A park access control system and method based on artificial intelligence

Abstract

The present invention discloses a park access control system and method based on artificial intelligence, wherein the method comprises: using a license plate recognition model based on an artificial intelligence algorithm to perform license plate recognition and color classification on visiting vehicles, obtaining the license plate number and vehicle type, and judging the access permission category of the visiting vehicle according to the license plate number; constructing a carbon emission calculation model according to the vehicle type, analyzing the carbon emissions of vehicles entering the park using the carbon emission calculation model, and obtaining carbon emission data; counting the total carbon emissions of vehicles entering the park within a preset period according to the carbon emission data, marking vehicles whose total carbon emissions and the preset carbon emission limit have a difference less than a preset threshold as restricted vehicles, and sending an early warning signal; making access decisions for external vehicles according to the vehicle type and the carbon emission capacity of the day. The present invention relates to the field of deep learning technology, and solves the technical problem that traditional access control systems cannot accurately identify vehicle types and make intelligent access decisions based on vehicle carbon emissions.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent access control and relates to deep learning technology, specifically to an artificial intelligence-based campus access control system and method. Background Art

[0002] Against the backdrop of the global "dual carbon" goals, industrial parks, as areas of dense industrial production and human activity, have seen vehicle carbon emissions become a significant source of pollution. Traditional access control systems rely solely on license plate character recognition for access control, failing to distinguish between vehicle types (e.g., fuel-powered vs. electric vehicles) and their carbon emission characteristics. This results in disorderly entry and exit of high-emission vehicles, making it difficult to meet low-carbon management requirements.

[0003] For example, the patent application with publication number "CN119445708A" discloses "An intelligent access control system for vehicle access control based on image recognition", which realizes license plate recognition and whitelist management through high-definition cameras and CNN models; the patent application with publication number "CN114283284A" discloses "A method and device for vehicle access control in a community", which proposes a multi-license plate adhesion segmentation method to improve the license plate recognition rate through color pixel analysis.

[0004] Therefore, existing technologies typically identify vehicles by license plate number and match them to access, lacking mechanisms for monitoring and managing vehicle carbon emissions. For example, existing technologies cannot accurately measure and compile statistics on carbon emissions from different types of vehicles, making it difficult to fully understand the total amount and distribution of carbon emissions from vehicles within the park. Furthermore, because carbon emissions are not factored into the vehicle access decision-making system, high-emission vehicles can easily enter and exit the park at will, increasing the park's carbon emissions burden. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes an artificial intelligence-based campus access control system and method to solve the technical problems that traditional access control systems are unable to accurately identify vehicle types, effectively manage vehicle carbon emissions, and make intelligent traffic decisions based on real-time carbon emissions.

[0006] To achieve the above objectives, the first aspect of the present invention provides an artificial intelligence-based campus access control system, comprising:

[0007] License Plate Recognition Module: This module uses an artificial intelligence-based license plate recognition model to identify and color-classify incoming vehicles, obtaining the license plate number and vehicle type. The module then determines the access permission category for incoming vehicles based on the license plate number. Vehicle types include electric vehicles, gasoline-powered vehicles, and hybrid vehicles, and access permission categories include in-park vehicles, pre-registered vehicles, and external vehicles.

[0008] Carbon emission analysis module: used to build a carbon emission calculation model based on vehicle type, use the carbon emission calculation model to analyze the carbon emissions of vehicles in the park and registered vehicles after entering the park, and obtain a number of carbon emission data;

[0009] Vehicle warning module: used to calculate the total carbon emissions of each vehicle entering the park within a preset period based on a number of carbon emission data. Vehicles whose total carbon emissions are less than the preset carbon emission limit are marked as restricted vehicles and a warning signal is sent to the restricted vehicles.

[0010] Traffic decision module: used to make traffic decisions for external vehicles based on vehicle type and daily carbon emission capacity.

[0011] It should be noted that the license plate recognition module, the carbon emission analysis module and the vehicle warning module of the present invention are in communication connection, and the carbon emission analysis module is in communication connection with the vehicle warning module and the traffic decision module respectively.

[0012] Furthermore, the license plate recognition model based on artificial intelligence algorithm is used to perform license plate recognition and color classification on visiting vehicles, including:

[0013] A1 uses a camera to capture an image of an incoming vehicle and uses a target detection model to locate the license plate in the image to obtain a license plate location frame. The target detection model is built based on a deep learning algorithm and deployed on the camera.

[0014] A2, crop the image of the visiting vehicle according to the license plate position frame to obtain the license plate image;

[0015] A3 uses a dual-branch license plate recognition model built based on a deep learning algorithm to perform license plate recognition and color classification on license plate images, and obtains the proportion of license plate numbers and several preset color types;

[0016] A4, obtaining the vehicle type of the visiting vehicle according to the proportion of several preset color types.

[0017] Furthermore, the dual-branch license plate recognition model includes: a shared feature network, a character recognition branch, and a color analysis branch; wherein,

[0018] The shared feature network is used to extract mixed features containing spatial semantic information and color distribution information from the input image to obtain character recognition guidance features and color analysis guidance features;

[0019] The character recognition branch is used to recognize license plate characters using character recognition guidance features to obtain the license plate number;

[0020] The color analysis branch is used to use the color analysis guidance feature to identify the background color of the license plate and quantify the color ratio to obtain the proportion of several preset color types.

[0021] The license plate recognition model extracts mixed features through a shared feature extraction network, and uses a dual-branch network architecture to simultaneously perform character recognition and color classification to achieve automatic classification of vehicle types; among them, the dual-branch network can simultaneously complete license plate character recognition and color classification tasks, reducing the time delay of traditional step-by-step processing and significantly improving recognition efficiency; the shared feature network splits the mixed features into character recognition-oriented features and color analysis-oriented features through a feature decoupling module, which not only retains spatial semantic information but also separates color distribution information, avoids feature interference, and can improve recognition accuracy.

[0022] Furthermore, the shared feature network includes a multi-level feature extraction module and a feature decoupling module; wherein,

[0023] The multi-level feature extraction module adopts an improved ConvNeXt-Tiny architecture, which includes the following connected in sequence:

[0024] Stem layer: It consists of a 4×4 convolution kernel, layer normalization, and GELU function to downsample the input image to 1 / 4.

[0025] A cascade module consisting of several Stage modules, each of which contains several dual-pass feature preservation units consisting of a main pass and a color preservation pass, where:

[0026] The main path includes depth-wise separable convolution, channel attention mechanism, and GELU function to extract spatial semantic features;

[0027] The color preservation path includes 3×3 average pooling and 1×1 convolution to compress and retain the original color information;

[0028] The outputs of the main path and the color preservation path are fused by element-by-element addition to obtain the mixed features of each Stage module, which are then input to the next Stage module.

[0029] The feature decoupling module is connected to the last Stage module of the multi-stage feature extraction module, and includes:

[0030] Feature channel splitting layer: splits the mixed features of the last Stage module into character recognition features and color analysis features;

[0031] Channel rearrangement layer: Performs random cross-channel permutation operations on character recognition features and color analysis features to obtain permuted character recognition-guided features and color analysis-guided features.

[0032] The multi-level feature extraction module is based on the improved ConvNeXt-Tiny architecture. Through deep separable convolution and channel attention mechanism, it can reduce computational complexity while ensuring feature extraction capabilities. It is suitable for deployment on edge devices such as cameras and can achieve real-time processing.

[0033] Furthermore, the character recognition branch includes a multi-scale feature fusion unit, a spatial attention enhancement module, a dynamic temporal modeling unit and a multi-head classification output layer; wherein,

[0034] The input of the multi-scale feature fusion unit comes from the output features of several Stage modules and the character recognition-oriented features. Through the 3×3 convolution kernel and bilinear upsampling operation, the output features of several Stage modules are unified to the same resolution, and the features are fused by channel splicing and 1×1 convolution cascade to obtain the fused features.

[0035] The spatial attention enhancement module uses SE Block to generate the channel weight C of the fusion feature through global average pooling, and uses the two-dimensional sine-cosine combination function to generate the initial parameter matrix P init :P init [i,j]=sin(i / H π)+cos(j / W π), the initial parameter matrix is ​​sequentially input into the 1×1 convolutional layer and the Sigmoid function to obtain the spatial position weight matrix S, and the channel weight C is multiplied element-by-element by the spatial position weight matrix S to obtain the spatial attention enhancement feature; where i and j represent the row index and column index in the initial parameter matrix respectively;

[0036] The dynamic time series modeling unit processes the spatial attention enhancement features using a bidirectional LSTM architecture to obtain time series features;

[0037] The multi-head classification output layer uses a fully connected network and a Softmax function to construct a digital recognition head, a letter recognition head, and a province abbreviation recognition head, and classifies and predicts the time series features to obtain the license plate number.

[0038] The character recognition branch enhances the model's adaptability to changes in license plate character size and position by fusing the output features of different Stage modules, making it especially robust to the recognition of blurred or tilted license plates. The SE Block generates a spatial position weight matrix to focus on the license plate character area, suppress background noise, and reduce the misrecognition rate. The bidirectional LSTM captures the contextual relationship of the character sequence, which can effectively solve the problem of character adhesion or occlusion, and improve the recognition accuracy of license plate numbers in complex scenarios.

[0039] Furthermore, the color analysis branch includes a shallow feature reuse unit, a hybrid color space conversion layer, a multi-receptive field classification unit and a color statistics output layer; wherein,

[0040] The shallow feature reuse unit performs channel splicing on the output features of the first Stage module of the multi-level feature extraction module through a jump connection to the color parsing guide feature of the feature decoupling module to obtain a spliced ​​feature;

[0041] The mixed color space conversion layer uses an HSV transformation subnetwork and a LAB transformation subnetwork, which are parallelized and include a 3×3 convolution kernel, batch normalization, and a ReLU activation function, to perform color conversion on the spliced ​​features, and splices the output of the HSV transformation subnetwork and the output of the LAB transformation subnetwork with the RGB3 channels of the input image to obtain a mixed color feature; wherein the HSV transformation subnetwork is optimized using a loss function that minimizes the classification error of the hue channel, and the LAB transformation subnetwork is optimized using a loss function that minimizes the regression error of the lightness channel;

[0042] The multi-receptive field classification unit processes the mixed color features using an expanded convolutional pyramid structure, and inputs the processed features into a 1×1 convolution layer for dimensionality reduction after channel splicing to obtain a color probability map; wherein the shape of the color probability map is [B, N, H, W], where B represents the batch size, N represents the preset number of color categories, and H and W are the length and width of the input image respectively;

[0043] The color statistics output layer includes:

[0044] Use the Softmax function to calculate along the category dimension N of the color probability map to obtain the color category probability distribution M of each pixel c,i,j ; Where c represents the color category, i and j represent the row index and column index of the pixel respectively;

[0045] The output features of the first Stage module of the multi-stage feature extraction module are averaged along the channel dimension to obtain a single-channel feature map F;

[0046] The single-channel feature map F is sequentially input into the 1×1 convolution and Sigmoid function to generate the spatial weight matrix S', and the color category probability distribution M c,i,j Multiply the spatial weight matrix S' element by element and sum along the spatial dimension to obtain the weighted probability sum P of each category c ;

[0047] The weighted probability sum P c Normalize and get the proportion of several preset color types.

[0048] The color analysis branch uses HSV and LAB color space conversions in parallel to optimize hue classification and brightness regression, respectively, which can enhance the accuracy of color classification and reduce the impact of lighting changes. At the same time, the expanded convolutional pyramid structure can capture color distribution characteristics at different scales, helping the model adapt to scenes with gradual changes in license plate background color or partial occlusion.

[0049] Furthermore, obtaining the vehicle type of the visiting vehicle according to the proportion of several preset color types includes:

[0050] Marking a vehicle with a green ratio greater than a first preset threshold, or a green ratio greater than a second preset threshold and a yellow ratio greater than a third preset threshold as an electric vehicle;

[0051] Marking a vehicle with a green ratio greater than a fourth preset threshold and a non-zero black ratio or a non-zero yellow ratio as a hybrid vehicle;

[0052] Vehicles with a blue ratio greater than a fifth preset threshold are marked as small fuel vehicles;

[0053] Vehicles with a yellow color ratio greater than a sixth preset threshold are marked as large fuel vehicles.

[0054] The method, which uses thresholds for different color ratios to determine vehicle type, is simple, direct, and scientific. It eliminates the need for additional sensors or complex detection methods; the vehicle type can be initially determined solely based on license plate color, reducing system cost and complexity.

[0055] Furthermore, the carbon emission calculation model is constructed according to the vehicle type, including:

[0056] The carbon emission calculation model for fuel vehicles is constructed as follows: C fuel =D×E base ×(1+αt emp )+β×t idle ; Among them, D represents the driving distance within the park, E base represents the baseline carbon emission coefficient, αt emp represents the temperature correction factor, β represents the idle emission rate, t idle Indicates idle time;

[0057] The carbon emission calculation model of electric vehicles is constructed as follows: C elec =E charge / η charge ×EF grid ×(1+α loss ); where E charge Indicates the charging capacity, η charge Indicates the charging efficiency of the charging pile, EF grid represents the grid emission factor, which is obtained from the State Grid regional emission factor database, αloss It represents the transmission loss compensation coefficient, which is obtained from the State Grid line loss rate report;

[0058] The carbon emission calculation model of hybrid vehicles is constructed as follows: C hybrid =γ×C fuel +(1-γ)×C elec ; Among them, γ represents the fuel driving weight, according to the formula Get, m represents the calibration coefficient, Indicates the preset power threshold.

[0059] Furthermore, the method for obtaining the baseline carbon emission coefficient and the idle emission rate includes:

[0060] When the visiting vehicle is an in-park vehicle or a pre-registered vehicle, the information is obtained through the built-in vehicle information database; wherein the built-in vehicle information database includes the license plate number, vehicle type, carbon emission standard, baseline carbon emission coefficient, idling carbon emission rate, preset carbon emission limit and vehicle photo;

[0061] When the visiting vehicle is a foreign vehicle, based on the vehicle type obtained by the license plate recognition model, several vehicle photos of the same vehicle type are screened out from the built-in information database, and the similarity between the several vehicle photos and the image of the visiting vehicle is calculated. The baseline carbon emission coefficient and idling carbon emission rate of the vehicle photo with the greatest similarity are screened out as the baseline carbon emission coefficient and idling carbon emission rate of the foreign vehicle.

[0062] Furthermore, the calculating of the similarity between the plurality of vehicle photos and the image of the visiting vehicle includes:

[0063] Using a pre-trained vehicle recognition model, features of a plurality of vehicle photos and images of visiting vehicles are respectively extracted to obtain a plurality of existing vehicle features and visiting vehicle features; wherein the pre-trained vehicle recognition model is constructed based on a deep learning algorithm, and the vehicle photos and the visiting vehicle images are obtained based on a camera at the same location;

[0064] The cosine similarity function is used to calculate the similarity between several existing vehicle features and the features of the visiting vehicle.

[0065] Different carbon emission calculation models are constructed for fuel vehicles, electric vehicles and hybrid vehicles, taking into account the carbon emission characteristics of different types of vehicles to make the calculation results more in line with actual conditions; and by matching the carbon emission parameters of external vehicles through vehicle image similarity, the problem of missing external vehicle data can be solved, thereby making more accurate communication decisions.

[0066] Furthermore, making a traffic decision for an incoming vehicle based on the vehicle type and the carbon emission capacity of the day includes:

[0067] D1, determine whether the vehicle type is an electric vehicle; if yes, send a pass control signal; if not, jump to D2;

[0068] D2, calculate the total carbon emissions of the park within a preset period based on a number of carbon emission data to obtain the used carbon emissions;

[0069] D3: Calculate the predicted carbon emissions from the parking location to the exit of the park by vehicles that did not leave the park the previous day using the carbon emission calculation model to obtain the reserved carbon emissions;

[0070] D4: Subtract the used carbon emissions and reserved carbon emissions from the park's carbon emission quota within the preset period, and divide the result by the remaining days in the preset period to obtain the estimated daily available carbon emissions.

[0071] D5, calculates the total carbon emissions for the day based on a number of carbon emission data, and compares whether the difference between the estimated daily available carbon emissions and the total carbon emissions for the day is greater than a preset reserved threshold; if yes, sends a pass-allowing control signal; if not, sends a pass-forbidden control signal.

[0072] The decision-making process comprehensively considers factors such as vehicle type and daily carbon emissions to develop reasonable access rules. Electric vehicles are directly allowed through, demonstrating support for environmentally friendly vehicles. For gasoline-powered vehicles, access is determined by calculating used carbon emissions, reserved carbon emissions, and estimated daily carbon emissions, and comparing them with the total carbon emissions for the day. This effectively controls the total carbon emissions within the park and enhances the intelligent and green nature of access control management.

[0073] A second aspect of the present invention provides an artificial intelligence-based campus access control method, comprising:

[0074] S1, using a license plate recognition model based on an artificial intelligence algorithm to perform license plate recognition and color classification on visiting vehicles, obtain the license plate number and vehicle type, and determine the access permission category of the visiting vehicle based on the license plate number; vehicle types include electric vehicles, fuel vehicles, and hybrid vehicles, and the access permission categories include in-park vehicles, pre-registered vehicles, and external vehicles;

[0075] S2, build a carbon emission calculation model based on vehicle type, use the carbon emission calculation model to analyze the carbon emissions of vehicles in the park and registered vehicles after entering the park, and obtain some carbon emission data;

[0076] S3, calculating the total carbon emissions of each vehicle entering the park within a preset period based on a number of carbon emission data, marking vehicles whose total carbon emissions are less than a preset threshold and the preset carbon emission limit as restricted vehicles, and sending a warning signal to the restricted vehicles;

[0077] S4 makes traffic decisions for incoming vehicles based on vehicle type and carbon emission capacity on that day.

[0078] Compared with the prior art, the present invention has the following beneficial effects:

[0079] Traditional access control systems rely on single license plate character recognition, are poorly adaptable to complex environments, and cannot distinguish between vehicle types. This invention utilizes a dual-branch network architecture to achieve parallel processing of license plate character and color classification. A shared feature network integrates spatial semantics and color information, while the character recognition branch employs multi-scale feature fusion and temporal modeling to address character adhesion and occlusion issues. The color parsing branch precisely quantifies the percentage of license plate colors through hybrid color space conversion and an expanded convolutional pyramid. Vehicle types are automatically classified based on the percentage of license plate colors, providing a reliable data foundation for carbon emission management.

[0080] Unlike traditional systems that only control the entry and exit of vehicles, the present invention constructs a differentiated carbon emission calculation model; among them, the fuel vehicle model introduces a temperature correction factor and an idling emission rate, and the electric vehicle model combines the grid emission factor and the transmission loss to achieve full life cycle carbon accounting. The data missing problem is solved by matching the parameters of external vehicles through vehicle image similarity; by real-time statistics of the total carbon emissions of a single vehicle, over-limit vehicles are dynamically marked and issued as warnings; at the same time, combined with the prediction of the exit of vehicles from the parking location to the day before that did not leave the park, carbon emission quotas are reserved to avoid waste of resources. On this basis, the traffic decision module gives priority to electric vehicles and intelligently adjusts the access rights of external fuel vehicles according to the remaining carbon emission capacity of the day, thereby achieving effective control of carbon emissions. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0082] Figure 1 A schematic diagram of the process of an artificial intelligence-based campus access control system provided by the present invention;

[0083] Figure 2 A schematic diagram of the framework of an artificial intelligence-based campus access control system provided by the present invention;

[0084] Figure 3 This is a schematic diagram of the architecture of the license plate recognition model based on artificial intelligence algorithm provided by the present invention. DETAILED DESCRIPTION

[0085] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0086] See also Figure 1-Figure 3 The first embodiment of the present invention provides an artificial intelligence-based campus access control system, including:

[0087] License Plate Recognition Module: This module uses an AI-based license plate recognition model to identify and color-classify incoming vehicles, obtaining the license plate number and vehicle type. The module then determines the access permission for incoming vehicles based on the license plate number. Vehicle types include electric vehicles, fuel vehicles, and hybrid vehicles, and access permission categories include in-park vehicles, pre-registered vehicles, and external vehicles.

[0088] Carbon emission analysis module: used to build a carbon emission calculation model based on vehicle type, use the carbon emission calculation model to analyze the carbon emissions of vehicles in the park and registered vehicles after entering the park, and obtain a number of carbon emission data;

[0089] Vehicle warning module: used to calculate the total carbon emissions of each vehicle entering the park within a preset period based on a number of carbon emission data. Vehicles whose total carbon emissions are less than the preset carbon emission limit are marked as restricted vehicles and a warning signal is sent to the restricted vehicles.

[0090] Traffic decision module: used to make traffic decisions for external vehicles based on vehicle type and daily carbon emission capacity.

[0091] In the existing access control system, the target detection model is mostly used to locate the license plate position, and then the convolutional neural network is used for character recognition, and then the vehicle's access rights are determined based on the recognized license plate number. When encountering foreign vehicles, on-site registration is usually adopted, and then they are allowed to enter. However, under the background of "dual carbon", the carbon emissions in the park come not only from industrial production and electricity consumption, but also from the exhaust gas of vehicles entering and leaving at will. Traditional systems cannot automatically distinguish between vehicle types (such as fuel vehicles and electric vehicles), and it is difficult to implement differentiated environmental protection control strategies. In order to achieve refined management of carbon emissions, the license plate recognition module of the present invention uses a dual-branch network architecture to add a vehicle type classification function on the basis of traditional license plate recognition, thereby providing data support for subsequent carbon emission calculations and traffic decisions. Specifically,

[0092] In this embodiment, a dual-branch deep learning network is used to simultaneously realize license plate character recognition and license plate background color classification. Its core architecture is as follows:

[0093] (1) Shared feature network: including multi-level feature extraction module and feature decoupling module;

[0094] The multi-level feature extraction module is based on the improved ConvNeXt-Tiny architecture. It consists of a Stem layer consisting of a 4×4 convolution kernel, layer normalization, and a GELU function (which downsamples the input image resolution to 1 / 4) and a feature-preserving cascade module consisting of several Stage modules. Each Stage module consists of two parallel paths:

[0095] Main path: Use depthwise separable convolution to extract spatial semantic features (such as license plate character outlines);

[0096] Color preservation path: compresses the original color information (such as the license plate background color) through 3×3 average pooling and 1×1 convolution;

[0097] The outputs of the main path and the color preservation path are fused by element-by-element addition to obtain the mixed features of each Stage module, which are then input to the next Stage module.

[0098] The feature decoupling module splits the output mixed features of the last stage module of the multi-level feature extraction module into character recognition features and color analysis features through the channel splitting layer. The channel rearrangement layer randomly permutes the channels of the character recognition features and color analysis features to eliminate the coupling interference between tasks. The permuted character recognition-oriented features and color analysis-oriented features are obtained and used as the input features of the two branches respectively.

[0099] The shared feature network adopts a two-stage architecture of shared feature extraction and task decoupling. It uses the ConvNeXt-improved DualPathBlock to achieve multi-path retention of underlying features. While the multi-level feature extraction module extracts common semantic features, it also preserves the original color distribution information through independent color paths. The feature decoupling layer uses channel splitting and rearrangement technology to separate mixed features into character recognition-oriented features and color analysis features. Compared with traditional single-task models, this can reduce repeated calculations and enable high-level semantic features to focus on character structure recognition, while shallow-level detail features support color analysis. This ensures the independence of the two tasks while achieving implicit feature interaction.

[0100] (2) Character recognition branch: including multi-scale feature fusion unit, spatial attention enhancement module, dynamic temporal modeling unit and multi-head classification output layer;

[0101] The input of the multi-scale feature fusion unit comes from the output features of several Stage modules and the character recognition-oriented features. Through the 3×3 convolution kernel and bilinear upsampling operation, the output features of several Stage modules are unified to the same resolution. The features are then fused using the cascade method of channel splicing and 1×1 convolution to obtain the fused features.

[0102] The spatial attention enhancement module uses SEBlock to generate the channel weight C of the fusion feature through global average pooling, and uses the two-dimensional sine-cosine combination function to generate the initial parameter matrix P init :P init [i,j]=sin(i / H π)+cos(j / W π), then the initial parameter matrix is ​​sequentially input into the 1×1 convolutional layer and the Sigmoid function to obtain the spatial position weight matrix S, and the channel weight C is element-wise multiplied with the spatial position weight matrix S to obtain the spatial attention enhancement feature; where i and j represent the row index and column index of the initial parameter matrix respectively;

[0103] The dynamic temporal modeling unit uses a bidirectional LSTM architecture to process the spatial attention enhancement features to obtain temporal features;

[0104] The multi-head classification output layer uses a fully connected network and a softmax function to construct a digit recognition head (containing 10 fully connected layers corresponding to the numbers 0-9), a letter recognition head (containing 24 fully connected layers, excluding the letters O / I that are easily confused with numbers), and a provincial abbreviation recognition head (containing 31 fully connected layers corresponding to the abbreviations of each province). The time series features are then classified and predicted to obtain the license plate number.

[0105] The character recognition branch uses a spatial attention enhancement module and dynamic LSTM sequence modeling, and embeds license plate format rules using a classification prediction head (province / letter / number). This can solve the recognition problem of easily confused characters such as "0 / O";

[0106] (3) Color analysis branch: including shallow feature reuse unit, mixed color space conversion layer, multi-receptive field classification unit and color statistics output layer; among them,

[0107] The shallow feature reuse unit reuses the high-resolution output features of the multi-stage feature extraction module, that is, the output features of the first stage module, and performs channel splicing of the features with the color parsing-guided features of the feature decoupling module through skip connections to obtain spliced ​​features;

[0108] The hybrid color space conversion layer sets up an HSV transformation subnetwork and a LAB transformation subnetwork in parallel. Each subnetwork contains a cascade structure of a 3×3 convolution kernel, batch normalization, and a ReLU activation function. The HSV transformation subnetwork and the LAB transformation subnetwork are optimized by minimizing the loss function of the hue channel classification error and the loss function of minimizing the lightness channel regression error respectively. The output features of the two subnetworks are then spliced ​​with the RGB3 channels of the input image to obtain a mixed color feature: original RGB (3 channels) + HSV (3 channels) + LAB (3 channels) = 9-channel mixed feature.

[0109] The multi-receptive field classification unit uses a dilated convolutional pyramid structure to process mixed color features using 3×3 convolutions with different dilation rates, expanding the receptive field to cover the color gradient area. The processed features are then channel-wise concatenated and input into a 1×1 convolution layer for dimensionality reduction to obtain a color probability map. The shape of the color probability map is [B, N, H, W], where B represents the batch size, N represents the preset number of color categories, and H and W represent the length and width of the input image, respectively.

[0110] Color statistics output layer, including:

[0111] Use the Softmax function to calculate along the category dimension N of the color probability map to obtain the color category probability distribution M of each pixel c,i,j ;

[0112] Then, the output features of the first Stage module of the multi-stage feature extraction module are averaged along the channel dimension to obtain a single-channel feature map F;

[0113] Input the single-channel feature map F into the 1×1 convolution and Sigmoid function in sequence to generate a spatial weight matrix , and the color category probability distribution M c,i,j With the spatial weight matrix Perform element-by-element multiplication and sum along the spatial dimension in sequence to obtain the weighted probability sum P of each category c , and its calculation formula is: Where c represents the color category, i and j represent the row index and column index in the color probability map, respectively, and h and w represent the height and width of the color probability map, respectively;

[0114] The weighted probability sum P c To standardize: , get the proportion of several preset color types; where N represents the number of preset color categories;

[0115] The color analysis branch uses the HSV-LAB hybrid space conversion layer, learns nonlinear color mapping through a two-stream convolutional network, and combines it with an expanded convolutional pyramid to capture gradient features. Compared with the traditional RGB space method, it can reduce the color recognition error rate in rainy weather; and the introduction of the spatial weighted statistical layer can quantitatively analyze the regional proportion of gradient colors on new energy vehicle license plates, which is more convenient for vehicle type classification.

[0116] In this example, a dual-branch deep learning network was built using the PyTorch framework. The CCPD license plate detection dataset (including license plate position annotations), the CRPD character recognition dataset (including character sequence labels), and several manually annotated license plate color classification datasets were used as experimental data. Enhancement strategies such as random rotation, illumination perturbation, and occlusion simulation were used to improve the robustness of the model.

[0117] During joint training of the two-branch task, the AdamW optimizer and cosine annealing scheduling strategy are used. The character recognition branch uses the CTC loss function (Connectionist Temporal Classification), and the color analysis branch uses the FocalLoss loss function to address the problem of color category imbalance.

[0118] After training, the resulting license plate recognition model is compressed through channel pruning and INT8 quantization, and ultimately deployed to a high-definition camera equipped with an NVIDIA Jetson Xavier NX embedded module for real-time single-frame processing.

[0119] In the license plate recognition module, a multi-stage deep learning algorithm is used to achieve intelligent recognition of vehicle identity and type. The specific steps are as follows:

[0120] A1 uses a camera to capture an image of an incoming vehicle and uses an object detection model to locate the license plate in the image, obtaining a license plate location frame. The object detection model is built based on a deep learning algorithm and deployed on the camera.

[0121] A2, crop the image of the visiting vehicle according to the license plate position frame to obtain the license plate image;

[0122] A3 uses a dual-branch license plate recognition model built based on a deep learning algorithm to perform license plate recognition and color classification on license plate images, and obtains the proportion of license plate numbers and several preset color types;

[0123] A4, automatic classification of vehicle types based on color ratio threshold rules:

[0124] Marking a vehicle with a green ratio greater than a first preset threshold, or a green ratio greater than a second preset threshold and a yellow ratio greater than a third preset threshold as an electric vehicle;

[0125] Marking a vehicle with a green ratio greater than a fourth preset threshold and a non-zero black ratio or a non-zero yellow ratio as a hybrid vehicle;

[0126] Vehicles with a blue ratio greater than a fifth preset threshold are marked as small fuel vehicles;

[0127] Vehicles with a yellow color ratio greater than a sixth preset threshold are marked as large fuel vehicles.

[0128] In this embodiment, the color ratio threshold rule may be as follows:

[0129] Electric vehicles: Green ratio > 90% (pure electric), or green > 60% and yellow > 20% (large new energy vehicles);

[0130] Hybrid vehicles: Green > 70% and black / yellow (hybrid or plug-in hybrid);

[0131] Small fuel vehicles: Blue>90% (ordinary blue license plate);

[0132] Large fuel vehicles: Yellow>80% (yellow-plate trucks / buses).

[0133] In the carbon emission analysis module, a carbon emission calculation model is constructed based on vehicle type, which serves as a statistical tool for carbon emissions from vehicles within the park. Specifically, it includes:

[0134] The carbon emission calculation model for fuel vehicles is constructed as follows: C fuel =D×E base ×(1+αt emp )+β×t idle ; Among them, D represents the driving distance within the park, E base represents the baseline carbon emission coefficient, αt emp represents the temperature correction factor, β represents the idle emission rate, t idle Indicates idle time;

[0135] The carbon emission calculation model of electric vehicles is constructed as follows: C elec =E charge / η charge ×EF grid ×(1+α loss ); where E charge Indicates the charging capacity, η charge Indicates the charging efficiency of the charging pile, EF grid represents the grid emission factor, which is obtained from the State Grid regional emission factor database, α loss It represents the transmission loss compensation coefficient, which is obtained from the State Grid line loss rate report;

[0136] The carbon emission calculation model of hybrid vehicles is constructed as follows: Chybrid =γ×C fuel +(1-γ)×C elec ; Among them, γ represents the fuel driving weight, according to the formula Obtained, m represents the calibration coefficient, obtained through practical experience, Indicates the preset power threshold;

[0137] The methods for obtaining the baseline carbon emission coefficient and idling emission rate include:

[0138] When the visiting vehicle is an in-park vehicle or a pre-registered vehicle, the information is obtained through the built-in vehicle information database; the built-in vehicle information database includes the license plate number, vehicle type, carbon emission standard, baseline carbon emission coefficient, idling carbon emission rate, preset carbon emission limit and vehicle photo;

[0139] When the visiting vehicle is a foreign vehicle, the vehicle type obtained by the license plate recognition model is used to filter out several vehicle photos of the same vehicle type from the built-in information database.

[0140] The pre-trained vehicle recognition model is then used to extract features from several vehicle photos and images of visiting vehicles, obtaining features of existing vehicles and visiting vehicles. The pre-trained vehicle recognition model is built based on a deep learning algorithm, and to ensure that the vehicles in the photos are viewed from the same angle as much as possible, both the vehicle photos and the visiting vehicle images are acquired from a camera in the same location.

[0141] Then, the cosine similarity function is used to calculate the similarity between several existing vehicle features and the visiting vehicle features, and the baseline carbon emission coefficient and idling carbon emission rate of the vehicle photo with the greatest similarity are selected as the baseline carbon emission coefficient and idling carbon emission rate of the foreign vehicle.

[0142] After each vehicle enters the park, the carbon emission analysis module uses the UWB ultra-wideband positioning system to obtain the vehicle's driving path in real time within the park. Combined with the digital twin map, it generates dynamic trajectory data and calculates the precise driving distance. Alternatively, it uses the park's cameras to capture multi-view video streams of the vehicle and uses the license plate number to match the trajectory of the same vehicle across cameras, forming a complete driving path and determining the driving distance.

[0143] For new energy vehicles, the module can exchange data with the on-board BMS system through the charging pile data interface, read the vehicle's charging power in the park in real time, and simultaneously obtain the charging efficiency parameters of the charging pile; for fuel vehicles, the module can collect idling time through the on-board OBD interface, or locate the vehicle through the target detection model of the camera in the park, and count the time when the vehicle is stationary and continuously emitting exhaust gas to obtain the idling time;

[0144] Then, the carbon emission output of each car is obtained according to the carbon emission calculation model and stored in the carbon emission database. The carbon emission data of vehicles in the park and registered vehicles are linked to the vehicle information database to obtain the vehicle's carbon footprint ledger and transmit it to the vehicle early warning module.

[0145] In the vehicle warning module, the total carbon emissions of each vehicle entering the park within a preset period are calculated based on a number of carbon emission data. Vehicles whose total carbon emissions are less than a preset threshold (the preset threshold can be set as 10% of the preset carbon emission limit) are marked as restricted vehicles and a warning signal is sent to the restricted vehicles.

[0146] For external vehicles, the traffic decision module is used to determine their traffic rights based on vehicle type and carbon emission capacity on that day. Specifically, it may include:

[0147] D1, determine whether the vehicle type is an electric vehicle; if yes, send a pass control signal; if not, jump to D2;

[0148] D2, calculate the total carbon emissions of the park within a preset period based on a number of carbon emission data to obtain the used carbon emissions;

[0149] D3: Calculate the predicted carbon emissions from the parking location to the exit of the park by vehicles that did not leave the park the previous day using the carbon emission calculation model to obtain the reserved carbon emissions;

[0150] D4: Subtract the used carbon emissions and reserved carbon emissions from the park's carbon emission quota within the preset period, and divide the result by the remaining days in the preset period to obtain the estimated daily available carbon emissions.

[0151] D5, calculates the total carbon emissions for the day based on a number of carbon emission data, and compares whether the difference between the estimated daily available carbon emissions and the total carbon emissions for the day is greater than a preset reserved threshold; if yes, sends a pass-allowing control signal; if not, sends a pass-forbidden control signal.

[0152] A second embodiment of the present invention provides an artificial intelligence-based campus access control method, comprising:

[0153] S1 uses a license plate recognition model based on an artificial intelligence algorithm to identify and color-classify visiting vehicles, obtain the license plate number and vehicle type, and determine the access permission category of the visiting vehicle based on the license plate number. Vehicle types include electric vehicles, fuel vehicles, and hybrid vehicles, and access permission categories include in-park vehicles, pre-registered vehicles, and external vehicles.

[0154] S2, build a carbon emission calculation model based on vehicle type, use the carbon emission calculation model to analyze the carbon emissions of vehicles in the park and registered vehicles after entering the park, and obtain some carbon emission data;

[0155] S3, calculating the total carbon emissions of each vehicle entering the park within a preset period based on a number of carbon emission data, marking vehicles whose total carbon emissions are less than a preset threshold and the preset carbon emission limit as restricted vehicles, and sending a warning signal to the restricted vehicles;

[0156] S4 makes traffic decisions for incoming vehicles based on vehicle type and carbon emission capacity on that day.

[0157] Some of the data in the above formula are calculated by removing the dimensions and taking their numerical values. The formula is a formula that is closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0158] Working principle of the present invention:

[0159] The present invention first uses the license plate recognition module to collect vehicle images through a camera, and uses the improved ConvNeXt-Tiny architecture and a dual-branch network to perform character recognition and color analysis of the license plate number, and then completes vehicle classification in combination with the license plate color ratio threshold;

[0160] The carbon emission analysis module then builds a differentiated model based on vehicle type and dynamically calculates the carbon emission data of vehicles entering the park;

[0161] The early warning module then counts the total carbon emissions of each vehicle within a preset period and sends early warnings to vehicles approaching the limit; the traffic decision module implements graded control for external vehicles: electric vehicles can be released directly, while fuel vehicles are dynamically evaluated based on the remaining carbon emission capacity of the day, and are allowed to pass if the remaining amount meets the standard.

[0162] The above embodiments are only used to illustrate the technical method 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 preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An artificial intelligence-based campus access control system, characterized in that: include: License Plate Recognition Module: This module uses a dual-branch license plate recognition model based on an artificial intelligence algorithm to identify and classify the license plates of incoming vehicles, obtaining the license plate number and vehicle type. The module then determines the access permission category of the incoming vehicle based on the license plate number. Vehicle types include electric vehicles, fuel vehicles, and hybrid vehicles, and access permission categories include in-park vehicles, pre-registered vehicles, and external vehicles. The dual-branch license plate recognition model based on artificial intelligence algorithm is used to perform license plate recognition and color classification on visiting vehicles, including: A1 uses a camera to capture an image of an incoming vehicle and uses a target detection model to locate the license plate in the image to obtain a license plate location frame. The target detection model is built based on a deep learning algorithm and deployed on the camera. A2, crop the image of the visiting vehicle according to the license plate position frame to obtain the license plate image; A3, using a dual-branch license plate recognition model to perform license plate recognition and color classification on the license plate image, and obtain the proportion of license plate numbers and several preset color types; A4, obtaining the vehicle type of the visiting vehicle according to the proportion of several preset color types; The dual-branch license plate recognition model includes: a shared feature network, a character recognition branch, and a color analysis branch; wherein, The shared feature network is used to extract mixed features containing spatial semantic information and color distribution information from the input image to obtain character recognition guidance features and color analysis guidance features; The character recognition branch is used to recognize license plate characters using character recognition guidance features to obtain the license plate number; The color analysis branch is used to use the color analysis guidance feature to identify the license plate background color and quantify the color ratio to obtain the proportion of several preset color types; Carbon emission analysis module: used to build a carbon emission calculation model based on vehicle type, use the carbon emission calculation model to analyze the carbon emissions of vehicles in the park and registered vehicles after entering the park, and obtain a number of carbon emission data; Vehicle warning module: used to calculate the total carbon emissions of each vehicle entering the park within a preset period based on a number of carbon emission data. Vehicles whose total carbon emissions are less than the preset carbon emission limit are marked as restricted vehicles and a warning signal is sent to the restricted vehicles. Traffic decision module: used to make traffic decisions for external vehicles based on vehicle type and daily carbon emission capacity.

2. The artificial intelligence-based campus access control system according to claim 1 is characterized in that: The shared feature network includes a multi-level feature extraction module and a feature decoupling module; wherein, The multi-level feature extraction module adopts an improved ConvNeXt-Tiny architecture, which includes the following connected in sequence: Stem layer: It consists of a 4×4 convolution kernel, layer normalization, and GELU function to downsample the input image to 1 / 4. A cascade module consisting of several Stage modules, each of which contains several dual-pass feature preservation units consisting of a main pass and a color preservation pass, where: The main path includes depth-wise separable convolution, channel attention mechanism, and GELU function to extract spatial semantic features; The color preservation path includes 3×3 average pooling and 1×1 convolution to compress and retain the original color information; The outputs of the main path and the color preservation path are fused by element-by-element addition to obtain the mixed features of each Stage module, which are then input to the next Stage module. The feature decoupling module is connected to the last Stage module of the multi-stage feature extraction module, and includes: Feature channel splitting layer: splits the mixed features of the last Stage module into character recognition features and color analysis features; Channel rearrangement layer: Performs random cross-channel permutation operations on character recognition features and color analysis features to obtain permuted character recognition-guided features and color analysis-guided features.

3. The artificial intelligence-based park access control system according to claim 2 is characterized in that: The character recognition branch includes a multi-scale feature fusion unit, a spatial attention enhancement module, a dynamic temporal modeling unit and a multi-head classification output layer; wherein, The input of the multi-scale feature fusion unit comes from the output features of several Stage modules and the character recognition-oriented features. Through the 3×3 convolution kernel and bilinear upsampling operation, the output features of several Stage modules are unified to the same resolution, and the features are fused by channel splicing and 1×1 convolution cascade to obtain the fused features. The spatial attention enhancement module uses SE Block to generate the channel weight C of the fusion feature through global average pooling, and uses the two-dimensional sine-cosine combination function to generate the initial parameter matrix P init :P init [i,j]=sin(i / H π)+cos(j / W π), the initial parameter matrix is ​​sequentially input into the 1×1 convolutional layer and the Sigmoid function to obtain the spatial position weight matrix S, and the channel weight C is multiplied element-by-element by the spatial position weight matrix S to obtain the spatial attention enhancement feature; where i and j represent the row index and column index in the initial parameter matrix respectively; The dynamic time series modeling unit processes the spatial attention enhancement features using a bidirectional LSTM architecture to obtain time series features; The multi-head classification output layer uses a fully connected network and a Softmax function to construct a digital recognition head, a letter recognition head, and a province abbreviation recognition head, and classifies and predicts the time series features to obtain the license plate number.

4. The artificial intelligence-based campus access control system according to claim 2 is characterized in that: The color analysis branch includes a shallow feature reuse unit, a mixed color space conversion layer, a multi-receptive field classification unit and a color statistics output layer; wherein, The shallow feature reuse unit performs channel splicing on the output features of the first Stage module of the multi-level feature extraction module through a jump connection to the color parsing guide feature of the feature decoupling module to obtain a spliced ​​feature; The mixed color space conversion layer uses an HSV transformation subnetwork and a LAB transformation subnetwork, which are parallelized and include a 3×3 convolution kernel, batch normalization, and a ReLU activation function, to perform color conversion on the spliced ​​features, and splices the output of the HSV transformation subnetwork and the output of the LAB transformation subnetwork with the RGB3 channels of the input image to obtain a mixed color feature; wherein the HSV transformation subnetwork is optimized using a loss function that minimizes the classification error of the hue channel, and the LAB transformation subnetwork is optimized using a loss function that minimizes the regression error of the lightness channel; The multi-receptive field classification unit processes the mixed color features using an expanded convolutional pyramid structure, and inputs the processed features into a 1×1 convolution layer for dimensionality reduction after channel splicing to obtain a color probability map; wherein the shape of the color probability map is [B, N, H, W], where B represents the batch size, N represents the preset number of color categories, and H and W are the length and width of the input image respectively; The color statistics output layer includes: Use the Softmax function to calculate along the category dimension N of the color probability map to obtain the color category probability distribution M of each pixel c,i,j ; Where c represents the color category, i and j represent the row index and column index of the pixel respectively; The output features of the first Stage module of the multi-stage feature extraction module are averaged along the channel dimension to obtain a single-channel feature map F; The single-channel feature map F is sequentially input into the 1×1 convolution and Sigmoid function to generate the spatial weight matrix S', and the color category probability distribution M c,i,j Multiply the spatial weight matrix S' element by element and sum along the spatial dimension to obtain the weighted probability sum P of each category c ; The weighted probability sum P c Normalize and get the proportion of several preset color types.

5. The artificial intelligence-based campus access control system according to claim 1 is characterized in that: The carbon emission calculation model is constructed based on vehicle type, including: The carbon emission calculation model for fuel vehicles is constructed as follows: C fuel =D×E base ×(1+αt emp )+β×t idle ; Among them, D represents the driving distance within the park, E base represents the baseline carbon emission coefficient, αt emp represents the temperature correction factor, β represents the idle emission rate, t idle Indicates idle time; The carbon emission calculation model of electric vehicles is constructed as follows: C elec =E charge / η charge ×EF grid ×(1+α loss ); where E charge Indicates the charging capacity, η charge Indicates the charging efficiency of the charging pile, EF grid represents the grid emission factor, which is obtained from the State Grid regional emission factor database, α loss It represents the transmission loss compensation coefficient, which is obtained from the State Grid line loss rate report; The carbon emission calculation model of hybrid vehicles is constructed as follows: C hybrid =γ×C fuel +(1-γ)×C elec ; Among them, γ represents the fuel driving weight, according to the formula Get, m represents the calibration coefficient, Indicates the preset power threshold.

6. The artificial intelligence-based campus access control system according to claim 5 is characterized in that: The method for obtaining the baseline carbon emission coefficient and the idle emission rate includes: When the visiting vehicle is an in-park vehicle or a pre-registered vehicle, the information is obtained through the built-in vehicle information database; wherein the built-in vehicle information database includes the license plate number, vehicle type, carbon emission standard, baseline carbon emission coefficient, idling carbon emission rate, preset carbon emission limit and vehicle photo; When the visiting vehicle is a foreign vehicle, based on the vehicle type obtained by the license plate recognition model, several vehicle photos of the same vehicle type are screened out from the built-in information database, and the cosine similarity function is used to calculate the similarity between the several vehicle photos and the image of the visiting vehicle. The baseline carbon emission coefficient and idling carbon emission rate of the vehicle photo with the greatest similarity are screened out as the baseline carbon emission coefficient and idling carbon emission rate of the foreign vehicle.

7. The artificial intelligence-based campus access control system according to claim 1 is characterized in that: The decision on the passage of external vehicles based on vehicle type and carbon emission capacity on the day includes: D1, determine whether the vehicle type is an electric vehicle; if yes, send a pass control signal; if not, jump to D2; D2, calculate the total carbon emissions of the park within a preset period based on a number of carbon emission data to obtain the used carbon emissions; D3: Calculate the predicted carbon emissions from the parking location to the exit of the park by vehicles that did not leave the park the previous day using the carbon emission calculation model to obtain the reserved carbon emissions; D4: Subtract the used carbon emissions and reserved carbon emissions from the park's carbon emission quota within the preset period, and divide the result by the remaining days in the preset period to obtain the estimated daily available carbon emissions. D5, calculates the total carbon emissions for the day based on a number of carbon emission data, and compares whether the difference between the estimated daily available carbon emissions and the total carbon emissions for the day is greater than a preset reserved threshold; if yes, sends a pass-allowing control signal; if not, sends a pass-forbidden control signal.

8. An artificial intelligence-based campus access control method, applied to an artificial intelligence-based campus access control system according to any one of claims 1 to 7, characterized in that: include: S1, using a license plate recognition model based on an artificial intelligence algorithm to perform license plate recognition and color classification on visiting vehicles, obtain the license plate number and vehicle type, and determine the access permission category of the visiting vehicle based on the license plate number; vehicle types include electric vehicles, fuel vehicles, and hybrid vehicles, and the access permission categories include in-park vehicles, pre-registered vehicles, and external vehicles; The dual-branch license plate recognition model based on artificial intelligence algorithm is used to perform license plate recognition and color classification on visiting vehicles, including: A1 uses a camera to capture an image of an incoming vehicle and uses a target detection model to locate the license plate in the image to obtain a license plate location frame. The target detection model is built based on a deep learning algorithm and deployed on the camera. A2, crop the image of the visiting vehicle according to the license plate position frame to obtain the license plate image; A3, using a dual-branch license plate recognition model to perform license plate recognition and color classification on the license plate image, and obtain the proportion of license plate numbers and several preset color types; A4, obtaining the vehicle type of the visiting vehicle according to the proportion of several preset color types; The dual-branch license plate recognition model includes: a shared feature network, a character recognition branch, and a color analysis branch; wherein, The shared feature network is used to extract mixed features containing spatial semantic information and color distribution information from the input image to obtain character recognition guidance features and color analysis guidance features; The character recognition branch is used to recognize license plate characters using character recognition guidance features to obtain the license plate number; The color analysis branch is used to use the color analysis guidance feature to identify the license plate background color and quantify the color ratio to obtain the proportion of several preset color types; S2, build a carbon emission calculation model based on vehicle type, use the carbon emission calculation model to analyze the carbon emissions of vehicles in the park and registered vehicles after entering the park, and obtain some carbon emission data; S3, calculating the total carbon emissions of each vehicle entering the park within a preset period based on a number of carbon emission data, marking vehicles whose total carbon emissions are less than a preset threshold and the preset carbon emission limit as restricted vehicles, and sending a warning signal to the restricted vehicles; S4 makes traffic decisions for incoming vehicles based on vehicle type and carbon emission capacity on that day.

Images