A curbstone vehicle license plate recognition method

Abstract

The application discloses a curbstone vehicle license plate recognition method, comprising the following steps: according to the extracted feature vector, a correlation coefficient between an environmental variable and a recognition performance is calculated by using a trend model to obtain a performance attenuation trend index, which is used as a reference basis for subsequent parameter adjustment; if the performance attenuation trend index exceeds a preset threshold, a sensitivity calibration operation is performed on a sensor array to obtain an adjusted sensor response curve as an intermediate result of dynamic optimization; from the adjusted sensor response curve, an update range of an algorithm threshold is determined, and through an iterative optimization method, environmental variable data is fused to obtain an optimized threshold set to adapt to a current dynamic environment; an intelligent system configuration is updated by using the determined threshold fine-tuning parameter to obtain an enhanced recognition module for processing image data streams of curbstone vehicle license plates in a complex scene.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, specifically a method for recognizing license plates on curbside vehicles. Background Technology

[0002] In smart parking scenarios, curb-mounted license plate recognition systems need to meet the high-precision recognition requirements in complex and dynamic environments. However, the core technical challenge lies in how to maintain the stability and accuracy of license plate recognition performance under real-time environmental variables such as changes in light intensity and vehicle speed fluctuations, while ensuring that the system can adaptively adjust to cope with diverse scenario challenges.

[0003] Drastic changes in light intensity (such as day-night cycles, shadows, or direct sunlight) can cause unstable contrast of license plate characters in the image data stream, increasing the risk of misidentification or missed identification. Meanwhile, rapid fluctuations in vehicle speed (such as instantaneous acceleration or deceleration at parking lot entrances) can lead to image blurring or license plate area positioning errors, reducing the reliability of feature extraction.

[0004] The dynamic nature of these environmental variables makes it difficult for traditional fixed-threshold recognition algorithms to adapt, causing performance degradation indicators to frequently exceed the preset threshold, thus affecting the robustness of the recognition module.

[0005] Furthermore, the sensitivity of the sensor array may become distorted due to environmental interference during long-term use, making it difficult for the response curve to accurately reflect the actual needs of the scenario, which further exacerbates the instability of the algorithm threshold optimization.

[0006] When dealing with complex scenarios (such as dense parking lots, low-light underground garages, or rainy / foggy weather), existing systems lack accurate modeling of the correlation between environmental variables and recognition performance, and cannot dynamically adjust the threshold set through feedback loop mechanisms, resulting in insufficient matching of recognition results with the applicability of smart parking scenarios. Summary of the Invention

[0007] The purpose of this invention is to provide a method for recognizing license plates on curb-mounted vehicles, in order to solve the problem in the prior art mentioned in the background that the dynamic nature of environmental variables makes it difficult for traditional fixed threshold recognition algorithms to adapt, resulting in performance degradation trends frequently exceeding the preset threshold and affecting the robustness of the recognition module.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A method for recognizing license plates on curbside motorcycles, the method comprising the following steps:

[0010] S101, by collecting real-time environmental variable data during the process of license plate recognition of curb motors, extracts feature vectors from the environmental variable data as the input basis for the trend model;

[0011] S102, Based on the extracted feature vector, the correlation coefficient between environmental variables and recognition performance is calculated using a trend model to obtain a performance degradation trend index, which serves as a reference for subsequent parameter adjustments.

[0012] S103, if the performance degradation trend index exceeds the preset threshold, then perform a sensitivity calibration operation on the sensor array and obtain the adjusted sensor response curve as an intermediate result of dynamic optimization.

[0013] S104. From the adjusted sensor response curve, determine the update range of the algorithm threshold, and fuse environmental variable data through iterative optimization method to obtain an optimized threshold set to adapt to the current dynamic environment.

[0014] S105. After obtaining the optimized threshold set, determine its matching degree with the recognition performance. If the matching degree is lower than the standard value, re-analyze the trend model output through the feedback loop mechanism to determine further threshold fine-tuning parameters.

[0015] S106, the intelligent system configuration is updated by fine-tuning parameters with a defined threshold to obtain an enhanced recognition module for processing image data streams of curb vehicles and license plates in complex scenarios;

[0016] S107: Obtain the processed image data stream from the enhanced recognition module, determine its applicability in the smart parking scenario, and if the applicability meets the requirements, output the final license plate recognition result to support traffic management applications.

[0017] According to the above technical solution, in S101, real-time environmental variable data is collected during the license plate recognition process of the curb vehicle, and feature vectors are extracted from the environmental variable data as the input basis for the trend model, including:

[0018] The raw dataset is obtained by collecting environmental variable data in real time through sensors, including light intensity and vehicle speed.

[0019] Preprocessing methods are used to denoise and standardize the original dataset to obtain a standardized dataset;

[0020] Feature vectors of light intensity and vehicle speed are extracted from the standardized dataset to obtain a feature vector set;

[0021] The feature vector set is evaluated. If the light intensity in the feature vector set is lower than a preset threshold, an image enhancement algorithm is applied to adjust the image brightness to obtain an enhanced feature vector set.

[0022] The enhanced feature vector set is classified using the support vector machine algorithm to determine the input category of the license plate recognition trend model.

[0023] The parameters of the trend model are adjusted based on the classification results to obtain an optimized trend model;

[0024] An optimized trend model is used to predict the trend of license plate recognition based on real-time collected environmental variable data.

[0025] According to the above technical solution, in S102, based on the extracted feature vector, a trend model is used to calculate the correlation coefficient between environmental variables and recognition performance, thereby obtaining performance degradation trend indicators, including:

[0026] From the obtained feature vectors and environmental variables, a preprocessing method is used to normalize the feature vectors to obtain standardized feature vectors;

[0027] Based on the standardized feature vector and environmental variables, the correlation coefficient between the two is calculated using the Pearson correlation coefficient method to obtain the correlation strength between environmental variables and recognition performance;

[0028] The obtained correlation strength is judged. If the correlation strength exceeds the preset threshold, a linear regression model is used to fit the relationship between environmental variables and recognition performance to obtain the trend model parameters.

[0029] Based on the trend model parameters, calculate the performance degradation trend index to determine the performance change trend;

[0030] Key points are extracted from the performance change trend, and the gradient descent method is used to optimize the parameter adjustment strategy to obtain the adjusted parameter set.

[0031] Based on the adjusted parameter set, the configuration of the recognition model is updated to obtain the optimized recognition performance;

[0032] By comparing the optimized recognition performance with the original performance, the effect of parameter adjustment is evaluated, and the performance improvement index is obtained.

[0033] According to the above technical solution, in S103, if the performance degradation trend index exceeds a preset threshold, a sensitivity calibration operation is performed on the sensor array to obtain the adjusted sensor response curve as an intermediate result of dynamic optimization, including:

[0034] If the performance degradation trend index exceeds the preset threshold, the data acquisition module will obtain real-time signal data from the sensor array to obtain the raw response dataset.

[0035] The original response dataset is filtered and denoised using signal processing techniques to obtain a smooth response dataset.

[0036] The support vector machine algorithm is used to extract features from the smooth response dataset to determine the key feature parameters of performance degradation;

[0037] If the deviation between the key feature parameters and the preset threshold exceeds the specified range, the sensitivity calibration module is triggered to obtain the calibration parameter set;

[0038] The sensitivity settings of the sensor array are adjusted by calibrating the parameter set to obtain the adjusted response curve;

[0039] The adjusted response curve was fitted and analyzed using a linear regression algorithm to obtain intermediate results of dynamic optimization.

[0040] The real-time monitoring module continuously tracks the intermediate results of dynamic optimization to determine whether the performance degradation trend has recovered to within the preset threshold.

[0041] According to the above technical solution, in S104, the update range of the algorithm threshold is determined from the adjusted sensor response curve, and environmental variable data is fused through an iterative optimization method to obtain an optimized threshold set to adapt to the current dynamic environment, including:

[0042] Raw response data is acquired from the sensor, and signal processing methods are used to denoise and normalize the data to obtain the adjusted response curve.

[0043] Based on the adjusted response curve, key feature points are extracted through curve analysis to determine the initial threshold range;

[0044] If the environmental variable data does not match the preset range, the environmental variables are integrated using a weighted average method to obtain comprehensive environmental parameters.

[0045] Based on the comprehensive environmental parameters and the initial threshold range, the gradient descent algorithm is used for iterative optimization to obtain the optimized threshold set.

[0046] If the dynamic environment changes, the threshold set is adjusted by monitoring and updating environmental parameters in real time for the optimized threshold set.

[0047] By comparing the adjusted threshold set with the sensor response data, it is determined whether the threshold is suitable for the current dynamic environment, and the final threshold set is obtained.

[0048] Based on the final threshold set, control parameters adapted to the dynamic environment are generated and output to the system execution module.

[0049] According to the above technical solution, in S105, after obtaining the optimized threshold set, its matching degree with the recognition performance is determined. If the matching degree is lower than the standard value, the trend model output is re-analyzed through a feedback loop mechanism to determine further threshold fine-tuning parameters, including:

[0050] After obtaining the threshold set, the matching degree between it and the recognition performance is calculated using the support vector machine algorithm to obtain the matching degree value;

[0051] If the matching degree value is lower than the preset threshold, the trend data output by the model is extracted through the feedback loop mechanism to obtain the trend analysis results;

[0052] Based on the trend analysis results, the gradient descent algorithm is used to optimize the threshold set and determine the fine-tuning parameter set.

[0053] The threshold set is adjusted by fine-tuning the parameter set to generate an updated threshold set.

[0054] For the updated threshold set, recalculate its matching degree with the recognition performance to obtain a new matching degree value;

[0055] If the new matching value is still lower than the preset threshold, repeat the feedback loop and parameter adjustment steps to obtain the final optimized threshold set.

[0056] By using the final optimized set of thresholds, a performance evaluation is performed to determine the stability of the system's recognition performance.

[0057] According to the above technical solution, in S106, the intelligent system configuration is updated using determined threshold fine-tuning parameters to obtain an enhanced recognition module, which is used to process the image data stream of curbside vehicle license plates in complex scenes, including:

[0058] By determining the parameter adjustment range through a preset threshold, the image data stream of curb motorcycle license plates is obtained from complex scenes to generate an initial image set.

[0059] If the initial image set contains noise, a Gaussian filtering algorithm is used to process the initial image set to obtain a denoised image set.

[0060] Based on the denoised image set, a convolutional neural network is used to extract the license plate features of curb motorcycles and generate a feature image set.

[0061] If the resolution of the feature image set is lower than a preset threshold, the feature image set is enhanced by a super-resolution algorithm to obtain an enhanced image set.

[0062] Based on the enhanced image set, an optical character recognition algorithm is used to parse the license plate characters and generate a character data stream;

[0063] By matching the character data stream with the preset license plate template, the consistency of the match is determined, and the license plate recognition result is obtained.

[0064] The intelligent system configuration is updated based on the license plate recognition results to generate an enhanced recognition module.

[0065] According to the above technical solution, in S107, the processed image data stream is obtained from the enhanced recognition module, and its applicability in the smart parking scenario is determined. If the applicability meets the requirements, the final license plate recognition result is output to support traffic management applications, including:

[0066] Image data streams are obtained from the enhancement recognition module, and preprocessing algorithms are used to denoise and enhance the images to obtain the first image data;

[0067] Based on the first image data, a convolutional neural network algorithm is used to extract the license plate area features to obtain the license plate location data;

[0068] For license plate location data, if the feature clarity exceeds a preset threshold, an optical character recognition algorithm is used to parse the license plate characters to obtain the license plate text data.

[0069] Based on the license plate text data, determine whether it meets the license plate format requirements of the smart parking scenario, and obtain the format verification result;

[0070] If the format verification result meets the requirements, the license plate text data will be matched with the preset parking lot vehicle database to obtain the vehicle identity information;

[0071] Based on the vehicle identification information, generate license plate recognition results and output them to the traffic management application interface to obtain the final recognition output;

[0072] The final identification output updates the parking management database, completing the data synchronization for traffic management applications.

[0073] According to the above technical solution, it also includes a high dynamic range image stream fusion and recognition process, specifically including:

[0074] Acquire multi-source sensor data and preprocess it for noise reduction to generate the first data stream;

[0075] Light characteristics are extracted by analyzing the real-time rate of change of light to obtain light adjustment parameters;

[0076] A high dynamic range image stream is generated by fusing the first data stream based on the light adjustment parameters.

[0077] If the image stream's clarity and data integrity meet a preset threshold, then the license plate features are extracted using a convolutional neural network, and the license plate character sequence is determined by combining it with an optical character recognition algorithm.

[0078] The character sequence is compared with the preset database, and the final license plate recognition result is output.

[0079] According to the above technical solution, it also includes an optimized recognition process for high dynamic range image streams, specifically including:

[0080] Raw image data is acquired from multiple source sensors, and a high dynamic range image stream is generated using data fusion technology.

[0081] If the image stream clarity is lower than a preset threshold, the contrast and brightness are optimized using an image enhancement algorithm to obtain an optimized image stream;

[0082] The light change rate of the image stream is analyzed and adjusted. If the integrity of the adjusted data meets the threshold, the license plate region features are extracted through a convolutional neural network.

[0083] The character sequence is parsed using an optical character recognition algorithm, its legality is verified by comparing it with a preset license plate format database, and the license plate recognition result is output.

[0084] Compared with the prior art, the present invention has the following advantages:

[0085] This invention discloses a dynamic optimization method for license plate recognition in complex scenarios, addressing the performance degradation caused by changes in light intensity and vehicle speed fluctuations. By collecting real-time environmental variable data and extracting feature vectors, this invention employs a trend model to calculate the correlation coefficient between environmental variables and recognition performance, generating a performance degradation trend index. When the index exceeds a threshold, this invention automatically triggers sensor array sensitivity calibration, obtains the optimized response curve, and then uses an iterative optimization method to fuse environmental data and dynamically update the algorithm's threshold set. If the threshold set does not adequately match the recognition performance, this invention re-analyzes the trend model through a feedback loop mechanism, determines fine-tuning parameters, updates the configuration, and generates an enhanced recognition module. Ultimately, this invention outputs license plate recognition results adapted to smart parking scenarios, effectively improving the accuracy and stability of recognition in complex environments and providing reliable support for traffic management. Attached Figure Description

[0086] Figure 1 This is a flowchart of a method for recognizing license plates on curb motorcycles according to the present invention. Detailed Implementation

[0087] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0088] Example 1

[0089] like Figure 1 As shown, a method for recognizing license plates on curb motorcycles in this embodiment may specifically include:

[0090] S101. By collecting real-time environmental variable data during the license plate recognition process of curb motors, including changes in light intensity and vehicle speed fluctuations, feature vectors are extracted from the data as the input basis for the trend model.

[0091] Environmental variable data, including light intensity and vehicle speed, are collected in real time by sensors to obtain a raw dataset. Preprocessing methods are used to denoise and standardize the raw dataset, resulting in a standardized dataset. Feature vectors for light intensity and vehicle speed are extracted from the standardized dataset to obtain a feature vector set. If the light intensity in the feature vector set is lower than a preset threshold, an image enhancement algorithm is applied to adjust the image brightness, resulting in an enhanced feature vector set. The enhanced feature vector set is then classified using a support vector machine algorithm to determine the input category for the license plate recognition trend model. The trend model parameters are adjusted based on the classification results to obtain an optimized trend model. This optimized trend model is then used to predict the trend of license plate recognition based on the real-time collected environmental variable data.

[0092] Specifically, by collecting real-time environmental variable data during the license plate recognition process at the curb, firstly, a light sensor collects light intensity data at a frequency of 10 times per second, measured in lux. It is assumed that dynamic data of 1000-5000 lux is collected during the day and 50-200 lux at night. The data is stored as a time series in CSV format. Subsequently, a Fast Fourier Transform (FFT) algorithm is used to perform frequency domain analysis on the light intensity sequence, extracting the dominant frequency component as a feature. For example, a dominant frequency of 0.5 Hz indicates a light change period of 2 seconds, reflecting the stability of ambient light. Next, vehicle speed is measured using onboard radar with millimeter-wave technology at a frequency of 20 Hz. Assuming a speed fluctuation range of 20-60 km / h, the data is smoothed using a Kalman filter algorithm to eliminate noise, obtaining the mean and variance of the speed per second. For example, a mean of 40 km / h and a variance of 5 km / h² are used as speed fluctuation features. The illumination frequency, mean, and variance of speed are combined to form a three-dimensional feature vector, for example, [0.5, 40, 5]. Principal Component Analysis (PCA) is then used to reduce the dimensionality to two dimensions, retaining 95% of the variance to ensure feature effectiveness. These feature vectors are input into a Support Vector Machine (SVM) tendency model, trained with radial basis function kernels, hyperparameters C=1.0, and gamma=0.1. The model is optimized based on historical data (assuming 1000 samples) to predict the accuracy trend of license plate recognition. Analysis shows that when the illumination frequency is below 0.3Hz, the recognition accuracy decreases by 10%, and when the speed variance is greater than 10km / h², image blurring leads to a 15% decrease in accuracy. Correlation analysis between the feature vectors and model output is used to adjust the acquisition frequency or algorithm parameters to improve system robustness. This method achieves environmental adaptability optimization through data-driven approaches, ensuring the stability of license plate recognition in complex scenarios.

[0093] S102. Based on the extracted feature vectors, the correlation coefficient between environmental variables and recognition performance is calculated using a trend model to obtain a performance degradation trend index, which serves as a reference for subsequent parameter adjustments.

[0094] Feature vectors and environmental variables are obtained from the data source. Preprocessing methods are used to normalize the feature vectors, resulting in standardized feature vectors. Based on the standardized feature vectors and environmental variables, the Pearson correlation coefficient method is used to calculate the correlation coefficient between them, obtaining the correlation strength between environmental variables and recognition performance. If the correlation strength exceeds a preset threshold, a linear regression model is used to fit the relationship between environmental variables and recognition performance, obtaining trend model parameters. Based on the trend model parameters, a performance degradation trend index is calculated to determine the performance change trend. Key points are extracted from the performance change trend, and the gradient descent method is used to optimize the parameter adjustment strategy, obtaining an adjusted parameter set. Based on the adjusted parameter set, the configuration of the recognition model is updated, resulting in optimized recognition performance. By comparing the optimized recognition performance with the original performance, the effect of parameter adjustment is evaluated, obtaining a performance improvement index.

[0095] Specifically, assuming the feature vector extracted from sensor data contains three environmental variables—temperature, humidity, and light intensity—at 25.5°C, 60.2%, and 1000 lux respectively, and the recognition performance is represented by the mean accuracy (mAP) of the target detection model, currently 0.85, the feature vector is dimensionality reduced using principal component analysis (PCA) while retaining 90% of the variance, resulting in principal component vectors [0.72, 0.45, 0.31]. First, the correlation coefficient between the environmental variables and mAP is calculated using the Pearson correlation coefficient, with the formula r = cov(X,Y) / (σ). X ·σ YLet X be an environmental variable vector and Y be the mAP sequence. Assume the historical data contains 10 samples: temperature sequence [24.0, 25.5, 26.2, ...], humidity sequence [58.0, 60.2, 61.5, ...], illumination sequence [950, 1000, 1050, ...], and mAP sequence [0.88, 0.85, 0.83, ...]. The calculated r values ​​for temperature and mAP are -0.78, humidity -0.65, and illumination -0.42, indicating that temperature has the greatest impact on performance, and a negative correlation means that increased temperature leads to a decrease in mAP. Next, a trend model was constructed, using linear regression to fit the relationship between mAP and environmental variables. The formula is mAP = β0 + β1·temperature + β2·humidity + β3·illuminance. Calculations using the least squares method yielded β0 = 1.25, β1 = -0.015, β2 = -0.008, and β3 = -0.0002. The performance degradation trend index was defined as the difference between the predicted mAP and the baseline mAP (0.9). Under the current conditions, the predicted mAP = 1.25 - 0.015·25.5 - 0.008·60.2 - 0.0002·1000 = 0.836, and the degradation index was 0.9 - 0.836 = 0.064. Analysis showed that for every 1°C increase in temperature, mAP decreased by 0.015, with humidity having a secondary effect. To optimize the parameters, the normalization range of the model input can be adjusted according to the degradation index. The upper limit of temperature normalization was adjusted from 30°C to 27°C to reduce the impact of high temperatures. All calculations were performed using Python scripts, calling NumPy to calculate correlation coefficients and sklearn to perform PCA and regression analysis, ensuring automated processing. The results can serve as a reference for subsequent model parameter optimization, prioritizing adjustments to temperature-related preprocessing strategies.

[0096] S103. If the performance degradation trend index exceeds the preset threshold, perform a sensitivity calibration operation on the sensor array and obtain the adjusted sensor response curve as an intermediate result of dynamic optimization.

[0097] If the performance degradation trend exceeds a preset threshold, real-time signal data is acquired from the sensor array via the data acquisition module to obtain the raw response dataset. Signal processing techniques are used to filter and denoise the raw response dataset, resulting in a smoothed response dataset. A support vector machine algorithm is employed to extract features from the smoothed response dataset, identifying key performance degradation parameters. If the key parameters deviate from the preset threshold by more than a specified range, a sensitivity calibration module is triggered to acquire a calibration parameter set. The sensitivity settings of the sensor array are adjusted using the calibration parameter set to obtain the adjusted response curve. A linear regression algorithm is used to fit and analyze the adjusted response curve, yielding an intermediate result for dynamic optimization. A real-time monitoring module continuously tracks the intermediate result of dynamic optimization to determine whether the performance degradation trend has recovered to within the preset threshold.

[0098] Specifically, when the performance degradation trend index exceeds a preset threshold, the system automatically triggers a sensitivity calibration operation for the sensor array, achieved through the following information technology means. First, the system collects the output data of the sensor array in real time, such as a voltage value sequence from 10 sensors. Assuming the currently collected voltage values ​​are [2.5, 2.7, 2.4, 2.6, 2.8, 2.3, 2.9, 2.5, 2.6, 2.7] volts, its mean is calculated to be 2.6 volts and its standard deviation to be 0.18. The performance degradation trend index is defined as the percentage deviation of the mean of 10 consecutive samples from the reference value (assumed to be 2.8 volts), calculated as |mean - reference value| / reference value × 100%. Currently, this is |2.6 - 2.8| / 2.8 × 100% = 7.14%. If the preset threshold is 5%, then 7.14% exceeds the threshold, triggering calibration. The calibration operation uses software to control the gain adjustment module of the sensor array, employing a gradient descent algorithm to optimize the gain parameters, aiming to make the average sensor output value close to the reference value of 2.8 volts. The algorithm's initial gain was set to 1.0, and the learning rate was 0.01. The iteration formula was gain(n+1) = gain(n) - learning rate × (mean - baseline). After 10 iterations, the gain was adjusted to 1.08, and the sensor output mean converged to 2.79 volts, close to the baseline value. The adjusted sensor response curve was generated through polynomial fitting, assuming the fitting function was y = 0.02x² + 1.08x + 0.1, where x is the input signal strength and y is the sensor output voltage. The system stored this curve as an intermediate result of dynamic optimization and analyzed the optimization effect by comparing it with the root mean square error (RMSE) of historical curves. Assuming the current RMSE is 0.05, which is below the acceptable threshold of 0.1, it indicates that the calibration is effective. Subsequently, the system uploaded the optimized curve parameters to a cloud database for subsequent real-time monitoring and equipment status prediction, ensuring the stable performance of the sensor array.

[0099] S104. From the adjusted sensor response curve, determine the update range of the algorithm threshold, and fuse environmental variable data through iterative optimization to obtain an optimized threshold set to adapt to the current dynamic environment.

[0100] Raw response data is acquired from sensors, and signal processing methods are used to denoise and normalize the data to obtain an adjusted response curve. Based on the adjusted response curve, key feature points are extracted through curve analysis to determine the initial threshold range. Environmental variable data is acquired; if the environmental variable data does not match the preset range, a weighted average method is used to fuse the environmental variables to obtain comprehensive environmental parameters. Based on the comprehensive environmental parameters and the initial threshold range, a gradient descent algorithm is used for iterative optimization to obtain an optimized threshold set. For the optimized threshold set, if the dynamic environment changes, the environmental parameters are updated in real time, and the threshold set is adjusted. By comparing the adjusted threshold set with the sensor response data, it is determined whether the thresholds are suitable for the current dynamic environment, resulting in a final threshold set. Based on the final threshold set, control parameters adapted to the dynamic environment are generated and output to the system execution module.

[0101] S105. After obtaining the optimized threshold set, determine its matching degree with the recognition performance. If the matching degree is lower than the standard value, re-analyze the trend model output through the feedback loop mechanism to determine further threshold fine-tuning parameters.

[0102] After obtaining the threshold set, the matching degree between it and the recognition performance is calculated using the support vector machine algorithm to obtain the matching degree value. If the matching degree value is lower than the preset threshold, the trend data output by the model is extracted through a feedback loop mechanism to obtain the trend analysis result. Based on the trend analysis result, the threshold set is optimized using the gradient descent algorithm to determine the fine-tuning parameter set. The threshold set is adjusted by fine-tuning the parameter set to generate an updated threshold set. For the updated threshold set, its matching degree with the recognition performance is recalculated to obtain a new matching degree value. If the new matching degree value is still lower than the preset threshold, the feedback loop and parameter adjustment steps are repeated to obtain the final optimized threshold set. Using the final optimized threshold set, performance evaluation is performed to determine the stability of the system's recognition performance.

[0103] Specifically, after obtaining the optimized threshold set, the system first determines the initial threshold set using a grid search algorithm. For example, it sets the probability thresholds for the classification model to 0.5, 0.6, and 0.7, and generates corresponding confidence thresholds for each category in a multi-class classification task. Assuming the dataset contains 10,000 samples, the model is a random forest classifier, and the training and test sets are divided in an 8:2 ratio, the system calculates the precision, recall, and F1 score for each threshold.

[0104] For example, with a threshold of 0.6, the precision is 0.85, the recall is 0.78, and the F1 score is 0.815. Next, the system assesses the match between the threshold set and the recognition performance. A weighted average method is used, with precision weighted at 0.4 and recall weighted at 0.6, to calculate the overall performance score, with a standard value of 0.8. If the score at a threshold of 0.6 is 0.79, lower than the standard value, a feedback loop mechanism is triggered. The system re-analyzes the trend model output, extracting feature importance. For example, the random forest model shows that features X1 (weight 0.35) and X2 (weight 0.25) have the greatest impact on classification. Based on this, the system fine-tunes the threshold using a gradient descent algorithm, setting the learning rate to 0.01, iterating 100 times, and optimizing to maximize the F1 score. After fine-tuning, the threshold is adjusted to 0.62, and the performance score is recalculated to 0.82, meeting the standard value. The entire process is implemented through automated scripts, with data flow processed in a Python environment, ensuring logical closure and performance optimization.

[0105] S106. The intelligent system configuration is updated by fine-tuning the parameters with a determined threshold to obtain an enhanced recognition module for processing image data streams of curbs, motorcycles, and license plates in complex scenarios.

[0106] The parameter adjustment range is determined by setting a preset threshold. Image data streams of curbside vehicles and license plates are acquired from complex scenes to generate an initial image set. If the initial image set contains noise, a Gaussian filtering algorithm is used to process it, resulting in a denoised image set. Based on the denoised image set, a convolutional neural network is used to extract features from the curbside vehicles and license plates, generating a feature image set. If the resolution of the feature image set is lower than a preset threshold, a super-resolution algorithm is used to enhance the feature image set, resulting in an enhanced image set. Based on the enhanced image set, an optical character recognition (OCR) algorithm is used to parse the license plate characters, generating a character data stream. The character data stream is matched against a preset license plate template to determine matching consistency, obtaining the license plate recognition result. Based on the license plate recognition result, the intelligent system configuration is updated, generating an enhanced recognition module.

[0107] Specifically, the parameters of the intelligent system are first fine-tuned by setting thresholds. An adaptive learning rate optimization algorithm (such as the Adam algorithm) is used with an initial learning rate of 0.001, momentum parameters β1=0.9, β2=0.999, and L2 regularization (regularization coefficient 0.01) is used to prevent overfitting. The system is trained on a dataset of 1000 images of curbs, motorcycles, and license plates in complex scenes (such as nighttime and rainy days), with a batch size of 32 and 100 epochs. The loss function (cross-entropy loss) on the validation set is monitored to converge to below 0.05 to ensure stable parameter updates. Next, the system configuration is updated by replacing the last layer of the convolutional neural network (CNN) model with a Softmax classifier, setting the output dimension to the size of the license plate character set (e.g., 37, including 0-9, 26 letters, and special characters), adjusting the convolutional kernel size to 3x3, the stride to 1, using 2x2 max pooling, and increasing the number of feature map channels to 256 to improve the feature extraction capability for complex scenes. The enhanced recognition module focuses on the license plate region by introducing an attention mechanism (SE module, compression ratio 16), combined with preprocessing steps (such as Gaussian blur, σ=1.5) to remove image noise, and enhances contrast through histogram equalization. This improves the recognition accuracy of license plate characters from 85% to 92% under low-light conditions (brightness below 50 lux). Data stream processing employs a real-time streaming framework (such as Apache Kafka), receiving image data at a rate of 10 frames per second. Multi-threaded parallel processing (8 threads) ensures latency of less than 100 milliseconds. The analysis process uses a confusion matrix to evaluate model performance, achieving a precision of 0.93, a recall of 0.91, and an F1 score of 0.92, indicating that the system can still stably recognize license plate characters in complex scenes (such as occlusion and 30-degree angle shift). The logic chain forms a closed loop from parameter optimization to model configuration adjustment, then to data stream processing and performance evaluation, ensuring that the output of each step serves as the input for the next step, ultimately achieving efficient license plate recognition.

[0108] S107. Obtain the processed image data stream from the enhanced recognition module, determine its applicability in the smart parking scenario, and if the applicability meets the requirements, output the final license plate recognition result to support traffic management applications.

[0109] Obtain the image data stream from the enhanced recognition module, and use a preprocessing algorithm to denoise and enhance the image to obtain the first image data. According to the first image data, use the convolutional neural network algorithm to extract the license plate area features and obtain the license plate location data. For the license plate location data, if the feature clarity exceeds the preset threshold, use the optical character recognition algorithm to parse the license plate characters and obtain the license plate text data. According to the license plate text data, judge whether it meets the license plate format requirements of the intelligent parking scenario to obtain the format verification result. If the format verification result meets the requirements, match the license plate text data with the preset parking lot vehicle database to obtain the vehicle identity information. According to the vehicle identity information, generate the license plate recognition result and output it to the traffic management application interface to obtain the final recognition output. Through the final recognition output, update the parking lot management database to complete the data synchronization of the traffic management application.

[0110] Specifically, obtain the processed image data stream from the enhanced recognition module. First, capture the parking lot vehicle image through a high-resolution camera. The image resolution is 1920x1080 pixels, and the frame rate is 30fps. Use the YOLOv5 object detection algorithm to preprocess the image, detect the license plate area, set the confidence threshold to 0.85, and剔除 low-confidence areas to reduce false detections. The detected license plate area is cropped to 256x128 pixels and input into a deep convolutional neural network (CNN) for feature extraction. The network includes 5 convolutional layers and 3 fully connected layers. The activation function uses ReLU, and a 64-dimensional feature vector is output. Then, perform cosine similarity matching between the feature vector and the pre-trained license plate character template library (including 36 characters, covering numbers 0-9 and 26 letters), and set the threshold to 0.9. Characters below this value are regarded as invalid. In the intelligent parking scenario, when analyzing the applicability, it is necessary to ensure that the recognition accuracy of the system under different lighting conditions (500-2000lux) and weather conditions (such as visibility ≥50m in rainy days) reaches more than 95%. By testing 1000 sample images, the actual accuracy is 96.2%, which meets the requirements. The final license plate recognition result is generated through the OCR algorithm. For example, the recognition result is "京A12345", and the result is output to the traffic management platform in JSON format, including the license plate number, the entry time (accurate to seconds), and the parking lot number (such as P001). The output data is transmitted to the cloud database through the RESTful API, supporting real-time query and illegal parking detection. The API response time is controlled within 200ms to ensure the efficiency of the traffic management application.

[0111] Obtain the high dynamic range image stream. Through the fusion of multi-source sensor data and real-time light change rate processing, if the image clarity and data integrity meet the threshold, use the deep learning algorithm to determine the license plate recognition result.

[0112] Multi-source sensor data is acquired, and noise is removed through preprocessing to generate a first data stream. Light features are extracted from the first data stream through real-time light change rate analysis to obtain light adjustment parameters. These parameters are then used to fuse the first data stream, generating a high dynamic range (HDL) image stream. If the sharpness of the HDL image stream reaches a preset sharpness threshold, image segmentation is performed to obtain a second image stream. If the data integrity of the second image stream reaches a preset integrity threshold, a convolutional neural network is used to extract license plate features. Based on these features, an optical character recognition (OCR) algorithm is employed to determine the license plate character sequence. The final license plate recognition result is obtained by comparing the character sequence with a preset database.

[0113] Specifically, when acquiring high dynamic range (HDR) image streams, multi-source sensors such as a CMOS camera and an infrared sensor can be fused to acquire visible light images (1920x1080 resolution, 30fps) and infrared images (640x480 resolution, 15fps). The timestamp synchronization error is controlled within 5ms. A weighted fusion algorithm (0.7 for visible light, 0.3 for infrared) is used to enhance details in highlight and low-light areas, generating HDR images. Real-time light change rate processing employs an optical flow algorithm to calculate the pixel brightness change rate per frame (threshold set at 0.1 lux / s). Kalman filtering is used to predict light trends, dynamically adjusting the camera exposure time (e.g., from 1 / 60s to 1 / 120s) to ensure the image is neither overexposed nor underexposed. Image sharpness is evaluated using the Laplacian variance method, with a sharpness threshold set at 100. Frames below this value are discarded. Data integrity is ensured by checking the pixel missing rate (required to be below 0.5%). Images meeting certain thresholds are input into the YOLOv5 model (training dataset contains 100,000 license plate images, mAP reaches 0.95). A convolutional neural network extracts license plate features, and an OCR algorithm (Tesseract, accuracy 0.92) is used to recognize characters, outputting the license plate number. Multi-source data fusion enhances robustness, optical flow and Kalman filtering ensure light adaptability, sharpness and integrity checks guarantee input quality, and a deep learning model accurately locates and recognizes the license plate, forming a complete logical chain from acquisition to recognition.

[0114] A high dynamic range image stream is obtained by fusing multi-source sensor data with real-time light change rate processing. If the image clarity and data integrity meet the threshold, a deep learning algorithm is used to determine the license plate recognition result.

[0115] Raw image data is acquired from multiple sensors, and data fusion technology is used to integrate the data from different sensors to obtain a high dynamic range (HDR) image stream. If the sharpness of the HDR image stream is lower than a preset threshold, an image enhancement algorithm is used to adjust the image contrast and brightness to obtain an optimized image stream. Real-time processing technology is used to analyze the light variation rate of the optimized image stream to obtain a light-adjusted image stream. If the data integrity of the light-adjusted image stream meets a preset threshold, a convolutional neural network algorithm is used to extract image features to obtain license plate region features. Based on the license plate region features, an optical character recognition (OCR) algorithm is used to parse the character information to obtain the license plate character sequence. By comparing the character sequence with a preset license plate format database, the legality of the license plate character sequence is determined, and the license plate recognition result is obtained.

[0116] Specifically, a high dynamic range image stream is obtained by fusing multi-source sensor data with real-time light change rate processing. The implementation method can employ a multi-sensor data fusion algorithm combined with adaptive illumination processing. First, multimodal data is acquired using an onboard camera, infrared sensor, and LiDAR. The camera captures RGB images at a resolution of 1920x1080 and a frame rate of 30fps; the infrared sensor provides thermal imaging under low-light conditions at a resolution of 640x480; and the LiDAR generates point cloud data with a point density of 1000 points per square meter. Data fusion uses a Kalman filter algorithm to align the three data sources to a unified timestamp (error <10ms), and a weighted average method (weights 0.5 for RGB, 0.3 for infrared, and 0.2 for point cloud) is used to generate a fused image. The light change rate is calculated in real time using luminance histogram analysis. A threshold of 0.1 is set for the luminance change rate; below this value, HDR synthesis is initiated. A logarithmic transformation algorithm is used to fuse multi-exposure images (exposure times of 1 / 100s, 1 / 500s, and 1 / 2000s) into a high dynamic range image stream, expanding the pixel grayscale range to 16 bits. Image sharpness is calculated using the Laplacian operator to determine edge intensity, with a threshold of 100. Data integrity is evaluated by detecting the proportion of missing pixels (<5%). If the threshold is met, a YOLOv5 deep learning model is used for license plate detection. The model input is 416x416 pixels, and the training dataset contains 100,000 license plate images, achieving an mAP of 0.95. License plate character recognition uses a CRNN network combined with the CTC loss function, achieving a recognition accuracy of 98%. If the light change rate is too high (>0.3), the fusion is optimized by increasing the infrared data weight (to 0.5) to ensure recognition stability under low light conditions. The entire process is handled in real time by an embedded GPU (such as NVIDIA Jetson TX2), with latency controlled within 50ms, ensuring system robustness and real-time performance.

[0117] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0118] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for recognizing license plates on curbside motorcycles, characterized in that: The method includes the following steps: S101, by collecting real-time environmental variable data during the process of license plate recognition of curb motors, extracts feature vectors from the environmental variable data as the input basis for the trend model; S102, Based on the extracted feature vector, the correlation coefficient between environmental variables and recognition performance is calculated using a trend model to obtain a performance degradation trend index, which serves as a reference for subsequent parameter adjustments. S103, if the performance degradation trend index exceeds the preset threshold, then perform a sensitivity calibration operation on the sensor array and obtain the adjusted sensor response curve as an intermediate result of dynamic optimization. S104, From the adjusted sensor response curve, determine the update range of the algorithm threshold, and fuse environmental variable data through an iterative optimization method to obtain an optimized threshold set to adapt to the current dynamic environment; including: Raw response data is acquired from the sensor, and signal processing methods are used to denoise and normalize the data to obtain the adjusted response curve. Based on the adjusted response curve, key feature points are extracted through curve analysis to determine the initial threshold range; If the environmental variable data does not match the preset range, the environmental variables are integrated using a weighted average method to obtain comprehensive environmental parameters. Based on the comprehensive environmental parameters and the initial threshold range, the gradient descent algorithm is used for iterative optimization to obtain the optimized threshold set; If the dynamic environment changes, the threshold set is adjusted by monitoring and updating environmental parameters in real time for the optimized threshold set. By comparing the adjusted threshold set with the sensor response data, it is determined whether the threshold is suitable for the current dynamic environment, and the final threshold set is obtained. Based on the final threshold set, control parameters adapted to the dynamic environment are generated and output to the system execution module; S105. After obtaining the optimized threshold set, determine its matching degree with the recognition performance. If the matching degree is lower than the standard value, re-analyze the trend model output through the feedback loop mechanism to determine further threshold fine-tuning parameters. S106, the intelligent system configuration is updated by fine-tuning parameters with a defined threshold to obtain an enhanced recognition module for processing image data streams of curb vehicles and license plates in complex scenarios; S107: Obtain the processed image data stream from the enhanced recognition module, determine its applicability in the smart parking scenario, and if the applicability meets the requirements, output the final license plate recognition result to support traffic management applications.

2. The method for recognizing license plates on curb motorcycles according to claim 1, characterized in that: In S101, real-time environmental variable data is collected during the license plate recognition process of the curb-mounted vehicle. Feature vectors are extracted from the environmental variable data as the input basis for the trend model, including: The raw dataset is obtained by collecting environmental variable data in real time through sensors, including light intensity and vehicle speed. Preprocessing methods are used to denoise and standardize the original dataset to obtain a standardized dataset; Feature vectors of light intensity and vehicle speed are extracted from the standardized dataset to obtain a feature vector set; The feature vector set is evaluated. If the light intensity in the feature vector set is lower than a preset threshold, an image enhancement algorithm is applied to adjust the image brightness to obtain an enhanced feature vector set. The enhanced feature vector set is classified using the support vector machine algorithm to determine the input category of the license plate recognition trend model. The parameters of the trend model are adjusted based on the classification results to obtain an optimized trend model; An optimized trend model is used to predict the trend of license plate recognition based on real-time collected environmental variable data.

3. The method for recognizing license plates on curb motorcycles according to claim 2, characterized in that: In S102, based on the extracted feature vectors, a trend model is used to calculate the correlation coefficient between environmental variables and recognition performance, obtaining performance degradation trend indicators, including: From the obtained feature vectors and environmental variables, a preprocessing method is used to normalize the feature vectors to obtain standardized feature vectors; Based on the standardized feature vector and environmental variables, the correlation coefficient between the two is calculated using the Pearson correlation coefficient method to obtain the correlation strength between environmental variables and recognition performance; The obtained correlation strength is judged. If the correlation strength exceeds the preset threshold, a linear regression model is used to fit the relationship between environmental variables and recognition performance to obtain the trend model parameters. Based on the trend model parameters, calculate the performance degradation trend index to determine the performance change trend; Key points are extracted from the performance change trend, and the gradient descent method is used to optimize the parameter adjustment strategy to obtain the adjusted parameter set. Based on the adjusted parameter set, the configuration of the recognition model is updated to obtain the optimized recognition performance; By comparing the optimized recognition performance with the original performance, the effect of parameter adjustment is evaluated, and the performance improvement index is obtained.

4. The method for recognizing license plates on curb motorcycles according to claim 3, characterized in that: In S103, if the performance degradation trend index exceeds a preset threshold, a sensitivity calibration operation is performed on the sensor array to obtain the adjusted sensor response curve as an intermediate result of dynamic optimization, including: If the performance degradation trend index exceeds the preset threshold, the data acquisition module will obtain real-time signal data from the sensor array to obtain the raw response dataset. The original response dataset is filtered and denoised using signal processing techniques to obtain a smooth response dataset. The support vector machine algorithm is used to extract features from the smooth response dataset to determine the key feature parameters of performance degradation; If the deviation between the key feature parameters and the preset threshold exceeds the specified range, the sensitivity calibration module is triggered to obtain the calibration parameter set; The sensitivity settings of the sensor array are adjusted by calibrating the parameter set to obtain the adjusted response curve; The adjusted response curve was fitted and analyzed using a linear regression algorithm to obtain intermediate results of dynamic optimization. The real-time monitoring module continuously tracks the intermediate results of dynamic optimization to determine whether the performance degradation trend has recovered to within the preset threshold.

5. The method for recognizing license plates on curb motorcycles according to claim 1, characterized in that: In S105, after obtaining the optimized threshold set, its matching degree with the recognition performance is determined. If the matching degree is lower than the standard value, the trend model output is re-analyzed through a feedback loop mechanism to determine further threshold fine-tuning parameters, including: After obtaining the threshold set, the matching degree between it and the recognition performance is calculated using the support vector machine algorithm to obtain the matching degree value; If the matching degree value is lower than the preset threshold, the trend data output by the model is extracted through the feedback loop mechanism to obtain the trend analysis results; Based on the trend analysis results, the gradient descent algorithm is used to optimize the threshold set and determine the fine-tuning parameter set. The threshold set is adjusted by fine-tuning the parameter set to generate an updated threshold set. For the updated threshold set, recalculate its matching degree with the recognition performance to obtain a new matching degree value; If the new matching value is still lower than the preset threshold, repeat the feedback loop and parameter adjustment steps to obtain the final optimized threshold set. By using the final optimized set of thresholds, a performance evaluation is performed to determine the stability of the system's recognition performance.

6. The method for recognizing license plates on curb motorcycles according to claim 5, characterized in that: In S106, the intelligent system configuration is updated using determined threshold fine-tuning parameters to obtain an enhanced recognition module, which is used to process the image data stream of curbside vehicle license plates in complex scenes, including: By determining the parameter adjustment range through a preset threshold, the image data stream of curb motorcycle license plates is obtained from complex scenes to generate an initial image set. If the initial image set contains noise, a Gaussian filtering algorithm is used to process the initial image set to obtain a denoised image set. Based on the denoised image set, a convolutional neural network is used to extract the license plate features of curb motorcycles and generate a feature image set. If the resolution of the feature image set is lower than a preset threshold, the feature image set is enhanced by a super-resolution algorithm to obtain an enhanced image set. Based on the enhanced image set, an optical character recognition algorithm is used to parse the license plate characters and generate a character data stream; By matching the character data stream with the preset license plate template, the consistency of the match is determined, and the license plate recognition result is obtained. The intelligent system configuration is updated based on the license plate recognition results to generate an enhanced recognition module.

7. The method for recognizing license plates on curb motorcycles according to claim 1, characterized in that: In S107, the processed image data stream is obtained from the enhanced recognition module, and its applicability in the smart parking scenario is determined. If the applicability meets the requirements, the final license plate recognition result is output to support traffic management applications, including: Image data streams are obtained from the enhancement recognition module, and preprocessing algorithms are used to denoise and enhance the images to obtain the first image data; Based on the first image data, a convolutional neural network algorithm is used to extract the license plate area features to obtain the license plate location data; For license plate location data, if the feature clarity exceeds a preset threshold, an optical character recognition algorithm is used to parse the license plate characters to obtain the license plate text data. Based on the license plate text data, determine whether it meets the license plate format requirements of the smart parking scenario, and obtain the format verification result; If the format verification result meets the requirements, the license plate text data will be matched with the preset parking lot vehicle database to obtain the vehicle identity information; Based on the vehicle identification information, generate license plate recognition results and output them to the traffic management application interface to obtain the final recognition output; The final identification output updates the parking management database, completing the data synchronization for traffic management applications.

8. The method for recognizing license plates on curb motorcycles according to claim 7, characterized in that: It also includes a high dynamic range image stream fusion and recognition process, specifically including: Acquire multi-source sensor data and preprocess it for noise reduction to generate the first data stream; Light characteristics are extracted by analyzing the real-time rate of change of light to obtain light adjustment parameters; A high dynamic range image stream is generated by fusing the first data stream based on the light adjustment parameters. If the image stream's clarity and data integrity meet a preset threshold, then the license plate features are extracted using a convolutional neural network, and the license plate character sequence is determined by combining it with an optical character recognition algorithm. The character sequence is compared with the preset database, and the final license plate recognition result is output.

9. The method for recognizing license plates on curb motorcycles according to claim 8, characterized in that: It also includes an optimized recognition process for high dynamic range image streams, specifically including: Raw image data is acquired through multiple source sensors, and a high dynamic range image stream is generated using data fusion technology. If the image stream clarity is lower than a preset threshold, the contrast and brightness are optimized using an image enhancement algorithm to obtain an optimized image stream; The light change rate of the image stream is analyzed and adjusted. If the integrity of the adjusted data meets the threshold, the license plate region features are extracted through a convolutional neural network. The character sequence is parsed using an optical character recognition algorithm, its legality is verified by comparing it with a preset license plate format database, and the license plate recognition result is output.

Images