A parking lot cloud seat remote management system

Abstract

This invention discloses a cloud-based remote parking management system. The system initially enhances the original image by introducing noise suppression and contrast enhancement algorithms. Then, a multi-feature fusion network is used to comprehensively extract features from the license plate, vehicle body contour, and motion trajectory to accurately locate the actual license plate area. Furthermore, adaptive image enhancement of the license plate area is performed based on the identified weather scene, providing more accurate image data for subsequent license plate recognition. This multi-optimization mechanism effectively overcomes the low recognition rate of traditional systems in adverse weather conditions. Simultaneously, the cloud service layer not only handles license plate recognition but also possesses abnormal event recognition capabilities. It can proactively detect abnormal situations within the lane and generate alarm information to push to the operator layer. Based on this, an intelligent upgrade from passive response to proactive early warning is achieved, significantly improving management efficiency and emergency response speed.

Description

Technical Field

[0001] This invention belongs to the field of unattended parking lot management technology, specifically relating to a parking lot cloud-based remote management system. Background Technology

[0002] The current mainstream unmanned remote parking management system has formed a technical architecture of "license plate recognition + cloud platform + remote operator". It is mainly used in scenarios such as commercial complexes, residential communities, and industrial parks, and can realize centralized management of multiple parking lots by one person, thereby reducing labor costs by 30%-70%. The core components of the aforementioned management system include: Vehicle self-service recognition module: Employs high-definition cameras and deep learning algorithms, achieving a recognition accuracy rate of 99.9%, and supports multiple license plate types including blue plates, green plates, and new energy vehicle plates; Cloud-based customer service management platform: Supports multi-terminal access, can manage 100+ parking lots simultaneously, and provides functions such as real-time monitoring, remote barrier lifting, and order confirmation; Remote intercom terminal: Integrates a camera and speaker to enable two-way voice communication between the vehicle owner and the customer service representative, with some models supporting dual-channel video feeds from the driver's seat and panoramic lane view; Video monitoring system: Records vehicle entry, exit, and parking in real time for safety management and dispute resolution.

[0003] Among them, the aforementioned management system has the following shortcomings: (1) Insufficient stability under extreme weather conditions: the recognition rate drops sharply under severe weather conditions such as rain, snow, fog, and strong light, leading to lane congestion; (2) Lack of intelligent early warning mechanism, that is, abnormal situations mostly rely on manual discovery, passive response rather than active early warning; Therefore, based on the aforementioned shortcomings, how to provide a parking lot cloud seat remote management system with high recognition accuracy and intelligent early warning has become an urgent problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a parking lot cloud-based remote management system to solve the problems of license plate recognition accuracy dropping sharply in extreme weather scenarios and lack of intelligent early warning mechanisms in existing technologies.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a parking lot cloud-based remote management system is provided, including: The front-end perception layer is used to acquire license plate images and vehicle video streams, and uses noise suppression and contrast enhancement algorithms to enhance the license plate images to obtain enhanced images. The network transport layer is used to transmit the enhanced images and vehicle video streams output by the front-end perception layer to the cloud service layer. The cloud service layer is used to extract the license plate features of vehicles in the enhanced image, as well as the vehicle body contour features and vehicle motion trajectory features in the vehicle video stream, using a multi-feature fusion network. This allows for feature fusion of the license plate features, vehicle body contour features, and vehicle motion trajectory features to obtain fused features, and outputs the vehicle's real license plate area based on the fused features. The cloud service layer is used to identify the weather scene where the vehicle is located based on the license plate image, and to perform adaptive image enhancement processing on the real license plate area according to the weather scene where the vehicle is located to obtain an enhanced license plate area. Different image enhancement methods are used for different weather scenes. The cloud service layer is also used to use multi-scenario license plate recognition algorithms to perform license plate recognition on the enhanced license plate area, obtain license plate recognition results, control the operation of the barrier gate control unit according to the license plate recognition results, and perform abnormal event recognition based on the enhanced image to generate abnormal alarm information and send it to the operator layer. The operator layer is used to receive and visualize the warning information, license plate recognition results and vehicle video stream pushed by the cloud service layer. The operator layer is also used to control the operation of the gate control unit through remote commands.

[0006] Based on the aforementioned disclosure, this invention introduces noise suppression and contrast enhancement algorithms into the front-end perception layer to initially enhance the original image. Then, a multi-feature fusion network is used to comprehensively extract features from the license plate, vehicle body contour, and motion trajectory to accurately locate the actual license plate area. Furthermore, after extracting the license plate area, adaptive image enhancement of the license plate area is performed based on the identified weather conditions (such as rain, snow, fog, and strong light), providing more accurate image data for subsequent license plate recognition. This multi-optimization mechanism effectively overcomes the low recognition rate of traditional systems in adverse weather conditions, ensuring stable operation of the system around the clock and preventing lane congestion caused by recognition failures. Simultaneously, the cloud service layer not only handles license plate recognition but also possesses abnormal event recognition capabilities, proactively detecting abnormal situations within the lane and generating alarm information to push to the operator layer. Based on this, an intelligent upgrade from passive response to proactive early warning is achieved, greatly improving management efficiency and emergency response speed, and reducing the risk of management oversights. Therefore, this invention is highly suitable for large-scale application and promotion.

[0007] In one possible design, the front-end perception layer includes a vehicle recognition module and an unmanned robot, wherein the vehicle recognition module integrates a high-definition camera, a multispectral imaging component and an image enhancement processing unit, and the unmanned robot is equipped with an intercom terminal and a lane camera. The vehicle recognition module is used to acquire license plate images based on a high-definition camera and a multispectral imaging component, and to enhance the license plate images using an image enhancement processing unit and noise suppression and contrast enhancement algorithms. Specifically, when using noise suppression and contrast enhancement algorithms to enhance the license plate image, the image enhancement processing unit is configured as follows: The license plate image is subjected to noise detection to obtain the noise intensity. Based on the noise intensity, a denoising algorithm is determined. Based on the determined denoising algorithm, the license plate image is subjected to noise suppression preprocessing to obtain a denoised license plate image. Adaptive histogram equalization is applied to the denoised license plate image to obtain the initial enhanced license plate image. The initial enhanced license plate image is subjected to post-denoising processing to obtain an enhanced image, which is then sent to the network transport layer. An unattended robot is used to collect vehicle video streams based on lane cameras and send the vehicle video streams to the network transmission layer. An intercom terminal is used to establish a voice communication link with the operator layer in response to human-machine interaction.

[0008] In one possible design, the front-end sensing layer also includes: a local emergency module and a barrier gate control unit; The barrier gate control unit is used to receive the release command sent by the cloud service layer to open the barrier gate based on the release command, or to receive the remote command sent by the operator layer to control the opening or closing of the barrier gate. The local emergency module includes a cache unit and a manual control button. The cache unit is used to store data collected by the front-end sensing layer when the network is interrupted, and the manual control button is used to control the gate control unit to perform an emergency lifting operation.

[0009] In one possible design, the network transport layer includes: a 4G / 5G communication module, a local cache module, and a multi-protocol adaptation module; The multi-protocol adaptation module is used to automatically identify the communication protocol type of the front-end device in the front-end perception layer, and call the corresponding protocol parsing module to parse the communication protocol type of the front-end device, complete the protocol conversion, and establish a communication link with front-end devices of different brands so as to receive license plate images and vehicle video streams sent by the front-end devices. The 4G / 5G communication module is used to transmit enhanced images and vehicle video streams to the cloud service layer in real time; The local caching module is used to store the operation data of the parking lot cloud seat remote management system within a preset historical period when the network is interrupted, and to incrementally synchronize the stored operation data to the cloud service layer after the network is restored.

[0010] In one possible design, the cloud service layer includes: an AI analysis unit, wherein the AI ​​analysis unit uses a GPU accelerator card to run the multi-feature fusion network to output the vehicle's real license plate area based on the enhanced image and the vehicle video stream; The multi-feature fusion network includes: an input layer, a feature extraction branch, a feature alignment layer, a learnable weight layer, a fusion layer, and an output layer. The feature extraction branch includes a license plate feature extraction branch, a vehicle body contour feature extraction branch, and a vehicle motion trajectory feature extraction branch. An input layer is used to receive raw data, wherein the raw data includes the enhanced image, vehicle video stream, and vehicle trajectory data, and the vehicle trajectory data is obtained based on the vehicle video stream; The license plate feature extraction branch is used to extract license plate features based on the enhanced image; The vehicle body contour feature extraction branch is used to extract the vehicle body contour mask based on the vehicle video stream, and encode the vehicle body contour mask to obtain the vehicle body contour features. The vehicle trajectory feature extraction branch is used to extract features from the vehicle trajectory data to obtain a trajectory time-series feature vector, which is then used as the vehicle trajectory feature. The feature alignment layer is used to perform 1×1 convolution processing on the license plate feature and the vehicle body contour feature to obtain aligned license plate feature and aligned vehicle body contour feature. The feature alignment layer is used to perform dimensionality increase or decrease processing on the vehicle motion trajectory feature to obtain aligned vehicle motion trajectory feature. The aligned license plate feature, aligned vehicle body contour feature and aligned vehicle motion trajectory feature have the same feature dimension. Learnable weight layers are used to generate fusion weights for aligned license plate features, aligned vehicle body contour features, and aligned vehicle motion trajectory features using the SE Block network; The fusion layer is used to perform weighted fusion processing on the aligned license plate features, aligned vehicle body contour features, and aligned vehicle motion trajectory features based on their respective fusion weights, so as to obtain fused features. The output layer is used to output the actual license plate area of ​​the vehicle based on the fused features.

[0011] In one possible design, the AI ​​analysis unit is configured to perform weather scene recognition and adaptive image enhancement processing on the real license plate area as follows: Extract the average brightness, contrast, noise intensity, and haze index of the license plate image; The average brightness, contrast, noise intensity, and haze index are input into a pre-trained scene classifier to determine the weather scene in which the vehicle is located. The weather scene includes rain and snow scene, fog scene, low light scene, and strong backlight scene. Based on the weather scene in which the vehicle is located, an enhancement method matching the weather scene is determined, and image enhancement processing is performed on the real license plate area according to the enhancement method matching the weather scene in which the vehicle is located to obtain the enhanced license plate area. The image enhancement method is different for different weather scenes.

[0012] In one possible design, when using a multi-scenario license plate recognition algorithm to perform license plate recognition on the enhanced license plate area, the AI ​​analysis unit is configured as follows: The license plate area is then corrected to obtain the corrected license plate area. The vertical projection and watershed algorithm are used to perform character segmentation on the corrected license plate area to obtain several initial character segmentation images. Morphological opening operation is then performed on these initial character segmentation images to obtain several candidate character segmentation images. Several candidate character segmentation images are processed to complete the characters, resulting in several pre-selected character segmentation images. Then, several pre-selected character segmentation images are processed to perform non-character filtering, resulting in several character segmentation images. Perform character recognition processing on several character segmentation images to obtain the character category and confidence level corresponding to each character segmentation image, and retain the character categories with confidence levels greater than or equal to the first threshold; Character segmentation images corresponding to character categories with confidence scores less than the first threshold are designated as low-confidence character images. Feature extraction is performed on low-confidence character images to obtain character features. Then, using the character features, fuzzy matching of characters is performed in a character template library to obtain the similarity between each low-confidence character image and each template character in the character template library. The low-confidence character images are divided into confused characters and unconfused characters. For any unconfused character, template characters with a similarity greater than or equal to a second threshold are selected from the character template library as candidate characters corresponding to any unconfused character. After traversing all unconfused characters, candidate characters corresponding to each unconfused character are obtained. The first threshold is less than the second threshold. For any obfuscated character, template characters with a similarity greater than or equal to the second threshold are selected from the character template library and used as the initial candidate characters corresponding to the obfuscated character. The initial candidate characters are subjected to secondary verification to select pre-selected characters from the initial candidate characters; The pre-selected characters of any of the obfuscated characters are subjected to obfuscated character expansion processing to obtain an expanded pre-selected character list; Using license plate filtering rules, the extended pre-selected character list is filtered to obtain the candidate character corresponding to any of the obfuscated characters. After traversing all obfuscated characters, the candidate character corresponding to each obfuscated character is obtained. Using character categories with confidence levels greater than or equal to the first threshold, candidate characters corresponding to each non-confusing character, and candidate characters corresponding to each confused character, several initial candidate license plates are generated. Using license plate filtering rules, at least one candidate license plate is selected from several initial candidate license plates, and the combined confidence of each candidate license plate is calculated. Based on the combined confidence of each candidate license plate, the license plate recognition result is determined from each candidate license plate.

[0013] In one possible design, when performing anomalous event recognition based on enhanced images, the AI ​​analysis unit is configured as follows: Based on the enhanced image, determine whether there is a real license plate area for the vehicle; If not, an abnormal event of no license plate is determined, and an abnormal alarm message corresponding to the no-license-plate lingering is generated; otherwise, it is determined whether the license plate in the enhanced license plate area is damaged. If so, an abnormal event of license plate defacement is determined, and an abnormal alarm message corresponding to the license plate defacement is generated; otherwise, the license plate recognition result is obtained based on the enhanced license plate area. After obtaining the license plate recognition result, the system receives the first barrier status data reported by the barrier control unit. The first barrier status data includes the barrier angle, status code, reporting timestamp, and license plate recognition result. From the first barrier gate status data, filter out data records where the barrier gate angle is greater than the preset angle and the status code is the specified status, and obtain the reporting timestamp corresponding to the data record; Starting from the reporting timestamp, the system continuously monitors the status data of the second barrier gate that is located after the reporting timestamp and within the preset time period, and determines whether there is any barrier gate closure status data in the status data of the second barrier gate. If not, it is determined that an abnormal event of the barrier gate not being lowered has occurred, and an abnormal alarm message corresponding to the barrier gate not being lowered is generated; Specifically, when the AI ​​analysis unit detects two consecutive vehicles with unpaid toll information and the gate is open, it determines that a following anomaly event has occurred and generates corresponding abnormal alarm information; and When a network interruption is detected, a network fault is determined, and an abnormal alarm message corresponding to the network fault is generated.

[0014] In one possible design, the front-end perception layer includes an unattended robot, and the cloud service layer also includes a cloud seat management platform. When the AI ​​analysis unit determines that an abnormal event of no license plate has occurred, it sends a dynamic QR code to the unattended robot so that the car owner can scan the dynamic QR code to generate a temporary license plate for the vehicle and feed it back to the cloud seat management platform. The cloud-based seat management platform is used to allocate parking spaces to vehicles and associate and store the parking spaces with the temporary license plates. The cloud-based management platform sends a release command to the gate control unit to allow unlicensed vehicles to enter the site.

[0015] In one possible design, the cloud service layer also includes a data center, wherein the data center adopts a master-slave cluster off-site disaster recovery storage architecture, the master cluster in the master-slave cluster stores all real-time data in the cloud service layer, the slave cluster in the master-slave cluster performs off-site backup of all data stored in the master cluster, and the data center uses the AES-256 encryption algorithm to encrypt data when transmitting and storing data.

[0016] In one possible design, the agent operation layer includes: an agent terminal and a mobile APP, and the agent terminal is equipped with a monitoring display screen; The operator terminal is used to receive warning information, license plate recognition results and vehicle video streams pushed by the cloud service layer, and to visualize the warning information, license plate recognition results and vehicle video streams on the monitoring display screen. The monitoring display screen supports the simultaneous display of multiple video streams. The operator terminal is also used to send remote commands to the barrier gate control unit to control the operation of the barrier gate control unit, and the mobile APP is used to remotely operate the barrier gate control unit from a mobile device.

[0017] Secondly, a method for remote management of parking lot cloud-based agents is provided, including: The front-end perception layer acquires license plate images and vehicle video streams, and uses noise suppression and contrast enhancement algorithms to enhance the license plate images to obtain enhanced images; The network transport layer transmits the enhanced images and vehicle video streams output by the front-end perception layer to the cloud service layer; The cloud service layer uses a multi-feature fusion network to extract the license plate features of vehicles in the enhanced image, as well as the vehicle body contour features and vehicle motion trajectory features in the vehicle video stream, so as to perform feature fusion on the license plate features, vehicle body contour features and vehicle motion trajectory features to obtain fused features, and output the real license plate area of ​​the vehicle based on the fused features. The cloud service layer identifies the weather scene where the vehicle is located based on the license plate image, and performs adaptive image enhancement processing on the real license plate area according to the weather scene to obtain the enhanced license plate area. Different weather scenes correspond to different image enhancement methods. The cloud service layer uses a multi-scenario license plate recognition algorithm to perform license plate recognition on the enhanced license plate area and obtain the license plate recognition result. Based on the license plate recognition result, the barrier gate control unit is controlled to operate, and abnormal events are identified based on the enhanced image to generate abnormal alarm information and send it to the operator layer. The operator layer receives and visualizes the warning information, license plate recognition results and vehicle video stream pushed by the cloud service layer. The operator layer also controls the gate control unit to operate via remote commands.

[0018] Thirdly, a parking lot cloud-based remote management device is provided. Taking the system as an electronic device as an example, it includes a memory, a processor, and a transceiver that are connected in sequence. The memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer programs and execute the parking lot cloud-based remote management method as described in the second aspect.

[0019] Fourthly, a storage medium is provided, on which instructions are stored, which, when executed on a computer, perform the parking lot cloud-based remote management method as described in the second aspect.

[0020] Fifthly, a computer program product containing instructions is provided, which, when executed on a computer, causes the computer to perform the parking lot cloud-based remote management method as described in the second aspect.

[0021] Beneficial effects: (1) This invention introduces noise suppression and contrast enhancement algorithms into the front-end perception layer to initially enhance the original image. Then, a multi-feature fusion network is used to comprehensively extract the features of the license plate, vehicle body outline and motion trajectory to accurately locate the real license plate area. After extracting the license plate area, adaptive image enhancement of the license plate area is further performed according to the identified weather scene (such as rain, snow, fog, strong light), thereby providing more accurate image data for subsequent license plate recognition. In this way, this multi-optimization mechanism effectively overcomes the defect of low recognition rate of traditional systems in bad weather, ensures the stable operation of the system around the clock, and avoids lane congestion caused by recognition failure from the source. At the same time, the cloud service layer is not only responsible for license plate recognition, but also has the ability to identify abnormal events. It can actively detect abnormal situations in the lane and generate alarm information to push to the operator layer. Based on this, an intelligent upgrade from passive response to active warning is realized, which greatly improves management efficiency and emergency response speed and reduces the risk of management oversight. Therefore, this invention is very suitable for large-scale application and promotion.

[0022] (2) Adopting a dual redundancy design of “front-end perception + local caching” to solve the problem of network failure in existing technologies, that is, local caching of running data within a preset historical time period, and matching with emergency barrier lifting function to ensure normal passage of vehicles in the state of network failure.

[0023] (3) The equipment compatibility has achieved a comprehensive breakthrough. The network transmission layer has added a multi-protocol adaptation module, which supports a variety of mainstream protocols and can be seamlessly connected to multiple mainstream brands of gates and cameras with a 100% compatibility rate. No additional custom interface development is required, which solves the compatibility limitation of the existing technology with a compatibility rate of ≤60%.

[0024] (4) Dual protection of security and reliability: The present invention adopts AES-256 encrypted transmission + “master-slave dual cluster” off-site disaster recovery storage architecture, which can realize end-to-end data security protection and incremental synchronization, effectively avoid the risk of cloud data leakage and loss. Therefore, it can ensure that core information such as toll records and access data are complete and traceable.

[0025] (5) Management efficiency and labor cost optimization: The four-layer distributed architecture realizes seamless data flow across the entire chain of "license plate recognition - video surveillance - remote intercom - gate control". The operator supports simultaneous display of multiple video feeds and can process multi-dimensional information synchronously without switching interfaces. Combined with the automated release process, the normal vehicle passage time is ≤3 seconds, which is more than 70% faster than the traditional system. At the same time, it reduces the workload of manual inspection and operation, and reduces the overall labor cost by 30%-70%.

[0026] (6) User experience and service quality have been fully upgraded: the unmanned robot and cloud seat can communicate in real time, and the car owner can establish communication after asking for help and solve the passage problem remotely and quickly; at the same time, the unlicensed car can complete the entire process of self-service entry and payment by scanning the code to generate a temporary license plate, without the need for manual registration, making the passage process more convenient and significantly reducing the complaint rate. Attached Figure Description

[0027] Figure 1 This is a structural diagram of the parking lot cloud-based remote control management system provided in an embodiment of the present invention; Figure 2 This is a specific structural diagram of the multi-feature fusion network provided in the embodiments of the present invention; Figure 3 This is a flowchart illustrating the steps of the parking lot cloud-based remote management method provided in an embodiment of the present invention. Detailed Implementation

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0029] It should be understood that although the terms first, second, etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit, without departing from the scope of the exemplary embodiments of the invention.

[0030] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0031] Example: See Figure 1As shown, the parking lot cloud-based remote management system provided in this embodiment may include, but is not limited to, a front-end perception layer, a network transmission layer, a cloud service layer, and a seat operation layer; wherein, through the coordinated linkage of the front-end perception layer, network transmission layer, cloud service layer, and seat operation layer, the system can realize intelligent management of unattended parking lots; optionally, the core components and connection relationships of each layer are as follows: In this embodiment, the front-end perception layer is used to acquire license plate images and vehicle video streams, and employs noise suppression and contrast enhancement algorithms to enhance the license plate images to obtain enhanced images. Its main function is data acquisition, enabling the acquisition of vehicle license plate images and lane panoramic video streams, while also enhancing the overall license plate image. This enhances local details while avoiding noise amplification, thus providing clean image data for subsequent license plate recognition.

[0032] In practical implementation, the core functional components of the front-end perception layer mainly include: a vehicle recognition module, an unmanned robot, a local emergency module, and a gate control unit. The vehicle recognition module integrates a high-definition camera, a multispectral imaging component, and an image enhancement processing unit. The unmanned robot is equipped with an intercom terminal and a lane camera. Thus, the functions of data acquisition and image enhancement are realized by the vehicle recognition module and the unmanned robot.

[0033] In practical applications, the license plate recognition module is responsible for the acquisition and enhancement of license plate images. That is, the vehicle recognition module is used to acquire license plate images based on high-definition cameras and multispectral imaging components, and to enhance the license plate images using image enhancement processing units and noise suppression and contrast enhancement algorithms to obtain enhanced images.

[0034] In one specific implementation, when using noise suppression and contrast enhancement algorithms to enhance a license plate image, the image enhancement processing unit is configured to perform the following operations: First, noise detection is performed on the license plate image to obtain the noise intensity. Based on the noise intensity, a denoising algorithm is determined. Then, based on the determined denoising algorithm, the license plate image is preprocessed to suppress noise, resulting in a denoised license plate image. In practical applications, on-site cameras often have different types of noise, such as Gaussian noise, salt-and-pepper noise, low-light noise, and compression noise. If these are not processed, they will seriously affect subsequent enhancement and recognition. Therefore, this embodiment first uses different denoising algorithms based on the noise intensity to suppress noise, so as to provide a "clean base" for subsequent contrast enhancement.

[0035] Furthermore, examples, but not limited to, can be used to perform variance and gradient statistics on local regions of a license plate image to determine noise intensity. The specific operation is as follows: the license plate image is divided into several image blocks, and the variance and gradient of each image block are calculated. Then, the ratio between the variance and the gradient is used as the noise intensity. That is, in noisy regions, because the variance is large and the gradient is small, the ratio is large, meaning the noise intensity is large; conversely, in edge regions, both variance and gradient are large, and the ratio is close to 1, meaning the noise intensity is small. Based on this, this embodiment sets an intensity threshold. When the noise intensity is greater than or equal to the intensity threshold, it is determined to be a high-noise region; otherwise, it is determined to be a low-noise region.

[0036] Thus, for high-noise areas, Gaussian filtering or non-local mean filtering can be used, while for low-noise areas, weak filtering or edge-preserving filtering (such as bilateral filtering or guided filtering) can be used. In this way, through precise local judgment, the system can achieve "noise reduction where it should be reduced and detail preservation where it should be preserved", thereby removing noise while preserving image details.

[0037] After noise suppression of each image block is completed based on the aforementioned operations, the image blocks are stitched together to obtain the denoised license plate image. Then, adaptive histogram equalization is performed on the denoised license plate image to obtain the initial enhanced license plate image. The main process of adaptive histogram equalization is as follows: (1) Image preprocessing and block division: the denoised license plate image is converted into a grayscale image and the grayscale image is evenly divided into tile grids (typically 8×8, 16×16, etc., the smaller the grid, the finer the details but the higher the computational load); (2) Local histogram calculation: 0-255 is statistically analyzed for each tile. (3) Contrast Limitation: Set clipLimit (which limits the contrast amplification ratio in histogram equalization, and specifies the maximum value of the brightness gain value, typically 2.0-5.0, the smaller the value, the stronger the suppression), and based on this, calculate the maximum allowed number of pixels for each gray level in each original histogram T=clipLimit×total number of pixels in the tile / 256, and then crop the part of the histogram that exceeds T, and redistribute the overflow pixels evenly to all gray levels to avoid excessive stretching of a single gray level leading to noise amplification; (4) Local Equalization and CDF Mapping: Calculate the cumulative distribution function (CDF) based on the cropped histogram, and establish the gray-level mapping relationship: new gray level = (CDF(original gray level)) CDF_min) × 255 / (total number of pixels in the tile) CDF_min), where CDF_min represents the minimum value of the cumulative distribution function. Thus, by using the aforementioned formula, it can be ensured that the grayscale coverage is complete within the dynamic range after mapping; (5) Bilinear interpolation stitching: Bilinear interpolation is performed on the mapping results of adjacent tiles to eliminate block artifacts and output a smooth and enhanced image.

[0038] Thus, based on the aforementioned adaptive histogram equalization processing, dark areas become clearer, textures become more pronounced, and target areas are easier to identify, thereby improving the readability of the license plate area in low-light, backlight, and blurred images. Simultaneously, residual noise may remain after histogram equalization suppression; therefore, this embodiment also includes post-noise reduction processing after equalization, namely: The image enhancement processing unit performs post-denoising processing on the initial enhanced license plate image to obtain an enhanced image, which is then sent to the network transport layer. In specific applications, post-denoising processing may, but is not limited to, performing initial adaptive denoising processing first, followed by secondary denoising processing, to obtain the enhanced image.

[0039] Optionally, the initial adaptive noise reduction dynamically adjusts the filter strength based on the local variance, thereby achieving noise reduction while preserving details. The specific processing steps are as follows: Step 1: For each pixel (x,y) in the initial enhanced license plate image, take a local sliding window (such as 3×3, 5×5, 5×5 is recommended for seated scenes to balance noise reduction and detail), denoted as W(x,y), covering m×m pixels centered at (x,y) (m is an odd number).

[0040] The second step is to calculate the local mean and local variance within the local sliding window. The local mean reflects the average brightness within the window and is the basic reference value for filtering. A large local variance (such as in text areas) indicates that the area contains "useful details" and the filtering intensity should be reduced. Conversely, a small local variance (such as in noisy areas or smooth backgrounds) indicates that the area is mainly noise and the filtering intensity should be increased.

[0041] Step 3: Calculate the filtering strength, that is, for pixel (x, y), the filtered output pixel g(x, y) is: ; In the formula, Represents the local mean. Represents local variance. This represents the noise variance (which can be selected as the minimum local variance among the local sliding windows corresponding to all pixels). This represents the original pixel value of the pixel (x, y).

[0042] Among them, the filter strength is determined by This coefficient, dynamically adjusted, is the filter weight (filter strength), denoted as... ; where the local variance σ 2 The effect of (x,y) on k(x,y) (i.e., the filter strength adjustment logic) is as follows: Detail area ( > ): k(x,y)≈1, at this time, g(x,y)≈μ(x,y)+1×[f(x,y) μ(x,y)]=f(x,y), meaning the filtering intensity is extremely low, almost preserving the original pixels without losing details (such as device interfaces and text).

[0043] Noise area ( ≈ ): k(x,y)≈0, at this time, g(x,y)≈μ(x,y)+0=μ(x,y), that is, the filtering intensity is extremely high, and the original pixels are replaced by the window mean, and the depth is reduced (such as smoothing the background).

[0044] Transition region (detail variance): k(x,y) is between 0 and 1, and the filtering intensity is dynamically compromised (such as slightly blurred device edges, which both reduce noise and preserve edges).

[0045] Step 4: Repeat steps 1 to 3 for each pixel in the initial enhanced license plate image until all pixels in the initial enhanced license plate image have been polled. Then the initial post-denoised image can be obtained.

[0046] In addition, in this embodiment, the initial post-denoised image can be mirrored or copied to avoid abnormal filtering of boundary pixels (there may be device borders at the edges in the seat scene, and it is necessary to ensure that the details are not affected after filling).

[0047] After the initial noise reduction process is completed, a second noise reduction process can be performed, which is as follows: (1) For any pixel in the initial post-denoised image, take that pixel as the center and obtain the local window of that pixel (such as a 5×5 or 7×7 local window).

[0048] (2) Calculate the spatial proximity weight of each local pixel in the local window. The calculation formula is as follows: ; In the formula, Let represent the spatial proximity weight of any local pixel, where i and j are the x and y coordinates of that local pixel, and x and y represent the x and y coordinates of that pixel. The spatial standard deviation is used to control the decay rate of the spatial weights; in this embodiment, the value is set to 5.

[0049] (3) Calculate the grayscale similarity weight of each local pixel in the local window. The calculation formula is as follows: ; This represents the grayscale similarity weight of any local pixel. This represents the grayscale value of any local pixel. This represents the grayscale value of any given pixel. This represents the grayscale standard deviation, used to control the decay rate of grayscale weights. In this embodiment, the value is set to 20.

[0050] (3) Based on the gray-scale similarity weight and spatial proximity weight of each local pixel, the final weight of each local pixel is calculated. In this embodiment, the gray-scale similarity weight of each local pixel is multiplied by its spatial proximity weight to obtain the final weight of each local pixel.

[0051] (4) Normalize the final weight of each local pixel to obtain the normalized weight of each local pixel; in this embodiment, divide each final weight by the sum of all final weights to obtain several normalized weights.

[0052] (5) The gray values ​​of each local pixel are weighted by the normalized weight of each local pixel to obtain the pixel value of any pixel after filtering. After all pixels in the initial post-denoised image have been polled, the enhanced image can be obtained.

[0053] By combining noise suppression and contrast enhancement, the image becomes clearer and details become more prominent. In this way, the module can enhance low-quality images acquired from a remote location in real time, thereby improving image clarity, contrast, and recognizability, ultimately enabling operators to identify the on-site status, equipment details, and license plate areas more quickly and accurately.

[0054] After explaining the license plate recognition module, the functions of the unmanned robot will be explained, namely: An unattended robot is used to collect vehicle video streams based on lane cameras and send the video streams to the network transmission layer. An intercom terminal is used to establish a voice communication link with the operator layer in response to human-machine interaction. In a specific implementation, for example, the unattended robot is also equipped with a "help" button. Therefore, when the car owner presses the "help" button on the unattended robot, the intercom terminal is activated, establishing a voice communication link with the operator layer, thereby enabling voice dialogue. In this way, the management personnel corresponding to the operator layer can guide the car owner through voice operations, thereby improving the user experience.

[0055] Of course, the aforementioned image enhancement processing unit can also perform image enhancement on video frames in the vehicle video stream. The processing procedure is the same as the aforementioned license plate image enhancement procedure, and will not be described again here.

[0056] Similarly, the barrier gate control unit is used to receive the release command sent by the cloud service layer to open the barrier gate based on the release command, or to receive the remote command sent by the operator layer to control the opening or closing of the barrier gate.

[0057] Meanwhile, in this embodiment, the local emergency module may include, but is not limited to, a cache unit and a manual control button. The cache unit is used to store data collected by the front-end perception layer (such as license plate images, vehicle video streams, etc.) when the network is interrupted, and the manual control button is used to control the gate control unit to perform an emergency lifting operation. The emergency lifting can be achieved by manually controlling the button, thereby improving the system's emergency handling capability.

[0058] Thus, after the front-end perception layer completes data collection, it can upload the data through the network transmission layer. That is, the network transmission layer is used to transmit the enhanced images and vehicle video streams output by the front-end perception layer to the cloud service layer.

[0059] In one specific implementation, the following discloses the core components of the network transport layer: The network transport layer, for example, may include, but is not limited to, a 4G / 5G communication module, a local caching module, and a multi-protocol adaptation module. Specifically, the multi-protocol adaptation module is used to automatically identify the communication protocol type of the front-end device in the front-end perception layer, and call the corresponding protocol parsing module to parse the communication protocol type of the front-end device, complete the protocol conversion, and establish a communication link with front-end devices of different brands so as to receive license plate images and vehicle video streams sent by the front-end devices. After completing the protocol conversion and adaptation, the 4G / 5G communication module is used to transmit the license plate images and vehicle video streams to the cloud service layer in real time, thereby completing the real-time data transmission.

[0060] Optionally, the multi-protocol adapter module supports mainstream protocols such as RS485, TCP / IP, and ONVIF, thus enabling seamless access to different brands of gates and cameras (i.e., the front-end device can be components such as gates and cameras in the front-end sensing layer) without the need for additional interface development.

[0061] Meanwhile, the local cache module is used to store the operation data of the parking lot cloud seat remote management system within a preset historical period when the network is interrupted, and to synchronize the stored operation data incrementally to the cloud service layer after the network is restored. In this embodiment, the local cache module can store operation data (such as license plates, orders, abnormal events, etc.) within 12 hours. In this way, the dual-link design of "wireless transmission + local redundancy" can save key data even when the system is offline, completely solving the pain points of strong network dependence and data loss when the network is offline in the existing technology. In addition, combined with the aforementioned emergency barrier lifting function, it can also ensure that vehicle passage is not affected.

[0062] Therefore, after the data is uploaded through the aforementioned network transmission layer, license plate and abnormal event identification can be performed within the cloud service layer; the specific working process of the cloud service layer is as follows: First, the cloud service layer is used to extract the license plate features of vehicles in the enhanced image, as well as the vehicle body contour features and vehicle motion trajectory features in the vehicle video stream, using a multi-feature fusion network. This allows for feature fusion of the license plate features, vehicle body contour features, and vehicle motion trajectory features to obtain fused features, and the true license plate area of ​​the vehicle is output based on the fused features.

[0063] Thus, after extracting the vehicle's actual license plate area, this embodiment further improves the accuracy of license plate recognition under extreme weather conditions by incorporating a customized image enhancement process based on weather scenarios. Specifically, a cloud service layer identifies the weather scenario in which the vehicle is located based on the license plate image. Then, it adaptively enhances the actual license plate area according to the weather scenario, resulting in an enhanced license plate area. Different weather scenarios correspond to different image enhancement methods. Finally, a multi-scenario license plate recognition algorithm is used to recognize the enhanced license plate area, obtaining a license plate recognition result. This result is used to control the operation of the gate control unit and to identify abnormal events based on the enhanced image, generating abnormal alarm information that is sent to the operator layer.

[0064] Furthermore, the detailed structure of the disclosed cloud service layer is as follows, wherein the cloud service layer may include, but is not limited to, an AI analysis unit, a cloud seat management platform, and a data center, and the license plate area segmentation, adaptive enhancement of the license plate area, and license plate recognition operations are all performed by the AI ​​analysis unit.

[0065] Furthermore, the cloud service layer, based on a distributed server cluster deployment, supports the parallel management of 100+ parking lots, and the AI ​​analysis unit uses a GPU acceleration card to realize real-time algorithm computation. That is, the AI ​​analysis unit uses a GPU acceleration card to run the multi-feature fusion network to output the real license plate area of ​​the vehicle based on the enhanced image and the vehicle video stream.

[0066] In one specific implementation, the AI ​​analysis unit copies the data to be processed (such as image frames and video streams) from CPU memory to GPU memory, prioritizing the use of page-locked memory to reduce copying time (this refers to avoiding additional memory copy operations during data transmission by using page-locked memory, thereby improving data transmission efficiency; page-locked memory refers to a physically fixed block of host memory that the operating system does not page (swap)); simultaneously, it converts the data to a GPU-supported format (such as NHWC→NCHW) to avoid GPU-side format conversion overhead; and for batch data (such as multiple frames of a video stream), it uses batch processing to transfer the data to the GPU to improve parallel efficiency. The process can be understood as: data copy → GPU computation → result return.

[0067] In practical applications, existing models (such as YOLO and DeepSORT) can only handle single or two types of features (e.g., YOLO handles visual features, while DeepSORT combines visual and trajectory features). There are no readily available, out-of-the-box models that simultaneously fuse license plate features (fine-grained visual), vehicle outline (coarse-grained visual), and motion trajectory (temporal geometry) and introduce learnable weights; custom development is required. Therefore, this implementation provides a multi-feature fusion method that automatically assigns optimal weights to the three types of features through a learnable weight network, thereby achieving high-precision feature fusion. Based on this, the true license plate region can be accurately extracted from the enhanced image.

[0068] Optionally, one specific network structure of the disclosed multi-feature fusion network is as follows: See Figure 2 As shown, the multi-feature fusion network described in the example may include, but is not limited to, an input layer, a feature extraction branch, a feature alignment layer, a learnable weight layer, a fusion layer, and an output layer, and the feature extraction branch includes a license plate feature extraction branch, a vehicle body contour feature extraction branch, and a vehicle motion trajectory feature extraction branch.

[0069] In a specific implementation, the working process of each layer of the model is as follows: The input layer receives raw data, which includes the enhanced image, vehicle video stream, and vehicle trajectory data. The vehicle trajectory data is obtained based on the vehicle video stream. In this embodiment, the vehicle trajectory data can be, but is not limited to, the vehicle coordinate sequence (x1, y1, x2, y2, t) between frames of the vehicle video stream, where x1 and y1 represent the coordinates in the first video frame of the license plate video stream, and t represents the frame time. Of course, the coordinates of the vehicle in the video frame can be obtained by performing target recognition and detection on the video frame, such as using a CNN network.

[0070] After obtaining the raw data, different feature extraction branches can be used to extract different types of features from the raw data. For example, the license plate feature extraction branch is used to extract license plate features based on the enhanced image. In specific implementation, for example, but not limited to, the MobileNetV2+ attention (focusing on the license plate region, anti-blurring / occlusion) mechanism can be used to extract license plate features, that is, output a 256-dimensional vector as the license plate features.

[0071] Meanwhile, the vehicle body contour feature extraction branch is used to extract the vehicle body contour mask based on the vehicle video stream and encode the vehicle body contour mask to obtain the vehicle body contour features. The vehicle motion trajectory feature extraction branch is used to extract features from the vehicle trajectory data to obtain the trajectory temporal feature vector, which is used as the vehicle motion trajectory feature. In specific implementation, for example, but not limited to, a Mask R-CNN network can be used to extract the contour mask and encode it into a 256-dimensional vector to obtain the vehicle body contour features. A bidirectional LSTM network is used to process the trajectory coordinate sequence to output a 128-dimensional temporal feature vector, thereby obtaining the vehicle motion trajectory features.

[0072] Thus, based on the aforementioned three feature branches, after independently extracting three types of heterogeneous features, feature alignment can be performed. Specifically, a feature alignment layer is used to perform 1×1 convolution processing on the license plate feature and the vehicle body contour feature to obtain aligned license plate features and aligned vehicle body contour features. A fully connected layer is used to perform dimensionality upscaling or dimensionality reduction processing on the vehicle motion trajectory feature to obtain aligned vehicle motion trajectory features. In this embodiment, the aligned license plate feature, aligned vehicle body contour feature, and aligned vehicle motion trajectory feature have the same feature dimensions, thus ensuring dimensionality consistency during fusion.

[0073] After completing feature alignment and eliminating dimensional differences between heterogeneous features, the weights of different features can be learned. The specific process is as follows: The learnable weight layer is used to generate fusion weights for aligned license plate features, aligned vehicle body contour features, and aligned vehicle motion trajectory features using the SE Block network.

[0074] It should be noted that: The SE Block (Squeeze-Activation Module) first performs global average pooling on the feature map before fusion to obtain global features; then, it performs activation operation, that is, it uses 2 layers of MLP to learn weights (the weights of the three are summed to 1), and automatically increases the contour / trajectory weights for key scenes (such as blurred license plates); finally, it performs weight reverse update: that is, it iteratively optimizes the weights through the loss function (classification loss + trajectory matching loss).

[0075] Thus, based on the aforementioned SE Block network, after determining the fusion weights of the three aligned heterogeneous features, feature fusion can be performed. Specifically, the fusion layer is used to perform weighted fusion processing on the aligned license plate features, aligned vehicle body contour features, and aligned vehicle motion trajectory features based on their respective fusion weights, to obtain fused features. Finally, the output layer can be used to output the actual license plate region of the vehicle based on the fused features.

[0076] It should be noted that the output layer contains a classification head and a tracking head. The classification head is used to identify license plate areas / vehicle IDs (i.e., on each cell of the feature map, the model pre-defines several anchor boxes of different sizes and proportions. For each anchor box, the network outputs the bounding box offset, target confidence, and class probability. Finally, the output offset is converted into actual bounding box coordinates). The tracking head is used to complete cross-frame vehicle matching based on fused features (adapting to real-time tracking with remote assistance).

[0077] Therefore, through the aforementioned multi-feature fusion network, the real license plate region in the enhanced image can be accurately extracted. Then, customized image enhancement processing based on the weather scene can be performed. Specifically, when performing weather scene recognition and adaptive image enhancement processing on the real license plate region, the AI ​​analysis unit is configured to execute the following adaptive image enhancement operations: (1) Extract the average brightness, contrast, noise intensity and haze index of the license plate image; In this embodiment, for example, but not limited to, the haze index can be obtained by using the dark channel prior calculation, and the noise intensity is calculated in the same way as the previous example, and will not be repeated here. The contrast can be calculated by the maximum and minimum gray level difference formula, that is, the ratio of the difference between the maximum gray level and the minimum gray level to the sum between the maximum gray level and the minimum gray level is used as the contrast.

[0078] After extracting the aforementioned four features, weather scene classification can be performed based on them, as shown in the following process: (2) Input the average brightness, the contrast, the noise intensity and the fog index into a pre-trained scene classifier to obtain the weather scene where the vehicle is located, wherein the weather scene includes rain and snow scene, fog scene, low light scene and strong backlight scene.

[0079] In specific implementation, the trained MLP model can be used as a scene classifier, but not limited to. After inputting the aforementioned features into the trained MLP model, the weather scene where the vehicle is located can be obtained. Then, the corresponding enhancement method can be used to enhance the image based on the identified weather scene. The process is shown in the following steps (3).

[0080] (3) Based on the weather scene where the vehicle is located, determine the enhancement method that matches the weather scene where the vehicle is located, and perform image enhancement processing on the real license plate area according to the enhancement method that matches the weather scene where the vehicle is located, so as to obtain the enhanced license plate area. The image enhancement methods are different for different weather scenes.

[0081] In practical implementation, for rainy and / or snowy days, examples include, but are not limited to, first using bilateral filtering for noise reduction, then performing morphological opening operations to remove raindrops, and finally using gamma correction to enhance contrast. The parameters for bilateral filtering are d (kernel diameter) = 5, sigmaColor (standard deviation of color space) = 30, and sigmaSpace (standard deviation of coordinate workpiece) = 30. The parameters for opening operations are: kernel size 3×3; and the gamma value for gamma correction is 1.5.

[0082] Similarly, for foggy scenes, examples include, but are not limited to, first using dark channel prior dehazing, and then using CLAHE adaptive histogram equalization; for low-light scenes, first use multi-scale Retinex enhancement (MSRCR) processing, and then perform adaptive threshold binarization; finally, for strong backlight scenes, first perform regional exposure correction (that is, divide the license plate into 4 regions and correct them separately); and then perform histogram equalization processing.

[0083] Thus, this embodiment employs a customized enhancement algorithm for different extreme weather conditions, which can solve the image quality problem under different weather scenarios, thereby providing finer images with less noise for subsequent license plate recognition.

[0084] After completing the adaptive image enhancement processing based on the weather scene, license plate recognition can be performed. When using a multi-scene license plate recognition algorithm to recognize license plates in the enhanced license plate area, the AI ​​analysis unit is configured to perform the following recognition steps: Step 1: Perform license plate correction processing on the enhanced license plate area to obtain the corrected license plate area. In specific implementation, for example, but not limited to perspective transformation correction (such as affine transformation based on the four corner points of the license plate), the enhanced license plate area can be corrected. After the corrected license plate area, the license plate character segmentation can be performed, as shown in Step 2 below.

[0085] The second step involves using vertical projection and the watershed algorithm to segment the corrected license plate area, obtaining several initial character segmentation images. Morphological opening is then performed on these initial character segmentation images to obtain several candidate character segmentation images. In practical applications, the corrected license plate area is binarized, and then pixel projection is performed on the binary image in the vertical direction to find the contiguous regions. The watershed algorithm is then used to segment the characters based on their contours, resulting in several initial character segmentation images. Of course, vertical projection and the watershed algorithm are commonly used techniques for character segmentation, and their principles will not be elaborated further.

[0086] During character segmentation, there may be adjoining characters (regions whose width exceeds the normal width of a single character are identified as adjoining characters, and their positions are recorded). Therefore, morphological opening operations (erosion followed by dilation) can be used to separate adjoining characters while preserving the character outlines. Then, a second character segmentation is performed to complete the segmentation of adjoining characters. Finally, the aforementioned initial character segmentation images can be combined to form candidate character segmentation images.

[0087] After the character segmentation is completed, character completion and filtering can be performed, as shown in step 3 below.

[0088] Step 3: Perform character completion processing on several candidate character segmentation images to obtain several pre-selected character segmentation images, and perform non-character filtering processing on several pre-selected character segmentation images to obtain several character segmentation images. In specific implementation, morphological closing operation (dilation followed by erosion) can be used, but is not limited to, to complete the missing character images in several candidate character segmentation images to obtain several pre-selected character segmentation images. Then, based on the aspect ratio of the characters (the aspect ratio of license plate characters is 2:1), images that do not conform to the aforementioned aspect ratio are filtered out from several pre-selected character segmentation images to obtain several character segmentation images.

[0089] After the license plate characters are segmented, preliminary character recognition can be performed, as shown in step four below.

[0090] Step 4: Perform character recognition processing on several character segmentation images to obtain the character category and confidence score corresponding to each character segmentation image, and retain the character categories with confidence scores greater than or equal to the first threshold. In specific implementation, for any character segmentation image, HOG features can be extracted first, and then the CNN shallow features of any character segmentation image can be extracted (that is, input the any character segmentation image into the CNN network and use the output of the convolutional layer in the CNN as the CNN shallow features). Then, the aforementioned HOG features and CNN shallow features are input into the MobileNetV2 network to output the character category (such as 0-9, AZ) and confidence score of the any character segmentation image.

[0091] After completing the character category and confidence score for each character segmentation image, character categories with a confidence score greater than or equal to the first threshold (e.g., 0.8) can be retained. Meanwhile, for character segmentation images corresponding to character categories with a confidence score less than the first threshold, the subsequent character-level fuzzy matching process needs to be entered, as shown in the following steps.

[0092] Step 5: The character segmentation images corresponding to character categories with confidence scores less than the first threshold are used as low-confidence character images.

[0093] Step 6: Extract features from the low-confidence character images to obtain character features. Then, use these character features to perform fuzzy matching on the character template library to obtain the similarity between each low-confidence character image and each template character in the library. In practice, a standard character template library (50+ templates for each character) can be pre-built to cover different fonts (license plate fonts), different angles, and different lighting conditions. Then, the fuzzy matching of the low-confidence character images is completed by calculating the similarity between the low-confidence character images and the template characters in the library.

[0094] The aforementioned character features are also HOG features + CNN shallow features, and their extraction process has been described above and will not be repeated here. Thus, the corresponding character features are also extracted for each template character. Then, the cosine similarity between the character features of the low-confidence character image and the character features of each template character is calculated, and the similarity between the low-confidence character image and each template character can be obtained.

[0095] Then, fuzzy matching can be performed based on similarity, as shown in step 7 below.

[0096] Step 7: Divide the low-confidence character images into confused characters and unconfused characters. For any unconfused character, select template characters from the character template library whose similarity to the unconfused character is greater than or equal to the second threshold as candidate characters for the unconfused character. After traversing all unconfused characters, obtain the candidate characters corresponding to each unconfused character. The first threshold is less than the second threshold.

[0097] In this embodiment, the obfuscated character pairs (such as 0 / O, 1 / I, 6 / G, 8 / B, 9 / q, Z / 2) are predefined. That is, the low-confidence character images have been preliminarily identified and the character categories have been obtained. Therefore, the low-confidence character images can be divided into obfuscated characters and non-obfuscated characters according to the character categories of the low-confidence character images. Therefore, for non-obfuscated characters, template characters with similarity greater than or equal to 0.85 are directly retained as candidate characters for non-obfuscated characters. For example, the similarity list of the non-obfuscated character "5" is 5 (0.92) and S (0.87). Then, the candidate list of the non-obfuscated character "5" is [5,S]. Similarly, the similarity list of the non-obfuscated character "9" is 9 (0.90). Then, its corresponding candidate list is [9].

[0098] After completing the fuzzy matching of non-obfuscated characters, the fuzzy matching of obfuscated characters can be performed, as shown in steps eight to eleven below.

[0099] Step 8: For any obfuscated character, select template characters from the character template library whose similarity to the obfuscated character is greater than or equal to the second threshold, and use them as the initial candidate characters corresponding to the obfuscated character. In this embodiment, for obfuscated characters, template characters with a similarity greater than or equal to 0.85 cannot be directly included in the candidates and need to be verified twice. The process is shown in Step 9 below.

[0100] Step 9: Perform secondary verification on the initial candidate characters to filter out pre-selected characters from the initial candidate characters. In specific implementation, for example, but not limited to, character structure features can be used to perform secondary verification on the initial candidate characters. That is, since the obfuscated character pairs have been predefined, any obfuscated character can be compared with its corresponding initial candidate character by character structure features, thereby further filtering characters from the initial candidate characters.

[0101] Optionally, character structure features can be: region filling ratio (i.e., the ratio of pixel area before and after filling holes in the character image), aspect ratio, vertical projection histogram of the character, etc. In this way, by extracting the aforementioned character structure features and comparing them with the characters in the confused character pair, the secondary verification of the initial candidate characters can be completed.

[0102] After completing the secondary verification, the easily confused characters can be expanded, as shown in step ten below.

[0103] Step 10: Perform an extended processing of easily confused characters on the preselected characters of any of the confused characters to obtain an extended preselected character list; in this embodiment, for any confused character, forced character extension is required, that is, when "0" is recognized, the candidate list must include the letter "o"; when "1" is recognized, "I" must be included, etc. After completing the character extension, for incomplete characters, character completion can also be performed. For example, in extreme weather, the upper half of "8" is missing, and the outline can be completed based on the structural characteristics of similar characters (such as "8" consisting of two upper and lower circles), and the similarity is recalculated.

[0104] In this way, after performing the character extension, the license plate filtering rules can be used to filter the license plates, and the process is as shown in the following Step 11.

[0105] Step 11: Use the license plate filtering rules to perform filtering on the extended preselected character list to obtain the candidate characters corresponding to any of the confused characters, and after traversing all the confused characters, obtain the candidate characters corresponding to each confused character; in specific implementation, the license plate filtering rules are preset. For example, a number cannot follow the province abbreviation. Then, according to this rule, filtering in the extended preselected character list is required. For example, if the candidate list for the second character of the license plate is ["8","B"], and since the second character must be a letter, "8" needs to be filtered and "B" is retained.

[0106] After obtaining the candidate characters corresponding to each confused character, several initial candidate license plates can be generated by combining the character categories with a confidence level greater than 0.8 described above and the candidate characters of each non-confused character, and the process is as shown in the following Step 12.

[0107] Step 12: Use the character categories with a confidence level greater than or equal to the first threshold, the candidate characters corresponding to each non-confused character, and the candidate characters corresponding to each confused character to generate several initial candidate license plates; in this embodiment, the candidate characters are combined in the order of character segmentation to obtain several initial candidate license plates, and then license plate filtering can be performed, and the process is as shown in the following Step 13.

[0108] Step 13: Use the license plate filtering rules to screen out at least one candidate license plate from several initial candidate license plates, and calculate the combined confidence level of each candidate license plate, so as to determine the license plate recognition result from each candidate license plate based on the combined confidence level of each candidate license plate.

[0109] In specific implementation, the recognized license plate must conform to the license plate format. For example, the license plate format is: province + letter + 5 letters / numbers. Therefore, the illegal combinations are filtered using the above license plate filtering rules. For example, if "Guangdong 812345" is recognized, since the second character must be a letter, it is automatically corrected to "Guangdong A12345" (combined with the candidate list).

[0110] After selecting at least one candidate license plate, the combined confidence score of the candidate license plates can be calculated. The process is as follows: the similarity of each character in the candidate vehicle is multiplied by its corresponding weight, and the results are summed to obtain the combined confidence score. Since each candidate character in the candidate license plate has a similarity score during the aforementioned fuzzy matching or initial recognition (for extended characters, character features can also be extracted to calculate their similarity with the original characters, while for character categories retained during initial recognition, their confidence scores are used as similarity scores), the combined confidence score of the candidate license plates can be obtained by weighted summing of the similarity scores of each character. If any character is a digit, its weight is 0.2; if any character is a number, its weight is set to 0.1.

[0111] Thus, after calculating the combined confidence of each candidate license plate, the candidate license plate with the highest combined confidence is selected as the license plate recognition result.

[0112] Thus, by using the aforementioned method, the AI ​​analysis unit can store the license plate recognition results obtained through multi-scenario license plate recognition algorithms and send a release command to the gate control unit, thereby enabling the vehicle to enter the site.

[0113] After license plate recognition is completed, abnormal events can be identified based on the enhanced image. In this embodiment, abnormal events may include, but are not limited to: no license plate abnormal event, license plate damage abnormal event, gate not lowered abnormal event, following vehicle abnormal event, and network failure abnormal event.

[0114] The detailed process of anomaly event identification is as follows: When anomaly event identification is performed based on enhanced images, the AI ​​analysis unit is configured to perform the following operations: First, based on the enhanced image, it is determined whether there is a real license plate area for the vehicle. In this embodiment, it is directly obtained whether the aforementioned multi-feature fusion network can output a real license plate area. That is, if it cannot output a real license plate area, it means that there is no license plate in the enhanced image. At this time, it is determined that an abnormal event of no license plate has occurred, and an abnormal alarm information corresponding to the no-license-plate lingering needs to be generated.

[0115] If the aforementioned multi-feature fusion network outputs a real license plate image, it is necessary to determine whether the license plate in the enhanced license plate region is damaged. For example, but not limited to, obtaining the segmented images of each character corresponding to the enhanced license plate region, then extracting the contour images of each character segmented image, then extracting the character features of the contour images (which can also be HOG features and shallow CNN features), and then calculating the similarity between the contour images and the character features of the standard contour images corresponding to the standard characters. If the similarity between any contour image and the standard contour image is less than the similarity threshold (which can be set to 0.9), then it is determined that the license plate is damaged, that is, a license plate damage anomaly event has occurred. At this time, it is necessary to generate an abnormal alarm information corresponding to the license plate damage.

[0116] Similarly, if the similarity between all contour images and the standard contour image is greater than or equal to the pixel threshold, then it is necessary to obtain the license plate recognition result based on the enhanced license plate area. Then, after obtaining the license plate recognition result, the first gate status data reported by the gate control unit is received. After obtaining the license plate recognition result, the gate will be controlled to open, thus generating gate status data. In this embodiment, the abnormal event of the gate not closing is identified by using the device status threshold + time series analysis.

[0117] For example, the status data of the first barrier gate includes the barrier gate angle, status code, reporting timestamp and license plate recognition result, and the status code can be, but is not limited to, open, closed, fault or manual.

[0118] After obtaining the first barrier gate status data, data filtering can be performed. Specifically, data records with barrier gate angles greater than a preset angle and status codes of a specified state are filtered from the first barrier gate status data, and the reporting timestamp corresponding to the data record is obtained. In this embodiment, records with "barrier gate angle ≥ 85°", "status code = open", and "no manual or lingering instruction markers" are filtered out and marked as "barrier gate events to be monitored". The reporting timestamp of this state is recorded as T0. Then, timing can be started from this reporting timestamp, and the second barrier gate status data located after the reporting timestamp and within a preset time period can be continuously monitored to determine whether there is barrier gate closed status data in the second barrier gate status data.

[0119] In practice, the timing starts from T0 and the system continuously monitors the subsequent reported data of the barrier gate. If, within T0+10 seconds, status data such as "barrier gate angle < 10°" (fully closed) or "barrier gate angle continuously decreasing from 85°" (normal closing process) is received (i.e., status data within the set time period, with the barrier gate angle continuously decreasing), it is determined to be a "normal event," the monitoring flag is cleared, and the process ends. If no barrier gate closing status data is received after T0+10 seconds, i.e., the barrier gate still maintains "angle ≥ 85°" and "status code = open," then the "threshold exceeded" flag is triggered. At this time, it is determined that an abnormal event of the barrier gate not closing has occurred, and an abnormal alarm message corresponding to the barrier gate not closing is generated.

[0120] Of course, after triggering the threshold over-limit flag, a secondary verification can be performed, that is, to determine whether the device is offline. If it is not offline, an abnormal alarm message corresponding to the gate not being lowered is generated. This message includes the gate number, over-limit duration, current status, etc.

[0121] In addition, in this embodiment, when the AI ​​analysis unit detects that two consecutive vehicles have unpaid information and the gate is open, it determines that a following abnormal event has occurred and generates an abnormal alarm message corresponding to the following abnormal event; at the same time, when a network interruption is detected, it determines that a network failure has occurred and generates an abnormal alarm message corresponding to the network failure.

[0122] Therefore, the core of the aforementioned equipment status threshold + time series analysis method is to make up for the limitations of static thresholds (which cannot cope with fluctuating scenarios and occasional noise) and achieve more accurate early warning or false alarm filtering by mining the time trend, periodicity and correlation of data (for example, the gate closing speed is abnormally slow, and the failure to close the gate can be predicted even if the 10-second threshold has not been reached).

[0123] Thus, the cloud service layer in this system is not only responsible for license plate recognition, but also has the ability to identify abnormal events. It can proactively detect abnormal situations in the lane and generate alarm information to push to the operator layer. Based on this, an intelligent upgrade from passive response to proactive early warning is achieved, which greatly improves management efficiency and emergency response speed.

[0124] Furthermore, for vehicles without license plates, this embodiment also includes a rapid entry process: when the AI ​​analysis unit determines that an abnormal event of no license plate has occurred, it sends a dynamic QR code to the unmanned robot so that the vehicle owner can scan the dynamic QR code to generate a temporary license plate for the vehicle, and then feeds it back to the cloud seat management platform; then, the cloud seat management platform is used to allocate parking spaces for the vehicle and associate the parking spaces with the temporary license plates for storage; finally, the cloud seat management platform sends a release command to the gate control unit to complete the entry of the vehicle without license plates.

[0125] Thus, the processing logic for vehicles without license plates is as follows: the vehicle owner scans the QR code on the gate → a temporary license plate is generated at the front end → the parking space information is linked in the cloud → payment is completed by matching the temporary license plate with the parking space when leaving the gate. No manual registration is required throughout the process, making the passage process more convenient.

[0126] Furthermore, a data center is also set up in the cloud service layer. The data center adopts a master-slave cluster off-site disaster recovery storage architecture. That is, the master cluster in the master-slave cluster stores all real-time data in the cloud service layer (such as license plate recognition results, received gate status data, identified abnormal data, payment information returned by the gate control unit of the front-end perception layer, etc.). The slave cluster in the master-slave cluster performs off-site backup of all data stored in the master cluster. Moreover, the data center uses the AES-256 encryption algorithm to encrypt data when transmitting and storing data. At the same time, when the system loses network access, the local cache module in the network transmission layer automatically takes over and synchronizes the cached data increments to the cloud after reconnecting to the network, thereby realizing redundant data storage and ensuring the integrity of data storage.

[0127] Therefore, after completing license plate recognition, abnormal event recognition, and barrier gate control, the cloud service layer can push information, namely: the operator layer, which is used to receive the warning information, license plate recognition results, and vehicle video stream pushed by the cloud service layer and display them visually. The operator layer is also used to control the operation of the barrier gate control unit through remote commands.

[0128] In specific applications, the seat operation layer described above may include, but is not limited to, a seat terminal and a mobile APP, and the seat terminal is equipped with a monitoring display screen and integrates an audio and video decoding card and a multi-window control module.

[0129] The operator terminal is used to receive warning information, license plate recognition results and vehicle video streams pushed by the cloud service layer, and to visualize the warning information, license plate recognition results and vehicle video streams on the monitoring display screen. The monitoring display screen supports the simultaneous display of multiple video streams (such as supporting the simultaneous display of 16 video streams).

[0130] Meanwhile, the operator terminal integrates a remote intercom function. When an abnormal event warning is triggered, the operator terminal automatically displays the real-time video, license plate information and historical records of the corresponding lane. The operator can issue control commands by clicking the "remotely raise the barrier" button. At the same time, when the car owner presses the help button, the operator terminal can also establish a voice communication link with the intercom terminal on the unmanned robot to achieve quick communication with the car owner.

[0131] In addition, the operator terminal is also used to send remote commands to the barrier gate control unit to control the operation of the barrier gate control unit, and the mobile APP is used to remotely operate the barrier gate control unit on a mobile device, thereby realizing remote mobile operation and increasing the convenience of operation.

[0132] Therefore, the workflow of the aforementioned parking lot cloud-based remote control management system is as follows: Data acquisition phase (front-end perception layer): The vehicle recognition module acquires images through cameras and multispectral imaging components, and the image enhancement processing unit removes interference from fog, rain, and strong light, outputting optimized license plate image data; at the same time, the unmanned robot acquires panoramic video of the lane.

[0133] During the data transmission phase (network transmission layer): the 4G / 5G module transmits license plate data and video stream to the cloud in real time, while the local cache module synchronously stores key data (license plate, timestamp, device status); if the network is interrupted, it automatically switches to the local emergency module, and the cached data is retained for 12 hours, supporting the emergency gate opening function.

[0134] Data processing stage (cloud service layer): After receiving the data, the AI ​​analysis unit first matches it with the license plate database (supports fuzzy matching, and a similarity of ≥85% is considered valid); if there is no recognition result, the process for vehicles without license plates is triggered, and at the same time, abnormal events are identified (such as two vehicles failing to pay to pass in a row are identified as following another vehicle, and the gate is not closed within 10 seconds after it is opened, which is identified as a malfunction), and an alert is sent to the operator layer.

[0135] Command feedback stage (seat operation layer): The seat terminal receives warning information and synchronous video stream, and displays license plate data, on-site video, and order information simultaneously through the multi-window control module; the seat can click on commands such as "remotely raise the barrier", and the operation is executed through the cloud → network transmission layer → front-end barrier control unit.

[0136] Therefore, through the detailed description of the aforementioned parking lot cloud-based remote control management system architecture, the beneficial effects of this invention are as follows: (1) This invention introduces noise suppression and contrast enhancement algorithms into the front-end perception layer to initially enhance the original image. Then, a multi-feature fusion network is used to comprehensively extract the features of the license plate, vehicle body outline and motion trajectory to accurately locate the real license plate area. After extracting the license plate area, adaptive image enhancement of the license plate area is further performed according to the identified weather scene (such as rain, snow, fog, strong light), thereby providing more accurate image data for subsequent license plate recognition. In this way, this multi-optimization mechanism effectively overcomes the defect of low recognition rate of traditional systems in bad weather, ensures the stable operation of the system around the clock, and avoids lane congestion caused by recognition failure from the source. At the same time, the cloud service layer is not only responsible for license plate recognition, but also has the ability to identify abnormal events. It can actively detect abnormal situations in the lane and generate alarm information to push to the operator layer. Based on this, an intelligent upgrade from passive response to active warning is realized, which greatly improves management efficiency and emergency response speed and reduces the risk of management oversight. Therefore, this invention is very suitable for large-scale application and promotion.

[0137] (2) The efficiency of abnormal response is exponentially improved: Through the intelligent early warning rules of "equipment status threshold + time series analysis", five core abnormal scenarios are preset. Compared with the existing manual discovery mode (average 30 seconds), the early warning response time of abnormal events is greatly shortened, and problems such as following vehicles and gates not being lowered can be dealt with quickly, reducing the risk of fault delay.

[0138] (3) Enhanced emergency response capability and system stability during network outages: The dual redundancy design of "front-end perception + local caching" ensures that the local caching module automatically takes over when the network is out of service, retains 12 hours of critical data and supports emergency gate lifting, completely solving the pain points of strong network dependence and failure during network outages in existing technologies, and ensuring that vehicle passage is not affected.

[0139] (4) Comprehensive breakthrough in equipment compatibility: The network transmission layer adds a multi-protocol adaptation module, which supports mainstream protocols such as RS485, TCP / IP, and ONVIF. It can seamlessly connect to ≥20 mainstream brands of gates and cameras with a 100% compatibility rate. No additional custom interface development is required, which solves the compatibility limitation of existing technologies with a compatibility rate of ≤60%.

[0140] (5) Dual protection of data security and reliability: AES-256 encrypted transmission + “master-slave dual cluster” off-site disaster recovery storage architecture is adopted to realize end-to-end data security protection and incremental synchronization, effectively avoid the risk of cloud data leakage and loss, and ensure that core information such as toll records and access data are complete and traceable.

[0141] (6) Optimization of management efficiency and human resource costs: The four-layer distributed architecture realizes seamless data flow across the entire chain of "license plate recognition - video surveillance - remote intercom - gate control". The operator can process multi-dimensional information simultaneously without switching interfaces. Combined with the automated release process, the normal vehicle passage time is ≤3 seconds, which is more than 70% faster than the traditional system. At the same time, it reduces the workload of manual inspection and operation, and reduces the overall human resource cost by 30%-70%.

[0142] (7) Significantly reduced operation and maintenance and deployment costs: The integrated design reduces the number of front-end devices deployed, and the multi-protocol plug-and-play feature reduces the difficulty of device access; it supports remote debugging and software upgrades in the cloud, eliminating the need for on-site maintenance, and the local cache reduces the burden on cloud storage, thus comprehensively reducing hardware procurement, operation and maintenance and data storage costs.

[0143] (8) User experience and service quality have been fully upgraded: the unmanned robot and cloud seat can communicate in real time, and the car owner can establish communication within 2 seconds to ask for help and solve the passage problem remotely and quickly; the car without license plate can complete the entire self-service payment by scanning the code to generate a temporary ID, without manual registration, making the passage process more convenient and the complaint rate significantly reduced.

[0144] (9) Enhanced data support for management decisions: The data center stores full data such as vehicle traffic, order transactions, and abnormal events, which can support subsequent intelligent analysis such as traffic flow prediction, parking space optimization, and fee statistics. It can provide accurate data basis for adjusting parking lot operation strategies and optimizing resource allocation, and can help improve parking space turnover rate and operating revenue.

[0145] (10) Versatility and compliance assurance across multiple scenarios: The system can be flexibly adapted to various parking lots such as commercial complexes, residential communities, industrial parks, airports, and high-speed rail stations. Whether it is a small single-entrance scenario or a large multi-area scenario, no customized modification is required; the functions of publicizing the charging standard, traceable transaction records, and offline data retransmission fully comply with the requirements of parking fee management regulations, avoiding compliance risks of missed or incorrect charges. like Figure 3 As shown, the second aspect of this embodiment provides a parking lot cloud seat remote management method, wherein the method is executed based on the parking lot cloud seat remote management system described in the first aspect of the embodiment, and the operation steps of the method may be, but are not limited to, the steps S1 to S6 below.

[0146] S1. The front-end perception layer acquires license plate images and vehicle video streams, and uses noise suppression and contrast enhancement algorithms to enhance the license plate images to obtain enhanced images.

[0147] S2. The network transmission layer transmits the enhanced image and vehicle video stream output by the front-end perception layer to the cloud service layer.

[0148] S3. The cloud service layer uses a multi-feature fusion network to extract the license plate features of vehicles in the enhanced image, as well as the vehicle body contour features and vehicle motion trajectory features in the vehicle video stream, so as to perform feature fusion on the license plate features, vehicle body contour features and vehicle motion trajectory features to obtain fused features, and output the real license plate area of ​​the vehicle based on the fused features.

[0149] S4. The cloud service layer identifies the weather scene where the vehicle is located based on the license plate image, and performs adaptive image enhancement processing on the real license plate area according to the weather scene where the vehicle is located to obtain an enhanced license plate area. Different image enhancement methods are corresponding to different weather scenes.

[0150] S5. The cloud service layer adopts a multi-scenario license plate recognition algorithm to perform license plate recognition on the enhanced license plate area and obtain the license plate recognition result. Based on the license plate recognition result, the barrier gate control unit is controlled to operate, and abnormal events are identified based on the enhanced image to generate abnormal alarm information and send it to the operator layer.

[0151] S6. The operator layer receives and visualizes the warning information, license plate recognition results and vehicle video stream pushed by the cloud service layer, and the operator layer also controls the gate control unit to operate via remote commands.

[0152] The working process, working details and technical effects of the method provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0153] The third aspect of this embodiment provides a parking lot cloud-based remote management device. Taking the device as an electronic device as an example, it includes: a memory, a processor, and a transceiver that are connected in sequence. The memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer programs and execute the parking lot cloud-based remote management method as described in the second aspect of this embodiment.

[0154] For specific examples, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or first-in-last-out (FILO) memory, etc.; specifically, the processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor may be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), PLA (Programmable Logic Array). The processor may also include a main processor and a coprocessor. The main processor, also known as the CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state.

[0155] In some embodiments, the processor may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. For example, the processor may not be limited to microprocessors of the STM32F105 series, reduced instruction set computer (RISC) microprocessors, x86 architecture processors, or processors with integrated neural network processing units (NPUs). The transceiver may be, but is not limited to, a Wi-Fi transceiver, a Bluetooth transceiver, a General Packet Radio Service (GPRS) transceiver, a ZigBee (a low-power LAN protocol based on the IEEE 802.15.4 standard) transceiver, a 3G transceiver, a 4G transceiver, and / or a 5G transceiver. Furthermore, the device may also include, but is not limited to, a power module, a display screen, and other necessary components.

[0156] The working process, working details and technical effects of the electronic device provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0157] The fourth aspect of this embodiment provides a storage medium that stores instructions containing the parking lot cloud-based remote management method described in the second aspect of the embodiment. That is, the storage medium stores instructions that, when executed on a computer, perform the parking lot cloud-based remote management method as described in the second aspect of the embodiment.

[0158] The storage medium refers to a carrier for storing data, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or memory sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0159] The working process, working details and technical effects of the storage medium provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0160] The fifth aspect of this embodiment provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform the parking lot cloud-based remote management method as described in the second aspect of this embodiment. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0161] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A parking lot cloud contact center remote management system, characterized by, include: The front-end perception layer is used to acquire license plate images and vehicle video streams, and uses noise suppression and contrast enhancement algorithms to enhance the license plate images to obtain enhanced images. The network transport layer is used to transmit the enhanced images and vehicle video streams output by the front-end perception layer to the cloud service layer. The cloud service layer is used to extract the license plate features of vehicles in the enhanced image, as well as the vehicle body contour features and vehicle motion trajectory features in the vehicle video stream, using a multi-feature fusion network. This allows for feature fusion of the license plate features, vehicle body contour features, and vehicle motion trajectory features to obtain fused features, and outputs the vehicle's real license plate area based on the fused features. The cloud service layer is used to identify the weather scene where the vehicle is located based on the license plate image, and to perform adaptive image enhancement processing on the real license plate area according to the weather scene where the vehicle is located to obtain an enhanced license plate area. Different image enhancement methods are used for different weather scenes. The cloud service layer is also used to use multi-scenario license plate recognition algorithms to perform license plate recognition on the enhanced license plate area, obtain license plate recognition results, control the operation of the barrier gate control unit according to the license plate recognition results, and perform abnormal event recognition based on the enhanced image to generate abnormal alarm information and send it to the operator layer. The operator layer is used to receive and visualize the warning information, license plate recognition results and vehicle video stream pushed by the cloud service layer. The operator layer is also used to control the operation of the gate control unit through remote commands.

2. The parking lot cloud contact center remote management system of claim 1, wherein, The front-end perception layer includes a vehicle recognition module and an unmanned robot. The vehicle recognition module integrates a high-definition camera, a multispectral imaging component, and an image enhancement processing unit, while the unmanned robot is equipped with an intercom terminal and a lane camera. The vehicle recognition module is used to acquire license plate images based on a high-definition camera and a multispectral imaging component, and to enhance the license plate images using an image enhancement processing unit and noise suppression and contrast enhancement algorithms. Specifically, when using noise suppression and contrast enhancement algorithms to enhance the license plate image, the image enhancement processing unit is configured as follows: The license plate image is subjected to noise detection to obtain the noise intensity. Based on the noise intensity, a denoising algorithm is determined. Based on the determined denoising algorithm, the license plate image is subjected to noise suppression preprocessing to obtain a denoised license plate image. Adaptive histogram equalization is applied to the denoised license plate image to obtain the initial enhanced license plate image. The initial enhanced license plate image is subjected to post-denoising processing to obtain an enhanced image, which is then sent to the network transport layer. An unattended robot is used to collect vehicle video streams based on lane cameras and send the vehicle video streams to the network transmission layer. An intercom terminal is used to establish a voice communication link with the operator layer in response to human-machine interaction.

3. The parking lot cloud contact center remote management system of claim 2, wherein, The front-end sensing layer also includes: a local emergency module and a barrier gate control unit; The barrier gate control unit is used to receive the release command sent by the cloud service layer to open the barrier gate based on the release command, or to receive the remote command sent by the operator layer to control the opening or closing of the barrier gate. The local emergency module includes a cache unit and a manual control button. The cache unit is used to store data collected by the front-end sensing layer when the network is interrupted, and the manual control button is used to control the gate control unit to perform an emergency lifting operation.

4. The parking lot cloud contact center remote management system of claim 1, wherein, The network transport layer includes: a 4G / 5G communication module, a local cache module, and a multi-protocol adaptation module; The multi-protocol adaptation module is used to automatically identify the communication protocol type of the front-end device in the front-end perception layer, and call the corresponding protocol parsing module to parse the communication protocol type of the front-end device, complete the protocol conversion, and establish a communication link with front-end devices of different brands so as to receive license plate images and vehicle video streams sent by the front-end devices. The 4G / 5G communication module is used to transmit enhanced images and vehicle video streams to the cloud service layer in real time; The local caching module is used to store the operation data of the parking lot cloud seat remote management system within a preset historical period when the network is interrupted, and to incrementally synchronize the stored operation data to the cloud service layer after the network is restored.

5. The parking lot cloud contact center remote management system of claim 1, wherein, The cloud service layer includes an AI analysis unit, wherein the AI ​​analysis unit uses a GPU acceleration card to run the multi-feature fusion network to output the vehicle's real license plate area based on the enhanced image and the vehicle video stream; The multi-feature fusion network includes: an input layer, a feature extraction branch, a feature alignment layer, a learnable weight layer, a fusion layer, and an output layer. The feature extraction branch includes a license plate feature extraction branch, a vehicle body contour feature extraction branch, and a vehicle motion trajectory feature extraction branch. An input layer is used to receive raw data, wherein the raw data includes the enhanced image, vehicle video stream, and vehicle trajectory data, and the vehicle trajectory data is obtained based on the vehicle video stream; The license plate feature extraction branch is used to extract license plate features based on the enhanced image; The vehicle body contour feature extraction branch is used to extract the vehicle body contour mask based on the vehicle video stream, and encode the vehicle body contour mask to obtain the vehicle body contour features. The vehicle trajectory feature extraction branch is used to extract features from the vehicle trajectory data to obtain a trajectory time-series feature vector, which is then used as the vehicle trajectory feature. The feature alignment layer is used to perform 1×1 convolution processing on the license plate feature and the vehicle body contour feature to obtain aligned license plate feature and aligned vehicle body contour feature. The feature alignment layer is used to perform dimensionality increase or decrease processing on the vehicle motion trajectory feature to obtain aligned vehicle motion trajectory feature. The aligned license plate feature, aligned vehicle body contour feature and aligned vehicle motion trajectory feature have the same feature dimension. Learnable weight layers are used to generate fusion weights for aligned license plate features, aligned vehicle body contour features, and aligned vehicle motion trajectory features using the SE Block network; The fusion layer is used to perform weighted fusion processing on the aligned license plate features, aligned vehicle body contour features, and aligned vehicle motion trajectory features based on their respective fusion weights, so as to obtain fused features. The output layer is used to output the actual license plate area of ​​the vehicle based on the fused features.

6. The parking lot cloud contact center remote management system of claim 5, wherein, When performing weather scene recognition and adaptive image enhancement processing on the real license plate area, the AI ​​analysis unit is configured as follows: Extract the average brightness, contrast, noise intensity, and haze index of the license plate image; The average brightness, contrast, noise intensity, and haze index are input into a pre-trained scene classifier to determine the weather scene in which the vehicle is located. The weather scene includes rain and snow scene, fog scene, low light scene, and strong backlight scene. Based on the weather scene in which the vehicle is located, an enhancement method matching the weather scene is determined, and image enhancement processing is performed on the real license plate area according to the enhancement method matching the weather scene in which the vehicle is located to obtain the enhanced license plate area. The image enhancement method is different for different weather scenes.

7. The parking lot cloud contact center remote management system of claim 5, wherein, When using a multi-scenario license plate recognition algorithm to perform license plate recognition on the enhanced license plate area, the AI ​​analysis unit is configured as follows: The license plate area is then corrected to obtain the corrected license plate area. The vertical projection and watershed algorithm are used to perform character segmentation on the corrected license plate area to obtain several initial character segmentation images. Morphological opening operation is then performed on these initial character segmentation images to obtain several candidate character segmentation images. Several candidate character segmentation images are processed to complete the characters, resulting in several pre-selected character segmentation images. Then, several pre-selected character segmentation images are processed to perform non-character filtering, resulting in several character segmentation images. Perform character recognition processing on several character segmentation images to obtain the character category and confidence level corresponding to each character segmentation image, and retain the character categories with confidence levels greater than or equal to the first threshold; Character segmentation images corresponding to character categories with confidence scores less than the first threshold are designated as low-confidence character images. Feature extraction is performed on low-confidence character images to obtain character features. Then, using the character features, fuzzy matching of characters is performed in a character template library to obtain the similarity between each low-confidence character image and each template character in the character template library. The low-confidence character images are divided into confused characters and unconfused characters. For any unconfused character, template characters with a similarity greater than or equal to a second threshold are selected from the character template library as candidate characters corresponding to any unconfused character. After traversing all unconfused characters, candidate characters corresponding to each unconfused character are obtained. The first threshold is less than the second threshold. For any obfuscated character, template characters with a similarity greater than or equal to the second threshold are selected from the character template library and used as the initial candidate characters corresponding to the obfuscated character. The initial candidate characters are subjected to secondary verification to select pre-selected characters from the initial candidate characters; The pre-selected characters of any of the obfuscated characters are subjected to obfuscated character expansion processing to obtain an expanded pre-selected character list; Using license plate filtering rules, the extended pre-selected character list is filtered to obtain the candidate character corresponding to any of the obfuscated characters. After traversing all obfuscated characters, the candidate character corresponding to each obfuscated character is obtained. Using character categories with confidence levels greater than or equal to the first threshold, candidate characters corresponding to each non-confusing character, and candidate characters corresponding to each confused character, several initial candidate license plates are generated. Using license plate filtering rules, at least one candidate license plate is selected from several initial candidate license plates, and the combined confidence of each candidate license plate is calculated. Based on the combined confidence of each candidate license plate, the license plate recognition result is determined from each candidate license plate.

8. The parking lot cloud contact center remote management system of claim 5, wherein, When identifying anomalous events based on enhanced images, the AI ​​analysis unit is configured as follows: Based on the enhanced image, determine whether there is a real license plate area for the vehicle; If not, an abnormal event of no license plate is determined, and an abnormal alarm message corresponding to the no-license-plate lingering is generated; otherwise, it is determined whether the license plate in the enhanced license plate area is damaged. If so, an abnormal event of license plate defacement is determined, and an abnormal alarm message corresponding to the license plate defacement is generated; otherwise, the license plate recognition result is obtained based on the enhanced license plate area. After obtaining the license plate recognition result, the system receives the first barrier status data reported by the barrier control unit. The first barrier status data includes the barrier angle, status code, reporting timestamp, and license plate recognition result. From the first barrier gate status data, filter out data records where the barrier gate angle is greater than the preset angle and the status code is the specified status, and obtain the reporting timestamp corresponding to the data record; Starting from the reporting timestamp, the system continuously monitors the status data of the second barrier gate that is located after the reporting timestamp and within the preset time period, and determines whether there is any barrier gate closure status data in the status data of the second barrier gate. If not, it is determined that an abnormal event of the barrier gate not being lowered has occurred, and an abnormal alarm message corresponding to the barrier gate not being lowered is generated; Specifically, when the AI ​​analysis unit detects two consecutive vehicles with unpaid toll information and the gate is open, it determines that a following anomaly event has occurred and generates corresponding abnormal alarm information; and When a network interruption is detected, a network fault is determined, and an abnormal alarm message corresponding to the network fault is generated.

9. A parking lot cloud-based remote control management system according to claim 8, characterized in that, The front-end perception layer includes an unattended robot, and the cloud service layer also includes a cloud seat management platform. When the AI ​​analysis unit determines that an abnormal event of no license plate has occurred, it sends a dynamic QR code to the unattended robot so that the car owner can scan the dynamic QR code to generate a temporary license plate for the vehicle and feed it back to the cloud seat management platform. The cloud-based seat management platform is used to allocate parking spaces to vehicles and associate and store the parking spaces with the temporary license plates. The cloud-based management platform sends a release command to the gate control unit to allow unlicensed vehicles to enter the site.

10. A parking lot cloud-based remote control management system according to claim 1, characterized in that, The cloud service layer also includes a data center, wherein the data center adopts a master-slave cluster off-site disaster recovery storage architecture, the master cluster in the master-slave cluster stores all real-time data in the cloud service layer, the slave cluster in the master-slave cluster performs off-site backup of all data stored in the master cluster, and the data center uses the AES-256 encryption algorithm to encrypt data when transmitting and storing data.

Images