A parking space number recognition method, device and computer readable storage medium

Abstract

The application discloses a parking space number recognition method and device and a computer readable storage medium, and the method comprises the following steps: acquiring video information of a parking lot by using a vehicle camera acquisition module, and acquiring image data according to the video information; preprocessing the image data; detecting and recognizing the preprocessed image data, using a single-stage deep learning model constructed by an OD detection model and an OCR recognition model to detect and recognize a parking space coding area, and outputting area detection data; converting the area detection data into ONNX model data, and calculating parking space number information according to the ONNX model data. The application can recognize the parking space number in an offline condition without the aid of a cloud system, and is suitable for various parking scenes in a broad sense, such as a ground open-air parking lot, a ground roadside parking space, an underground parking lot and the like. The application can be suitable for different weather conditions, and can recognize parking space numbers of various fonts through the deep learning model, and has high flexibility.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle assistance technology, specifically to a parking space number identification method, device, and computer-readable storage medium. Background Technology

[0002] As the management of parking spaces becomes more standardized in governments, schools, residential communities, and commercial complexes, the method of managing parking spaces by number will become increasingly common. Compared to general management methods, parking space numbering management has the following advantages:

[0003] 1) Facilitates the orderly and standardized management of public places. Parking space codes within the management area are unique and have comprehensive application features. Uniqueness is reflected in the fact that each space has a unique code, allowing for location identification based on the code. Comprehensive application is reflected in the fact that parking space codes can be designed to include address code segments, spatial characteristic code segments, and attribute code segments, allowing for the interpretation of attributes such as the size of the parking space and the number of vehicles that can be parked based on the code.

[0004] 2) Facilitates parking fee management in public places. Parking time can be recorded through parking space coding, enabling unmanned charging and saving manpower and financial resources for parking space management. It can also be further expanded to include long-term rental spaces, prime location parking spaces, shared or private parking spaces, and special parking spaces, facilitating categorized management.

[0005] 3) It is beneficial to public and personal safety. Under the trend of unified management of urban parking spaces, the uniqueness of parking space numbers can enable the traceability of parking events.

[0006] Currently, existing parking management systems on the market primarily identify the property management systems of shopping malls and parking lots. These systems include entrance gates, control terminals, information displays above parking spaces, and cameras. The entrance gate identifies the license plate information of entering vehicles. After entering the parking lot, drivers follow the parking space availability information displayed on the control terminal system and are guided to park in their designated spaces. The cameras above the parking spaces identify license plate information, record parking duration, and display this information on the screen. When leaving the parking lot, drivers can locate their vehicles using the information terminal at the parking lot elevator entrance or the shopping mall's mobile app and navigate to their cars using the provided directions.

[0007] However, current mainstream parking management systems have the following drawbacks:

[0008] 1) The mainstream market solution relies on terminal camera devices in shopping malls or parking lots for identification. The design logic involves pre-recording parking space numbers and matching the identified license plate with the fixed parking space number. This requires deploying a large number of cameras within the parking lot. Furthermore, after successful identification, users need to obtain their vehicle's location through a dedicated system.

[0009] 2) The existing parking space numbering management system has limited scope and functionality. After changing a parking space, the number needs to be re-entered, and it does not have an automatic recognition function. Summary of the Invention

[0010] In view of the above problems, embodiments of the present invention provide a parking space number identification method to solve the problems of complex structure, cumbersome operation, and limited scope of use and function of existing parking space number management systems.

[0011] According to one aspect of the present invention, a parking space number identification method is provided, the method comprising:

[0012] The vehicle camera acquisition module is used to acquire video information of the parking lot, and image data is obtained by extracting frames from the video information.

[0013] Preprocess the image data;

[0014] The preprocessed image data is detected and identified, and a single-stage deep learning model constructed from the OD detection model and the OCR end-to-end recognition model is used to detect and identify the parking space coding region and output the region detection data.

[0015] The area detection data is converted into ONNX model data, and the parking space number information is calculated based on the ONNX model data.

[0016] In one alternative approach, the step of preprocessing the image data includes:

[0017] Linearly project points in three-dimensional space onto a virtual unit sphere;

[0018] The point lens on the virtual unit sphere refracts the image onto the image plane to form an image.

[0019] In an alternative approach, the step of preprocessing the image data further includes:

[0020] Obtain a point P in the camera's 3D coordinate system C (x c ,y c ,z c );

[0021] Point P C Normalized to a point P on the surface of a predefined sphere S (x s ,y s ,z s );

[0022] Calculate point P S Longitude in the preset sphere and latitude θ;

[0023] The distorted latitude θ is obtained by applying distortion to the latitude θ using the KB distortion model. d ;

[0024] Point P is calculated based on the equidistant projection model. C The pixel coordinates (u, v) in the plane.

[0025] In one alternative approach, the step of using a single-stage deep learning model constructed from an OD detection model and an OCR end-to-end recognition model to detect and identify parking space coding areas and output area detection data includes:

[0026] By developing a deep learning model, a recognition framework is formed;

[0027] The aforementioned recognition framework is used to simultaneously perform character region detection and character recognition;

[0028] The initial localization area is obtained by using character recognition, and character region recognition is performed in the initial localization area to finally obtain and output the region detection data.

[0029] In one alternative approach, forming a recognition framework by developing a deep learning model includes the following sub-steps:

[0030] The preprocessed image data is then subjected to upsampling and downsampling to extract multi-level features, resulting in multiple feature maps of different sizes.

[0031] The feature maps of the same size selected from the feature maps of the above multiple feature maps of different sizes are added together to further fuse the features;

[0032] Each scale feature map is interpolated to the same size and then concatted to obtain the final segmentation map, forming a recognition framework.

[0033] In one alternative approach, the parking space number information includes: the coordinates of the vertex of the smallest bounding rectangle of the character region, the detected character content, and the target confidence level.

[0034] According to another aspect of the present invention, a parking space number identification device is provided, the device comprising:

[0035] The vehicle camera acquisition module acquires video information from the parking lot and extracts image data by frame extraction based on the video information.

[0036] The image preprocessing module is used to perform distortion correction preprocessing on the image data;

[0037] The parking space coding area detection and recognition module is used to detect and recognize parking space coding areas and output area detection data; and

[0038] The output module is used to convert the area detection data into ONNX model data and calculate the parking space number information based on the ONNX model data.

[0039] In one alternative embodiment, the parking space coding area detection and identification module includes:

[0040] A lightweight backbone network capable of performing depthwise separable convolutions and spatially separable convolutions;

[0041] The feature extraction module is used to extract multi-level features through upsampling and downsampling, forming multiple feature maps of different sizes; and

[0042] The feature fusion module is used to perform concat feature fusion on the feature map and output the final parking space coded area data.

[0043] According to another aspect of the present invention, a parking space number identification device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus;

[0044] The memory is used to store at least one executable instruction that causes the processor to perform the operation of the parking space number identification method as described above.

[0045] According to another aspect of the present invention, a computer-readable storage medium is provided, the storage medium storing at least one executable instruction, which, when executed on a parking space number identification device, causes the parking space number identification device to perform the operation of the parking space number identification method as described above.

[0046] This invention utilizes the vehicle's own camera to identify the vehicle's parking location, enabling a more direct acquisition of the parking space number. Furthermore, by combining the vehicle's high-definition camera with a deep learning model, it offers the following advantages:

[0047] 1) It can identify parking space numbers offline without relying on a cloud system, and is applicable to a wide range of parking scenarios, such as ground-level open-air parking lots, ground-level roadside parking spaces, underground parking lots, etc.

[0048] 2) Due to the flexibility of the vehicle's own mobility, the present invention can be applied to different weather conditions, such as sunny days, rainy days, daytime, evening, etc.

[0049] 3) This invention can recognize parking space numbers in various fonts, such as printed and handwritten fonts, through a deep learning model; and it can also recognize characters that are obscured, blurred, reflective, or incomplete.

[0050] 4) This invention is not affected by the position and direction of the parking space number in the field of vision, and is highly flexible.

[0051] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0052] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0053] Figure 1 A flowchart of a first embodiment of the parking space number identification method of the present invention is shown;

[0054] Figure 2 This invention illustrates a flowchart of the parking space number recognition method for detecting and recognizing the parking space coding area and outputting area detection data.

[0055] Figure 3 The flowchart illustrating the formation of the identification framework in the parking space number identification method of the present invention is shown.

[0056] Figure 4 A schematic diagram of the parking space number identification device provided by the present invention is shown;

[0057] Figure 5 A schematic diagram of an embodiment of the parking space number recognition device provided by the present invention is shown. Detailed Implementation

[0058] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0059] Figure 1 A flowchart of a first embodiment of the parking space number recognition method of the present invention is shown. This method is executed by a parking space number recognition device, which is primarily a vehicle equipped with a surround-view camera. Figure 1 As shown, the method includes the following steps:

[0060] Step 110: Use the vehicle camera acquisition module to acquire video information of the parking lot, and extract image data by frame extraction based on the video information.

[0061] Among these features, it can acquire video information in parking lots while vehicles are moving slowly, and perform frame extraction on the video information to obtain relatively clear raw image data.

[0062] Step 120: Preprocess image data.

[0063] The vehicle camera can be a surround-view camera, typically a fisheye lens. Utilizing the wide field of view of a fisheye lens, it can capture richer image information. Image preprocessing primarily involves anti-distortion correction of the distorted image acquired by the vehicle camera, ensuring all pixels conform to a planar effect in frontal space, thereby improving the recognition rate. Distortion is inherent in the imaging process for cameras. Because distortion causes objects to appear curved, fisheye images are unsuitable for direct application of deep learning models for object detection and recognition. Therefore, distortion correction is necessary. Distortion correction involves finding a remapping function to rearrange the pixels in the fisheye image onto the image plane, and different distortion correction schemes need to be applied to different scenes. For fisheye cameras, radial distortion is dominant, while tangential distortion can be ignored.

[0064] For example, planar information of an image can be formed through spatial coordinate projection. Different projection models can be used, such as pinhole projection, isometric projection, equal solid angle projection, orthographic projection, and stereoscopic projection. By using one or more combinations of these models for back projection, the mapping relationship between the coordinate information of the original image and the coordinate information projected onto the planar space can be obtained.

[0065] In some embodiments, an equidistant projection model can be used to linearly project points in three-dimensional space onto a virtual unit sphere. Points on the virtual unit sphere are then refracted by lenses onto an image plane to form an image.

[0066] Specifically, the steps include the following:

[0067] Obtain a point P in the camera's 3D coordinate system C (x c ,y c ,z c );

[0068] Point P C Normalize the projection onto a predefined spherical surface, forming point P. S (x s ,y s ,z s );

[0069] in ;

[0070] Calculate point PS Longitude in the preset sphere And latitude θ. This can be achieved using inverse trigonometric functions:

[0071]

[0072]

[0073] The distorted latitude θ is obtained by applying distortion to the latitude θ using the Kannala-Brandt distortion model. d .

[0074] Among them, the KB distortion model is a general polynomial approximation model for fisheye cameras proposed by Kannala-Brandt.

[0075] Where θ d It is an odd function of θ. Expanding these expressions using a Taylor series, we find that θ d Expressed in terms of odd-degree polynomials of θ, it can be specifically:

[0076]

[0077] Taking the first 5 terms, we have enough degrees of freedom to approximate various projection models well.

[0078]

[0079] Point P is calculated based on the equidistant projection model. C The pixel coordinates (u, v) in the plane.

[0080] Specifically:

[0081]

[0082]

[0083]

[0084] Among them, f x f y c x c y It's the intrinsic parameter matrix of the fisheye lens, used for scaling and translation between two two-dimensional planes. Scaling converts physical units to pixel units, and translation aligns the coordinate origin.

[0085] Step 130: Detect and recognize the preprocessed image data, and use a single-stage deep learning model constructed by the OD detection model and the OCR recognition model to detect and recognize the parking space coding area and output the area detection data.

[0086] The preprocessed image data is fed into an end-to-end recognition model consisting of a self-developed OD detection model and an OCR recognition model. At the same time, the mechanism of the deep learning model is used to detect and recognize the parking space coding area and output the area detection data.

[0087] The end-to-end model, also known as the single-stage model, integrates character region detection and character recognition into a single framework, compared to the existing two-stage OD detection model + CRNN recognition model solution in the technology center. This allows for single-stage detection and recognition, making it easier to train and deploy on edge chips.

[0088] In addition, the deep learning model is a pre-trained deep learning model that can be used to perform inference model inference results tensor. The specific process is as follows:

[0089] Test vehicles equipped with AVM surround-view cameras arrived at designated ground / underground parking lots and roadside parking spaces, and automatically collected video data.

[0090] Video data is extracted by frame extraction to obtain image data, which is then processed by LDC to obtain cylindrically corrected image data. The data is then labeled according to pre-defined rules to obtain the ground truth labels required by the model.

[0091] The corrected image data and the labeled ground truth values ​​are fed into the detection / recognition network to train the model.

[0092] Set the acceptance metrics as accuracy P, recall R, mean precision AP, and harmonic mean Hmean, and select the optimal model.

[0093] A more optimized recognition model can be obtained through the deep learning model described above.

[0094] Step 140: Convert the area detection data into ONNX model data, and calculate the parking space number information based on the ONNX model data.

[0095] ONNX is an open file format designed for machine learning, used to store trained models. It allows different deep learning frameworks (such as PyTorch and MXNet) to store model data in the same format. In short, ONNX is an intermediate representation format that facilitates the transfer of models across various mainstream deep learning frameworks. Converted ONNX model data is more conducive to data output, meeting different post-processing requirements.

[0096] For example, you can directly obtain the parking space number, or you can highlight the parking spaces in the video in the visualization video information.

[0097] In some embodiments, the parking space number information includes: the coordinates of the vertex of the smallest bounding rectangle of the character region, the detected character content, and the target confidence level.

[0098] In summary, this invention, by utilizing the vehicle's own camera to identify the vehicle's parking location, can more directly obtain the parking space number. Furthermore, by combining the vehicle's high-definition camera with a deep learning model, it offers the following advantages:

[0099] 1) It can identify parking space numbers offline without relying on a cloud system, and is applicable to a wide range of parking scenarios, such as ground-level open-air parking lots, ground-level roadside parking spaces, underground parking lots, etc.

[0100] 2) Due to the flexibility of the vehicle's own mobility, the present invention can be applied to different weather conditions, such as sunny days, rainy days, daytime, evening, etc.

[0101] 3) This invention can recognize parking space numbers in various fonts, such as printed and handwritten fonts, through a deep learning model; and it can also recognize characters that are obscured, blurred, reflective, or incomplete.

[0102] 4) This invention is not affected by the position and direction of the parking space number in the field of vision, and is highly flexible.

[0103] Figure 2 The flowchart illustrates the parking space number recognition method of the present invention, which involves detecting and recognizing the parking space coding area and outputting area detection data. This method is executed by a parking space number recognition device, primarily a vehicle equipped with a surround-view camera. Figure 2 As shown, based on implementation 1, this method further optimizes the steps of detecting and recognizing parking space coding regions and outputting region detection data using a single-stage deep learning model constructed from the OD detection model and OCR recognition model on the preprocessed image data. Specifically, it includes the following steps:

[0104] Step 210: Form a recognition framework by developing a deep learning model.

[0105] The deep learning model extracts images of regions where parking space characters might appear by using feature extraction and feature enhancement. The recognition framework built using a single-stage model architecture can find and accurately identify the image region containing the parking space number in a very short time. Furthermore, this algorithm architecture is relatively simple, has low computational load, and is suitable for implementation in vehicles and other devices.

[0106] Specific step 210 may include the following sub-steps, such as Figure 3 As shown, Figure 3The flowchart illustrating the formation of the recognition framework is shown below:

[0107] Step 310: The preprocessed image data is upsampled and downsampled to extract multi-level features, resulting in multiple feature maps of different sizes.

[0108] This process involves extracting multi-level features through upsampling and downsampling, and continuously fusing adjacent feature maps. To reduce the number of parameters, depthwise separable convolution and spatially separable convolution are employed. The final output will contain multiple feature maps of different sizes. These feature maps contain patches corresponding to the parking space numbers.

[0109] Step 320: Add the feature maps of the same size selected from the feature maps of the above multiple feature maps of different sizes to further fuse the features.

[0110] Since the feature maps extracted in step 310 have different sizes, in order to make the features in the feature maps more prominent, it is necessary to select feature maps of the same size from all the feature maps and add them together to make the features more prominent.

[0111] Step 330: Interpolate the feature maps of each scale to the same size and perform concat fusion to obtain the final segmentation map, forming a recognition framework.

[0112] In this process, some feature maps may be too small or misaligned. To address this, feature map interpolation can be used to ensure all feature maps have the same size, thus forming a unified recognition framework.

[0113] Step 220: Use the recognition framework to simultaneously perform character region detection and character recognition to obtain a preliminary localization region.

[0114] In step 210, feature extraction results in multiple recognition frameworks that may contain text. Character region detection primarily employs text detection methods, which can parse out the text region matrix, text kernel region matrix, and similarity matrix. Preferably, connected component analysis can be used to determine the positions of all possible text kernels. Further, a text recognition model is used to perform text detection on the text regions obtained from the text detection model. If text is found, the confidence level of the detected character regions can be further increased. This leads to the preliminary location of areas where parking space numbers may exist.

[0115] Step 230: Perform character region recognition in the initial localization area, and finally obtain and output the region detection data.

[0116] The initially located area may contain parking space numbers, but since the confidence level for their presence is already high based on previous calculations, identification is performed. Once it is confirmed that the identified data is a parking space number, the coordinates of the initially located area in the image can be output simultaneously.

[0117] In some embodiments, the format of the identified parking space is also verified. For example, the parking space number format of the current area is obtained by using information such as the current vehicle location and vehicle surround view image through a cloud database, and is compared with the identified parking space number. If they are the same, the parking space number is considered to have been successfully identified.

[0118] For example, if the vehicle's location information indicates that the parking space number in the current parking lot is in the format "BX-XXXX", and the identified parking space number is "B1-0157", then the identification data can be considered accurate. Another example is that all parking space numbers in the current vehicle surround view image information can be identified, and their formats can be validated. This ensures the accurate identification of the current parking space number.

[0119] Example 3:

[0120] Figure 4 A schematic diagram of an embodiment of the parking space number identification device of the present invention is shown. Figure 4 As shown, the device 400 includes: a vehicle camera acquisition module 410, an image preprocessing module 420, a parking space coding area detection and recognition module 430, and an output module 440.

[0121] The vehicle camera acquisition module 410 is used to acquire video information from the parking lot and extract image data by frame extraction based on the video information.

[0122] The image preprocessing module 420 is used to perform distortion correction preprocessing on the image data.

[0123] The parking space coding area detection and recognition module 430 is used to detect and recognize parking space coding areas and output area detection data.

[0124] The output module 440 is used to convert the area detection data into ONNX model data and calculate the parking space number information based on the ONNX model data.

[0125] In some embodiments, the parking space coding area detection and recognition module 410 includes: a lightweight backbone network, a feature extraction module, and a feature fusion module.

[0126] A lightweight backbone network capable of performing depthwise separable convolutions and spatially separable convolutions;

[0127] The feature extraction module is used to extract multi-level features through upsampling and downsampling, forming multiple feature maps of different sizes; and

[0128] The feature fusion module is used to perform concat feature fusion on the feature map and output the final parking space coded area data.

[0129] The parking space number recognition device disclosed in this embodiment

[0130] Example 4:

[0131] Figure 5 The diagram shows a structural schematic of an embodiment of the parking space number recognition device of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the parking space number recognition device.

[0132] As shown in Figure 5, the parking space number identification device may include: a processor 502, a communication interface 504, a memory 506, and a communication bus 508.

[0133] The processor 502, communication interface 504, and memory 506 communicate with each other via communication bus 508. Communication interface 504 is used to communicate with other network elements such as clients or other servers. The processor 502 executes program 510, specifically performing the relevant steps described in the embodiment of the parking space number identification method.

[0134] Specifically, program 510 may include program code, which includes computer-executable instructions.

[0135] Processor 502 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The parking space number identification device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0136] Memory 506 is used to store program 510. Memory 506 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0137] Specifically, program 510 can be called by processor 502 to cause the parking space number identification device to perform the following operations:

[0138] The vehicle camera acquisition module is used to acquire video information of the parking lot, and image data is obtained by extracting frames from the video information.

[0139] Preprocess the image data;

[0140] The preprocessed image data is detected and identified, and a single-stage deep learning model constructed from the OD detection model and the OCR recognition model is used to detect and identify the coded parking space region and output the region detection data.

[0141] The area detection data is converted into ONNX model data, and the parking space number information is calculated based on the ONNX model data.

[0142] Example 5:

[0143] This invention provides a computer-readable storage medium storing at least one executable instruction that, when executed on a parking space number identification device, causes the parking space number identification device to perform the parking space number identification method in any of the above method embodiments.

[0144] Specifically, the executable instructions can be used to cause the parking space number identification device / mechanism to perform the following operations:

[0145] The vehicle camera acquisition module is used to acquire video information of the parking lot, and image data is obtained by extracting frames from the video information.

[0146] Preprocess the image data;

[0147] The preprocessed image data is detected and identified, and a single-stage deep learning model constructed from the OD detection model and the OCR recognition model is used to detect and identify the coded parking space region and output the region detection data.

[0148] The area detection data is converted into ONNX model data, and the parking space number information is calculated based on the ONNX model data.

[0149] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Furthermore, the embodiments of this invention are not directed to any particular programming language.

[0150] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. Similarly, for the sake of brevity and to aid in understanding one or more aspects of the invention, in the description of exemplary embodiments of the invention above, various features of the embodiments are sometimes grouped together in a single embodiment, figure, or description thereof. The claims, which follow the detailed description, are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of the invention.

[0151] Those skilled in the art will understand that the modules in the device of the embodiment can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiment can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components, except that at least some of such features and / or processes or units are mutually exclusive.

[0152] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

Claims

1. A method for identifying parking space numbers, characterized in that, The method includes: The vehicle camera acquisition module acquires video information of the parking lot, and extracts fisheye image data by frame extraction based on the video information; The fisheye image data is subjected to distortion correction preprocessing, wherein the preprocessing includes: using the KB distortion model to map points in the camera's three-dimensional coordinate system to the image plane to generate a distortion-free image for recognition model; The preprocessed image data is detected and identified. A single-stage deep learning model, constructed from an OD detection model and an OCR end-to-end recognition model, is used to detect and identify the coded parking space region and output the region detection data. The single-stage end-to-end deep learning model uses a lightweight backbone network composed of depthwise separable convolutions and spatially separable convolutions for feature extraction. The area detection data is converted into ONNX model data, and the parking space number information is calculated based on the ONNX model data to achieve offline recognition.

2. The parking space number identification method according to claim 1, characterized in that, The step of preprocessing the image data includes: Linearly project points in three-dimensional space onto a virtual unit sphere; The point lens on the virtual unit sphere refracts the image onto the image plane to form an image.

3. The parking space number identification method according to claim 1, characterized in that, The step of preprocessing the image data includes: Obtain a point P in the camera's 3D coordinate system C (x c ,y c ,z c ); Point P C Normalized to a point P on the surface of a predefined sphere S (x s ,y s ,z s ); Calculate point P S Longitude in the preset sphere and latitude θ; The distorted latitude θ is obtained by applying distortion to the latitude θ using the KB distortion model. d ; Point P is calculated based on the equidistant projection model. C The pixel coordinates (u, v) in the plane.

4. The parking space number identification method according to claim 1, characterized in that, The steps of using a single-stage deep learning model constructed from an OD detection model and an OCR end-to-end recognition model to detect and identify parking space coding areas and output area detection data include: By developing a deep learning model, a recognition framework is formed; The aforementioned recognition framework is used to simultaneously perform character region detection and character recognition to obtain a preliminary localization region. Character region recognition is performed in the initial localization area, and finally the region detection data is obtained and output.

5. The parking space number identification method according to claim 4, characterized in that, The process of forming a recognition framework by developing a deep learning model includes the following sub-steps: The preprocessed image data is then subjected to upsampling and downsampling to extract multi-level features, resulting in multiple feature maps of different sizes. The feature maps of the same size selected from the feature maps of the above multiple feature maps of different sizes are added together to further fuse the features; Each scale feature map is interpolated to the same size and then concatted to obtain the final segmentation map, forming a recognition framework.

6. The parking space number identification method according to claim 1, characterized in that, The parking space number information includes: the coordinates of the vertex of the smallest bounding rectangle of the character area, the detected character content, and the target confidence level.

7. A parking space number identification device, characterized in that, The device includes: The vehicle camera acquisition module acquires video information from the parking lot and extracts fisheye image data by frame extraction based on the video information. An image preprocessing module is used to perform distortion correction preprocessing on the fisheye image data. The preprocessing includes: using the KB distortion model to map points in the camera's three-dimensional coordinate system to the image plane to generate a distortion-free image for model recognition. A parking space coding region detection and recognition module is used to detect and recognize the preprocessed image data. It utilizes a single-stage deep learning model constructed from an OD detection model and an OCR end-to-end recognition model to perform parking space coding region detection and recognition and output region detection data. The single-stage end-to-end deep learning model employs a lightweight backbone network composed of depthwise separable convolutions and spatially separable convolutions for feature extraction. The output module is used to convert the area detection data into ONNX model data and calculate the parking space number information based on the ONNX model data to achieve offline recognition.

8. The parking space number identification device according to claim 7, characterized in that, The parking space coding area detection and recognition module includes: A lightweight backbone network capable of performing depthwise separable convolutions and spatially separable convolutions; The feature extraction module is used to extract multi-level features through upsampling and downsampling, forming multiple feature maps of different sizes; and The feature fusion module is used to perform concat feature fusion on the feature map and output the final parking space coded area data.

9. A parking space number identification device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation of the parking space number identification method as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one executable instruction, which, when executed on the parking space number identification device, causes the parking space number identification device to perform the operation of the parking space number identification method as described in any one of claims 1-6.

Images