Intelligent parking management system using artificial intelligence image recognition

Abstract

The invention relates to the field of intelligent traffic systems, and discloses an intelligent parking management system using artificial intelligence image recognition, and the system comprises an image collection module which is used for collecting a parking space image in a parking lot in real time, and is combined with at least one sensor unit which comprises a high-definition camera, an infrared thermal imager and a laser radar sensor; the image preprocessing module is used for preprocessing the acquired parking space image, removing noise in the image and enhancing the image; the geometric transformation module is used for converting the parking space image from the current view angle into a standard parking space model and eliminating view angle distortion; and the illumination optimization module is used for carrying out illumination correction on the parking space image and reducing the influence of illumination change on the parking space image. Through an image acquisition technology based on multi-sensor data fusion and in combination with a high-definition camera, an infrared thermal imager and a laser radar sensor, a parking space image can be accurately acquired under different illumination and complex environment conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent transportation systems, and specifically to an intelligent parking management system using artificial intelligence image recognition. Background Art

[0002] With the continuous acceleration of the urbanization process, the urban traffic pressure is gradually increasing, and the problem of parking difficulty is becoming increasingly serious. Especially in densely populated areas such as large commercial areas, residential areas, and hospitals, the parking space resources are tense, and car owners need to spend a lot of time looking for parking spaces. To solve this problem, an intelligent parking management system has emerged. By using modern information technologies such as sensors, image recognition, and artificial intelligence, the system can monitor the usage of parking spaces in the parking lot in real time, help car owners quickly find available parking spaces, and greatly improve the parking efficiency.

[0003] Currently, intelligent parking management systems generally use sensor technology, video surveillance, and image recognition technology for parking space monitoring. By collecting real-time data of the parking lot through high-precision cameras and sensors, the system can judge whether a parking space is occupied and provide relevant information to car owners. At the same time, combined with image processing technology, the system can clearly identify the status of parking spaces, avoiding the waste of time and energy caused by manual inspections. However, existing technologies mostly rely on single sensors or simple image processing methods. Although these solutions have solved the basic needs of parking management to a certain extent, their accuracy and stability in complex environments still need to be improved.

[0004] The main problem of the existing technology is that its image acquisition and processing technology has not been able to effectively adapt to complex environmental conditions. The reliance on a single sensor is likely to result in poor image acquisition quality in low light, strong light, and occlusion situations, making it impossible to accurately identify the occupancy status of parking spaces. In addition, traditional image processing methods still have limitations in noise removal, image clarity improvement, and perspective distortion correction, resulting in insufficient accuracy of parking space recognition. These problems limit the application of intelligent parking systems in complex parking environments and cannot provide more efficient and accurate parking management services. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides an intelligent parking management system using artificial intelligence image recognition, which solves the problem that the existing system's reliance on a single sensor is likely to result in poor image acquisition quality in low light, strong light, and occlusion situations, making it impossible to accurately identify the occupancy status of parking spaces.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: An intelligent parking management system using artificial intelligence image recognition, comprising:

[0007] The image acquisition module is used to collect the parking space images in the parking lot in real time, and in combination with at least one sensor unit, including a high-definition camera, an infrared thermal imager, and a lidar sensor, to obtain parking space images from different perspectives and environmental conditions. The images and sensor data are fused by a data fusion processing module;

[0008] The image preprocessing module is used to preprocess the collected parking space images, remove the noise in the images and perform image enhancement. The image preprocessing module denoises and enhances the images through a deep learning network, using a deep adaptive denoising network;

[0009] The geometric transformation module is used to transform the parking space image from the current perspective to a standard parking space model, eliminating perspective distortion. The geometric transformation module uses a combined training method of projective geometric four-point transformation and a deep convolutional neural network for geometric correction;

[0010] The illumination optimization module is used to perform illumination correction on the parking space images, reducing the impact of illumination changes on the parking space images. The illumination optimization module performs illumination correction on the parking space images through a physical modeling method based on the radiative transfer equation;

[0011] The deep learning module is used to identify the occupancy status of the parking space based on the collected images and sensor data, using a multi-modal learning network, combining the hybrid structure of a convolutional neural network and a long short-term memory network;

[0012] The real-time feedback module is used to real-time feedback the parking space occupancy status results to the vehicle owners and the parking lot management system, and provide personalized recommendations based on the vehicle owners' historical parking data.

[0013] Preferably, the image acquisition module includes a high-definition camera, an infrared thermal imager, and a lidar sensor. The images and sensor data are processed by a deep fusion algorithm, and the algorithm optimizes the recognition ability of the parking space occupancy status through an adaptive learning method.

[0014] Preferably, the image preprocessing module uses a generative adversarial network for image data enhancement. The method includes improving the model adaptability by generating a diverse training dataset of parking space images. The enhancement method trains a deep convolutional neural network model to remove the noise in the images.

[0015] Preferably, the geometric transformation module uses projective geometric four-point transformation to perform geometric correction on the parking space images, and combines with a deep convolutional neural network for joint training. The training process uses the method of augmented dataset, by increasing the training samples of different parking space layouts and perspectives, to optimize the correction of the geometric shape of the parking space images.

[0016] Preferably, the deep learning module includes a hybrid network structure of a convolutional neural network and a long short-term memory network. The convolutional neural network is used to extract spatial features in the parking space image, and the long short-term memory network is used to analyze the sequence information of the parking space occupancy status changing over time. The hybrid network structure can dynamically identify the parking space occupancy status.

[0017] Preferably, the lighting optimization module calculates the lighting information of the parking space in real time through a physical modeling method based on the radiative transfer equation. The method includes correcting uneven lighting in the parking space image and correcting the lighting changes in the image by automatically adjusting the lighting coefficient.

[0018] Preferably, the deep learning module enhances the diversity of training data through data augmentation techniques. The data augmentation techniques include operations such as rotating, scaling, translating, mirror-inverting, and lighting changes on the parking space image, enhancing the adaptability of the model in complex parking lot environments.

[0019] Preferably, the system further includes a data sharing module for uploading real-time parking lot data and parking space occupancy status information to the cloud platform. The cloud platform conducts unified management, scheduling, and optimization of parking space resources in multiple parking lots through big data analysis, and adjusts the parking space allocation strategy in real time.

[0020] Preferably, the real-time feedback module pushes the parking space occupancy status information in real time through the vehicle owner's terminal device and provides personalized recommendations based on the vehicle owner's parking history data. The recommendation system combines the vehicle owner's parking habits, real-time parking space occupancy, and the vehicle owner's historical preferences for parking space recommendations.

[0021] Preferably, the system predicts the parking demand by integrating an adaptive scheduling algorithm, combining real-time traffic flow and parking space occupancy status. The scheduling algorithm automatically adjusts the parking space allocation strategy according to the real-time data of the parking lot, optimizing the use of parking spaces in the parking lot.

[0022] The present invention provides an intelligent parking management system using artificial intelligence image recognition.

[0023] It has the following beneficial effects:

[0024] 1. The present invention adopts an image acquisition technology based on multi-sensor data fusion, combined with a high-definition camera, an infrared thermal imager, and a lidar sensor, which can accurately acquire parking space images under different lighting and complex environmental conditions. It can still stably collect high-quality parking space data in low light, strong light, and occlusion situations. Compared with the image acquisition scheme that relies on a single sensor in the prior art, the technical solution of the present invention solves the problems of poor accuracy and insufficient stability of the traditional scheme in complex parking lot environments.

[0025] 2. The present invention performs denoising and enhancement processing on the parking space image through a deep adaptive denoising network, which can retain key features and effectively remove noise in an environment with poor image quality. It improves the clarity of the image, and can still accurately identify the parking space especially in low-light and high-noise environments. Compared with the solutions using traditional image processing methods in the prior art, the present invention can process images in more complex environments, solving the problem that traditional methods cannot effectively remove noise and maintain details.

[0026] 3. The present invention adopts a joint training method combining projective geometric four-point transformation and a deep convolutional neural network to perform geometric correction on the parking space image, eliminating perspective distortion and ensuring the standardized presentation of the parking space area in the image. It achieves accurate geometric restoration of the parking space and improves the accuracy of parking space recognition. Compared with the solutions in the prior art that only rely on projective geometry or traditional geometric correction methods, the technical solution of the present invention can more effectively process image distortion under complex perspectives, significantly improving the geometric accuracy of the image.

[0027] 4. The present invention combines a deep learning module with a convolutional neural network and a long short-term memory network to achieve accurate recognition of the parking space occupancy status, and conducts personalized recommendations through historical data analysis. It achieves efficient and accurate parking space occupancy status recognition and intelligent recommendation functions, optimizing the parking experience. Compared with the single image recognition solutions in the prior art, the present invention can provide higher accuracy in the time series analysis of changes in the parking space occupancy status and provide personalized recommendations for car owners, solving the deficiency that traditional methods cannot make full use of historical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is the main framework diagram of the present invention;

[0029] Figure 2 is the schematic flow diagram of the image acquisition module of the present invention;

[0030] Figure 3 is the schematic flow diagram of the image preprocessing module of the present invention;

[0031] Figure 4 is the schematic flow diagram of the geometric transformation module of the present invention;

[0032] Figure 5 is the schematic flow diagram of the deep learning module of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0033] Next, in combination with the accompanying drawings of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0034] Please refer to the attached Figure 1 - attached Figure 5 , the embodiment of the present invention provides an intelligent parking management system using artificial intelligence image recognition, including:

[0035] An image acquisition module, configured to collect parking space images in the parking lot in real time, and in combination with at least one sensor unit, including a high-definition camera, an infrared thermal imager, and a lidar sensor, for obtaining parking space images from different perspectives and environmental conditions. The images and sensor data are fused through a data fusion processing module;

[0036] An image preprocessing module, configured to preprocess the collected parking space images, remove noise in the images and perform image enhancement. The image preprocessing module denoises and enhances the images through a deep learning network, using a deep adaptive denoising network;

[0037] A geometric transformation module, configured to transform the parking space images from the current perspective to a standard parking space model, eliminate perspective distortion. The geometric transformation module performs geometric correction by means of a combined training method of projective geometric four-point transformation and a deep convolutional neural network;

[0038] A lighting optimization module, configured to perform lighting correction on the parking space images, reduce the influence of lighting changes on the parking space images. The lighting optimization module performs lighting correction on the parking space images through a physical modeling method based on the radiative transfer equation;

[0039] A deep learning module, configured to identify the occupancy status of the parking space based on the collected images and sensor data, using a multi-modal learning network, combining a hybrid structure of a convolutional neural network and a long short-term memory network;

[0040] A real-time feedback module, configured to real-time feedback the parking space occupancy status results to the vehicle owner and the parking lot management system, and provide personalized recommendations based on the vehicle owner's historical parking data.

[0041] The image acquisition module includes a high-definition camera, an infrared thermal imager, and a lidar sensor. The images and sensor data are processed through a deep fusion algorithm, and the algorithm optimizes the recognition ability of the parking space occupancy status through an adaptive learning method.

[0042] The image preprocessing module uses a generative adversarial network for image data augmentation. The methods include enhancing the model adaptability by generating a diverse parking space image training dataset, and the enhancement method trains a deep convolutional neural network model to remove noise from the images.

[0043] The geometric transformation module uses projective geometric four-point transformation to perform geometric correction on the parking space images, and conducts joint training in combination with a deep convolutional neural network. The training process adopts the method of augmented dataset. By increasing the training samples of different parking space layouts and perspectives, the correction of the geometric shape of the parking space images is optimized.

[0044] The deep learning module includes a hybrid network structure of a convolutional neural network and a long short-term memory network. The convolutional neural network is used to extract the spatial features in the parking space images, and the long short-term memory network is used to analyze the sequential information of the parking space occupancy status changing over time. The hybrid network structure can dynamically identify the parking space occupancy status.

[0045] The illumination optimization module calculates the illumination information of the parking space in real time through a physical modeling method based on the radiative transfer equation. The methods include correcting the uneven illumination in the parking space images and automatically adjusting the illumination coefficient to correct the illumination changes in the images.

[0046] The deep learning module enhances the diversity of training data through data augmentation techniques. The data augmentation techniques include operations such as rotating, scaling, translating, mirror-inverting, and illumination changing of the parking space images, which enhance the adaptability of the model in complex parking lot environments.

[0047] The system further includes a data sharing module for uploading the real-time data of the parking lot and the parking space occupancy status information to the cloud platform. The cloud platform conducts unified management, scheduling, and optimization of the parking space resources of multiple parking lots through big data analysis, and adjusts the parking space allocation strategy in real time.

[0048] The real-time feedback module pushes the parking space occupancy status information to the vehicle owner's terminal device in real time, and provides personalized recommendations based on the vehicle owner's parking history data. The recommendation system combines the vehicle owner's parking habits, the real-time parking space occupancy situation, and the vehicle owner's historical preferences for parking space recommendations.

[0049] The system predicts the parking demand by integrating an adaptive scheduling algorithm, combining the real-time traffic flow and the parking space occupancy status. The scheduling algorithm automatically adjusts the parking space allocation strategy according to the real-time data of the parking lot, and optimizes the use of parking spaces in the parking lot.

[0050] The image acquisition module includes multiple sensor devices, specifically including high-definition cameras, infrared thermal imagers, and lidar sensors. Each sensor has different advantages and can provide complementary data under different environmental conditions. The working methods and functions of these sensors are as follows:

[0051] High-definition camera: The high-definition camera is responsible for collecting the images of parking spaces in the parking lot and providing high-resolution image data. Generally, the high-definition camera can clearly capture the parking space information during the day or in an environment with good lighting conditions.

[0052] Infrared thermal imager: The infrared thermal imager is suitable for low-light environments or night-time environments and can identify the status of parking spaces by capturing heat differences. In low-light or completely dark environments, the infrared thermal imager can effectively provide parking space information, especially when other sensors cannot work, it can supplement the information.

[0053] LiDAR sensor: The LiDAR sensor is responsible for obtaining the depth information of the parking lot in real time and can measure the spatial distance of the parking space, the boundaries of the parking space, and the distance to surrounding obstacles with high precision. The high-precision ability of LiDAR helps the system accurately identify the outline of the parking space and vehicle occupancy.

[0054] By using these devices in combination, the image acquisition module can provide stable parking space image data under different lighting, weather, and parking lot layout conditions. Specifically, the data from LiDAR and infrared thermal imagers can help the system correctly identify the occupancy status of parking spaces even when the parking spaces are blocked by vehicles or in low-light conditions.

[0055] After the data acquisition is completed, the sensor data will be processed by the data fusion module. The function of the data fusion module is to integrate the data from the high-definition camera, infrared thermal imager, and LiDAR sensor. Specifically, the image data and sensor data will be processed into a unified format for subsequent image preprocessing and deep learning model use. The work of the data fusion processing module can be achieved through various algorithms, and common ones include weight-based fusion algorithms, Kalman filtering algorithms, or convolutional neural networks, etc.

[0056] As an option, the algorithm for data fusion can use the following formula to calculate the fusion weights:

[0057]

[0058] Where:

[0059] D f is the fused data;

[0060] D i represents the data provided by the i-th sensor;

[0061] w i is the weight value of the i-th sensor

[0062] n is the number of sensors.

[0063] Specifically, the weight w of the sensor i can be automatically adjusted according to the signal-to-noise ratio, accuracy, and working environment of the sensor to improve the accuracy of data fusion. In this way, the image acquisition module can achieve effective fusion of multi-source data, avoid the defects of a single sensor, and thus provide more stable and accurate data.

[0064] In some embodiments, the image acquisition module can further combine an automatic exposure control algorithm to adjust the exposure parameters of the camera in real time to cope with strong light or low light conditions. In this case, the exposure control algorithm adjusts the shutter speed and aperture size of the camera so that the image always remains within the optimal brightness range, avoiding overexposure or underexposure.

[0065] The image acquisition module is connected to other modules in the system (such as the image preprocessing module, geometric transformation module, etc.) through a data stream. After collecting the original image data, the image acquisition module transmits the data to the image preprocessing module for denoising and enhancement. In the image preprocessing module, first, the noise in the image is removed and the clarity of the image is enhanced, and then the processed image is passed to the geometric transformation module for geometric correction.

[0066] Image preprocessing module: During image preprocessing, the collected image will be denoised through a depth adaptive denoising network, and the network will dynamically adjust the denoising strategy according to different regions and noise levels of the image to ensure the integrity of image details while removing image noise.

[0067] Geometric transformation module: The geometric transformation module receives the preprocessed image and performs standardized correction on the image through the joint training of projective geometric four-point transformation and a deep convolutional neural network. The goal of this module is to correct the distortion caused by different perspectives in the image to obtain a standardized parking space image.

[0068] In a possible implementation, the image acquisition module can further add more types of sensors, such as millimeter-wave radar sensors or ultrasonic sensors, to cope with more complex parking environments. For example, in an underground parking lot or a multi-story parking lot, millimeter-wave radar can penetrate walls and other obstacles to provide additional spatial information, further improving the coverage and accuracy of parking space detection.

[0069] In addition, the deep learning data fusion module mentioned in this embodiment can be further optimized. By introducing a deep fusion technology that combines a deep convolutional neural network (CNN) and a recurrent neural network (RNN), the image and sensor data can be efficiently integrated at multiple levels, thereby enhancing the system's adaptability to complex parking environments.

[0070] The main task of the image preprocessing module is to remove noise, enhance, and optimize the images obtained from the image acquisition module. During the image acquisition process, the image data may be affected by different lighting conditions, noise interference, and vehicle occlusion, resulting in low image quality. Therefore, the process of image preprocessing is crucial for improving the accuracy of subsequent parking space recognition.

[0071] In this embodiment, a deep learning method is used for image preprocessing. Specifically, a deep adaptive denoising network is used to remove noise from the acquired images. Compared with traditional image processing techniques, this denoising method can more accurately process image data in low-light and high-noise environments.

[0072] In the image preprocessing module, first, the image undergoes denoising. Generally, during the image acquisition process, noise mainly comes from factors such as insufficient lighting and environmental interference. Through the deep adaptive denoising network, high-frequency noise in the image can be effectively removed while retaining the details in the image, avoiding over-smoothing and detail loss of the image.

[0073] To further enhance the image quality, the preprocessing module also includes the step of image enhancement. Specifically, image enhancement mainly enhances the readability of the image and improves the clarity of parking space markings and vehicle contours by adjusting parameters such as contrast, brightness, and sharpness. During the image enhancement process, the brightness and contrast of the image are adaptively adjusted according to the specific situation of the scene to adapt to parking space images under different lighting conditions.

[0074] In some embodiments, the image preprocessing module uses a deep convolutional neural network (CNN) to learn and automatically optimize denoising and enhancement operations. By training an adaptive denoising convolutional neural network, the network can adaptively adjust the denoising strategy according to different features of the image, so that each image can be processed most appropriately. During training, the input image to the network is the acquired original image, and the output of the network is the image after noise removal and enhancement. The goal of this process is to remove noise to the maximum extent while retaining the important features in the image.

[0075] In specific implementations, the loss function of the deep adaptive denoising network usually includes the following parts:

[0076] L total =L data +λL smoothness +αL perceptual

[0077] Where:

[0078] L data is the reconstruction loss of the image data, and usually the mean square error (MSE) is used to measure the difference between the denoised image and the real image;

[0079] L smoothness is the smoothness loss, which is used to maintain the texture and details in the image and avoid over - denoising;

[0080] L perceptual is the perceptual loss, which is used to ensure that the image is similar to the real image at the perceptual level;

[0081] λ and α are the weight hyperparameters of the smoothness loss and the perceptual loss respectively.

[0082] The design of the loss function can ensure that while the network is denoising, it retains important information in the image, such as parking space markings and vehicle contours, thereby improving the accuracy of parking space recognition.

[0083] As an option, the image pre - processing module can also use a generative adversarial network (GAN) for data augmentation. Through the generative adversarial network, a diverse training dataset of parking space images can be generated, thereby increasing the diversity of training data and improving the robustness of the deep - learning model. Specifically, the generative adversarial network consists of a generator and a discriminator. The generator is responsible for generating pseudo - images similar to real parking space images, and the discriminator is responsible for judging the authenticity of the images. During the training process, the generator and the discriminator play against each other, and the generator learns to generate more realistic parking space images.

[0084] In a possible implementation, the generative adversarial network is optimized by the following formula:

[0085]

[0086] Where:

[0087] G is the generator, which is responsible for generating pseudo - images;

[0088] D is the discriminator, which is responsible for judging the authenticity of the images;

[0089] x represents the real image, and G(z) represents the generated pseudo - image;

[0090] p data (x) is the true distribution of the data, and p z (z) is the noise distribution of the generator input.

[0091] Through this method, the image pre - processing module can generate more diverse training images, further improving the accuracy of parking space occupancy status recognition and the adaptability of the model.

[0092] The image preprocessing module is closely connected with the aforementioned image acquisition module and the subsequent geometric transformation module and deep learning module. In the image acquisition module, after the raw parking space images collected by the sensor device are processed through data fusion, they are transmitted to the image preprocessing module for noise removal and image enhancement. The denoised and enhanced image data will be transmitted to the geometric transformation module for further geometric correction to remove perspective distortion and make the parking space images more standardized.

[0093] After geometric transformation, the image is sent to the deep learning module for recognition of the occupancy status of the parking space. The deep learning module uses a combination of convolutional neural network (CNN) and long short-term memory network (LSTM) to extract spatial and temporal features from the image and classify according to the occupancy status of the parking space.

[0094] After the image acquisition module and the image preprocessing module complete the preliminary image acquisition and denoising and enhancement processing, the parking space images will be transmitted to the geometric transformation module for processing. Generally, since the parking space images are collected from different angles, distortion often appears in the images, such as the inclination of the parking space markings and the irregular shape of the parking space area. Therefore, the goal of the geometric transformation module is to convert the image from any perspective to a standard parking space perspective and eliminate these distortions, so that the parking space area in the image conforms to the real geometric shape.

[0095] Specifically, the geometric transformation module uses the joint training of projective geometric four-point transformation and deep convolutional neural network (CNN) to optimize the geometric correction of the image. The projective geometric four-point transformation method is based on four calibration points in the image (usually the four corners of the parking space markings), and uses projective geometric algorithms to map these points to the standard parking space model.

[0096] During implementation, the four calibration points can be the four corners of the parking space markings or other significant geometric feature points within the parking space. By selecting these points and applying projective geometric transformation, the parking space image can be effectively converted from any shooting angle to a standard parking space model. This conversion process is not just simple transformations such as translation and rotation of the image, but a correction of the perspective of the image, making the parking space markings and the parking space area present the real geometric shape.

[0097] In a possible implementation, the core formula of the geometric transformation is as follows:

[0098]

[0099] Where:

[0100] (x, y) are the coordinate points in the original image;

[0101] (x′, y′) are the standard coordinate points after projective transformation;

[0102] h 1 , h 2 , …, h 8 are the parameters of the transformation matrix, which determine how to map from the original image to the standard perspective.

[0103] The transformation matrix is obtained by solving four known calibration points in the image. The optimization process of this transformation matrix usually uses the least squares method or other optimization algorithms for calculation to ensure that the transformed image is as close as possible to the standard parking space model.

[0104] In some embodiments, the geometric transformation module also combines with a deep convolutional neural network (CNN) for joint training to further optimize the accuracy of geometric correction. Specifically, the CNN is used to automatically extract the spatial features in the parking space image to help determine how to select the appropriate transformation matrix.

[0105] Through the training of the convolutional neural network, the network can learn the features of different perspectives and geometric forms in the parking space image, and then guide the parameter adjustment of the projective geometry four-point transformation. During the training process of the CNN, through a large number of image data with different angles and parking space layouts, its feature extraction ability is continuously optimized, and finally more accurate geometric correction is achieved.

[0106] In practical applications, the training process of the deep convolutional neural network usually uses the following loss function:

[0107]

[0108] Where:

[0109] L georm is the loss function of geometric transformation, representing the difference between the transformed image and the image of the standard parking space model;

[0110] is the predicted image after transformation;

[0111] is the target image, that is, the image of the standard parking space model;

[0112] N is the number of training samples;

[0113] represents the square of the Euclidean distance, which is used to measure the difference between images.

[0114] By minimizing this loss function, the CNN can gradually learn the transformation relationship between the parking space image and the standard perspective, thus providing a more accurate geometric transformation.

[0115] In a possible implementation, the geometric transformation module can also dynamically adapt to different types of parking space layouts. For example, for angled parking spaces or parking spaces with special shapes, the geometric transformation module automatically identifies the geometric features of the parking space and adjusts the parameters of the projective geometric transformation, so that even in an irregular parking lot, the parking space image can be accurately converted into a standard model.

[0116] Specifically, the geometric transformation module automatically calculates a transformation matrix suitable for this type of parking space based on the characteristics of the parking space markings and the parking space area in the input image. In this process, the deep convolutional neural network extracts features from the parking space markings and the parking space area, determines the shape and angle of the parking space, and then selects a suitable geometric transformation method for each type of parking space.

[0117] The illumination optimization module first conducts a preliminary analysis of the parking space image to identify uneven illumination areas in the image. Specifically, the image acquisition module and the image preprocessing module have passed the original image to the illumination optimization module, and the uneven illumination, shadows, or highlight areas that may exist in the image need to be processed. By calculating the illumination distribution in the image, the illumination optimization module can accurately identify overexposed or underexposed areas and adjust them according to the physical model.

[0118] Generally, the uneven illumination in the image may be due to strong light or shadows caused by obstacles. Through the radiative transfer equation, the module can simulate the illumination changes in the parking space area and adjust these uneven areas by correcting the illumination coefficient. This process ensures that the illumination distribution in the parking space image is as uniform as possible, thereby improving the accuracy of subsequent parking space recognition.

[0119] The illumination optimization module optimizes the parking space image using physical modeling methods. Specifically, the model based on the radiative transfer equation simulates the illumination path from the light source to the parking space surface and then reflected to the camera. By modeling these illumination paths, the module can calculate and adjust the brightness, contrast, and color balance of the illumination in the image, eliminating errors caused by different illumination conditions.

[0120] In a possible implementation, the illumination optimization module corrects the image using the following formula:

[0121] I corrected (x,y) = I original (x,y) × α(x,y)

[0122] Where:

[0123] I corrected (x,y) represents the corrected image pixel value;

[0124] I original (x,y) represents the original image pixel value;

[0125] α(x, y) is the illumination correction coefficient, representing the adjustment made to the brightness of each pixel point.

[0126] The illumination correction coefficient α(x, y) is calculated through physical modeling and is adjusted according to the change in illumination intensity at the location of each pixel point. Generally, the correction coefficient will be adaptively adjusted according to the brightness distribution of the image to adapt to different illumination conditions.

[0127] In some embodiments, the illumination optimization module also uses local illumination compensation technology to perform local adjustments according to the illumination conditions in different regions of the image. This method takes into account the different illumination conditions that may exist in different parking space regions, locally adjusts the brightness and contrast of the image, and makes the brightness and contrast of each parking space image reach the optimal state.

[0128] Specifically, the local illumination compensation method determines whether to perform local compensation by calculating the local brightness of each small region of the image and comparing it with the global brightness. This local optimization method can avoid the loss of image details caused by global adjustment and retain important information such as parking space markings and vehicle outlines.

[0129] In terms of formula implementation, the adjustment process of local illumination compensation can be expressed by the following formula:

[0130] I corrected (x, y) = I original (x, y) × (1 + β(L local (x, y) - L global ))

[0131] Where:

[0132] L local (x, y) is the illumination intensity of the local region of the image;

[0133] L global is the global illumination intensity of the image;

[0134] β is the adjustment factor that controls the intensity of local illumination compensation.

[0135] This local compensation method can dynamically adjust the brightness of each region according to the difference in illumination intensity in each region of the image, making the parking space image more uniform under different illumination conditions.

[0136] Collaboration between the illumination optimization module and other modules

[0137] There is a close cooperation relationship among the light optimization module, the image acquisition module, the image preprocessing module, the geometric transformation module, and the deep learning module. The images captured by the image acquisition module often need to be optimized due to reasons such as uneven lighting. The image preprocessing module first denoises and enhances the images, removing the noise in the images and improving the image quality. The preprocessed image data is transmitted to the light optimization module for light correction.

[0138] The images after light optimization will be passed as input data to the geometric transformation module for geometric correction. The geometric transformation module performs a standardized transformation on the parking space images, making the parking space markings and parking space areas in the images present the correct geometric shapes. Next, the corrected images will be transmitted to the deep learning module for the recognition of the parking space occupancy status.

[0139] In this embodiment, the deep learning module is the core component of the intelligent parking management system, responsible for performing in-depth analysis on the parking space images after image acquisition, preprocessing, geometric transformation, and light optimization to identify the occupancy status of the parking spaces. The deep learning module combines the convolutional neural network (CNN) and the long short-term memory network (LSTM), and through the combination of spatial feature extraction and time series analysis, realizes the accurate prediction of the parking space occupancy status.

[0140] There is a close connection relationship between this module and the aforementioned image preprocessing and geometric transformation modules. The original images obtained by the image acquisition module through the sensors, after being denoised and enhanced by the image preprocessing module, and then through the standardized transformation by the geometric transformation module, the image data after light optimization will be transmitted to the deep learning module for the judgment of the parking space occupancy status.

[0141] In this embodiment, the working principle of the deep learning module is as follows:

[0142] The deep learning module includes a convolutional neural network (CNN) and a long short-term memory network (LSTM), which are respectively used to process the spatial features and time series features of the images. The CNN is used to extract spatial features from the images, while the LSTM processes the time series features and analyzes the trend of the parking space occupancy status changing over time.

[0143] The convolutional neural network (CNN) is a common network structure for image processing in deep learning. Through multiple layers of convolution, pooling, and fully connected operations, it automatically extracts effective spatial features from the parking space images. These features include the shapes of parking space markings, vehicle contours, and parking space areas.

[0144] Specifically, the convolution operation can be described by the following formula:

[0145]

[0146] Where:

[0147] $y_{i,j}$ represents the output feature map after the convolution operation;

[0148] $x$ i,j is the pixel value of the input image;

[0149] $w$ m,n is the weight of the convolution kernel;

[0150] $b$ is the bias term of the convolution operation;

[0151] $M$ and $N$ represent the height and width of the convolution kernel respectively.

[0152] Through multiple layers of convolution and pooling operations, the CNN can extract local and global features in the parking space image layer by layer, and finally output a feature map containing parking space information. This feature map is the key data input for parking space occupancy status recognition.

[0153] The long short-term memory network (LSTM) is used to process time series data, especially the dynamic information of the parking space occupancy status changing over time. In the parking space status recognition of a parking lot, the occupancy status of a parking space is not only related to the current image but also affected by the past status. The LSTM effectively captures these long-term dependencies through a special gating mechanism.

[0154] The basic structure of the LSTM network includes an input gate, a forget gate, and an output gate, which can determine whether a parking space is occupied based on the context information of time series data. The calculation process of the LSTM can be expressed by the following formula:

[0155] $f$ t $= \sigma(W$ f $\cdot [h$ t-1 , $x$ t $] + b$ f )$

[0156] $i$ t $= \sigma(W$ i $\cdot [h$ t-1 , $x$ t $] + b$ i )$

[0157]

[0158] $o$ t $= \sigma(W$ o $\cdot [h$ t-1 , $x$ t $] + b$ o )$

[0159] $h$ t $= o$ t $* \tanh(C$ t )$

[0160] Among them:

[0161] f t is the forget gate, which is used to control the degree of retention of past information;

[0162] i t is the input gate, which is used to control the writing of current information;

[0163] is the candidate cell state, representing the potential impact of the current input;

[0164] C t is the current cell state, which stores long-term information;

[0165] o t is the output gate, which controls the output at the current moment;

[0166] h t is the output of the LSTM and serves as the final output of the network.

[0167] Through these gating mechanisms, the LSTM can effectively capture the changing trend of the parking space occupancy status and transmit this information to the downstream decision-making layer. Through the temporal analysis of the parking space status, the LSTM can determine whether the parking space will be occupied or idle in the future for a period of time.

[0168] In some embodiments, the deep learning module also combines data augmentation techniques to increase the diversity of training data. Data augmentation includes operations such as rotating, scaling, translating, and mirror-inverting the parking space images. These operations enable the model to adapt to image changes in different scenarios and improve the robustness of the model. Specifically, data augmentation can be achieved through the following formula:

[0169] I augmented (x, y) = α · I(x, y) + β

[0170] Among them:

[0171] I augmented (x, y) are the pixels of the enhanced image;

[0172] I(x, y) are the pixels of the original image;

[0173] α is the contrast adjustment factor, which adjusts the contrast of the image;

[0174] β is the brightness adjustment factor, which adjusts the brightness of the image.

[0175] Through data augmentation, the deep learning module can better adapt to complex environments and handle the impacts of factors such as lighting changes and different parking space layouts on images.

[0176] There is a close cooperation relationship between the deep learning module and the aforementioned image acquisition module, image preprocessing module, geometric transformation module, and lighting optimization module. The raw images captured by the image acquisition module are denoised and enhanced by the preprocessing module, then the images are standardized by the geometric transformation module, the lighting in the images is further adjusted by the lighting optimization module, and finally passed to the deep learning module for identifying the occupancy status of parking spaces.

[0177] The working process of the real-time feedback module is first based on the occupancy status recognition results of the deep learning module. The deep learning module determines the occupancy status of each parking space in the parking lot through multiple image processings. Specifically, the status may be "occupied" or "idle", and may include information such as parking space number, location, and remaining time.

[0178] Under normal circumstances, the vehicle owner's terminal device receives real-time parking space occupancy information through a wireless communication connection with the parking lot management system. When the vehicle owner enters the parking lot, the system determines the vehicle owner's current location through location-based services or other positioning technologies, and displays the specific information of nearby available parking spaces.

[0179] To ensure the timeliness and accuracy of information, the real-time feedback module also provides updates on the parking space status in the parking lot to vehicle owners through push notifications. For example, when a parking space changes from an idle state to an occupied state, the system will immediately push a notification to nearby vehicle owners to help them avoid unnecessary parking attempts.

[0180] In some embodiments, the real-time feedback module also adopts a personalized recommendation function. This function recommends the best parking spaces to vehicle owners based on their historical parking data, parking preferences, and real-time parking space status. The vehicle owner's parking historical data includes information such as parking time, frequently parked locations, and the size of the parking space. The system uses this data to analyze the vehicle owner's parking habits and pushes personalized recommendation results based on the real-time parking space status. The recommendation algorithm is optimized through the following formula:

[0181]

[0182] Where:

[0183] R recommend represents the final score of the personalized recommendation;

[0184] w i is the weight of each feature, representing the importance of various factors in the parking historical data;

[0185] f i are the various features in the vehicle owner's historical data (such as parking time, parking frequency, etc.);

[0186] N is the number of features.

[0187] Through this recommendation algorithm, the system can dynamically adjust the weight of each feature and update the recommendation results in real time. Specifically, when the parking preference of the car owner changes, the system can timely adjust the recommendation strategy, thereby improving the accuracy of the recommendation and the satisfaction of the car owner.

[0188] As an option, the real-time feedback module also supports multiple feedback methods. In addition to pushing notifications through the mobile application, it can also feedback the parking space occupancy status information to the car owner through in-vehicle devices, electronic displays in the parking lot, etc. In the parking lot, the electronic display can display the number and location of available parking spaces in real time, helping the car owner make a quick parking decision. The in-vehicle device synchronizes the parking space information in real time through the Bluetooth or Wi-Fi connection with the car owner's mobile phone, so that the car owner can accurately understand the parking space status regardless of their location in the parking lot.

[0189] In some embodiments, the real-time feedback module can also combine with a dynamic scheduling algorithm to intelligently allocate the parking space resources in the parking lot. This scheduling algorithm is based on the real-time data of the parking lot (such as parking space occupancy rate, car owner's parking habits, etc.), and automatically adjusts the parking space allocation strategy to improve the utilization efficiency of the parking lot. The scheduling algorithm can be modeled by the following formula:

[0190]

[0191] Where:

[0192] S optimal is the parking space allocation score of the parking lot, representing the best allocation plan for the parking space;

[0193] α i and β i are the weights of the occupancy status and the distance respectively;

[0194] Occupancy i is the occupancy status of parking space i (for example, 1 means occupied, 0 means available);

[0195] Distance i is the distance between parking space i and the current position of the car owner.

[0196] This scheduling algorithm can adjust the parking space allocation strategy according to the real-time occupancy situation of the parking lot, thereby optimizing the overall utilization efficiency of the parking lot. For example, during peak demand periods, the system can preferentially recommend the relatively distant available parking spaces to the car owner to balance the utilization rate of the parking spaces.

[0197] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent parking management system using artificial intelligence image recognition, characterized in that: include: An image acquisition module, used for real-time acquisition of parking space images in the parking lot, and combined with at least one sensor unit, including a high-definition camera, an infrared thermal imager and a laser radar sensor, for acquiring parking space images from different viewing angles and environmental conditions, and the image and sensor data are fused through a data fusion processing module; An image preprocessing module is used to preprocess the collected parking space images, remove noise from the images and perform image enhancement. The image preprocessing module denoises and enhances the images through a deep learning network, using a deep adaptive denoising network; A geometric transformation module is used to transform the parking space image from the current perspective into a standard parking space model to eliminate perspective distortion. The geometric transformation module uses a projective geometry four-point transformation and a deep convolutional neural network joint training method to perform geometric correction; An illumination optimization module is used to perform illumination correction on the parking space image to reduce the influence of illumination changes on the parking space image. The illumination optimization module performs illumination correction on the parking space image by a physical modeling method based on the radiation transfer equation; A deep learning module for identifying parking space occupancy status based on collected images and sensor data using a multimodal learning network, combining a hybrid structure of a convolutional neural network and a long short-term memory network; The real-time feedback module is used to provide real-time feedback of parking space occupancy status to car owners and parking lot management systems, and provide personalized recommendations based on car owners' historical parking data.

2. The intelligent parking management system using artificial intelligence image recognition according to claim 1 is characterized in that: The image acquisition module includes a high-definition camera, an infrared thermal imager and a lidar sensor. The image and sensor data are processed by a deep fusion algorithm, and the algorithm optimizes the recognition capability of parking space occupancy status by adaptive learning.

3. The intelligent parking management system using artificial intelligence image recognition according to claim 1 is characterized in that: The image preprocessing module uses a generative adversarial network to enhance image data. The method includes improving model adaptability by generating a diverse parking space image training data set. The enhancement method removes noise in the image by training a deep convolutional neural network model.

4. The intelligent parking management system using artificial intelligence image recognition according to claim 1 is characterized in that: The geometric transformation module uses projective geometry four-point transformation to perform geometric correction on the parking space image, and combines it with a deep convolutional neural network for joint training. The training process uses an augmented data set to optimize the correction of the parking space image geometry by adding training samples with different parking space layouts and perspectives.

5. The intelligent parking management system using artificial intelligence image recognition according to claim 1 is characterized in that: The deep learning module includes a hybrid network structure of a convolutional neural network and a long short-term memory network. The convolutional neural network is used to extract spatial features in parking space images, and the long short-term memory network is used to analyze sequence information of parking space occupancy status changing over time. The hybrid network structure can dynamically identify the parking space occupancy status.

6. The intelligent parking management system using artificial intelligence image recognition according to claim 1 is characterized in that: The illumination optimization module calculates the illumination information of the parking space in real time through a physical modeling method based on the radiation transfer equation. The method includes correcting the uneven illumination in the parking space image and correcting the illumination changes in the image by automatically adjusting the illumination coefficient.

7. The intelligent parking management system using artificial intelligence image recognition according to claim 1 is characterized in that: The deep learning module enhances the diversity of training data through data enhancement technology, which includes operations such as rotation, scaling, translation, mirror reversal and lighting changes on parking space images, thereby enhancing the adaptability of the model in complex parking lot environments.

8. The intelligent parking management system using artificial intelligence image recognition according to claim 1 is characterized in that: The system further includes a data sharing module for uploading real-time parking lot data and parking space occupancy status information to a cloud platform. The cloud platform uniformly manages, dispatches and optimizes parking space resources in multiple parking lots through big data analysis, and adjusts parking space allocation strategies in real time.

9. The intelligent parking management system using artificial intelligence image recognition according to claim 1, characterized in that: The real-time feedback module pushes parking space occupancy status information in real time through the car owner's terminal device, and provides personalized recommendations based on the car owner's parking history data. The recommendation system recommends parking spaces based on the car owner's parking habits, real-time parking space occupancy status and the car owner's historical preferences.

10. The intelligent parking management system using artificial intelligence image recognition according to claim 1, characterized in that: The system predicts parking demand by integrating an adaptive scheduling algorithm, combining real-time traffic flow and parking space occupancy status. The scheduling algorithm automatically adjusts the parking space allocation strategy according to the real-time data of the parking lot to optimize the use of parking spaces in the parking lot.

Images