An unattended parking lot AI auxiliary guarding method based on a cloud brain map large model

Abstract

The application relates to the technical field of image recognition, and discloses an unmanned parking lot AI auxiliary guarding method based on a cloud brain map large model, which comprises the following steps: taking images when vehicles enter and exit the parking lot and performing fuzzy processing, and automatically identifying and matching the vehicles by using the similarity between the images. The application can identify the vehicles entering and exiting the parking lot and make release or prohibition decisions under different illumination, angles and environmental changes. The application can realize automatic management under high privacy protection, improve the efficiency, accuracy and safety of parking lot management, and simultaneously reduce the misjudgment rate and operation cost of a traditional identity recognition system.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to an AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model. Background Technology

[0002] With the acceleration of urbanization, the number and usage frequency of parking lots are constantly increasing, significantly raising the demand for intelligent vehicle access management. Traditional parking lot gate systems generally rely on license plate recognition technology, using license plate number detection and character recognition to automatically record vehicle entry and exit and manage fees. However, in practical use, these character recognition-based systems have the following problems: 1. Environmental factors such as damaged, obscured, or reflective mud on license plates can significantly reduce recognition rates; 2. Uneven lighting at night, strong light reflection, or rain and snow weather can result in insufficient image clarity from the camera, further affecting the license plate detection performance. 3. With increasingly stringent privacy protection regulations, license plate numbers are considered sensitive personal information, and their collection, storage, and transmission must meet higher security and compliance requirements.

[0003] To address the aforementioned issues, vehicle recognition solutions that do not rely on license plate numbers have gradually emerged in recent years. These solutions analyze vehicle appearance elements such as overall shape, outline, color distribution, lighting structure, and grille layout to re-identify or match vehicles. However, due to the limited differences in vehicle appearance, vehicles of the same brand and color are extremely similar in appearance, making it easy for traditional visual recognition methods to misidentify different vehicles.

[0004] In addition, the structure and operation of parking lot gates also introduce new technical challenges: First, the gate's operation is event-triggered data acquisition; the system only captures an image the instant a vehicle triggers the gate to open, lacking continuous frame information. Therefore, highly reliable matching and judgment are required for single-frame images. Second, the installation angle, lighting conditions, and reflection environment of cameras at parking lot entrances and exits typically differ, leading to significant differences in images of the same vehicle at the entrance and exit. Third, to meet privacy protection requirements, images need to be blurred or de-identified after acquisition, further weakening directly usable salient visual features in the images.

[0005] In existing solutions, some systems attempt to match vehicle entry and exit using low-resolution or blurry images. However, most methods rely solely on simple color, edge, or area comparisons, lacking stable geometric constraints and illumination reflection modeling. This makes it difficult to achieve highly reliable recognition in parking lots, where there is a fixed viewing angle, short-term parking, and strong glare. On the other hand, parking management systems need to ensure that vehicle entry and exit records match. Any matching error could lead to gate erroneous access, billing errors, or security risks. Summary of the Invention

[0006] This invention provides an AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model, which solves the technical problems mentioned in the background.

[0007] A cloud-based mind map model-based AI-assisted monitoring method for unattended parking lots includes: S1. Establish the mapping relationship between the gate scene and the image plane, convert the acquired entrance image into a normalized image with the intersection of the parking lines as the positioning reference, and perform detexturing and blurring processing. S2, establish a coordinate system with the positioning reference as the pole, scan and aggregate the radial photometric distribution of the vehicle area along the angular direction for the normalized image to form an anti-blurring feature, and perform frequency domain transformation on the anti-blurring feature to generate a vehicle entrance angular feature spectrum characterizing the vehicle front structure. S3, constructing a dynamic database based on the vehicle entrance angular feature spectrum; S4, perform AND with S1 and S2 on the acquired exit image to obtain the vehicle exit angular feature spectrum; S5, calculate the similarity distance between the vehicle exit angular feature spectrum and the database record to obtain the matching confidence level, and generate a gate release or prohibition command based on the confidence level and confidence threshold.

[0008] The beneficial effects of this invention include: It enables vehicle entry and exit matching and automatic release without recognizing sensitive information such as license plate numbers, significantly improving the reliability and security of parking lot gate management. By performing structured analysis on blurred images from a fixed camera perspective, this invention preserves key appearance differences of vehicles under complex conditions such as lighting, reflection, and angle changes, ensuring stable image comparison capabilities even with high privacy protection requirements. This invention can automatically determine vehicle correspondence based on a single blurred image frame, reducing system maintenance costs and the frequency of manual intervention, ensuring the real-time and deterministic nature of gate release decisions, and is applicable to various closed or semi-closed intelligent parking management scenarios. Attached Figure Description

[0009] Figure 1 This is a flowchart of an AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model, according to the present invention. Figure 2 This is a comparison diagram of the implementation effects of the present invention. Detailed Implementation

[0010] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0011] like Figure 1 As shown, an AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model includes: S1. Establish the mapping relationship between the gate scene and the image plane, convert the acquired entrance image into a normalized image with the intersection of the parking lines as the positioning reference, and perform detexturing and blurring processing. S2, establish a coordinate system with the positioning reference as the pole, scan and aggregate the radial photometric distribution of the vehicle area along the angular direction for the normalized image to form an anti-blurring feature, and perform frequency domain transformation on the anti-blurring feature to generate a vehicle entrance angular feature spectrum characterizing the vehicle front structure. S3, constructing a dynamic database based on the vehicle entrance angular feature spectrum; S4, perform AND with S1 and S2 on the acquired exit image to obtain the vehicle exit angular feature spectrum; S5, calculate the similarity distance between the vehicle exit angular feature spectrum and the database record to obtain the matching confidence level, and generate a gate release or prohibition command based on the confidence level and confidence threshold.

[0012] In one embodiment of the present invention, establishing the mapping relationship between the gate scene and the image plane includes: The camera's intrinsic parameter matrix is ​​obtained through calibration. and the rotation matrix relative to the ground plane Translation vector Calculate the homography matrix from the ground plane to the image plane. ; Using formula Pixel coordinates in the original image Mapped to pixel coordinates in a normalized image ; in, The homography matrix represents homogeneous coordinate normalization. This is used to forcibly map the intersection of the stop line and the lane center line in the spatial plane to a pre-defined fixed anchor point in the normalized image. Its coordinates are .

[0013] Camera internal parameters are a set of parameters describing the camera's optical characteristics and imaging geometry. They can be obtained using camera calibration techniques such as the Zhang Zhengyou calibration method.

[0014] Camera external attitude parameters describe the camera's position and attitude relative to the ground coordinate system. The rotation matrix represents the camera's rotation state, and the translation vector represents the camera's translation state. These parameters can be obtained through camera calibration combined with measurements from ground reference points.

[0015] The ground plane information where the stop line is located refers to the relevant data that defines the ground spatial plane where the stop line is situated. This information can be obtained by measuring the spatial coordinates of multiple reference points on the ground or by using plane fitting techniques.

[0016] The intrinsic parameter matrix is ​​a matrix composed of the camera's internal parameters, used to establish the mapping relationship between the camera's pixel coordinates and the imaging plane coordinates.

[0017] The homography transformation matrix is ​​a matrix that describes the projection transformation relationship between a spatial plane and an image plane, and is used to realize the coordinate transformation of points between the two planes.

[0018] The actual intersection of the stop line and the lane centerline in a spatial plane is the point where the actual extension of the stop line on the ground intersects with the actual extension of the lane centerline; it serves as the spatial prototype of the positioning reference. This can be obtained through manual measurement and marking on the ground or by using a semantic segmentation model to detect the two lines and then calculating the intersection.

[0019] The preset fixed pixel coordinates in the image coordinate system are fixed points, or fixed anchor points, pre-defined in the image coordinate system for aligning spatial intersections. Preferably, they are the pixel coordinates corresponding to 50% of the image width and 85% of the image height, ensuring that this position guarantees that most of the subsequent fan-shaped area falls within the image, adapting to different vehicle front-end ranges.

[0020] The original image pixel coordinates are the position coordinates of each pixel in the original input image in the image coordinate system.

[0021] Normalized image pixel coordinates are the position coordinates in the normalized image after the original image pixel coordinates have been mapped by the homography transformation matrix.

[0022] The logic of associating camera intrinsic and extrinsic parameters with ground plane information is as follows: Camera intrinsic parameters determine the internal geometric relationships of the image, extrinsic parameters determine the camera's position and attitude relative to the ground, and ground plane information defines the spatial equation of the plane containing the stop line. When these three are combined, points on the spatial plane are transformed into points on the image plane using projection geometry principles, thereby solving for the homography transformation matrix. For example, given the spatial coordinates of four non-collinear reference points on the ground and their corresponding image pixel coordinates, the homography matrix can be solved using the direct linear transformation method, combining the intrinsic and extrinsic parameters.

[0023] The technical design logic of the positioning reference includes: selecting the actual intersection of the stop line and the lane center line as the positioning reference because this point has a fixed position in the parking lot scenario and is strongly correlated with the vehicle entry and exit paths, ensuring that the imaging of different vehicles at the entrance has a unified reference benchmark. Mapping it to fixed pixel coordinates can eliminate image offset caused by differences in the parking positions of different vehicles, allowing subsequent feature extraction to be based on a unified coordinate system and improving matching accuracy.

[0024] The parameter adaptation rules for the homography transformation matrix include: In a turnstile scenario, the homography transformation matrix needs to adapt to the camera's installation height and angle, typically requiring coverage of the main areas for vehicle entry and exit. For example, when the camera's installation height is 3 meters, the solution to the homography matrix must ensure that points within 5 meters of the stop line on the ground can be accurately mapped onto the image plane, avoiding excessive distortion in edge areas.

[0025] Ground planar information includes: Ground planar information can be defined by a plane equation, which is a linear combination of three coefficients and spatial coordinates equal to a constant. The three coefficients constitute the plane normal vector, and the constant term is the distance from the plane to the origin. To obtain this information, four non-collinear points can be selected on the ground where the stop line is located. The spatial coordinates of each point are measured, and the coefficients of the plane equation are calculated using a plane fitting algorithm, thus obtaining the ground planar information.

[0026] The actual intersection points can be obtained in several ways: when obtained manually, the coordinates of the intersection of the ground stop line and the lane center line can be measured with a tape measure and recorded as spatial coordinates; when detected automatically, the stop line and lane center line in the image can be extracted using a line segment detection algorithm, converted into ground spatial coordinates, and the intersection point can be calculated. If the detection confidence is less than 80%, the system will automatically switch to manual calibration results.

[0027] The selection principles for fixed anchor points include: in addition to prioritizing values, the anchor point coordinates should be adjusted proportionally when the image resolution changes. For example, if the image width changes from 1920 pixels to 1280 pixels, the horizontal coordinate of the anchor point should be adjusted from 960 pixels to 640 pixels. When selecting anchor points, avoid placing them near the image edges to ensure that the subsequent polar coordinate sector does not exceed the image range.

[0028] The calculation of the homography transformation matrix includes: first, obtaining the camera's intrinsic and extrinsic parameters, rotation matrix, and translation vector through calibration; then, deriving the projection formula from spatial plane points to image plane points based on the ground plane equation; next, selecting four non-collinear reference points on the ground and recording their spatial coordinates and corresponding image pixel coordinates; finally, substituting these corresponding points into the direct linear transformation formula to solve for the eight unknown parameters of the homography transformation matrix.

[0029] Homogeneous coordinate normalization is the process of converting homogeneous coordinates obtained from homography transformation into two-dimensional image coordinates. Specifically, it involves dividing the first two components of the transformed homogeneous coordinates by the third component; the results are the normalized image pixel coordinates. For example, if the homogeneous coordinates consist of three values, dividing the first two values ​​by the third value yields the corresponding two pixel coordinates.

[0030] In one embodiment of the present invention, the acquired entrance image is converted to a normalized image based on the intersection of parking lines, and then de-textured and blurred to obtain a normalized image, including: For normalized images A blurred image is obtained by performing a two-dimensional Gaussian convolution operation. The calculation formula is: ; in, For integration variables, The Gaussian kernel function has the following formula: ; In the formula, The pre-defined Gaussian kernel standard deviation is used to control the degree of privacy protection and contour preservation. Pi; Represented by natural constant An exponential function with base 0; and is the distance variable relative to the center of the kernel.

[0031] The Gaussian kernel standard deviation is the core parameter controlling the intensity of Gaussian blur, used to adjust the balance between privacy protection and vehicle outline preservation. It is preferably set to 2 to 6 in 1080p image scenes, or at a ratio of 0.002 to 0.006 of the image width, to ensure that this range effectively filters out high-frequency textures such as license plate characters without destroying the overall vehicle outline features.

[0032] A normalized image is an image obtained by perspective correction of the original entry image through a homography transformation matrix, and it has a unified coordinate reference.

[0033] A blurred image is an image generated by performing a two-dimensional Gaussian convolution operation on a normalized image, which has filtered out high-frequency texture information.

[0034] The two-dimensional Gaussian kernel function is a mathematical function used to achieve image blurring, and its shape is determined by the standard deviation of the Gaussian kernel.

[0035] The integral variable is used in Gaussian convolution operations to iterate through image pixels and calculate the blurred brightness value of each target pixel.

[0036] Pi is a mathematical constant with a value of approximately 3.1416, used in the calculation of the two-dimensional Gaussian kernel function.

[0037] The natural constant is a mathematical constant with a value of approximately 2.7183, used in the exponential operation of the two-dimensional Gaussian kernel function.

[0038] The matching logic between the Gaussian kernel standard deviation and high-frequency texture filtering for license plates includes: license plate character details belong to high-frequency textures, with pixel stroke widths typically ranging from 2 to 5 pixels. The magnitude of the Gaussian kernel standard deviation directly determines the degree of blurring; a smaller value can only weaken slight textures, while a larger value will destroy the vehicle outline. The preferred value range corresponds precisely to the license plate character stroke width, accurately filtering out character details while preserving low-frequency features such as the vehicle's front outline and lighting layout, achieving a balance between privacy protection and recognition requirements.

[0039] The selection of the order of perspective transformation followed by blurring is based on the following: The purpose of perspective transformation is to correct image distortion and establish a unified coordinate reference. If blurring is performed first, the distorted areas in the image will be mixed and blurred with the normal areas, making it difficult to retain the complete contour after subsequent correction. Correcting first and then blurring ensures that the blurring is based on a distortion-free standardized image, so that the vehicle contour features remain consistent after blurring, providing a reliable foundation for subsequent feature extraction.

[0040] The specific rules for determining the Gaussian kernel standard deviation include: in addition to the preferred values, when the image resolution is 720p, the value can range from 1.5 to 4.5; when the camera is installed at a low height and the shooting distance is close, the value can be biased towards the lower limit of this range to avoid excessive blurring. The calibration method involves acquiring 10 sets of entrance images containing different license plates, testing the character indistinguishability rate under different σ values, and selecting values ​​where the character indistinguishability rate reaches 90% or higher while the outline remains intact.

[0041] The rules for determining the fuzzy kernel size include: the fuzzy kernel size is calculated by multiplying 2 by 3 times the standard deviation of the Gaussian kernel, rounded up to the nearest integer, and then adding 1, ensuring coverage of the main energy region of the Gaussian function. For example, when the standard deviation of the Gaussian kernel is 4, 3 times the standard deviation is 12, rounded up to 12, and the fuzzy kernel size is 25.

[0042] The criteria for achieving the blur intensity include: the quantitative criterion is that the edge energy or gradient amplitude of the license plate area decreases to 20% to 40% of that before blurring; the qualitative criterion is that in 10 manually sampled blurred images, the license plate characters are indistinguishable while the vehicle outline is clearly discernible.

[0043] Maintaining a fixed Gaussian kernel standard deviation means keeping the standard deviation constant for all cameras pointing at the same parking lot after deployment. If the camera is replaced, the installation height is adjusted, or the image resolution is modified, the Gaussian kernel standard deviation must be recalibrated.

[0044] The specific implementation of Gaussian convolution includes: setting the convolution stride to 1 pixel to ensure that the image size is not reduced; using mirror filling for boundary padding to avoid black edges or information loss in the image edge areas due to convolution operations; and implementing convolution operations through the 2D convolution interface in the image processing library without the need for custom complex calculation logic.

[0045] In one embodiment of the present invention, establishing a coordinate system with the positioning reference as the pole includes: With the fixed anchor point Establish a system with the origin of polar coordinates as the reference point. To normalized rectangular coordinates The transformation relationship is expressed by the formula: ; ; in, This represents the radial distance extending from the anchor point along the vehicle direction, and its value range is... It can cover the height of the front of the vehicle; This represents the scanning angle covering the width of the vehicle's front end, and its value range is... ; It is a cosine function. It is a sine function.

[0046] The angular range is the angular interval covering the width of the vehicle's front end in the polar coordinate system, defined by the minimum and maximum angles. It is preferably between -30 degrees and +30 degrees to ensure that this range covers the front end width of most small and medium-sized vehicles while avoiding exceeding the image boundaries.

[0047] The radial distance range is the distance interval extending from the stop line to the top of the vehicle in polar coordinates, defined by zero and the maximum radial distance. It is preferably 0.35 to 0.55 times the normalized image height to ensure that this range fully covers the structure of the vehicle's front end from bottom to top, accommodating vehicles of different heights.

[0048] Polar coordinates is a coordinate system that uses a positioning reference as the origin and describes the position of a pixel using radial distance and angular direction.

[0049] The cosine function is a mathematical function used for converting between polar and rectangular coordinates, and for calculating the coordinate components in the horizontal direction.

[0050] The sine function is a mathematical function used for converting between polar and rectangular coordinates, and for calculating the coordinate components in the vertical direction.

[0051] The selection of the polar coordinate origin includes choosing a fixed anchor point corresponding to the intersection of parking lines as the polar coordinate origin. This is because the position of this point is fixed in the parking lot scene and its relative position to the vehicle's entry and exit is stable. Regardless of slight deviations in the vehicle's parking position, using this point as the origin ensures that the vehicle's front structure maintains a relatively consistent distribution in the polar coordinate system, providing a unified benchmark for subsequent feature extraction. For example, when different vehicles are parked at the stop line, the relative position of the front of the vehicle to the origin is relatively small, making the polar coordinate distribution of the front structure comparable after conversion.

[0052] The matching of angle and radial ranges with the vehicle's front end includes: The angle range is set based on the width of the vehicle's front end. The width of a typical small car's front end is approximately 1.6 to 1.8 meters, corresponding to a polar coordinate angle of approximately 30 degrees from the camera's shooting angle. Therefore, the range from -30 degrees to +30 degrees is fully covered. The radial range is based on the height of the vehicle's front end. The height of a small car's front end is approximately 1.2 to 1.5 meters, corresponding to 0.35 to 0.55 times the normalized image height, ensuring that the structure from the stop line to the top of the front of the car is included in the polar coordinate system.

[0053] The design of the coordinate transformation and alignment of the fan-shaped area includes converting Cartesian coordinates to polar coordinates, which maps the rectangular area of ​​the vehicle's front face into a fan-shaped area. This allows the radial structure of the vehicle's front face (such as the contour changes from the bottom to the top of the front) and the angular structure (such as the lighting layout from left to right) to be presented separately. This alignment method can highlight the inherent structural features of the vehicle's front face, reduce feature offset caused by slight changes in the vehicle's parking angle, and improve feature stability.

[0054] The zero-angle direction includes the direction along the lane centerline pointing in the direction of vehicle travel, consistent with the extension direction of the lane centerline in the normalized image. Angle values ​​increase counter-clockwise, meaning the right side of zero angle is positive and the left side is negative, ensuring that the angle range division conforms to the conventional logic of image processing.

[0055] The specific methods for determining the angle and radial ranges include: For the angle range, in addition to the preferred values, if the parking lot mainly contains large vehicles, it can be adjusted to -45 degrees to +45 degrees; the radial range can be determined through calibration. A standard vehicle model is parked at the stop line, and the radial distance of the top of the vehicle's hood in the normalized image is measured. This distance is used as the maximum radial distance. For example, after calibration, the maximum radial distance for a standard vehicle is 0.4 times the image height, and this value is used for all subsequent vehicles.

[0056] The precision control rules for coordinate transformation include: during coordinate transformation, bilinear interpolation is used to calculate the rectangular pixel brightness value corresponding to the polar coordinate point to ensure that no details are lost in the transformed image. During interpolation calculation, the brightness values ​​of the four pixels surrounding the target point are weighted and averaged. The weights are determined based on the distance between the target point and each pixel; the closer the distance, the greater the weight.

[0057] The pixel-level determination rules for the boundary of the sector region include: if the transformed polar coordinate point exceeds the pixel range of the normalized image, a mirror filling method is used, that is, the brightness value of the pixel at the edge of the image is taken as the filling value. For example, when the radial distance of the polar coordinate point exceeds the maximum radial distance, the pixel brightness value corresponding to the maximum radial distance is taken; when the angle exceeds the set range, the pixel brightness value corresponding to the angle boundary is taken.

[0058] In one embodiment of the present invention, scanning and aggregating the radial photometric distribution of a vehicle region along an angular direction in a normalized image to form an anti-blurring feature includes: Set within the angular range of the polar coordinate system Discrete sampling angle and setting within the radial range Discrete sampling radius ,in , ; For each sampling angle Calculate in this direction First radial moment The calculation formula is: ; in, To normalize the image at polar coordinates The pixel brightness value at that location; Represents radius of Power; Radial sampling interval; The order of the moments is predetermined, and its value is [value missing]. to ; Using a pre-defined fuzzy invariant mapping function Radial moments of different orders Combining into fuzzy parameters Robust radial description vector .

[0059] The number of discrete sampling angles is a set number of discrete angles within the polar coordinate system angle range, used to cover features in the width direction of the vehicle's front end. Preferably, it is 64, 128, or 256, to satisfy the requirement that this range is a power of 2, adapting to the fast calculation of the discrete Fourier transform, and balancing recognition accuracy and computational efficiency.

[0060] The number of discrete sampling radii is the number of discrete radii set within the radial range of the polar coordinate system, used to cover features in the height direction of the vehicle's front end. Preferably, it is the calculated result of dividing the maximum radial distance by the radial sampling interval to ensure that no radial feature sampling is missed, adapting to the front end structures of vehicles of different heights.

[0061] The radial sampling interval is the distance between two adjacent discrete sampling radii, used to control the fineness of radial sampling. It is preferably 1 pixel or 2 pixels to meet the requirements of high precision (1 pixel) and lightweight scenarios (2 pixels), avoiding excessive computation due to overly dense sampling or feature loss due to overly sparse sampling.

[0062] The order of the moment is the power of the radius in the radial moment calculation, used to characterize the radial luminous distribution features at different levels. It is preferably 0 to 3 or 0 to 4, as this range can effectively describe the brightness distribution pattern of the vehicle's front face. If the order is too low, the discrimination power is insufficient, and if it is too high, it is easily affected by noise interference.

[0063] The highest order of a moment is the maximum value of its order, used to limit the calculation range of the radial moment. It is preferably 3 or 4 to match the range of moment orders, ensuring a moderate feature dimension while balancing discriminability and stability.

[0064] The p-order radial moment is a feature quantity obtained by accumulating the pixel brightness value by power-law of radius at each discrete sampling angle, and is used to characterize the radial luminance distribution characteristics in that angular direction.

[0065] Pixel brightness values ​​are the brightness data of a normalized image at a specific polar coordinate point, used to construct a radial luminance distribution sequence. They can be obtained by querying the pixel after polar-to-Cartesian coordinate conversion; if no pixel is found at the corresponding location, they are calculated using bilinear interpolation.

[0066] Fuzzy invariant mapping functions combine radial moments of different orders into anti-fuzzy features, used to eliminate the impact of fuzzing on features. Ideally, they should be normalized moment vectors or ratio invariants, as both forms can reduce feature fluctuations caused by changes in fuzziness intensity, adapting to fuzzing scenarios requiring privacy protection.

[0067] The radial description vector is a vector obtained after processing by the fuzzy invariant mapping function, used to characterize the anti-fuzziness features under a single discrete sampling angle.

[0068] Feature extraction through angular scanning and radial luminosity aggregation includes: scanning sequentially along discrete sampling angles in a polar coordinate system, extracting pixel brightness sequences in the radial direction at each angle, and converting these sequences into radial moment features through weighted accumulation. This method decomposes the two-dimensional structure of a vehicle's front end into two one-dimensional features: angle and radial, highlighting structural differences in different directions. For example, the radial luminosity distribution of a small car's headlights at a specific angle exhibits a unique pattern, and this logic can accurately capture this feature.

[0069] Applications of fuzzy invariant moment theory include: blurring smooths the brightness distribution of an image, but the proportional relationships of radial moments of different orders or the normalized values ​​remain stable. Fuzzy invariant mapping functions eliminate feature shifts caused by variations in blur intensity by combining or normalizing the radial moments. For example, after different degrees of blurring, the normalized radial moment vectors of the same vehicle show relatively small differences, ensuring the comparability of entrance and exit features.

[0070] The radial moment weighted accumulation design includes: weighting the accumulation by using the radius raised to the power of p as the weight, so that the brightness value at different radial positions contributes differently to the feature. Positions closer to the origin of the polar coordinates have lower weights, while positions farther from the origin have higher weights as the order increases, which can highlight the structural features of the upper and middle parts of the vehicle's front, such as the air intake grille and the layout of the lights. These features vary more significantly between different vehicles.

[0071] The constraints on the number of discrete sampling angles include: 64 when the parking lot mainly consists of small cars and high real-time requirements are needed; 128 or 256 when higher recognition accuracy is required or the vehicle types are complex. If computational resources are limited, 64 can be selected with a radial sampling interval of 2 pixels to balance efficiency and accuracy.

[0072] The selection criteria for the order and the highest order of the moment include: in practical applications, 20 sets of images of different vehicles can be collected for testing. When the highest order is 3, the feature discrimination can meet the needs of most scenarios. If there are a large number of vehicles of the same brand and color, it can be increased to 4 to enhance the discrimination.

[0073] The specific forms of fuzzy invariant mapping functions include: Form 1 is a normalized moment vector, which divides each order of radial moment by the zeroth order radial moment plus the minimum value, and the minimum value is taken as 1e-6 or 1e-8 to avoid division by zero; Form 2 is a ratio invariant, which combines the three adjacent orders of radial moments by dividing the product of the previous order and the next order by the square of the middle order, which can be further logarithmized to improve stability.

[0074] Radial moment normalization methods include: for each discrete sampling angle, first calculating the mean and standard deviation of the brightness values ​​of all pixels at that angle, then subtracting the mean from the brightness value and dividing by the standard deviation; or directly dividing each order radial moment by the zeroth order radial moment plus the minimum value to weaken the influence of illumination changes on features.

[0075] The rules for extracting pixel brightness values ​​include: after converting polar coordinates to rectangular coordinates, if the corresponding position is an integer pixel, the brightness value of that pixel is directly taken; if it is a non-integer pixel, bilinear interpolation is used to calculate the brightness value of the surrounding four integer pixels by distance-weighted average.

[0076] The aggregation rules for the radial photometric distribution sequence include: each discrete sampling angle corresponds to a radial description vector, and the radial description vectors of all angles are arranged in the order of the sampling angles to form a one-dimensional sequence that varies with the angle. This sequence is the input signal for subsequent frequency domain transformation.

[0077] In one embodiment of the present invention, frequency domain transformation is performed on the anti-fuzzy features to generate a vehicle entrance angular feature spectrum characterizing the vehicle's front-end structure, including: Calculate the radial description vector The L2 norm is used as the scalar energy value of that angle. ; For discrete angle sequences Perform a Discrete Fourier Transform (DFT) to obtain the angular spectral coefficients. The calculation formula is: ; in, The imaginary unit (satisfying) ); The frequency order has a range of values. to ; This represents the total number of sampling angles; Pi; Calculate the complex spectral coefficients modulus The vehicle entrance angle feature spectrum vector is constructed from the modulus sequence arranged from low frequency to high frequency. : .

[0078] The highest frequency order is the upper limit of the frequency components retained in the frequency domain transformation, used to limit the dimension of the feature spectrum. It is preferably 16 or 32, so that this range can retain the key low-frequency components characterizing the front structure of the vehicle, balance feature discrimination and computational efficiency, and avoid high-frequency noise interference.

[0079] The scalar energy value is the result of the L2 norm calculation of the radial description vector at each discrete sampling angle, used to transform multidimensional features into a one-dimensional angular domain signal.

[0080] The 2-norm is a mathematical operator used to calculate the magnitude of a vector, and is used to quantify the energy of a radially describing vector.

[0081] Discrete Fourier transform coefficients are complex coefficients obtained by mapping an angle-domain signal to the frequency domain, containing amplitude and phase information of the frequency components.

[0082] The imaginary unit is a mathematical constant used to construct complex number operations in the discrete Fourier transform; its squared value is negative one.

[0083] The frequency order is the sequence number of the frequency component, used to distinguish signal components of different frequencies, and it increases from 0.

[0084] The amplitude of a frequency component is the magnitude of the discrete Fourier transform coefficients, used to characterize the intensity of the corresponding frequency component, but does not contain phase information.

[0085] The entrance angular feature spectrum vector is a vector composed of different frequency component amplitudes in sequence, used to characterize the angular structural features of the vehicle's front face.

[0086] Applications of the L2 norm in converting multidimensional vectors to one-dimensional signals include: Radial description vectors are multidimensional data, and direct frequency domain transformation increases computational complexity and makes them susceptible to noise. The L2 norm can compress multidimensional vectors into a single scalar, preserving the overall energy characteristics of the vector while simplifying the signal dimension. For example, the radial description vectors of different vehicles at the same sampling angle may differ, and their corresponding L2 norms (scalar energy values) will also exhibit stable differences, providing effective input for subsequent frequency domain analysis.

[0087] The parameter adaptation for the Discrete Fourier Transform (DFT) includes: the effectiveness of the DFT is closely related to the length of the input signal (i.e., the number of discrete sampling angles). Preferably, powers of 2 are used as the number of sampling angles to adapt to the Fast Fourier Transform (FFT) algorithm and improve computational efficiency. The highest frequency order is set to 16 or 32 because the key features of the vehicle's front structure are concentrated in the low-frequency region, which covers the main structural information while discarding high-frequency noise.

[0088] The technical choice to forgo phase preservation and instead retain amplitude information includes the following: phase information is sensitive to slight deviations in vehicle parking angle and lighting fluctuations during image acquisition, which can easily lead to feature instability. Amplitude information, on the other hand, reflects the signal intensity distribution and is robust to these disturbances. For example, for the same vehicle, the phase information may vary significantly between images taken at the entrance and exit due to different lighting conditions, but the amplitude information remains consistent, ensuring the comparability of the feature spectra.

[0089] The rules for determining the highest frequency order are as follows: when the number of discrete sampling angles is 64, the highest frequency order is 16; when the number of sampling angles is 128, it is 32, ensuring that the frequency domain covers the main characteristics of the angle domain signal. If the parking lot has a single type of vehicle, 16 can be selected to simplify the calculation; if the vehicle types are complex and it is necessary to improve the differentiation, 32 should be selected.

[0090] Normalization of the characteristic spectrum vector includes: dividing the amplitude of all frequency components by the DC component (zero-order frequency amplitude) and adding the minimum value, where the minimum value is 1e-6 or 1e-8 to avoid division by zero. Normalization reduces the influence of overall brightness variation on the characteristic spectrum, making the characteristics under different lighting conditions comparable.

[0091] The specific implementation of the Discrete Fourier Transform includes: using a Hanning window to preprocess the angle domain signal to reduce spectral leakage; retaining only the components from zero to the highest order of frequency after the transform, and discarding high-frequency components above that order; and automatically adapting the number of iterations of the Fast Fourier Transform algorithm according to the number of sampling angles to ensure computational efficiency.

[0092] The arrangement rules for frequency components include: arranging them in order of frequency from low to high, with the zero-order component being the DC component, corresponding to the overall brightness characteristics of the vehicle's front face, and the higher-order components corresponding to local detail features. The arrangement order remains fixed to ensure the dimensional alignment of the inlet and outlet feature spectra.

[0093] Optimization of the feature spectrum dimension includes: if the parking lot hardware computing resources are limited, the highest frequency order can be reduced to 8 to reduce the amount of computation during feature matching; if the recognition accuracy requirement is high, it can be increased to 64, but more sampling angles are needed to avoid feature redundancy.

[0094] In one embodiment of the present invention, a dynamic database is constructed based on the vehicle entrance angular feature spectrum, including: For each entrance gate opening event Generate a record containing a system-unique identifier. timestamp and the vehicle entrance angular feature spectrum vector Data records; Record the data Store in feature database ; The feature database Do not store raw acquired images Or an image after Gaussian blurring. .

[0095] The system's unique identifier is a dedicated identifier used to distinguish different entrance gate opening events and their corresponding vehicles, ensuring the uniqueness of each data record. Preferably, it uses a universally unique identifier, an ID generated by the snowflake algorithm, or a combination of an auto-incrementing ID and a camera ID to ensure global uniqueness and adaptability to parking lot scenarios with multiple lanes and multiple cameras.

[0096] An event timestamp is a system time record of the time when the entrance gate opens, used to mark the time of vehicle entry. It can be obtained from the clock module of the gate control system or a server to ensure time synchronization.

[0097] The entrance gate opening event is a trigger signal generated when a vehicle triggers the entrance gate to raise its arm; it is the trigger condition for generating data records. This event can be obtained through the gate's infrared sensor, inductive loop, or image recognition trigger signal.

[0098] Data records are structured data units containing a unique system identifier, an event timestamp, and a vehicle entrance angular feature spectrum, used for database storage and retrieval.

[0099] The feature database is a storage medium used to store all vehicle entry data records, specifically adapted to the need for rapid matching of vehicle feature spectra.

[0100] The privacy-preserving design logic of not storing raw or blurred images: Both raw and blurred images contain sensitive information such as vehicle appearance and license plates, and storing them would pose compliance risks. Only storing structured feature spectra, identifiers, and timestamps preserves the core information needed for vehicle matching while avoiding the leakage of sensitive data, thus complying with privacy regulations. For example, the feature spectrum vector of the same vehicle may contain only a few dozen values ​​and lacks identifiable vehicle appearance information.

[0101] The structured record design includes: three core elements of the data record each fulfilling their respective functions; a system-unique identifier for precise record location; an event timestamp for expiration cleanup and time range filtering; and an entry-angle feature spectrum as the core basis for matching. This structure ensures the completeness of the key information required for matching while minimizing data volume and improving database storage and retrieval efficiency.

[0102] The maintenance of the dynamic database includes: The dynamic database does not permanently store all records, but maintains data validity through mechanisms such as expiration cleanup and capacity control. Given the limited dwell time of vehicles in parking lots, only recent entry records are retained to avoid invalid data consuming storage resources and ensure fast retrieval during matching.

[0103] The generation of the system's unique identifier includes: when using a universal unique identifier, a 32-bit string is randomly generated using the system's built-in algorithm; when using the snowflake algorithm, a 64-bit integer is generated by combining the timestamp, data center ID, machine ID, and serial number; when using an auto-incrementing ID plus a camera ID, the auto-incrementing ID increments by lane dimension, and the camera ID is a preset integer from 1 to N, which together form a unique identifier.

[0104] The database expiration policy includes setting the lifespan of each record to 24 or 72 hours, calculated from the time corresponding to the event timestamp. Expired records are automatically deleted by a scheduled database task. For example, a small parking lot might use 24 hours, while a large shopping mall parking lot might use 72 hours, adapting to the vehicle dwell time in different scenarios.

[0105] The database capacity limit and eviction rules include: the capacity limit is set to twice the number of parking spaces in the parking lot; when the number of records stored in the database reaches the limit, the oldest records are evicted in the order of event timestamps to ensure that the database always has enough space to store new records.

[0106] The rules for handling abnormal entry scenarios include: when the same vehicle repeatedly triggers the entrance gate opening event without having exited, a strategy of overwriting old records is adopted, replacing the old records with the system's unique identifier, timestamp, and feature spectrum of the new event. This rule can prevent multiple invalid records for the same vehicle from consuming resources and ensure that the latest vehicle features are used during matching.

[0107] Data record retrieval optimizations include: the database is stored in buckets by lane ID, with each lane corresponding to an independent data bucket. During exit matching, only records within the same lane bucket are retrieved, or records from the past 24 to 72 hours are limited to the event timestamp, significantly reducing the search scope and improving matching speed.

[0108] In one embodiment of the present invention, the same operations as S1 and S2 are performed on the acquired exit image to obtain the vehicle exit angular feature spectrum, including: Calculate the homography matrix on the exit side using the calibration parameters of the exit camera. The original export image Transform into a normalized image This aligns the intersection of the exit stop lines with the fixed anchor point. ; Use the same fuzzy parameters as the inlet processing stage. right Gaussian blurring is performed to obtain ; Based on anchor points Establish a polar coordinate system, using the same sampling angle. The radial description vector is extracted using a radial sampling mechanism, and the amplitude spectrum is calculated using a discrete Fourier transform to finally generate the vehicle exit angular feature spectrum vector. : ; in, The export image is shown in the first... The amplitude coefficient of the first frequency component.

[0109] The exit homography transformation matrix is ​​a matrix that describes the relationship between the projection transformation of the exit scene space plane and the exit image plane, and is used for spatial correction of the exit image.

[0110] The original exit image is the raw image captured by the exit camera when a vehicle enters the parking lot, including information about the vehicle's front and surrounding environment. It can be captured by a high-definition camera deployed at the exit, which takes a picture when the vehicle triggers the exit gate's detection signal.

[0111] The normalized exit image is the image obtained after perspective correction of the original exit image through the exit homography transformation matrix, and has the same coordinate reference as the normalized entrance image.

[0112] The blurred exit image is generated by performing a two-dimensional Gaussian convolution operation on the normalized exit image, which has filtered out high-frequency texture information such as license plate characters.

[0113] The exit angular feature spectral vector is a vector composed of the frequency component amplitudes of the exit image in sequence. It is used to characterize the angular structural features of the vehicle's front face and has the same dimension as the entrance angular feature spectral vector.

[0114] The amplitude of the exit frequency component is the magnitude of the coefficients of each frequency component after the exit image has undergone discrete Fourier transform, and is used to construct the exit angular feature spectrum vector.

[0115] The core logic of isomorphic processing for entrances and exits is that feature matching for vehicle entry and exit relies on the comparability of feature spectra. Isomorphic processing requires that the correction, blurring, and feature extraction processes for the exit image be completely consistent with those for the entrance. For example, if the entrance uses Gaussian blur with σ=4, and the exit uses a different σ value, the degree of blurring of the vehicle contour features will be different, and the feature spectra cannot be directly compared. Isomorphic processing can avoid this problem.

[0116] Adaptive design to address differences in entrance and exit cameras: Entrance and exit cameras may differ in installation angle, resolution, and imaging parameters. By using an exit homography transformation matrix to align the intersection of the exit stop lines to the same fixed anchor point as the entrance, feature offset caused by spatial location differences can be eliminated. Simultaneously, standardized sampling parameters and transformation rules ensure that images acquired by different devices are converted into a unified standard feature spectrum.

[0117] The consistency guarantee logic for the exit feature extraction process is as follows: from image correction to blurring, and then to polar coordinate transformation, radial moment calculation, and frequency domain transformation, the algorithms and parameters of each step are strictly consistent with those of the inlet. This end-to-end consistency design can minimize the impact of equipment and environmental differences on features and ensure the effectiveness of the matching between the inlet and outlet feature spectra.

[0118] Mandatory constraint on consistency of inlet and outlet parameters: The fixed anchor point of the outlet must be exactly the same as that of the inlet. All custom parameters, such as Gaussian kernel standard deviation, number of discrete sampling angles, highest order of frequency, and radial sampling interval, must use the inlet settings and cannot be adjusted separately. For example, if the inlet K=128 and N=32, the outlet must use the same values.

[0119] Matching requirements for export camera calibration parameters: The calibration methods for the export camera's intrinsic and extrinsic parameters must be consistent with those for the import camera, employing the same techniques such as the Zhang Zhengyou calibration method. After calibration, the correction effect of the homography transformation matrix must be verified to ensure that the distortion level of the export image after correction is comparable to that of the import image.

[0120] Fallback strategy for outgoing image processing anomalies: If the outline of the outgoing image is severely lost after blurring, or the polar coordinate transformation exceeds the image range, the system will automatically trigger a second capture; if the second capture still fails, a prohibition command will be output and a manual review will be prompted.

[0121] Adaptation method when entrance and exit image resolutions are inconsistent: When the exit camera resolution differs from the entrance resolution, first scale the original exit image to the resolution of the entrance image, then perform subsequent correction and feature extraction steps. For example, if the entrance is 1080p and the exit is 720p, scale the exit image to 1920×1080 pixels before processing.

[0122] Verification rules for the exit feature spectrum: After extracting the exit angular feature spectrum, calculate the mean and variance of its frequency component amplitudes. If the mean is lower than 30% or higher than 200% of the mean of the entrance feature spectrum, the feature is deemed invalid and a second capture is triggered. If the variance is within the range of 50% to 150% of the variance of the entrance feature spectrum, the feature is deemed valid.

[0123] In one embodiment of the present invention, calculating the similarity distance between the vehicle exit angular feature spectrum and the database record to obtain the matching confidence score includes: Traversing the feature database For each record in the database, calculate the exit feature vector. With the entry feature vector Weighted Euclidean distance between The calculation formula is: ; in, For the first Preset weighting coefficients for the first-order frequency components; The minimum distance calculated from all records is selected as the optimal matching distance. ; Using the Sigmoid function to find the optimal matching distance Mapped to normalized matching confidence The calculation formula is: ; in, To control the sensitivity parameter of the conversion slope, As the reference parameter for distance determination, It is an exponential function.

[0124] The frequency weighting coefficient is a weight value assigned to different frequency components to adjust the contribution of each frequency component in the similarity distance calculation. Preferably, it is the result of dividing by the frequency order or an exponential negative lambda multiplied by the frequency order. The lambda is preferably 0.05 to 0.2 to ensure that low-frequency components better characterize the inherent structure of the vehicle. This design can highlight key features and suppress high-frequency noise interference.

[0125] The sensitivity parameter of the Sigmoid function controls the slope of the distance-to-confidence conversion, adjusting how sensitive the confidence level is to changes in distance. Ideally, it should be a value divided by the standard deviation of the matching pair distances, ensuring that the confidence level exhibits a suitable gradient within a reasonable distance range, balancing matching sensitivity and stability.

[0126] The distance benchmark parameter of the Sigmoid function is a reference value for the distance, used to determine the midpoint of the confidence level. It is preferably the mean or median of the distances between matching pairs, ensuring that the confidence level distribution closely matches the actual matching scenario and that the confidence levels of most matching pairs are within a reasonable range.

[0127] The weighted Euclidean distance is a measure of the difference in the characteristic spectrum calculated by fusing frequency weighting coefficients, used to quantify the similarity between the characteristic spectra of the exit and the inlet.

[0128] The optimal matching distance is the minimum of all weighted Euclidean distances, used to filter entry records in the database that are most similar to the exit feature spectrum.

[0129] Match confidence is a normalized value obtained by mapping the best match distance. It is used to intuitively represent the reliability of vehicle matching, and its value ranges from zero to one.

[0130] The amplitude of the exit frequency component is the intensity value of each frequency component in the exit angular characteristic spectrum, which is used to participate in the similarity distance calculation.

[0131] The amplitude of the inlet frequency component is the intensity value of each frequency component in the inlet angular characteristic spectrum, which is used to compare with the outlet characteristics.

[0132] The exponential function is a mathematical function used to calculate the sigmoid function. It has a base of the natural constant and is used to achieve a non-linear mapping from distance to confidence.

[0133] The design of frequency weighting coefficients includes the following: Key features of the vehicle's front structure are concentrated in low-frequency components, while high-frequency components are susceptible to noise and localized occlusion. The weighting coefficients decrease as the frequency order increases, allowing low-frequency components to contribute more to distance calculations and high-frequency components to contribute less, thus highlighting stable structural features and reducing the impact of interference factors. For example, the zeroth-order frequency corresponds to overall brightness characteristics and has the highest weight, while higher-order frequencies correspond to detail fluctuations and have lower weights.

[0134] Applications of weighted Euclidean distance include: Ordinary Euclidean distance treats all frequency components equally, failing to distinguish feature importance. Weighted Euclidean distance, through weight allocation, ensures that differences in critical frequency components have a greater impact on the total distance, while differences in non-critical components have a smaller impact, making the distance calculation more closely match the actual needs of vehicle matching. For example, for the same vehicle, changes in lighting can cause differences in high-frequency components, but after weighting, the total distance change is minimal, allowing for correct matching.

[0135] The parameter adaptation of the Sigmoid function includes: the Sigmoid function can map distance values ​​with uncertain ranges to a confidence level of zero to one, making it easy to set a uniform threshold. The values ​​of the sensitivity parameter and the baseline parameter are related to the distance distribution of actual matched pairs. The parameters are determined by statistically analyzing the distance mean and standard deviation of the samples, which allows the confidence level to effectively distinguish between matched and unmatched scenarios. For example, the distances of matched pairs are concentrated in a small range, while the distances of unmatched pairs are scattered over a large range. After parameter adaptation, the confidence level will show obvious stratification.

[0136] The frequency weighting coefficient includes two forms: Form 1 is one divided by one plus the frequency order. For example, the weight is one when the frequency order is zero, and 0.5 when the order is one, decreasing sequentially as the order increases. Form 2 is an exponential negative lambda multiplied by the frequency order. When the lambda is 0.1, the weight is one when the order is zero, and approximately 0.606 when the order is five. The higher the order, the faster the weight decays. The appropriate form can be chosen based on the complexity of the vehicle types in the parking lot: Form 1 for simpler types, and Form 2 for more complex types.

[0137] The calibration of sensitivity and baseline parameters includes: collecting 50 matching pairs of vehicles entering and exiting the same vehicle and 50 non-matching pairs of different vehicles, calculating the mean and standard deviation of the distance between the matching pairs, taking the mean as the baseline parameter, and taking one divided by the standard deviation as the sensitivity parameter; or traversing parameter combinations through grid search and selecting the parameter value that has the highest confidence differentiation between the matching pairs and non-matching pairs.

[0138] The weighted Euclidean distance formula includes: the sum of the squares of the weighting coefficients of all frequency components multiplied by the squares of the differences in frequency amplitude between the outlet and inlet, and then the square root of the sum.

[0139] The selection rules for the optimal matching distance include: traversing all valid records in the feature database, calculating the weighted Euclidean distance between each record and the exit feature spectrum, sorting all distance values ​​from smallest to largest, taking the first value after sorting as the optimal matching distance, and the corresponding entry record as the candidate matching record.

[0140] The numerical stability handling of the Sigmoid function includes: when the best matching distance is much larger than the benchmark parameter, the exponent part may be too large, causing the confidence to approach zero. In this case, the confidence is directly set to zero. When the best matching distance is much smaller than the benchmark parameter, the exponent part may be too small, causing the confidence to approach one. In this case, the confidence is directly set to one. The minimum value is taken as 1e-6 or 1e-8 to avoid numerical overflow during the calculation process.

[0141] In one embodiment of the present invention, generating a gate passage or prohibition command based on a confidence level and a confidence threshold includes: Pre-set a system confidence threshold ,in ; Constructing decision indicator variables ,when hour, ;otherwise ; according to Value output final control command: like Output the release command to control the gate to open, and retrieve data from the database. Delete the corresponding entry record; like The system outputs a prohibition command to keep the gate closed.

[0142] The system confidence threshold is a critical value used to determine whether a vehicle match is successful, and it serves as the basis for deciding whether to issue a pass or prohibition order. The preferred value is the 99.9th quantile of the confidence level between a non-match and a match, as this value keeps the false release rate low, balancing traffic efficiency and management safety.

[0143] The decision indicator variable is a binary variable that represents the matching result and is used to directly associate the gate control command.

[0144] Setting the confidence threshold involves considering the following: the threshold needs to be determined based on the distribution of matching data in the actual scenario. The confidence levels of non-matching pairs are typically low and dispersed, while the confidence levels of matching pairs are high and concentrated. Selecting the high percentile of non-matching pairs as the threshold can minimize the probability of mistakenly allowing non-matching vehicles to pass, while ensuring that most matching vehicles can pass smoothly. For example, if 1000 sets of non-matching pair data are collected, and the 99.9th percentile of the confidence level is 0.75, setting this value as the threshold would mean that only 0.1% of non-matching pairs might be mistakenly allowed to pass.

[0145] The design of deleting entry records after a successful match includes: each entry record corresponds to only one vehicle's entry once; deleting the record after a successful match prevents it from being repeatedly matched by subsequent vehicles, ensuring the uniqueness and validity of database records. It also reduces database storage pressure, improves the retrieval speed of subsequent matches, and avoids matching confusion caused by redundant records.

[0146] The application of binary decision variables includes: transforming the comparison results of confidence level and threshold into simple binary variables, making the generation logic of gate control commands more concise and clear, eliminating the need for complex calculations, improving command response speed, and adapting to the real-time release requirements of parking lot gates.

[0147] Determining the system confidence threshold involves: collecting 50 matched pairs from the same vehicle entering and exiting the same area and 200 unmatched pairs from different vehicles; calculating the matching confidence for all samples; statistically analyzing the distribution of confidence for unmatched pairs; and using the 99.9 quantile as the threshold. Alternatively, the threshold can be determined by using the receiver operating characteristic (ROC) curve, selecting the confidence level corresponding to a false release rate below 0.1%. For example, if the maximum confidence level for an unmatched pair is 0.8 and the 99.9 quantile is 0.72, then 0.72 can be selected as the threshold.

[0148] The handling of ambiguous scenarios includes: setting a distance difference threshold of 0.1; when the difference between the best matching distance and the second-best matching distance is less than this value, it is determined to be an ambiguous scenario. At this time, a prohibition command is output, and the exit camera is triggered to capture a second time to re-extract the feature spectrum for matching; if the second matching is still ambiguous, manual review is prompted.

[0149] The rules for deleting entry records include: the entry record to be deleted is the record corresponding to the best matching distance, and the record is accurately located through the system's unique identifier. Before deletion, it is necessary to verify that the timestamp of the record is within its valid lifespan and is logically consistent with the passage time of exit vehicles to avoid mistakenly deleting valid records.

[0150] The follow-up processing mechanism after a no-entry order is triggered includes: After the no-entry order is issued, the system will automatically provide a voice prompt to the driver: "Matching failed, please wait for manual review." Simultaneously, the exit feature spectrum, the best matching record, and relevant images (if any) will be pushed to the management backend. Management personnel can view the matching details in the backend and manually determine whether to allow passage. The determination result will be synchronized to the gate control system.

[0151] The dynamic adjustment rules for the threshold include: statistical analysis of system matching data every 30 days; if the false release rate is higher than 0.1%, the threshold will be increased by 0.05; if the false rejection rate (successful match but blocked) is higher than 5%, the threshold will be decreased by 0.03; after adjustment, it must be verified through 10 sets of test samples to ensure that the false release rate is controlled within the target range.

[0152] like Figure 2 As shown, Figure 2 The system displays vehicle identification performance under twenty different test scenarios (horizontal axis, numbered 1 to 20). The vertical axis represents the matching accuracy rate, expressed as a percentage (%). Figure 2 The light gray bar chart represents the test results of existing technical solutions based on traditional license plate recognition and conventional vehicle re-identification technologies, while the dark gray bar chart represents the test results of the parking lot gate management method proposed in this invention. Specifically, in all test scenarios, the matching accuracy of existing technologies fluctuated between 78.7% and 85.6%, indicating that they are relatively sensitive to environmental changes and lack stability. In contrast, the matching accuracy of the proposed solution exceeded 90% at all test points, reaching a maximum of 96.5%.

[0153] Example 2: Full-process operation of an unattended parking lot.

[0154] I. Pre-configuration and entry monitoring process: The parking lot is equipped with one high-definition camera each at the entrance and exit, supporting 1080p resolution capture. A dynamic database and mind map model are deployed in the cloud, with preset core parameters: Gaussian kernel standard deviation σ=4, polar coordinate angle range [-30°, +30°], number of discrete sampling angles K=128, highest frequency order N=32, confidence threshold τ=0.75, and database recording TTL=72 hours. The cameras obtain intrinsic and extrinsic parameters (rotation matrix, translation vector) using the Zhang Zhengyou calibration method. Ground plane information is obtained by fitting four reference points, with the fixed anchor point O set as the pixel coordinates corresponding to 50% of the image width and 85% of the height. After a vehicle enters the entrance lane, the ground inductive loop triggers a capture signal, and the camera captures the original entrance image. If no valid image is captured after three consecutive captures, the gate will prompt a retry, the cloud model will record the anomaly and push it to the management personnel. If the retry fails, the gate will be prohibited from opening and a log will be generated. If the capture is successful, non-1080p images need to be scaled to the corresponding resolution. Then, S1 is executed, which corrects the image based on the homography transformation matrix. If the distortion rate exceeds 5% after correction, backup calibration parameters are used. If this is still ineffective, manual review is triggered. During this process, manual gate raising is allowed and recorded. After normal correction, Gaussian convolution blur is performed. If the edge energy of the license plate area does not drop to 20%-40% of the pre-blurring level, σ is adjusted to 5 and blurred again. If this is still not satisfactory, relevant information is recorded and passage is allowed normally. Next, S2 is executed, which establishes a polar coordinate system, combines anti-blurring features through normalized moment vectors, and generates an entrance angular feature spectrum through DFT. If the feature spectrum dimension is missing, K is increased to 256 and re-extracted. If it fails twice, the feature is marked as low quality and stored in the database, and optimization suggestions are pushed to the cloud model. Finally, S3 is executed, which generates a snowflake algorithm ID and an entry timestamp, which, together with the feature spectrum, form a record and are stored in the cloud database. If the database reaches its capacity limit, the oldest record is discarded. If the same vehicle re-enters without having exited, the old record is overwritten. After completion, the gate raises to allow passage, and closes after the vehicle enters.

[0155] II. Export Monitoring Process and Branch Decisions: When a vehicle enters the exit lane, the infrared sensor triggers a capture signal, and the camera acquires the original exit image. If the image has motion blur, the shutter speed is automatically increased for a second capture. If it is still blurry, σ is adjusted to 3 and relevant information is marked during execution S1. Subsequently, execution S4 completes image correction, blur processing, and feature extraction using the same parameters as the entrance, generating an exit angular feature spectrum. If the dimensions of the exit and entrance feature spectra are inconsistent, the entrance parameters are automatically synchronized and re-extracted. If a camera malfunction causes parameter loss, the cloud model sends out default parameters and records the fault. Next, S5 is executed, iterating through the database of records for the same lane within the past 72 hours, calculating the weighted Euclidean distance and mapping it to the matching confidence level c. If no valid record is found, the gate provides a voice prompt and pushes vehicle information to the management personnel, supporting manual verification and release or guidance to the manual lane. If the difference between the best and second-best matching distances is <0.1, a second capture and rematch is triggered. If the second match is still ambiguous, the cloud model adjusts the parameters and performs a third match. If this fails, the vehicle is prohibited from entering, and a manual review is prompted. If c ≥ 0.75, the match is considered successful, the gate opens, and the corresponding entry record is deleted. If c < 0.5, the match is directly deemed a failure, and the vehicle is prohibited from entering. If 0.5 ≤ c < 0.75, the cloud model dynamically adjusts the threshold τ to 0.7. If the threshold is met after adjustment, the vehicle is released; otherwise, a remote manual review is triggered. After the vehicle leaves the gate, the cloud database updates the record status, and the mind map model statistically analyzes the matching time and confidence level distribution. If the confidence level is below 0.8 for five consecutive matches, the frequency weight coefficient w is automatically adjusted. n It is exp(-0.1n).

[0156] III. Anomaly Handling and System Optimization: The system automatically performs device self-checks at 2 AM daily, including camera parameter calibration and database connection testing. If camera calibration drift is detected, the cloud model sends calibration parameters to complete automatic calibration. If the database connection fails, it switches to local temporary storage; data is synchronized upon reconnection. In extreme environments, such as rain or snow, cameras automatically activate rain covers and supplementary lights. The cloud model analyzes image brightness variance in real time; if the variance exceeds a threshold, it switches to radial moment normalization to improve feature robustness. In case of emergencies such as fires in the parking lot, management personnel issue emergency mode commands through the cloud backend, automatically raising all exit gates and suspending the feature matching process. Normal operation resumes after the event. Through end-to-end branch judgments and anomaly fallback designs, the system ensures stable operation in complex scenarios. Furthermore, dynamic parameter optimization through the cloud model continuously improves matching accuracy and operational efficiency.

[0157] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for AI-assisted monitoring of unattended parking lots based on a cloud-based mind map model, characterized in that: include: S1. Establish the mapping relationship between the gate scene and the image plane, convert the acquired entrance image into a normalized image with the intersection of the parking lines as the positioning reference, and perform detexturing and blurring processing. S2, establish a coordinate system with the positioning reference as the pole, scan and aggregate the radial photometric distribution of the vehicle area along the angular direction for the normalized image to form an anti-blurring feature, and perform frequency domain transformation on the anti-blurring feature to generate a vehicle entrance angular feature spectrum characterizing the vehicle front structure. S3, constructing a dynamic database based on the vehicle entrance angular feature spectrum; S4, perform AND with S1 and S2 on the acquired exit image to obtain the vehicle exit angular feature spectrum; S5, calculate the similarity distance between the vehicle exit angular feature spectrum and the database record to obtain the matching confidence level, and generate a gate release or prohibition command based on the confidence level and confidence threshold.

2. The AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model as described in claim 1, characterized in that, Establish the mapping relationship between the gate scene and the image plane, including: The camera's internal parameters and external attitude parameters relative to the ground are obtained, and the homography transformation matrix from the spatial plane to the image plane is calculated by combining the ground plane information where the stop line is located. Using the homography transformation matrix, the actual intersection point of the stop line and the lane center line in the spatial plane is mapped to a preset fixed pixel coordinate in the image coordinate system, which serves as the positioning reference.

3. The AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model as described in claim 2, characterized in that, The acquired entrance image is converted to a normalized image based on the intersection of parking lines, and then de-textured and blurred to obtain a normalized image, including: The original entrance image is subjected to perspective projection transformation using the homography transformation matrix to generate a spatially corrected image. A fixed Gaussian kernel standard deviation parameter is set, and a two-dimensional Gaussian function is used to perform convolution operation on the corrected image to filter out high-frequency texture information containing license plate character details, thereby generating the normalized image.

4. The AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model as described in claim 3, characterized in that, Establishing a coordinate system with the positioning reference as the pole includes: Using the positioning reference in the normalized image as the origin of polar coordinates, the angular range covering the width of the vehicle's front face and the radial distance range extending from the stop line to the top of the vehicle are set; the pixel positions in the normalized image are converted from rectangular coordinates to polar coordinates determined by radial distance and angular direction, thereby aligning the optical structure of the vehicle's front face to a fan-shaped area centered on the origin of polar coordinates.

5. The AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model according to claim 4, characterized in that, To form anti-blurring features, the radial luminosity distribution of the vehicle region is aggregated along the angular direction of the normalized image, including: Multiple discrete sampling angles are set within the angular range covered by the polar coordinate system. For each sampling angle, a sequence of pixel brightness values ​​distributed along the radial direction is extracted. The brightness value sequence is calculated using a weighted accumulation method to obtain radial moment features. Based on the fuzzy invariant moment theory, the radial moment features are combined into a radial descriptive scalar that is stable to the degree of fuzziness, thereby forming a radial luminance distribution sequence that varies with the angle.

6. The AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model as described in claim 5, characterized in that, To combat fuzziness, frequency domain transformation is performed on the features to generate a vehicle entrance angular feature spectrum characterizing the vehicle's front-end structure, including: The radial photometric distribution sequence that varies with angle is taken as a one-dimensional angle domain signal, and a discrete Fourier transform operation is performed to map the radial photometric distribution sequence that varies with angle to the frequency domain; the amplitude value of each frequency component in the frequency domain is calculated, and the vector composed of the amplitude values ​​of each frequency component is taken as the angular feature spectrum of the vehicle entrance.

7. The AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model according to claim 6, characterized in that, A dynamic database is constructed based on the vehicle entrance angular feature spectrum, including: For each entrance gate opening event, a data record is generated containing an internal system identification number, an event timestamp, and the vehicle entrance angular feature spectrum; the data record is stored in the feature database to form a vehicle index library.

8. The AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model as described in claim 7, characterized in that, Performing AND operations S1 and S2 on the acquired exit image yields the vehicle exit angular feature spectrum, including: S1 is performed on the exit image to obtain the corresponding normalized image. S2 is then performed on the normalized image to generate the vehicle exit angular feature spectrum characterizing the vehicle's front structure.

9. A method for AI-assisted monitoring of unattended parking lots based on a cloud-based mind map model, as described in claim 8, is characterized in that... Calculate the similarity distance between the vehicle's exit angular feature spectrum and the database records to obtain the matching confidence score, including: For each data record in the feature database, the numerical differences between the corresponding vehicle entrance angular feature spectrum and the vehicle exit angular feature spectrum at each frequency component amplitude are calculated. The numerical differences of all frequency components are squared and multiplied by a preset frequency weighting coefficient, then summed and squared to obtain the weighted Euclidean distance. Using a monotonically decreasing nonlinear mapping function, the weighted Euclidean distance is converted into a matching confidence value between zero and one, such that the smaller the distance, the higher the confidence value.

10. The AI-assisted monitoring method for unattended parking lots based on a cloud-based mind map model according to claim 9, characterized in that, Generate gate access or access control commands based on confidence level and confidence threshold, including: A system confidence threshold is preset to determine whether a match is successful. The calculated match confidence value is compared with the system confidence threshold. When the match confidence value is greater than or equal to the system confidence threshold, the vehicle is determined to be successfully matched, a release command to open the exit gate is output, and the corresponding entry record is deleted from the feature database. When the match confidence value is less than the system confidence threshold, the vehicle is determined to be unmatched, and a prohibition command to keep the exit gate closed is output.

Images