained in step 5: Based on the correlation information obtained in step 5, construct the device recognition feature vector and use the cluster analysis method to classify the devices according to their ability to recognize different characters. Then, use radar charts to display the recognition capabilities of different types of devices in a multidimensional way and make targeted improvements.

[0047] The beneficial effects of this invention include:

[0048] 1. Multimodal data fusion: Integrate checkpoint data, floating car trajectories, parking lot entry and exit records, and real-time congestion index to build an evaluation system covering the entire vehicle lifecycle, reducing dependence on a single data source.

[0049] 2. Multi-level mechanism modeling: Construct a progressive analysis framework of "spatiotemporal continuity correction - travel chain integrity analysis - individual pattern matching", and combine road network topology, real-time traffic conditions and user profiles to improve the ability to identify hidden wrong shots (such as instantaneous crossing of districts and violation of patterns).

[0050] 3. In-depth diagnostic system for incorrect captures: Establish a classification model for incorrect capture types (such as character misidentification, misjudgment of counterfeit license plates, and missed detection due to environmental interference), correlate environmental interference, and form a closed-loop feedback mechanism by combining algorithm output features with hardware status parameters. Attached Figure Description

[0051] Figure 1 This is a flowchart of a closed-loop evaluation method for checkpoint equipment driven by multi-source data fusion and traffic mechanism, which is involved in an embodiment of this application.

[0052] Figure 2 shows the device identification accuracy curves involved in the embodiments of this application. Figure 2(a) shows the identification accuracy curve of random devices, and Figure 2(b) shows the identification accuracy curve of periodic devices.

[0053] Figure 3 is an explanation diagram of the device shap involved in the embodiments of this application. Figure 3(a) is an explanation diagram of the periodic device shap, and Figure 3(b) is an explanation diagram of the random device shap.

[0054] Figure 4 This is a clustering diagram of OCR recognition capabilities involved in the embodiments of this application.

[0055] Figure 5 shows the daily variation of OCR recognition capability clusters involved in the embodiments of this application. Figure 5(a) shows the daily variation of OCR recognition capability cluster 1, Figure 5(b) shows the daily variation of OCR recognition capability cluster 2, Figure 5(c) shows the daily variation of OCR recognition capability cluster 3, and Figure 5(d) shows the daily variation of OCR recognition capability cluster 4. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0057] Example 1

[0058] A closed-loop evaluation method for checkpoint equipment driven by multi-source data fusion and traffic mechanism, such as... Figure 1 As shown, it includes the following steps:

[0059] Step 1: Data Acquisition and Preprocessing

[0060] 1.1 Obtain vehicle record data, including the time the vehicle passed through the checkpoint, license plate number, checkpoint equipment number, and other information.

[0061] 1.2 Obtain equipment data, including the mapping relationship between equipment number and road segment number.

[0062] 1.3 Obtain road data, including the starting point node, ending point node, road segment number, and road length.

[0063] 1.4 Obtain equipment location data, including equipment number, longitude, and latitude.

[0064] 1.5. Obtain license plate rule data and load the license plate verification rules for subsequent verification of the license plate format. The rules are stored in the form of a dictionary, where the key is the length of the license plate content and the value is a list of corresponding regular expressions.

[0065] 1.6 Constructing a road network map and equipment mapping: Construct a directed road network map using the acquired road data, where nodes are the start and end points of roads, edges represent road segments, and the attributes of the edges include the segment number and road length.

[0066] 1.7 Load device data, establish a mapping from device number to road segment number, and construct a reverse mapping for subsequent lookup.

[0067] 1.8 Constructing an upstream device mapping table: By traversing the road network map in reverse, search for the upstream road segments that each device can reach, calculate the path distance between the current road segment and the upstream road segment, filter out the upstream devices whose distance is within the maximum search distance, construct an upstream device mapping table, and record the upstream device and distance corresponding to each device.

[0068] 1.9 Obtain the floating car's GPS trajectory data, retaining fields such as license plate number, time, and latitude / longitude coordinates.

[0069] Step 2, Real-time Assessment

[0070] 2.1 Verification based on historical traffic patterns at checkpoints: First, aggregate the historical traffic data obtained by the checkpoint equipment itself at the hourly granularity, construct a time series, and calculate the fluctuation threshold of the historical data; monitor the current traffic in real time and compare it with the distribution of the historical values ​​for the same period. If the deviation exceeds the threshold for multiple consecutive periods, the equipment may be marked as abnormal (i.e., there is a problem with the equipment, and maintenance personnel need to conduct on-site inspection).

[0071] The specific implementation process is as follows:

[0072] 2.1.1 Time Series Construction (Historical Data Aggregation)

[0073] Historical traffic data at checkpoints is aggregated at the hourly granularity to construct a time series:

[0074] Historical time series: ,in Indicates the first Hourly traffic volume This represents the total number of hours in the historical data.

[0075] Synchronous time series: for the current moment The set of its historical data for the same period (such as the same working day or hour) is as follows: .

[0076] 2.1.2 Fluctuation threshold calculation (based on historical distribution)

[0077] Based on historical data from the same period Calculate the dynamic threshold based on the statistical distribution of the data.

[0078] Definition of statistic:

[0079] Mean:

[0080] Standard deviation:

[0081] in This represents the sample size for data from the same period.

[0082] Dynamic threshold:

[0083]

[0084] in, The upper limit of the threshold, This is the lower limit of the threshold. This is the sensitivity parameter (taken as 3).

[0085] 2.1.3 Real-time monitoring and anomaly detection

[0086] Current traffic: (t represents the current hour)

[0087] Continuous over-limit condition: If k consecutive periods (e.g., k=3) satisfy the following:

[0088]

[0089] Then mark the device as faulty.

[0090] 2.2 Accuracy assessment based on floating car trajectory data: Statistical analysis is performed hourly, comparing the license plate numbers of floating cars passing through the checkpoint with the license plate numbers identified by the checkpoint, and calculating the real-time accuracy rate of the checkpoint equipment. The real-time accuracy rate is an indicator of the normality of equipment operation and is used to judge the normality of equipment operation. The higher the accuracy rate, the more normal the equipment is. If the accuracy rate is lower than the minimum standard, the equipment is considered abnormal. The minimum standard value is set based on expert experience.

[0091] The specific implementation is as follows:

[0092] 2.2.1 Coordinate System Transformation Model

[0093] The original trajectory data of the floating car (WGS84 coordinate system) needs to be converted to the GCJ02 coordinate system suitable for map matching:

[0094]

[0095]

[0096]

[0097]

[0098] in: , It is a nonlinear transformation function (including trigonometric series expansion); The radius of the Earth; The value is in radians representing latitude. Let be the eccentricity of the ellipsoid.

[0099] 2.2.2 Mathematical Model for Trajectory Matching

[0100] 1. Road network topology modeling

[0101] Node set The geometric coordinates are geom;

[0102] Road Section Collection The geometric path is

[0103] LineString(from_node,to_node).

[0104] 2. Hidden Markov Model (HMM) Matching

[0105] Observation probability: GPS point Projected onto candidate road segment points The probability of:

[0106]

[0107] State transition probability: continuous projection points Path rationality:

[0108]

[0109] Optimal path solution: Maximize using Viterbi's algorithm or bidirectional Dijkstra's algorithm:

[0110]

[0111] 2.2.3 Assessment of Checkpoint Recognition Accuracy

[0112] 1. Spatiotemporal matching conditions

[0113] floating car trajectory points With checkpoint data

[0114]

[0115] 2. Accuracy Calculation

[0116]

[0117] in: The number of floating cars identified by the checkpoint; It is the total number of floating cars passing through the road segment.

[0118] 2.3 Traffic Conservation Verification: Traffic flow for each device is calculated hourly. For each device, traffic information from its upstream devices is obtained from the upstream device mapping table, and the expected traffic flow is calculated considering traffic propagation. By comparing the actual traffic flow of the current device with the expected traffic flow, it is determined whether the traffic flow is conserved. If the difference exceeds the allowable fluctuation range, the traffic flow is considered abnormal. Abnormal traffic flow indicates a potential problem with the device, requiring on-site personnel intervention.

[0119] The three verification results in step 2 are parallel, representing a real-time evaluation of the device from three different perspectives.

[0120] If any of the three assessment results is not met, it indicates an equipment malfunction, requiring manual inspection. Subsequent factor analysis was conducted to assess the accuracy of the floating car trajectory data, in order to achieve a closed-loop process.

[0121] The specific implementation is as follows:

[0122] 2.3.1 Screening of flow conservation verification equipment

[0123] Each flow conservation verification device is paired with It is an ordered pair ,in It is an entrance device. It is exported equipment. It meets the requirements from... arrive The path segment (composed of a sequence of road segments) has no additional inflow or outflow paths (i.e., the path segment is a closed system, and vehicles can only flow from...). Enter and from (Leaving). Specifically: Traverse the road network graph G in reverse, using depth-first search, and calculate the value of each device. Upstream equipment set And path distance. Filter by path distance within the maximum search distance. Upstream equipment within ( (This is a preset threshold) for the device If there is an upstream device Make from arrive If the path is unique (i.e., there are no branch nodes in the path), then To construct a flow conservation verification device .

[0124] 2.3.2 Hourly traffic flow statistics

[0125] Based on the vehicle record data (step 1.1), we statistically analyzed the impact of each flow conservation verification device on... The inbound and outbound traffic within each hour. First, define the time period, dividing time into hour intervals, t=1,2,…,T (for example, t represents the t-th hour of the day).

[0126] Inbound traffic : Passing through the inlet device within hour t Number of vehicles

[0127]

[0128] Export flow : Passing through the inlet device within hour t Number of vehicles

[0129]

[0130] in:

[0131] VehicleRecords is a dataset of vehicle records (provided in step 1.1). Each record contains the following fields: time, plate_number, and device_id.

[0132] [t, t+1) represents the time interval from the t-th hour to the t+1-th hour.

[0133] 2.3.3 Determining Flow Conservation and Anomaly Detection

[0134] For each flow conservation verification device Calculate the flow difference for each hour t:

[0135]

[0136] This represents the absolute difference between inflow and outflow. If the difference exceeds the allowable fluctuation range, it is considered abnormal. The allowable fluctuation range is defined by the threshold δ.

[0137] Step 3: License Plate Type Check

[0138] The GA-36-2018 industry standard is used to verify the format of license plate numbers. The verification includes the length of the license plate number, the abbreviation of the province, whether it contains prohibited characters, and whether it conforms to the rules for ordinary license plates or new energy vehicle license plates, which are divided into normal license plate numbers and abnormal license plate numbers.

[0139] The specific verification rules are as follows:

[0140] I. General Rules

[0141] Province abbreviation requirements: The first character of the license plate must be the province abbreviation specified in the "Administrative Division Code of the People's Republic of China" (Beijing, Tianjin, Hebei, Shanxi, Inner Mongolia, Liaoning, Jilin, Heilongjiang, Shanghai, Jiangsu, Zhejiang, Anhui, Fujian, Jiangxi, Shandong, Henan, Hubei, Hunan, Guangdong, Guangxi, Hainan, Chongqing, Sichuan, Guizhou, Yunnan, Tibet, Shaanxi, Gansu, Qinghai, Ningxia, Xinjiang).

[0142] Character restrictions: The letters O and I are prohibited (to avoid confusion with the numbers 0 and 1); Chinese characters, letters AH / JN / PZ, and numbers 0-9 are allowed; only the middle dot character is allowed as a separator.

[0143] II. Ordinary vehicle license plates

[0144] Basic structure: [Province abbreviation][License issuing authority code][Serial number]

[0145] Total length: 7 characters (4 for the sequence number) or 8 characters (5 for the sequence number)

[0146] Numbering rules (4 digits):

[0147] Serial Number Type Regular expressions Example Pure numbers ^\d{4}$ 1234 1 letter + 3 numbers ^[A-HJ-NP-Z]\d{3}$ A123 2 letters + 2 numbers ^[A-HJ-NP-Z]{2}\d{2}$ AB12 1 number + 1 letter + 2 numbers ^\d[A-HJ-NP-Z]\d{2}$ 1A23 2 numbers + 1 letter + 1 number ^\d{2}[A-HJ-NP-Z]\d$ 12A3 3 numbers + 1 letter ^\d{3}[A-HJ-NP-Z]$ 123A Cross combination 1 ^[A-HJ-NP-Z]\d[A-HJ-NP-Z]\d$ A1B2 Cross combination 2 ^\d[A-HJ-NP-Z]{2}\d$ 1AB2 Cross combination 3 ^[A-HJ-NP-Z]\d{2}[A-HJ-NP-Z]$ A12B Cross combination 4 ^\d[A-HJ-NP-Z]\d[A-HJ-NP-Z]$ 1A2B 2 numbers + 2 letters ^\d{2}[A-HJ-NP-Z]{2}$ 12AB

[0148] Numbering rules (5 digits):

[0149] Serial Number Type Regular expressions Example Pure numbers ^\d{5}$ 12345 1 letter + 4 numbers ^[A-HJ-NP-Z]\d{4}$ A1234 2 letters + 3 numbers ^[A-HJ-NP-Z]{2}\d{3}$ AB123 1 number + 1 letter + 3 numbers ^\d[A-HJ-NP-Z]\d{3}$ 1A234 2 numbers + 1 letter + 2 numbers ^\d{2}[A-HJ-NP-Z]\d{2}$ 12A34 3 numbers + 1 letter + 1 number ^\d{3}[A-HJ-NP-Z]\d$ 123A4 4 numbers + 1 letter ^\d{4}[A-HJ-NP-Z]$ 1234A 1 letter + 3 numbers + 1 letter ^[A-HJ-NP-Z]\d{3}[A-HJ-NP-Z]$ A123B 3 numbers + 2 letters ^\d{3}[A-HJ-NP-Z]{2}$ 123AB Cross combination 1 ^[A-HJ-NP-Z]\d[A-HJ-NP-Z]\d{2}$ A1B23 Cross combination 2 ^\d[A-HJ-NP-Z]{2}\d{2}$ 1AB23 Cross combination 3 ^[A-HJ-NP-Z]\d{2}[A-HJ-NP-Z]\d$ A12B3 Cross combination 4 ^\d[A-HJ-NP-Z]\d[A-HJ-NP-Z]\d$ 1A2B3 Cross combination 5 ^\d[A-HJ-NP-Z]\d{2}[A-HJ-NP-Z]$ 1A23B Cross combination 6 ^\d{2}[A-HJ-NP-Z]{2}\d$ 12AB3 Cross combination 7 ^\d{2}[A-HJ-NP-Z]\d[A-HJ-NP-Z]$ 12A3B

[0150] III. License Plates for Large New Energy Vehicles

[0151] Basic structure [Province Abbreviation][License Issuing Authority Code][5-Digit Serial Number][Type Letter] The total length is 8 characters, with the last character indicating the type letter position. pure electric vehicles Type letters: D, A, B, C, E Numbering rules: The number part must be 5 digits. Non-pure electric vehicles Type letters: F, G, H, J, K Numbering rules: The number part must be 5 digits.

[0152] IV. License Plates for Small New Energy Vehicles

[0153] Basic structure:

[0154] [Province Abbreviation][License Issuing Authority Code][Type Letter][5-Digit Serial Number]

[0155] The total length is 8 characters, with the 3rd character indicating the type letter position.

[0156] Type letters: D, A, B, C, E (pure electric vehicles), F, G, H, J, K (non-pure electric vehicles)

[0157] Numbering rules (two formats):

[0158] Format type Regular rule Example Pure numeric type ^Type letter\d{5}$ Beijing AD12345 Letter + numeric type ^Type letter[A-HJ-NP-Z]\d{4}$ Beijing ADA1234

[0159] V. Special license plates

[0160]

[0161] Step 4: Error Analysis

[0162] 4.1 Consistency judgment of travel time and traffic status, handling objects including: abnormal license plate numbers appearing multiple times, and normal license plate numbers appearing multiple times. Since a single license plate number cannot be analyzed using cross-checkpoint spatiotemporal rules, a single abnormal license plate number is directly regarded as a mistakenly captured license plate number. The specific implementation is as follows:

[0163] a. Travel route calculation and checkpoint matching, including: calculating the shortest path between checkpoints based on the road network topology using Dijkstra's or A* algorithm. Input parameters include road segment length and real-time traffic condition weights (such as congestion index). The output is the theoretical travel path for the vehicle. Checkpoint data matching: License plate recognition technology is used to obtain the time it takes for a vehicle to pass through an upstream checkpoint. and downstream checkpoint time ,

[0164] Calculate the actual travel time.

[0165]

[0166] b. Calculation of travel time based on congestion index, including:

[0167] Free-flow time of road segment ( ):

[0168]

[0169] in For the length of the road segment, Free-flow speed (ideal vehicle speed when there is no congestion)

[0170] Congestion Index ( ) and predicted travel time ( ):

[0171]

[0172] c. Consistency assessment of travel time differences, including:

[0173] Calculate the time difference ( This reflects the deviation between the actual travel time and the model's predicted time.

[0174]

[0175] Error distribution modeling: Collect historical data on all vehicles passing through the same checkpoint. Construct a normal distribution ,like Exceeding The range was determined to be an abnormal trip.

[0176] 4.2 Path reversal judgment, the processing objects include: abnormal license plate numbers that appear multiple times, and normal license plate numbers that appear multiple times; specifically, license plate numbers that appear only once cannot be analyzed for cross-checkpoint spatiotemporal rules, and abnormal license plate numbers that appear only once are directly regarded as wrongly captured license plate numbers.

[0177] The specific implementation is as follows:

[0178] a. Formula for calculating the direction angle

[0179] For three consecutive points on the path :

[0180] The direction angle of line segment AB :

[0181]

[0182] The direction angle of line segment BC :

[0183]

[0184] in:

[0185] arctan2(Δy,Δx) is the four-quadrant arctangent function, with an output range of (−π,π] radians, ensuring accurate direction representation.

[0186] b. Angle of directional change Calculation

[0187] Difference formula:

[0188]

[0189] Normalization

[0190] In order to make If it falls within [−180∘, 180∘], adjustments are needed:

[0191]

[0192] c. Determination of directional change

[0193] Threshold condition: If If | (β is a preset threshold, taken as 90) it is determined to be a directional change.

[0194] 4.3 Determine travel patterns and process all normal license plate numbers and abnormal license plate numbers that appear multiple times.

[0195] a. Spatiotemporal feature extraction

[0196] Time distribution probability:

[0197]

[0198] Spatial distribution probability:

[0199]

[0200] b. Bayesian probabilistic fusion

[0201]

[0202] like It is then marked as a suspected misshot.

[0203] Step 5: Mistake Photo Restoration. For all suspected mistaken license plates from Step 4, calculate their similarity to upstream and downstream vehicles as possible real license plate numbers, and associate them with algorithm and hardware information, etc.

[0204] To achieve a closed-loop business logic at the algorithm level, it's essential to find the actual license plate number to analyze which characters were misrecognized. For example, if the license plate number is mistakenly captured as AD123456, finding the actual license plate number as A0123456 is crucial to identifying the incorrect recognition types of "0" and "D," allowing for targeted optimization of the license plate recognition algorithm.

[0205] 5.1 Data Preprocessing: Merge the vehicle record data that is considered to be mistakenly captured with the device-road segment mapping data obtained in step 1.2, and convert the time field to datetime type for subsequent processing.

[0206] 5.2 Searching for candidate license plates: For each incorrectly captured license plate record, search for upstream and downstream related road segments based on the road segment it is located on. Combine the road segment length and time buffer to set a time window, and filter out vehicle records that pass through the upstream and downstream related road segments within the time window and were not captured by the current device as candidate license plates.

[0207] 5.3 Calculate Similarity: For each incorrectly photographed license plate, calculate its similarity score with candidate license plates. The similarity score considers the overall similarity of the license plates and the matching of the province abbreviation. Select candidate license plates with a similarity score greater than or equal to the threshold, sort them in descending order of score, and take the top 5 as possible real license plate numbers. The similarity calculation formula is as follows.

[0208]

[0209] in It is to put the string Convert to string The minimum number of single-character edits (insert, delete, replace) required. , Representing strings respectively and Length, This indicates taking the maximum value.

[0210] 5.4. Related Information: Record information such as wrongly captured license plates, candidate license plates, similarity scores, and device numbers, and associate them with information such as environmental interference, algorithm and hardware conditions to facilitate targeted improvements in the future.

[0211] Identifying incorrectly captured license plate numbers allows for the calculation of the error rate of checkpoint equipment, a key performance indicator. Correcting incorrectly captured license plate numbers can reveal common OCR recognition error types, such as D0 and S5. Further analysis, by considering meteorological factors, equipment hardware information (e.g., age), and equipment software information (e.g., OCR algorithm version), can reveal the impact of these factors on the final recognition accuracy. This allows for consideration of whether to address the issue at the operational level.

[0212] 1. Replace equipment that has been in use for a long time;

[0213] 2. Install supplementary lighting for devices that have poor nighttime recognition performance;

[0214] 3. Install rain covers on devices that do not perform well in the rain;

[0215] 4. Update the OCR recognition algorithm to address common OCR recognition errors.

[0216] Step 6: Closed-loop analysis

[0217] 6.1 Analyze the real-time evaluation results of step 2.

[0218] (1) Data preprocessing and standardization

[0219] MinMaxScaler is used to normalize the data to the [0,1] interval, eliminating dimensional differences and preserving the shape characteristics of the time series.

[0220] (2) Clustering algorithm and parameter optimization

[0221] K-Shape clustering is used to cluster time series based on the cosine distance metric for shape similarity. The number of clusters is determined by the Elbow Method to balance intra-cluster distance and computational complexity; as shown in Figure 2.

[0222] Figure 2 is a visualization of two typical clusters of the device accuracy curves, where the horizontal axis represents time (two days in total) and the vertical axis represents the device accuracy; each curve in the figure represents a device. It can be seen that the device accuracy in Figure 2(a) is similar to random fluctuation, while the device accuracy in Figure 2(b) shows obvious time periodic fluctuations (higher accuracy during the day and lower accuracy at night).

[0223] (3) Classification processing based on clustering results

[0224] Figure 3 shows the interpretable results of the accuracy fluctuations of the two types of equipment in Figure 2. This was achieved by correlating equipment accuracy with factors such as time and weather, fitting the relationships using a random forest model, and finally interpreting the results using the SHAPA method.

[0225] The results show that for periodic equipment, i.e. equipment whose accuracy exhibits obvious time-period fluctuations (Figure 2(b)), solar radiation is the main factor affecting its accuracy. Strong solar radiation results in high accuracy, while low solar radiation results in low accuracy. Solar radiation is strongly dependent on lighting conditions, and it can be assumed that the lighting conditions for periodic equipment are insufficient. Therefore, supplementary lighting lamps need to be installed at night.

[0226] For randomized devices, where the accuracy of the device resembles random fluctuations (Figure 2(a)), total_cases (the number of floating cars passing through the device) is the main influencing factor, which may be due to statistical errors caused by the sample size. In addition, wind_speed is the second largest factor causing fluctuations in the accuracy of randomized devices. Higher wind speeds result in lower device accuracy, while lower wind speeds result in higher accuracy, indicating that randomized devices require additional fixed installations.

[0227] 6.2 Analysis of OCR error types obtained in step 5

[0228] (1) Construction of device identification feature vector

[0229] Group by device ID and character, calculate the total number of recognitions (size) and the number of correct recognitions (sum), and generate the recognition rate; generate a [database name] for each device. dimensional vector ( (This represents the number of high-frequency characters), and missing values ​​are filled with 0 (indicating that the character was not recognized).

[0230] (2) Cluster analysis method

[0231] Min-Max scaling is used to map the recognition rate to the [0,1] interval, eliminating the impact of dimensional differences on distance calculation;

[0232] The k-means++ algorithm is used to initialize the centroids to avoid local optima caused by random initialization.

[0233] The devices are divided into k clusters based on Euclidean distance, and the objective function is to minimize the in-cluster sum of squares (WCSS).

[0234]

[0235] Where S represents the device partitioning method, Sᵢ represents the i-th cluster (i-th type of device) partitioning; x represents the identification feature vector of each device; μᵢ is the centroid of cluster Sᵢ, and is the mean of the identification feature vectors of all devices in the cluster.

[0236] (3) Multi-dimensional display of radar chart

[0237] Polar coordinate transformation: Mapping the d-dimensional feature space to a circular coordinate system, where each axis represents the recognition rate of a character.

[0238] Cluster center rendering: Calculate the average recognition rate of each cluster, generate closed polygons and fill them with semi-transparent color bands, and visually compare the capability patterns of different clusters: (Specific details are as follows...) Figure 4 As shown in Figure 5.

[0239] Figure 4 The results indicate that the devices can be categorized into four types based on their character recognition capabilities. The second type of devices (cluster 2) shows significantly insufficient recognition capability for the character 'T', while the third type lacks sufficient recognition capability for characters '6', '8', and '5', requiring targeted fine-tuning. Figure 5 further illustrates the daily character recognition capabilities of representative devices from each category to understand the stability of the recognition capabilities for each type. The results show that the recognition capabilities of each category do not change significantly over time.

[0240] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A closed-loop evaluation method for checkpoint equipment driven by multi-source data fusion and traffic mechanism, characterized in that, Includes the following steps: Step 1: Construct an upstream device mapping table for the checkpoint devices based on road network information and checkpoint device location information; Step 2: Verify the historical traffic pattern of the checkpoint based on historical traffic data, calculate the real-time accuracy of the checkpoint equipment based on floating car trajectory data, and finally compare the traffic situation of the checkpoint equipment to be evaluated with that of the upstream checkpoint equipment through the upstream equipment mapping table to determine whether the traffic is conserved; if all the above verifications pass, skip to step 3; if they fail, directly determine that the equipment is abnormal and skip to step 6. Step 3: Verify the format of the license plate numbers collected by the checkpoint equipment according to the license plate verification rules to obtain normal license plate numbers and abnormal license plate numbers; Step 4: Combine spatiotemporal rules and travel patterns to further determine the wrongly photographed license plate number by judging normal and abnormal license plate numbers; Step 5: Determine the real license plate number of the wrongly photographed license plate number by using similar vehicles in the upstream and downstream vehicles, and associate the algorithm and hardware information; Step 6: Analyze the real-time evaluation results from Step 2, and then analyze the OCR error types obtained in Step 5; analyze the degree of influence of external factors on the device's recognition effect, and take targeted improvement measures.

2. The method for closed-loop evaluation of checkpoint equipment driven by multi-source data fusion and traffic mechanism as described in claim 1, characterized in that, Step 1 includes: Step 1.1: Construct a road network map and equipment mapping: Construct a directed road network map using road data. Nodes in the directed road network map are the starting and ending points of roads, and edges represent road segments. The attributes of the edges include the segment number and the road length. Step 1.2: Load device data, establish a mapping from checkpoint device number to road segment number, and construct a reverse mapping for subsequent lookup; Step 1.3: Construct an upstream device mapping table: By traversing the road network map in reverse, search for the upstream road segments that each checkpoint device can reach, calculate the path distance between the current road segment and the upstream road segment, filter out the upstream devices whose distance is within the maximum search distance, construct an upstream device mapping table, and record the upstream device and distance corresponding to each device.

3. The method for closed-loop evaluation of checkpoint equipment based on multi-source data fusion and traffic mechanism driven according to claim 1, characterized in that, Step 2 includes: Step 2.1, Verification based on historical traffic patterns at checkpoints: First, aggregate historical traffic data by hourly granularity to construct a time series and calculate the fluctuation threshold of historical data; monitor the current traffic in real time and compare it with the distribution of historical values ​​for the same period. If the deviation exceeds the threshold for multiple consecutive periods, the checkpoint equipment may be abnormal. Step 2.2, Accuracy assessment based on floating car trajectory data: Statistically analyze the data hourly, compare the license plate numbers of floating cars passing through the checkpoint equipment with the license plate numbers identified by the checkpoint equipment, calculate the real-time accuracy rate of the checkpoint equipment, and judge the corresponding checkpoint equipment to be abnormal if the real-time accuracy rate is lower than the minimum standard. Step 2.3, Traffic Conservation Verification: Count the traffic flow of each checkpoint device on an hourly basis. For each checkpoint device, obtain the traffic information of its upstream device according to the upstream device mapping table, calculate the expected traffic considering traffic propagation, and judge whether the traffic is conserved by comparing the actual traffic of the current device with the expected traffic. If the difference exceeds the allowable fluctuation range, the checkpoint device is considered to be abnormal.

4. The method for closed-loop evaluation of checkpoint equipment driven by multi-source data fusion and traffic mechanism as described in claim 1, characterized in that, Step 3 includes: Step 3.1, License Plate Format Verification: Use industry standards to verify the format of license plate numbers. The verification includes the length of the license plate number, the abbreviation of the province, whether it contains prohibited characters, and whether it conforms to the rules for ordinary license plates or new energy vehicle license plates, which are divided into normal license plate numbers and abnormal license plate numbers. Step 3.2, Abnormal License Plate Number Statistics: Statistically analyze the license plates with abnormal verification results and record the number of times each abnormal license plate number was captured by different checkpoints; for abnormal license plates that appear only once, they are directly judged as miscaptured license plates; for abnormal license plates that appear multiple times, further judgment is needed based on spatiotemporal rules and travel patterns. Step 3.3, Processing of Normal License Plate Numbers: Statistically record the number of times each normal license plate number is captured by different checkpoints for license plate numbers that are verified as normal. For normal license plate numbers that appear only once, judge them based on travel patterns. If the travel patterns are not met, they are considered as mistakenly captured license plate numbers. For normal license plate numbers that appear multiple times, judge them based on spatiotemporal rules and travel patterns. If the conditions are not met, they are considered as mistakenly captured license plate numbers.

5. The method for closed-loop evaluation of checkpoint equipment driven by multi-source data fusion and traffic mechanism as described in claim 4, characterized in that, Step 4 includes: Step 4.1, Consistency judgment of travel time and traffic conditions: Introduce real-time congestion index data, establish a travel time distribution model, combine the time of the vehicle between the two checkpoints, and use the 3σ principle to judge whether there is an anomaly. If there is an anomaly, it is considered that the travel time is inconsistent with the traffic conditions, and it is regarded as a wrong license plate number. Step 4.2, Route Reversal Judgment: Check the nodes passed by the vehicle during the journey to determine if there is a route reversal. If the nodes passed by the vehicle during the journey are repeated, it is considered that there is an unreasonable route reversal. Step 4.3, Determining Travel Patterns: Based on parking data and vehicle travel patterns, analyze vehicle travel time, routes, and stopping locations, and suggest spatiotemporal probability distributions; determine if the probability of a vehicle currently appearing is greater than 0.05; otherwise, consider the license plate number to be an incorrectly registered number.

6. The method for closed-loop evaluation of checkpoint equipment based on multi-source data fusion and traffic mechanism driven according to claim 1, characterized in that, Step 5 includes: Step 5.1, Data Preprocessing: Merge the vehicle record data considered as mis-captured with the device-road segment mapping data, and convert the time field to datetime type for subsequent processing; Step 5.2: Find candidate license plates: For each wrongly captured license plate record, find the upstream and downstream related road segments according to the upstream device mapping table, and set a time window based on the road segment length and time buffer. Select vehicle records that pass through the upstream and downstream related road segments within the time window and are not captured by the current device as candidate license plates. Step 5.3: Calculate similarity: For each wrongly photographed license plate, calculate its similarity score with the candidate license plates. The similarity score takes into account the overall similarity of the license plates and the matching of the province abbreviation. Select the candidate license plates with similarity scores greater than or equal to the threshold, sort them in descending order of score, and take the top 5 as possible real license plate numbers. Step 5.4: Associate Information: Record the wrongly captured license plate, candidate license plate, similarity score, and device number, and associate them with environmental interference, algorithm, and hardware conditions.

7. The method for closed-loop evaluation of checkpoint equipment based on multi-source data fusion and traffic mechanism driven according to claim 1, characterized in that, Step 6 includes: 6.1 Analysis of the real-time evaluation results in step 2: The time series data of the real-time evaluation results obtained in step 2 are preprocessed and standardized, and then clustered into periodic and random devices using a clustering algorithm. The device accuracy is correlated with time and external environmental factors, and the relationship is fitted by a random forest model. Finally, the SHAPA method is used to interpret and determine the degree of influence of external factors on the device accuracy, and targeted improvements are made. 6.2 OCR error type analysis obtained in step 5: Based on the correlation information obtained in step 5, construct the device recognition feature vector and use the cluster analysis method to classify the devices according to their ability to recognize different characters. Then, use radar charts to display the recognition capabilities of different types of devices in a multidimensional way and make targeted improvements.

Images