Automatic replenishment method for abnormal parking orders based on Lagrange interpolation

Abstract

The present invention relates to the technical field of parking management, and in particular to a method for automatically filling abnormal parking orders based on Lagrange interpolation. The method uses vehicle structured data and camera-captured images to simulate driving trajectories through wavelet transform, Lagrange interpolation and Kalman filtering, detects and records license plate information through a license plate correction and recognition system, analyzes parking locations, parking duration, entry and exit times and image evidence, so as to fill in or update missing parking data in an existing database. The method has low calculation cost and high speed, and improves the accuracy and stability of trajectory simulation. Complete parking data is obtained through trajectory simulation and license plate detection technology, and is synchronously updated to a database in real time, so as to realize automatic filling and ensure the integrity and accuracy of parking records.

Description

Technical Field

[0001] The present invention relates to the technical field of parking management, and in particular to an automatic order filling method for abnormal parking orders based on Lagrange interpolation. Background Art

[0002] With the surge in the number of motor vehicles, the complexity of parking management continues to rise. Although intelligent parking payment systems have significantly improved efficiency and reduced labor costs through automated processes, they still encounter many challenges in actual operation. Especially during peak hours, the concentrated entry and exit of a large number of vehicles may cause the system to be overloaded, making it impossible to collect, analyze and store all parking records in real time, resulting in data omissions. In addition, technical failures such as hardware damage, software defects or network interruptions can also interfere with the normal data collection and writing process, resulting in incomplete or lost parking records, that is, lost orders. If these problems are not properly resolved, they will seriously affect the reliability and user experience of the parking management system.

[0003] In order to automatically detect lost orders and make up for them in real time, we designed an automatic parking order making-up system based on Lagrange interpolation. First, we simulated the vehicle's driving trajectory by analyzing the structured data of passing vehicles, such as speed, acceleration, angular velocity, direction, etc. Since the existing trajectory simulation technology has poor anti-interference ability, we used wavelet transform, Lagrange interpolation and Kalman filtering to filter out outliers and noise during trajectory simulation. At the same time, combined with the camera capture data, the system realizes license plate detection, monitors the entry and exit of vehicles in real time, and retains image evidence. Since license plate detection is easily affected by angles or bad weather, we use license plate correction and recognition algorithms to improve the accuracy of license plate number recognition. Finally, combined with vehicle trajectory and captured images, this system generates credible parking records. Then a comprehensive difference score is designed to measure the difference between the credible record and the existing parking record. If the difference exceeds the threshold, it is considered as a lost order, and finally the existing parking record is corrected or completed. Summary of the invention

[0004] In order to solve the above problems such as poor anti-interference and low license plate detection accuracy, the present invention proposes an automatic parking order filling system based on Lagrange interpolation. The present invention uses vehicle structured data and camera snapshot images to simulate a more accurate driving trajectory through wavelet transform, Lagrange interpolation and Kalman filtering, and detects and records license plate information through a license plate correction and recognition system. Based on the above information, the parking location, parking duration, entry and exit time and image evidence can be analyzed to complete or update the missing parking data in the existing database.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A method for automatically filling abnormal parking orders based on Lagrange interpolation includes the following steps:

[0007] S1. Data collection: Collect the structured data of vehicles through GPS equipment and ground sensing coils, and capture images of vehicles entering and leaving the checkpoints through cameras installed at the checkpoints;

[0008] S2, trajectory simulation: First, the structured data is analyzed in the time and frequency domain by wavelet transform to identify and eliminate outliers, and then the Lagrange interpolation method is used to estimate and fill in the missing data caused by signal interruption or delay. Then, the trajectory data is denoised and smoothed by Kalman filter to obtain the trajectory simulation result;

[0009] S3, license plate detection: using automatic license plate recognition system to identify the license plate in the vehicle image;

[0010] S4. Automatic correction of parking records: A credible parking record is obtained based on the trajectory simulation results and the license plate detection results. By measuring the degree of difference between the existing parking record and the credible record, it is determined whether the parking record in the database needs to be supplemented or updated, so as to correct the parking record.

[0011] As a further technical solution of the present invention, the structured data of the vehicle in step S1 includes speed, acceleration, direction and position.

[0012] As a further technical solution of the present invention, step S2 performs time-frequency domain analysis on structured data by wavelet transform, and the process of identifying and eliminating outliers is as follows:

[0013] S211, Wavelet transform uses frequency domain transform to extract time series features ,in , , is the original time domain signal, It represents the frequency domain function, represents the position coefficient, represents the wavelet function, represents the frequency coefficient, is the estimated coefficient, represents the detailed coefficient;

[0014] S212, discrete continuous wavelet transform detects abnormal values ​​in structured data according to the decision threshold, the decision threshold , where the decision threshold is the limit of structured data, and Represent the mean and standard deviation of structured data respectively, It represents the coefficient of the standard deviation, and the data points beyond the judgment threshold are considered as outliers.

[0015] As a further technical solution of the present invention, the process of using the Lagrange interpolation method to estimate and fill in the missing data caused by signal interruption or delay in step S2 is as follows: given a set of time series Speed ​​data points below , , use Lagrange interpolation to estimate the missing velocity data: ,

[0016] in, Represents the Lagrangian interpolation basis function of degree n. Similarly, Lagrangian interpolation is performed on other structured data.

[0017] As a further technical solution of the present invention, the process of performing denoising and smoothing on the trajectory data by using a Kalman filter in step S2 is as follows:

[0018] S231. The Kalman filter establishes a state space model of a discrete random system through the state equation and the observation equation, and estimates the vehicle movement state according to the joint distribution of the observation data at different times: , ,

[0019] in, and Respectively indicate time The vectors of the state system and the observation system, , , and H represent the state transfer matrix, control coefficient matrix, input noise coefficient matrix and observation coefficient matrix respectively. It is input control, the initial value is set to 0, is the input white noise, represents the observation noise;

[0020] S232, alternately perform trajectory prediction and trajectory update to achieve denoising and optimize state estimation, wherein the trajectory prediction predicts the state at the next time point through a dynamic model to obtain the predicted value of the state at the next time point and the forecast error covariance , , ,in, represents the variance matrix of the process noise;

[0021] Trajectory update uses the observation data at the current time point to optimize the state prediction obtained by trajectory prediction. The Kalman gain is calculated to quantify the credibility of the observation value relative to the prediction error, and the prediction value is dynamically adjusted to obtain the optimal prediction value. Specifically:

[0022] ,

[0023] ,

[0024] ,

[0025] in, represents the Kalman gain, represents the optimal prediction value, and R is the variance matrix of the observation noise. represents the error covariance;

[0026] S233, obtaining the optimal driving trajectory simulation by introducing the optimal filtering parameter model, specifically:

[0027] ,

[0028] ,

[0029] ,

[0030] in, represents the signal energy ratio, and They represent the state values ​​before and after filtering, Q and R represent the covariance matrices of process noise and prediction noise, respectively. The vehicle entry time and exit time are estimated based on the denoised driving trajectory. For the entry time, the corresponding vehicle speed is reduced to 0 and then accelerated, and the direction is positive; while the exit time is negative, and the direction pointing to the parking lot is positive.

[0031] As a further technical solution of the present invention, the specific process of step S3 is:

[0032] S31. Input the vehicle image captured at the checkpoint , the target detection model YOLOv2 is used to detect vehicles, and the bounding box represents the position of the vehicle in the image, where , , They represent the center position and length and width of the vehicle bounding box respectively, and the category corresponding to the bounding box is "car";

[0033] S32. Use the distorted plane object detection network WPOD-Net to detect the area where the license plate is located ,in , , , It is the coordinates of the four vertices of the license plate bounding box expressed in the clockwise direction. When the license plate is tilted, the area where the license plate is located is adjusted to a rectangle in the front view state through affine transformation. Specifically, the license plate area feature map output by the feature extractor in WPOD-Net is first divided into a matrix grid, and then the grids with high target probability are found. The affine coefficients are calculated according to the positions of these grids, and the original license plate area is converted into a rectangular license plate in the front view;

[0034] S33. Use the character detection and recognition model OCR-Net to transcribe the license plate number in the license plate area into a string and save it.

[0035] As a further technical solution of the present invention, the specific process of step S4 is:

[0036] S41. Obtaining credible parking records based on trajectory simulation and license plate detection results ,in is the license plate number, is the entry time, For departure time, To enter the image, To exit the image, existing parking records Stored in the non-relational database MongoDB, the data format is exactly the same as the trusted parking record;

[0037] S42. By measuring the difference between the existing parking record and the credible record, it is determined whether the parking record in the database needs to be completed or updated. First, ensure that the license plate number in the database record and the credible record are completely consistent. If the license plate number does not match, the difference is directly considered to be infinite without further comparison. If any one of the license plate number, entry time, exit time, entry image, and exit image is missing, the difference is also considered to be infinite. At the same time, the corresponding fields in the database are corrected and completed according to the credible parking record:

[0038] ;

[0039] S43, comparing and analyzing the credible parking record and the existing image record:

[0040] For the entry time, directly find the difference and take the absolute value, that is, , the departure time is calculated in the same way as the entry time ,choose as a time difference;

[0041] For entry / exit images, perceptual hashing is used to evaluate the images separately. and and The similarity between As the image similarity, represents the similarity of the incoming image between the trusted record and the existing record, represents the similarity of the exit image; S44, calculating the comprehensive difference score, the comprehensive difference score ,in and are the weights of the time field and image field respectively and ; If the comprehensive difference score exceeds the set threshold, the record in the database is updated according to the trusted parking record. For missing fields, the value of the trusted parking record is directly used to fill in; for the time field, if the error exceeds the acceptable range, it is updated; for the image, if the image quality of the trusted parking record is significantly better than the record in the database, it is updated. Detailed logs need to be recorded for each order replenishment operation, including the difference, data before and after the update, and operation time, to achieve automated order replenishment.

[0042] As a further technical solution of the present invention, in step S43, the perceptual hash algorithm is used to evaluate the image and and The process of similarity between them is:

[0043] S431, scale the image using bilinear interpolation, with each pixel represented by (x, y), and convert the scaled image into a grayscale image, with the grayscale value of each pixel represented by g(x, y);

[0044] S432, perform discrete cosine transform (DCT) on the grayscale image to convert the image from the spatial domain to the frequency domain to obtain a The coefficient matrix of is:

[0045] ,

[0046] Where K represents the width and height of the scaled image, and u and v represent the index of the discrete cosine transform (DCT) coefficients;

[0047] S433, feature extraction and hash value generation, extract the upper left corner of the coefficient matrix The region represents the main features of the image, where n is a positive integer. Then the median d of the coefficients in the region is calculated, and the coefficients are thresholded to obtain the thresholded binary value. :

[0048] ,

[0049] The thresholded binary values ​​are then concatenated to get the final hash value :

[0050] ;

[0051] S433, quantify the image similarity by calculating the Hamming distance between the corresponding hash values ​​of the two images. The more similar the discrete cosine transform (DCT) coefficients of the two images are, the smaller the Hamming distance is, and the closer the image contents are. ,

[0052] in, is the indicator function, when Returns 1 if yes, otherwise returns 0. Represents a trustworthy image The discrete cosine transform (DCT) coefficient matrix of This means that the existing image The discrete cosine transform (DCT) coefficient matrix of .

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. Robust trajectory reconstruction: In the face of common interference factors such as sensor hardware limitations, environmental interference, data transmission defects and internal noise, the present invention effectively filters out outliers and noise in vehicle structure data by applying wavelet transform, Lagrange interpolation and Kalman filtering technology in sequence, thereby achieving accurate simulation and reconstruction of vehicle driving trajectory. Compared with deep learning algorithms, this method has low computational cost and high speed, and improves the accuracy and stability of trajectory simulation, showing excellent anti-interference performance.

[0055] 2. Adaptive license plate recognition: The present invention uses three core links, namely adaptive vehicle detection, license plate positioning and character recognition, to ensure that the license plate information in the camera image can be accurately captured and recognized in various complex scenarios, such as when the license plate is tilted or there is background interference.

[0056] 3. Fast parking record correction: The present invention uses a high-capacity, high-performance MongoDB non-relational database to store key parking records including license plate numbers, entry and exit times, and corresponding images. In the face of missing records caused by communication failures or signal interference, complete parking data is obtained through trajectory simulation and license plate detection technology, and updated to the database in real time, realizing automatic ticket replenishment and ensuring the integrity and accuracy of parking records. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present disclosure and do not constitute improper limitations on the present invention.

[0058] Figure 1 A schematic diagram of the automatic parking order replenishment process based on Lagrange interpolation provided by the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0060] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0061] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, device (device), product or equipment that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0062] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0063] Example:

[0064] like Figure 1 As shown, this embodiment provides a parking order automatic filling method based on Lagrange interpolation, including the following steps:

[0065] S1. Data collection:

[0066] The collected data consists of two parts: vehicle structured data and images captured at the checkpoint. The vehicle structured data includes speed, acceleration, direction, and position. The vehicle structured data is mainly captured by GPS equipment. However, since GPS is easily interfered by obstacles in underground parking lots, resulting in poor signal reception or failure, this embodiment additionally introduces ground sensing coils as auxiliary data sources. Ground sensing coils are an induction coil buried under the road surface, which are used to detect the presence, speed and direction of passing vehicles. When a vehicle passes through the coil, the vehicle status can be accurately detected. Moreover, ground sensing coils are not easily affected by weather conditions or signal interference, and have low deployment and maintenance costs. High-definition cameras are responsible for capturing vehicles entering and exiting the checkpoint, thereby obtaining images of vehicles entering / leaving. The camera can adapt to lighting conditions under different times and weather conditions, which is beneficial for license plate recognition to retain the image evidence chain when entering / leaving.

[0067] S2. Trajectory simulation:

[0068] In vehicle trajectory simulation, many challenges are often faced, such as interruption of trajectory recording caused by loss of GPS signal; signal interference of electronic equipment introduces errors in vehicle structured data; and the rapid movement or turning of the vehicle is missed due to the low sampling rate of the GPS device, resulting in incoherent trajectory. To address these problems, this embodiment uses wavelet transform, Lagrange interpolation and Kalman filtering in sequence to detect and remove outliers and noise in the trajectory data, and simulates a vehicle driving trajectory with more continuity, accuracy and reliability, specifically:

[0069] S21. Discrete continuous wavelet transform is a signal processing method based on wavelet analysis, which is used to detect abnormal values ​​between vehicle speed and acceleration in time series. Wavelet transform uses frequency domain transformation to extract time series features to reduce the dimension of feature space and obtain:

[0070] ,

[0071] ,

[0072] ,

[0073] in, is the original time domain signal, It represents the frequency domain function, represents the position coefficient, represents the wavelet function, represents the frequency coefficient, is the estimated coefficient, represents the detailed coefficient,

[0074] Then the discrete continuous wavelet transform detects outliers in the structured data according to the decision threshold. , where the decision threshold That is, the limit of structured data, i.e., the threshold; and Represent the mean and standard deviation of structured data respectively, It represents the coefficient of the standard deviation. Any data point that exceeds the judgment threshold range is considered an outlier.

[0075] S22. After wavelet transform detection and outlier removal, the vehicle's structured data (such as speed and acceleration) needs to be smoothed to fill in missing data points. This embodiment uses Lagrange interpolation to estimate and fill in missing data caused by signal interruption or delay. Given a set of speed data points in a time series, , use Lagrange interpolation to estimate the missing velocity data:

[0076] ,

[0077] in, Represents the Lagrangian interpolation basis function of degree n. Similarly, Lagrangian interpolation is performed on other vehicle structured data such as acceleration data.

[0078] S23, Kalman filter filters out noise in the prediction and update trajectory stage to obtain a smoother and more reliable trajectory curve. First, Kalman filter establishes the state space model of discrete random system through state equation and observation equation, and estimates the vehicle movement state according to the joint distribution of observation data at different times, and obtains:

[0079] ,

[0080] ,

[0081] in, and Respectively indicate time The vectors of the state system and the observation system, , , and H represent the state transfer matrix, control coefficient matrix, input noise coefficient matrix and observation coefficient matrix respectively. It is input control, the initial value is set to 0, is the input white noise, represents the observation noise;

[0082] Secondly, the two steps of trajectory prediction and trajectory update are performed alternately to achieve denoising and optimize state estimation. Specifically, the trajectory prediction step predicts the state at the next time point based on the dynamic model of the system. This process does not use observed data and only relies on the estimate and system model of the previous moment. The trajectory prediction step generates a predicted value of the state at the next time point. and the forecast error covariance , , ,in, represents the variance matrix of the process noise;

[0083] The trajectory update step uses the observation data at the current time point to optimize the state prediction obtained in the prediction step. This process calculates a Kalman gain to quantify the credibility of the observation value relative to the prediction error and dynamically adjusts the prediction value to obtain the optimal prediction value:

[0084] ,

[0085] ,

[0086] ,

[0087] in, represents the Kalman gain, represents the optimal prediction value, and R is the variance matrix of the observation noise. The Kalman filter repeats the prediction and update steps at each iteration, and sets the prediction value according to the final estimate of the previous step to gradually improve the continuity, consistency, and accuracy of the state estimation. In this way, the Kalman filter can effectively remove noise and track changes in the system state.

[0088] Then, in order to effectively reduce the impact of noise, the optimal filtering parameter model is introduced to obtain the optimal driving trajectory simulation:

[0089] ,

[0090] ,

[0091] ,

[0092] in, represents the signal energy ratio, and They represent the state values ​​before and after filtering, Q and R represent the covariance matrices of process noise and prediction noise, respectively. Finally, based on the denoised driving trajectory, we can estimate the entry time and exit time. For the entry time, the corresponding vehicle speed is reduced to 0 and then accelerated, and the direction is positive, while the exit time is negative, where the direction pointing into the parking lot is positive.

[0093] S3, license plate detection:

[0094] The automatic license plate recognition system is used to identify the license plate in the vehicle image. The specific process is as follows:

[0095] S31. Input the vehicle image captured at the checkpoint , use the target detection model YOLOv2 to detect the vehicle, S31, input the vehicle image captured at the checkpoint , the target detection model YOLOv2 is used to detect vehicles, and the bounding box represents the position of the vehicle in the image, where , , They represent the center position and length and width of the vehicle bounding box respectively, and the category corresponding to the bounding box is "car";

[0096] S32. Use the distorted plane object detection network WPOD-Net to detect the area where the license plate is located ,in , , , It is the coordinates of the four vertices of the license plate bounding box expressed in the clockwise direction. Taking into account the problem of license plate tilt, such as the license plate distortion into a trapezoid or parallelogram, the area where the license plate is located is adjusted to a rectangle in the front view state through affine transformation. Specifically, the license plate area feature map output by the feature extractor in WPOD-Net is first divided into a matrix grid, and then the grids with high target probability are found. The affine coefficients are calculated according to the positions of these grids, and the original license plate area is converted into a rectangular license plate in the front view;

[0097] S33. Use the character detection and recognition model OCR-Net to transcribe the license plate number in the license plate area into a string and save it.

[0098] S4. Automatic correction of parking records:

[0099] According to the results of trajectory simulation in step S2 and license plate detection in step S3, a reliable parking record is obtained. , which consists of five fields: license plate number, entry time, exit time, entry image and exit image. The data is stored in the non-relational database MongoDB, which has the same data format as the trusted parking record. The difference between the existing parking record and the trusted parking record is measured to determine whether the parking record in the database needs to be completed or updated. First, the license plate number is ensured to be completely consistent in the database record and the trusted record. If the license plate number does not match, the difference is directly considered to be infinite without further comparison. In addition, if any of the five fields is missing, the difference will be considered to be infinite. At the same time, the corresponding fields in the database are corrected and completed according to the trusted parking record:

[0100] ,

[0101] Then, the credible parking records and the existing parking records are compared and analyzed in detail. For the entry time field, the difference can be directly calculated and the absolute value can be taken, that is, The same method is used for the departure time, and the calculation is ,choose As a time difference.

[0102] For images of vehicles entering / exiting, perceptual hashing is used to evaluate the images separately. and and This is because the image evidence chain in the existing parking records may be blurred, overexposed or underexposed, or the image may be damaged during transmission, or even cause frame loss and incomplete vehicle body due to data delay. The process of the perception hash algorithm is as follows:

[0103] Image scaling and grayscale conversion: Use bilinear interpolation to scale the image to a smaller size, with each pixel represented by (x, y). Then convert the scaled image to a grayscale image, and use g(x, y) to represent the grayscale value of each pixel.

[0104] The discrete cosine transform (DCT) is performed on the grayscale image. DCT converts the image from the spatial domain to the frequency domain. The result of the DCT transform is a The coefficient matrix of

[0105] ),

[0106] Among them, K represents the width and height of the scaled image, and u, v represent the index of the DCT coefficient;

[0107] Feature extraction and hash value generation, extract the upper left corner of the coefficient matrix The main features of the image are represented by the region, where n is a positive integer. Then the median d of the coefficients in the region is calculated to perform thresholding on the coefficients.

[0108] ,

[0109] Finally, concatenate the thresholded binary values ​​to form the final hash value:

[0110] ;

[0111] Calculate image similarity. The image similarity is quantified by calculating the Hamming distance between the corresponding hash values ​​of two images. The more similar the DCT coefficients of two images are, the smaller the Hamming distance is, and the closer the image contents are. In image hashing and comparison, the Hamming distance is a fast and effective method to measure image similarity. ,in, is the indicator function, when Returns 1 if yes, otherwise returns 0; Represents a trustworthy image The DCT coefficient matrix, This means that the existing image The DCT coefficient matrix; finally, take As the image similarity, represents the similarity of the incoming image between the trusted record and the existing record, It represents the similarity of the exit image;

[0112] After measuring the differences between each field of the credible parking record and the existing parking record, the final comprehensive difference score is obtained. , set weights for time fields and image fields separately and , to weigh their importance in the parking records. If the comprehensive difference score exceeds the set threshold, the records in the database are updated according to the trusted parking records. For missing fields, the values ​​of the trusted parking records are directly used to fill in; for time fields, if the error exceeds the acceptable range, they are updated; for images, if the image quality of the trusted parking records is significantly better than that of the records in the database, they are updated; each order-making operation needs to record detailed logs, including the difference, data before and after the update, and operation time, for subsequent review and problem tracking. Through this method, while ensuring data accuracy, the possible missing and errors of the fields can be taken into account to achieve automated order-making operations.

[0113] The algorithms, systems, and network structures not described in detail in the present invention are all general technologies in the art.

[0114] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0115] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for automatically filling abnormal parking orders based on Lagrange interpolation, characterized in that: The following steps are involved: S1. Data collection: Collect the structured data of vehicles through GPS equipment and ground sensing coils, and capture images of vehicles entering and leaving the checkpoints through cameras installed at the checkpoints; S2, trajectory simulation: First, the structured data is analyzed in the time and frequency domain by wavelet transform to identify and eliminate outliers, and then the Lagrange interpolation method is used to estimate and fill in the missing data caused by signal interruption or delay. Then, the trajectory data is denoised and smoothed by Kalman filter to obtain the trajectory simulation result; S3, license plate detection: using automatic license plate recognition system to identify the license plate in the vehicle image; S4, automatic correction of parking records: A credible parking record is obtained based on the trajectory simulation results and the license plate detection results. By measuring the difference between the existing parking record and the credible record, it is determined whether the parking record in the database needs to be supplemented or updated, so as to correct the parking record. The specific process is as follows: S41. Obtaining credible parking records based on trajectory simulation and license plate detection results ,in is the license plate number, is the entry time, For departure time, To enter the image, To exit the image, existing parking records Stored in the non-relational database MongoDB, the data format is exactly the same as the trusted parking record; S42. By measuring the difference between the existing parking record and the credible record, it is determined whether the parking record in the database needs to be completed or updated. First, ensure that the license plate number in the database record and the credible record are completely consistent. If the license plate number does not match, the difference is directly considered to be infinite without further comparison. If any one of the license plate number, entry time, exit time, entry image, and exit image is missing, the difference is also considered to be infinite. At the same time, the corresponding fields in the database are corrected and completed according to the credible parking record: ； in They are the license plate number, entry time, exit time, entry image, and exit image of the existing parking record; S43, comparing and analyzing the credible parking record and the existing image record: For the entry time, directly find the difference and take the absolute value, that is, , the departure time is calculated in the same way as the entry time ,choose as a time difference; For entry / exit images, perceptual hashing is used to evaluate the images separately. and and The similarity between As the image similarity, represents the similarity of the incoming image between the trusted record and the existing record, It represents the similarity of the exit image; S44. Calculate the comprehensive difference score. ,in and are the weights of the time field and image field respectively and ; If the comprehensive difference score exceeds the set threshold, the record in the database is updated according to the trusted parking record. For missing fields, the value of the trusted parking record is directly used to fill in; for the time field, if the error exceeds the acceptable range, it is updated; for the image, if the image quality of the trusted parking record is significantly better than the record in the database, it is updated. Detailed logs need to be recorded for each order replenishment operation, including the difference, data before and after the update, and operation time, to achieve automated order replenishment.

2. According to claim 1, the method for automatically filling abnormal parking orders based on Lagrange interpolation is characterized in that: The structured data of the vehicle in step S1 includes speed, acceleration, direction and position.

3. The method for automatically filling abnormal parking orders based on Lagrange interpolation according to claim 2 is characterized in that: Step S2 uses wavelet transform to analyze the structured data in the time and frequency domains, and the process of identifying and eliminating outliers is as follows: S211, Wavelet transform uses frequency domain transform to extract time series features ,in , , is the original time domain signal, It represents the frequency domain function, represents the position coefficient, represents the wavelet function, represents the frequency coefficient, is the estimated coefficient, represents the detailed coefficient; S212, discrete continuous wavelet transform detects abnormal values ​​in structured data according to the decision threshold, the decision threshold , where the decision threshold is the limit of structured data, and Represent the mean and standard deviation of structured data respectively, It represents the coefficient of the standard deviation, and the data points beyond the judgment threshold are considered as outliers.

4. According to claim 3, the method for automatically filling abnormal parking orders based on Lagrange interpolation is characterized in that: Step S2 uses the Lagrange interpolation method to estimate and fill in the missing data caused by signal interruption or delay. The process is as follows: Given a set of time series The speed data points below , , use Lagrange interpolation to estimate the missing velocity data: ,in, Represents the Lagrangian interpolation basis function of degree n. Similarly, Lagrangian interpolation is performed on other structured data.

5. The method for automatically filling abnormal parking orders based on Lagrange interpolation according to claim 4 is characterized in that: The process of denoising and smoothing the trajectory data by using the Kalman filter in step S2 is as follows: S231. The Kalman filter establishes a state space model of a discrete random system through the state equation and the observation equation, and estimates the vehicle movement state according to the joint distribution of the observation data at different times: , , in, and Respectively indicate time The vectors of the state system and the observation system, , , and H represent the state transfer matrix, control coefficient matrix, input noise coefficient matrix and observation coefficient matrix respectively. It is input control, the initial value is set to 0, is the input white noise, represents the observation noise; S232, alternately perform trajectory prediction and trajectory update to achieve denoising and optimize state estimation, wherein the trajectory prediction predicts the state at the next time point through a dynamic model to obtain the predicted value of the state at the next time point and the forecast error covariance , , ,in, represents the variance matrix of the process noise; Trajectory update uses the observation data at the current time point to optimize the state prediction obtained by trajectory prediction. The Kalman gain is calculated to quantify the credibility of the observation value relative to the prediction error, and the prediction value is dynamically adjusted to obtain the optimal prediction value. Specifically: ， ， ， in, represents the Kalman gain, represents the optimal prediction value, and R is the variance matrix of the observation noise. represents the error covariance; S233, obtaining the optimal driving trajectory simulation by introducing the optimal filtering parameter model, specifically: ， ， ， in, represents the signal energy ratio, and They represent the state values ​​before and after filtering, Q and R represent the covariance matrices of process noise and prediction noise, respectively. The vehicle entry time and exit time are estimated based on the denoised driving trajectory. For the entry time, the corresponding vehicle speed is reduced to 0 and then accelerated, and the direction is positive; while the exit time is negative, and the direction pointing to the parking lot is positive.

6. The method for automatically filling abnormal parking orders based on Lagrange interpolation according to claim 5 is characterized in that: The specific process of step S3 is: S31. Input the vehicle image captured at the checkpoint , the target detection model YOLOv2 is used to detect vehicles, and the bounding box represents the position of the vehicle in the image, where , , They represent the center position and length and width of the vehicle bounding box respectively, and the category corresponding to the bounding box is "car"; S32. Use the distorted plane object detection network WPOD-Net to detect the area where the license plate is located ,in , , , It is the coordinates of the four vertices of the license plate bounding box expressed in the clockwise direction. When the license plate is tilted, the area where the license plate is located is adjusted to a rectangle in the front view state through affine transformation. Specifically, the license plate area feature map output by the feature extractor in WPOD-Net is first divided into a matrix grid, and then the grids with high target probability are found. The affine coefficients are calculated according to the positions of these grids, and the original license plate area is converted into a rectangular license plate in the front view; S33. Use the character detection and recognition model OCR-Net to transcribe the license plate number in the license plate area into a string and save it.

7. The method for automatically filling abnormal parking orders based on Lagrange interpolation according to claim 6 is characterized in that: In step S43, the image is evaluated using the perceptual hashing algorithm. and and The process of similarity between them is: S431, scale the image using bilinear interpolation, with each pixel represented by (x, y), and convert the scaled image into a grayscale image, with the grayscale value of each pixel represented by g(x, y); S432, perform discrete cosine transform on the grayscale image, convert the image from the spatial domain to the frequency domain, and obtain a The coefficient matrix of is: ， Among them, K represents the width and height of the scaled image, and u and v represent the index of the discrete cosine transform coefficient; S433, feature extraction and hash value generation, extract the upper left corner of the coefficient matrix The region represents the main features of the image, where n is a positive integer. Then the median d of the coefficients in the region is calculated, and the coefficients are thresholded to obtain the thresholded binary value. : ， The thresholded binary values ​​are then concatenated to get the final hash value : ； S434, quantify the image similarity by calculating the Hamming distance between the corresponding hash values ​​of the two images. The more similar the discrete cosine transform coefficients of the two images are, the smaller the Hamming distance is, and the closer the image contents are. , in, is the indicator function, when Returns 1 if yes, otherwise returns 0. Represents a trustworthy image The discrete cosine transform coefficient matrix, This means that the existing image The discrete cosine transform coefficient matrix.

Images