Parking lot patrol robot license plate recognition method and system based on dynamic visual compensation

Abstract

The application discloses a parking lot patrol robot license plate recognition method and system based on dynamic visual compensation, and relates to the technical field of data processing.The method comprises the following steps: predicting the transient stationary point of chassis vibration according to the motion posture data sequence to trigger license plate image acquisition; combining angular velocity data and imaging internal parameters to calculate the motion blur vector and generate the spatial variation degradation kernel; using the kernel to perform non-uniform deblurring on the license plate candidate area, extracting and fusing the restored texture features and original topological features, and outputting the license plate recognition result and confidence; calculating the visual degradation parameter according to the blur vector length and confidence, and triggering deceleration and starting the hysteresis maintenance period control when the parameter is greater than the threshold value.The application solves the technical problems of non-uniform image motion blur caused by vibration and rotation in dynamic inspection and frequent control oscillation caused by single feedback, and effectively improves the license plate recognition accuracy and operation stability of the robot.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for license plate recognition in parking lot patrol robots based on dynamic visual compensation. Background Technology

[0002] In existing parking lot inspection operations, due to differences in the flatness of the ground in underground or open-air parking garages, the chassis of the inspection robot experiences high-frequency, small-amplitude mechanical vibrations as it continuously moves and traverses parking spaces. Existing license plate recognition systems often use a fixed frame rate for image acquisition. This acquisition method is prone to shutter activation at moments of high robot chassis acceleration or speed, resulting in physical motion blur superimposed on the license plate image at the source.

[0003] Meanwhile, when the robot avoids obstacles, turns, or one wheel drives over bumpy surfaces, the vehicle body generates rotational angular velocities, including roll and yaw. This three-dimensional spatial rotation, projected onto a two-dimensional image plane, causes the pixel linear velocity at one end of the license plate to be greater than that at the other end, resulting in non-uniform spatial variation blur. Conventional algorithms use a globally uniform blur kernel for image restoration, which may lead to some characters on the license plate being clear while the other end shows ghosting or overcompensation artifacts. Furthermore, existing solutions, when evaluating recognition quality and issuing speed control commands to the robot chassis, typically rely solely on the character recognition confidence score output by the neural network, lacking an objective measure of physical blur. This not only easily leads to high-confidence misrecognition but also causes control signal jumps due to frame-by-frame image quality fluctuations, resulting in frequent acceleration and deceleration of the chassis motors, increasing hardware wear and affecting the robot's motion stability. This further degrades the image acquisition environment, making it difficult to meet the stable recognition requirements in complex dynamic scenes. Summary of the Invention

[0004] In response to the technical problems described in the background art, the present invention provides a license plate recognition method and system for parking lot patrol robots based on dynamic visual compensation, which solves the problem of non-uniform image motion blur caused by high-frequency vibration of the chassis and multi-dimensional rotation in space during the dynamic inspection process of parking lot patrol robots, as well as the problem of frequent chassis control oscillation caused by single recognition confidence feedback.

[0005] A license plate recognition method for a parking lot patrol robot based on dynamic visual compensation includes: acquiring a motion posture data sequence; predicting a transient stationary point where the vibration direction reverses based on the motion posture data sequence; and triggering an image acquisition command for the imaging component to acquire a license plate image when the transient stationary point is predicted; acquiring angular velocity data during the imaging process; calculating the motion blur vector of the coordinate point in the image within a single exposure time of the imaging component based on the imaging intrinsic parameters of the imaging component to generate a spatial variation degradation kernel; extracting a license plate candidate region from the license plate image; performing non-uniform deblurring processing on the license plate candidate region using the spatial variation degradation kernel; extracting the restored texture features and the undeblurred topological features; fusing the texture features and topological features to output the license plate number and average recognition confidence; calculating a visual degradation parameter based on the length of the synthesized motion blur vector corresponding to the center point of the license plate candidate region and the average recognition confidence; and triggering a deceleration command and initiating a hysteresis hold period with control locking when the visual degradation parameter is greater than a set threshold.

[0006] Optionally, the transient stationary point of vibration direction reversal is predicted based on the motion posture data sequence, and when the transient stationary point is predicted, an image acquisition command is triggered for the imaging component to obtain a license plate image, including: acquiring the synthetic acceleration modulus sequence in the motion posture data sequence; performing a first-order backward difference operation on the synthetic acceleration modulus sequence to calculate the difference between the current sampling time and the previous sampling time; when the difference value of the previous sampling time is detected to be negative and the difference value of the current sampling time is non-negative, the current sampling time is determined as the transient stationary point and the image acquisition command is triggered.

[0007] Optionally, the motion blur vector of the image coordinate point within the imaging component during a single exposure time of the imaging component is calculated based on the imaging intrinsic parameters of the imaging component to generate a spatial variation degradation kernel. This includes: obtaining the equivalent focal length of the imaging component and the single exposure time; determining the lateral and longitudinal coordinates of the image coordinate point relative to the optical center of the image; calculating the lateral and longitudinal vector length components of the image coordinate point in the image plane based on the lateral and longitudinal coordinates, the equivalent focal length, and the single exposure time, combined with the instantaneous angular velocity components of the angular velocity data in the three axes of the spatial coordinate system, according to the projection mapping relationship; and assigning restoration operators of corresponding scales to pixels at different coordinate positions within the license plate candidate area based on the lateral and longitudinal vector length components corresponding to each pixel coordinate point, thus forming a spatial variation degradation kernel.

[0008] Optionally, a license plate candidate region is extracted from the license plate image, and a non-uniform deblurring process is performed on the license plate candidate region using a spatial variation degradation kernel to extract the restored texture features and the un-deblurred topological features. This includes: inputting the non-uniformly deblurred license plate candidate region into a first feature extraction branch to extract texture features representing the restored state of character boundaries; and inputting the un-deblurred license plate candidate region into a second feature extraction branch to extract topological features representing the overall semantics of the license plate.

[0009] Optionally, the textural features and topological features are fused to output the license plate number and the average recognition confidence, including: calculating the response weights of the textural features and topological features in each channel dimension through a channel attention mechanism; performing weighted fusion of the textural features and topological features based on the response weights to generate a joint feature map; and inputting the joint feature map into the sequence recognition decoder to output the license plate number and the average recognition confidence.

[0010] Optionally, the visual degradation parameters are calculated based on the length of the synthetic motion blur vector corresponding to the center point of the license plate candidate region and the average recognition confidence. This includes: obtaining the vector length components of the center point of the license plate candidate region in the horizontal and vertical directions, and calculating the length of the synthetic motion blur vector; adding a constant compensation term used to establish the reference offset to the length of the synthetic motion blur vector to obtain the corrected blur amount; and dividing the corrected blur amount by the average recognition confidence to obtain the visual degradation parameters.

[0011] Optionally, when the visual degradation parameter is greater than a set threshold, a deceleration command is triggered and a hysteresis holding period with control lock is initiated, including: when the visual degradation parameter is greater than the set threshold, a deceleration command is triggered to reduce the current moving speed of the inspection robot to a target operating speed; during the initiated hysteresis holding period, a maximum speed limit is set for the drive motor that drives the inspection robot, and acceleration commands exceeding the maximum speed limit are rejected; when the hysteresis holding period ends and the visual degradation parameter is lower than a recovery threshold within a continuously set time period, the maximum speed limit is released and the initial speed command before the deceleration command was executed is restored.

[0012] A parking lot patrol robot license plate recognition system based on dynamic visual compensation is also provided, comprising: a synchronous perception module for acquiring motion posture data sequences, predicting transient stationary points of vibration direction reversal based on the motion posture data sequences, and triggering an image acquisition command for the imaging component to acquire a license plate image when the transient stationary point is predicted; a kernel parameter generation module for acquiring angular velocity data during the imaging process, and calculating the motion blur vector of the coordinate points in the image within a single exposure time of the imaging component based on the imaging intrinsic parameters of the imaging component to generate a spatial variation degradation kernel; a joint recognition module for extracting license plate candidate regions in the license plate image, performing non-uniform deblurring processing on the license plate candidate regions using the spatial variation degradation kernel, extracting the restored texture features and the undeblurred topological features, and fusing the texture features and topological features to output the license plate number and average recognition confidence; and a servo control module for calculating visual degradation parameters based on the length of the synthetic motion blur vector corresponding to the center point of the license plate candidate region and the average recognition confidence, and triggering a deceleration command and initiating a hysteresis hold period with control locking when the visual degradation parameters are greater than a set threshold.

[0013] Optionally, the synchronous sensing module is also used to: acquire the synthetic acceleration magnitude sequence in the motion attitude data sequence; perform a first-order backward difference operation on the synthetic acceleration magnitude sequence to calculate the difference between the current sampling time and the previous sampling time; when the difference value of the previous sampling time is detected to be negative and the difference value of the current sampling time is non-negative, determine the current sampling time as a transient stationary point and trigger an image acquisition command.

[0014] Optionally, the servo control module is provided with a degradation calculation unit, which is used to: obtain the vector length components of the center point of the license plate candidate area in the horizontal and vertical directions, and calculate the synthetic motion blur vector length; add a constant compensation term used to establish the reference offset to the synthetic motion blur vector length to obtain the corrected blur amount; divide the corrected blur amount by the average recognition confidence to obtain the visual degradation parameters.

[0015] The beneficial effects of this invention are reflected in: In the entire license plate recognition method for parking lot patrol robots based on dynamic visual compensation, firstly, the system predicts the transient stationary point of chassis vibration direction reversal by analyzing motion posture data and triggers exposure, reducing the initial image motion blur from the physical source. Secondly, to address the non-uniform degradation problem caused by robot movement, this scheme calculates the motion blur vector of different coordinate points within the license plate candidate area within a single exposure time based on angular velocity and imaging intrinsic parameters, generates a spatial variation degradation kernel, and performs non-uniform deblurring, improving the inconsistent blurring degree of characters at different positions of the license plate due to rotation components. Simultaneously, by fusing the restored edge texture features with the unprocessed original topological features, the system combines local detail restoration with global structural information, improving the recognition accuracy under complex working conditions. Finally, this solution combines the fuzzy vector length of the physical dimension with the recognition confidence of the visual dimension to quantify the visual degradation parameter, reducing the probability of misjudgment caused by relying solely on confidence. When the parameter exceeds the set threshold, it triggers deceleration and a hysteresis holding cycle with control lock, suppressing the frequent acceleration and deceleration phenomenon of the chassis motor caused by the fluctuation of single-frame image quality, and maintaining the motion stability and perception stability of the inspection robot in uneven road environment. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0017] Figure 1 This is a schematic diagram illustrating the steps of the license plate recognition method for parking lot patrol robots based on dynamic visual compensation according to the present invention. Figure 2 This is a schematic diagram of a portion of step S1 in the license plate recognition method for parking lot patrol robots based on dynamic visual compensation of the present invention; Figure 3 This is a schematic diagram of a portion of step S2 in the license plate recognition method for parking lot patrol robots based on dynamic visual compensation of the present invention; Figure 4 This is a schematic diagram of part S3 in the license plate recognition method for parking lot patrol robots based on dynamic visual compensation of the present invention; Figure 5 This is a schematic diagram of part of step S4 in the license plate recognition method for parking lot patrol robots based on dynamic visual compensation of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0020] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] This invention provides a license plate recognition method for parking lot patrol robots based on dynamic visual compensation, such as... Figure 1 As shown, in one specific embodiment, the method includes: S1: Data Acquisition and Transient Stationary Point Trigger Acquisition; By acquiring the motion posture data sequence, the transient stationary point of vibration direction reversal is predicted based on the chassis vibration characteristics, thereby triggering the image acquisition command. Here, the motion posture data sequence refers to the continuous data set collected by sensors that reflects the real-time displacement and rotation state of the robot chassis in three-dimensional space; while the transient stationary point refers to the millisecond-level instantaneous velocity zero point generated by the velocity vector switching due to the reciprocating vibration motion of the chassis.

[0022] S2: Spatial variability degradation kernel generation and non-uniform deblurring; combining angular velocity data and imaging intrinsic parameters, pixel-level motion blur vectors are calculated, and refined restoration is performed on the license plate candidate region. The spatial variability degradation kernel here is a mathematical model that describes the difference in the degree of motion blur influence at different coordinate points of the image, which breaks the limitations of existing global uniform compensation.

[0023] S3: Dual-branch feature extraction and license plate sequence recognition; extracting restored texture features and unblurred topological features, and outputting recognition results and average recognition confidence through feature fusion. The restored texture features focus on the clarity details of character edges, while the unblurred topological features retain the original global structural information of the license plate.

[0024] S4: Recognition quality assessment and chassis servo closed-loop control; calculates visual degradation parameters based on the length of the motion fuzzy vector and the average recognition confidence, and adjusts the robot's moving speed accordingly, executing a hysteresis hold cycle. The visual degradation parameters are a comprehensive indicator used to quantify the uncertainty of physical fuzziness and algorithm recognition, while the hysteresis hold cycle is an anti-vibration mechanism that prevents frequent switching of control commands through a time buffer.

[0025] In this embodiment, it should be noted that in S1, the synchronous acquisition of multi-sensor data and the triggering exposure of transient stationary points are mainly based on the coupling mechanism of chassis kinematics and visual imaging. When the parking lot inspection robot travels on an uneven road surface at an inspection speed of 2.0 m / s, the chassis will generate high-frequency mechanical vibration due to ground excitation. To avoid the problem that existing fixed frame rate acquisition, which lacks a triggering mechanism, tends to image at the point of maximum amplitude, the internal six-axis inertial measurement unit is used to acquire the robot's motion posture data sequence in real time at a sampling frequency of 200 Hz. Specifically, the changing trend of the synthetic acceleration modulus is continuously tracked. For example, the acceleration modulus is monitored to be 9.850 m / s², 9.810 m / s², and 9.825 m / s² in three consecutive sampling moments. By performing a first-order backward difference operation on this sequence, the difference value of the previous sampling moment is calculated to be -0.040 m / s², while the difference value of the current sampling moment is 0.015 m / s². The physical phenomenon of the difference value turning from negative to positive indicates that the acceleration modulus of the chassis has undergone a process of transformation from decreasing to increasing, which mechanically corresponds to the millisecond-level transient stationary point when the vibration direction reverses. After accurately determining this physical trough, an exposure command is sent to the hardware trigger pin of the binocular camera module. By binding the physical vibration state with the camera shutter speed, image capture can be completed at moments when the robot's relative motion trend is small, reducing the initial motion blur of the license plate image from the physical source and providing high-quality raw input data for subsequent algorithmic reconstruction.

[0026] In S2, after acquiring the initial license plate image, the process enters the stage of generating the spatial variation degradation kernel and non-uniform deblurring. When the robot avoids obstacles or drives over bumpy surfaces, it makes slight steering movements, resulting in roll, yaw, and other rotational angular velocities. The camera's equivalent focal length is known to be 1200 pixels, with a single exposure time of 0.01 s. The inertial measurement unit provides real-time feedback of the three-axis instantaneous angular velocity components in the imaging coordinate system, which are 0.10 rad / s, 0.20 rad / s, and 0.30 rad / s, respectively. Taking the pixel coordinates (400, 200) of a character edge on the left side of the license plate image as an example, combining the aforementioned intrinsic parameters and angular velocity data, the objective physical displacement of this point during a single exposure time is calculated using a differential optical flow model. The calculation shows that the motion blur vector length components of this pixel in the horizontal and vertical directions are -2.00 pixels and -0.1 pixels, respectively, indicating that this coordinate point experiences significant lateral trailing deformation. This pixel-by-pixel vector calculation based on physical projection mapping is performed on all coordinate points within the license plate candidate area to construct a spatial variation degradation kernel matrix that reflects the differences in damage across different regions of the image. This technique overcomes the problem of local overcompensation or undercompensation that is easily caused by conventional algorithms using a globally uniform blur kernel. By assigning restoration operators of corresponding scales to pixels at different coordinate positions, it improves the problem of non-uniform motion blur of license plates caused by rotation components.

[0027] In S3, the generated spatial variation degradation kernel matrix is ​​used to perform dual-branch feature extraction and license plate sequence recognition tasks. For the specific coordinate point (400, 200) obtained above, the corresponding restoration operator is called to perform inverse filtering on the lateral offset of about 2.06 pixels around it. After non-uniform deblurring, the image enters the feature fusion network. The first feature extraction branch is responsible for processing the restored license plate image, specifically extracting texture features representing the restored state of character boundaries, so that the edges of strokes stuck together due to jitter can be clearly presented; the second feature extraction branch receives the original license plate image without deblurring, extracts topological features representing the overall semantics of the license plate, so as to preserve the global geometric structure in the original image that has not been changed by the restoration algorithm. Subsequently, the response weights of these two types of features in each channel dimension are calculated using the channel attention mechanism, and weighted fusion is performed to generate a joint feature map that contains both physical deblurring gain and original semantic boundaries. After the joint feature map is input into the sequence recognition decoder, the license plate number "XA88888" is output, and the average recognition confidence of this recognition is calculated simultaneously as 0.92. This feature extraction method, which combines restored texture with original topology, balances the relationship between local detail restoration and global structure preservation, thereby improving the accuracy of license plate recognition for patrol robots under complex vibration conditions.

[0028] In S4, the evaluation results based on both physical and visual dimensions are comprehensively quantified, and chassis servo hysteresis closed-loop control is executed accordingly. The motion blur vector length component at the center point of the license plate candidate area is retrieved. Assuming its horizontal and vertical components are 1.25 pixels and 0.35 pixels respectively, the calculated synthetic motion blur vector length is approximately 1.298 pixels. Combining the constant compensation term of 1.0 pixels and the average recognition confidence of 0.92 output from S3, the current visual degradation parameter is calculated to be 2.498 pixels. Since this value is below the set safety threshold of 8 pixels, the robot continues its inspection at a speed of 2.0 m / s. When the robot enters an area with sudden changes in lighting and increased bumps, if the average recognition confidence drops to 0.35 and the horizontal blur component at the center point increases to 3.5 pixels, the synthetic physical blur reaches 3.517 pixels, and the recalculated visual degradation parameter increases to 12.906 pixels. Given that this parameter exceeds the 8-pixel threshold, a deceleration command is triggered to the chassis, reducing the travel speed to 1.5 m / s, and simultaneously initiating a 5-second hysteresis hold period. During this period, the maximum speed limit of the drive motor is locked, and acceleration is refused until the observation period ends and the visual degradation parameter stably falls below the 5-pixel recovery threshold within 1 second, at which point the speed lock is released. This technique improves upon the frequent speed adjustments caused by relying solely on confidence feedback, suppresses high-frequency oscillations in the chassis motors, and maintains the stability of the robot's movement.

[0029] In summary, the entire license plate recognition method for parking lot patrol robots based on dynamic visual compensation firstly reduces the initial image motion blur by analyzing motion posture data to predict the transient stationary point of chassis vibration direction reversal and triggering exposure. Secondly, addressing the non-uniform degradation problem caused by robot movement, this scheme calculates the motion blur vector of different coordinate points within the license plate candidate area during a single exposure time based on angular velocity and imaging intrinsic parameters, generates a spatial variation degradation kernel, and performs non-uniform deblurring, improving the inconsistent blurring degree of characters at different positions of the license plate due to rotation components. Simultaneously, by fusing the restored edge texture features with the unprocessed original topological features, combining local detail restoration and global structural information, the recognition accuracy under complex working conditions is improved. Finally, this solution combines the fuzzy vector length of the physical dimension with the recognition confidence of the visual dimension to quantify the visual degradation parameter, reducing the probability of misjudgment caused by relying solely on confidence. When the parameter exceeds the set threshold, it triggers deceleration and a hysteresis holding cycle with control lock, suppressing the frequent acceleration and deceleration phenomenon of the chassis motor caused by the fluctuation of single-frame image quality, and maintaining the motion stability and perception stability of the inspection robot in uneven road environment.

[0030] like Figure 2As shown, in one specific embodiment, S1 includes: S11: acquiring a sequence of motion posture data of the patrol robot during its movement. This sequence is obtained through an inertial sensing device integrated on the chassis, capable of capturing minute physical bumps.

[0031] S12: Predict the transient stationary point of vibration direction reversal based on the motion attitude data sequence, and trigger an image acquisition command for the imaging component to obtain the license plate image when the transient stationary point is predicted. Specifically, this involves acquiring the synthetic acceleration modulus sequence and performing a first-order backward difference operation. When the... A point is considered a transient stationary point when the following discrete difference condition is met at each sampling time:

[0032] In the above formula, Indicates the first The synthesized acceleration modulus at the sampling time and the first sampling time are related to the second sampling time. The difference in the magnitude of the synthesized acceleration at each sampling time; Indicates the first The synthesized acceleration modulus at the sampling time and the first sampling time are related to the second sampling time. The difference in the magnitude of the composite acceleration at each sampling moment. This first-order backward difference operation can accurately capture the inflection point where the acceleration trend changes from decreasing to increasing, i.e., the physical static transient, by calculating the numerical difference between adjacent sampling points.

[0033] In this embodiment, it should be noted that in S11, the main task is to synchronously acquire raw data from multiple sensors based on the coupling mechanism between chassis kinematics and visual imaging. When the parking lot inspection robot continuously travels at a preset speed of 2.0 m / s on the underground parking garage surface with minor cracks or unevenness deviations, the chassis will generate high-frequency mechanical vibrations due to ground excitation. In order to accurately capture these minute vibration details and provide a physical basis for subsequent image intervention, a six-axis inertial measurement unit integrated on the robot chassis is used to acquire the robot's motion posture data sequence in real time at a high-frequency sampling rate of 200 Hz.

[0034] In practical deployment, the preset inspection speed of 2.0 m / s and the sampling frequency of the inertial measurement unit (IMU) of 200 Hz were not randomly selected, but rather determined jointly based on historical measured data of the robot's weight and the typical underground parking garage road surface spectrum. Specifically, through multiple test runs in a test garage with deviations in standard flatness, it was found that when the inspection speed was set to 2.0 m / s, it could meet the inspection efficiency requirement of covering the specified number of parking spaces per hour, while also converging the main frequency of chassis mechanical vibration within the 10 Hz to 30 Hz range. According to the Nyquist sampling theorem, in order to reconstruct and accurately capture the millisecond-level transient stationary point of vibration direction reversal within this frequency band without distortion, the system reserved a reasonable safety redundancy, calibrating the lower limit of the IMU's sampling frequency to 200 Hz, thereby ensuring the optimal balance between the accuracy of physical vibration sampling and the computational power consumption of the microcontroller.

[0035] This sequence contains high-precision three-dimensional acceleration and angular velocity information, reflecting the real-time displacement and rotation of the chassis in three-dimensional space. This technique, which uses high-frequency sensors to directly read the physical state, changes the previous passive mode of license plate recognition that involved indiscriminate image acquisition detached from the vehicle's motion environment. It provides fundamental data support for establishing the spatiotemporal correlation between physical motion and visual imaging, improving the objectivity and accuracy of the perception process.

[0036] In S12, the transient stationary point of chassis vibration direction reversal is predicted in real time using the acquired motion attitude data sequence, and an image acquisition command is triggered. The synthetic acceleration modulus at three consecutive sampling times is detected to be 9.850 m / s², 9.810 m / s², and 9.825 m / s², respectively. A first-order backward difference operation is performed, and the difference value at the previous sampling time is calculated to be -0.040 m / s², and the difference value at the current sampling time is 0.015 m / s². At this time, the discrete difference condition is met, that is, the difference value at the previous sampling time is negative and the difference value at the current sampling time is non-negative. In terms of calculation logic, this represents that the chassis acceleration modulus is experiencing a physical trough of transition from decreasing to increasing. At this time, the chassis is in a relatively static millisecond-level transient state due to the change in vibration direction. When this transient stationary point is determined, an exposure command is immediately sent to the hardware trigger pin of the binocular camera module. This technique anchors the camera's exposure at the moment of minimum physical motion displacement, reducing the initial motion blur of the license plate image from the hardware physical source of imaging.

[0037] like Figure 3 As shown, in one specific embodiment, S2 includes: S21: acquiring angular velocity data during the imaging process. The original rotational rate sensed by the chassis is mapped to the imaging coordinate system through coordinate transformation to obtain the instantaneous angular velocity components around the three axes of the imaging coordinate system.

[0038] S22: Calculate the motion blur vector of the coordinate points within the image during a single exposure time of the imaging component based on the imaging intrinsic parameters of the imaging component, to generate a spatial variation degradation kernel. The imaging intrinsic parameters here include the equivalent focal length and optical center coordinates, which are fundamental parameters determining the optical projection relationship. The calculation is as follows:

[0039] In the above formula, and These represent the vector length components of the coordinate points within the image plane in the horizontal and vertical directions, respectively; that is, the components of the motion blur vector in each axis, with the unit being pixels. Indicates the single exposure time of the imaging component; Indicates the equivalent focal length of the imaging component; and This represents the horizontal and vertical coordinates of a point within the image relative to the image's optical center. This represents the instantaneous angular velocity components along the three axes.

[0040] S23: Extract the license plate candidate region from the license plate image, and perform non-uniform deblurring on the license plate candidate region using a spatial variation degradation kernel. Based on the vector length components of each pixel obtained in S22, assign corresponding restoration operators to pixels at different coordinate positions within the license plate candidate region. The restoration operator is a mathematical compensation operator designed for a specific displacement, and this operation achieves non-uniform texture restoration.

[0041] In this embodiment, it should be noted that in S21, after image acquisition is completed, it is necessary to obtain the angular velocity data during the imaging process and unify it to the reference coordinate system for image processing. The original angular velocity measured by the chassis inertial measurement unit reflects the rotational state of the robot body, while the blurred shape of the image directly depends on the three-dimensional spatial rotation in the optical center coordinate system of the camera lens. Therefore, by performing spatial coordinate transformation through a pre-calibrated external parameter matrix, the original rotational rate sensed by the chassis is accurately mapped to the imaging coordinate system, obtaining the instantaneous angular velocity components around the three axes of the imaging coordinate system. In specific operating conditions, the instantaneous angular velocity components obtained and mapped are roll angular velocity 0.10 rad / s, pitch angular velocity 0.20 rad / s, and yaw angular velocity 0.30 rad / s. This data mapping process opens up the information channel between chassis mechanical dynamics and optical imaging geometry, enabling the subsequent quantification of the spatial rotational distortion degradation caused by the slight deflection of the vehicle body in different areas of the image using objective physical parameters from the camera's perspective.

[0042] In S22, the motion blur vector of the coordinate points within the image during a single exposure time of the imaging component is calculated based on the imaging intrinsic parameters of the imaging component, in order to generate a spatial variation degradation kernel. Specifically, the expression for calculating the motion blur vector of the coordinate points within the image during a single exposure time is as follows: as well as The calculation logic strictly follows the differential projection geometry of a rigid body rotating in three-dimensional space onto a two-dimensional pinhole imaging plane.

[0043] In this expression, the part within the parentheses represents the coordinates of a point on the image. The instantaneous pixel linear velocity at a given point is generated by the coupling of instantaneous angular velocities along three axes. Specifically, taking the horizontal expression as an example, It is determined by yaw rate The main lateral translation term caused by this, the nonlinear factor here This reveals a key optical phenomenon: when the camera rotates around the Y-axis, because the image sensor is planar, light rays pass through the lens optical center and are projected onto the sensor edge (i.e., In regions with larger values, the tangent angle decreases, causing edge pixels to physically displace faster than center pixels under the same angle of rotation. This is the geometric root of spatial variation blur. This is due to the pitch angular velocity. Cross-coupling deformation terms generated in regions of the image far from the central axis; It is purely determined by roll angular velocity. The resulting two-dimensional plane tangential velocity is approximated. These instantaneous velocities are then combined and multiplied by the single exposure time of the imaging element. The objective trailing length of a pixel during shutter opening can be approximated by integration.

[0044] In addition, the single exposure time is a core variable that determines the initial physical blur. The baseline parameters are determined in advance by the system through joint calibration of typical lighting environments in underground parking garages and chassis amplitude ranges. During system initialization, 1000 sets of illuminance data were collected in different areas, such as strong backlight entrances and exits and low-light depths, and optical flow was extrapolated by combining this with the peak maximum angular velocity when the chassis travels at a preset speed. Data processing results show that if the exposure time is longer than 0.015s, the tail length of non-uniform motion blur will likely exceed the effective deconvolution compensation radius of the subsequent spatial variation degradation kernel; while if it is less than 0.005s, it will result in an excessively low image signal-to-noise ratio, severely damaging the semantic integrity of topological features. Therefore, the system optimizes the single exposure time of the baseline through interval optimization. Set to 0.01s.

[0045] Based on the specific data in the application scenario, it is known , For the coordinates of the license plate area And substitute the angular velocity components mapped by the inertial sensor. , , First, calculate the horizontal vector length component. : ; ; Combining all terms yields the horizontal pixel linear velocity as follows: multiplied by Later obtained Similarly, calculate the longitudinal vector length component. : ; ; The vertical pixel linear velocity is multiplied by Later obtained .

[0046] This computational process overcomes the limitations of manually preset fixed parameters. By introducing physical measurement data, it effectively calculates the non-uniform degradation of each local area in the two-dimensional image when the sensor rotates in three-dimensional space due to its own flatness limitations. This provides reliable mathematical input for the subsequent allocation of differentiated restoration operators and improves the problem of local feature loss caused by inconsistent blur at both ends of the license plate.

[0047] In step S23, the calculated spatial variation degradation kernel is used to perform non-uniform deblurring on the license plate candidate region in the license plate image. Since the specific motion blur vector length component has been solved for each pixel coordinate point in the license plate candidate region in step S22, a restoration operator of corresponding scale is assigned to pixels at different coordinate positions. For example, for the region at coordinate point (400, 200), the called restoration operator will specifically perform inverse filtering or deconvolution compensation for the horizontal offset of about 2.00 pixels and the vertical offset of about 0.1 pixels, while for the pixel region on the right side of the license plate with a smaller offset, a restoration operator with a smaller kernel size is assigned. This processing method, which performs pixel-by-pixel or region-by-region adjustment based on the spatial variation vector array, ensures that the deblurring operation intensity matches the actual physical degradation at that position, improves the local overcompensation artifacts or undercompensation trailing problems that are easily caused when using a globally uniform blur kernel to process rotational blur, and improves the clarity of the overall character region of the license plate.

[0048] Specifically, the spatial variation degradation kernel is implemented by constructing a local point spread function (PSF). Since different coordinate positions within the license plate candidate region are affected by non-uniform rotation blurring, let's assume a certain pixel coordinate point within the region... The corresponding motion blur vector is The system then generates a parameterized linear motion blur point spread function for that local region. Its mathematical model is defined as: when a certain coordinate offset is along the vector... The set direction and distance are within the vector length When the range is met, the kernel weight of that point is Other position weights are When performing non-uniform deblurring, a block-overlapping processing strategy is adopted to divide the license plate image into multiple local image blocks. The center coordinates of each image block are determined based on the generated... A mutation degradation matrix is ​​constructed, and a non-blind deconvolution algorithm (such as Richardson-Lucy deconvolution algorithm or Wiener filtering) is used to solve for the clear texture of the corresponding image block. Finally, weighted fusion is used to eliminate the stitching gaps between blocks, thereby completing the restoration of non-uniform mutation degradation caused by spatial three-dimensional rotation projection.

[0049] like Figure 4 As shown, in one specific embodiment, S3 includes: S31: Extracting the restored texture features and the un-deblurred topological features. The deblurred license plate candidate region is processed through the first feature extraction branch to extract texture features; the original, un-restored license plate candidate region is processed through the second feature extraction branch to extract topological features.

[0050] S32: Fusion of texture and topological features. The response weights of the two types of features in each channel dimension are calculated using a channel attention mechanism, and then weighted and fused to generate a joint feature map that includes physical gain and original semantics.

[0051] S33: Input the joint feature map into the sequence recognition decoder and output the license plate number and the average recognition confidence.

[0052] In this embodiment, it should be noted that in S31, a dual-branch network is used to perform feature extraction on the processed image. Specifically, both the first and second feature extraction branches use a deep residual network (such as ResNet-18) with fully connected layers removed as the backbone network, and perform multi-scale feature mapping through stacked residual blocks. The residual blocks adopt existing basic mathematical expressions, specifically:

[0053] In the above formula, Indicates the first The input feature tensor of the layer; Indicates the first The output feature tensor of the layer; This represents the existing residual mapping function, which typically includes convolutional layers, batch normalization, and activation functions; This represents the convolution weight parameters within the residual block.

[0054] The first feature extraction branch receives the license plate candidate region image after S23 non-uniform deblurring and focuses on extracting texture features representing the restored state of character boundaries. This branch can effectively capture the sharpened edges of Chinese characters and letter strokes after restoration by the restoration operator. Simultaneously, the second feature extraction branch receives the original license plate candidate region image without deblurring and focuses on extracting topological features representing the overall semantics of the license plate. These features preserve the true geometric structure of the original optical imaging, such as the license plate border ratio and character spacing, which have not been altered by the deconvolution algorithm. This technique of processing the restored image and the original image separately utilizes the high-frequency detail gain brought by non-uniform deblurring while avoiding the risk of local structural distortion that may be introduced by the restoration algorithm in a single route, achieving a complementary advantage of edge sharpness and global topological stability at the feature level.

[0055] In S32, texture features and topological features are fused to construct high-quality input for subsequent recognition. A channel attention mechanism is used to calculate response weights for the texture features output from the first feature extraction branch and the topological features output from the second feature extraction branch across each channel dimension. This channel attention mechanism first compresses spatial dimensional features into channel descriptors using a global average pooling layer.

[0056] In the above formula, Represents the first element of the input feature map (texture feature or topological feature). One channel; and These represent the height and width of the feature map, respectively. Indicates the compressed number of... Channel global descriptor.

[0057] Then, The response weight coefficients for each channel are calculated using a multilayer perceptron (MLP) network. The complete attention weight calculation model is as follows:

[0058] In the above formula, Represents a vector set containing all channel descriptors; and These represent the weight matrices of the dimension-reduced fully connected layer and the dimension-upper fully connected layer in a multilayer perceptron, respectively. It is the ReLU activation function; The Sigmoid activation function is used to map the output to... interval; This refers to the generated channel response weights.

[0059] Finally, based on the response weight, the two feature paths are fused element-wise by weighting and concatenating (or adding) to generate a joint feature map. This attention-based feature weighting fusion method enables the adaptive selection of information combinations that are more conducive to subsequent character parsing in the multi-dimensional feature space, suppresses interfering background noise channels, and enhances robustness to complex lighting and residual deformation interference. Thus, it provides a stable feature representation for sequence recognition models that is mutually verified between physical gain and original semantics.

[0060] In S33, the generated joint feature map is input to the sequence recognition decoder, which outputs the license plate number and the average recognition confidence. The sequence recognition decoder employs an architecture combining a bidirectional long short-term memory network (BiLSTM) with a connectionist temporal classification (CTC) loss function. First, BiLSTM is used to capture the sequence context dependencies between characters in the joint feature map. For each time step... Its hidden layer state is represented as the forward hidden state. With backward hidden state splicing: Subsequently, the hidden state sequence passes through a fully connected layer and a Softmax activation function, outputting the probability distribution for each character category at each time step. The CTC layer is used to solve the misalignment problem between the input feature sequence and the output character sequence, ultimately outputting the effective character sequence distribution with the highest probability.

[0061] Simultaneously, the decoder also calculates and outputs the average recognition confidence level corresponding to this recognition result. This average recognition confidence level... The calculation model is as follows: ,in, Indicates the effective time step Below, the model predicts that the character is the first character in the character set. The probability values ​​of each category; This represents the final set and number of valid character time steps determined after CTC deduplication and blank label removal. This confidence score reflects the probability assessment of the accuracy of the output license plate string by the neural network model given the current quality of extracted features. This technique unifies the front-end physical triggering intervention, the mid-end mutation kernel optical restoration, and the back-end dual-path feature fusion effect into intuitive text information and probability indicators. This not only provides the recognized content but also simultaneously offers a measure of the reliability of the current recognition result at the algorithm's subjective level, providing data input for subsequent quality monitoring and chassis control closed-loop systems.

[0062] like Figure 5As shown, in one specific embodiment, S4 includes: S41: Calculating visual degradation parameters based on the length of the synthetic motion blur vector corresponding to the center point of the license plate candidate region and the average recognition confidence. The calculation model is as follows:

[0063] In the above formula, Indicates the visual degradation parameters; and This represents the vector length components of the center point of the license plate candidate area in the horizontal and vertical directions; To establish a constant compensation term for the reference offset; This represents the average recognition confidence level. The correction ambiguity (i.e., the numerator) represents the degree of objective physical degradation, while the confidence level reflects the subjective recognition certainty.

[0064] S42: When the visual degradation parameter D index When the speed exceeds a set threshold, a deceleration command is triggered and a hysteresis hold cycle with control lock is initiated. The maximum speed limit of the drive motor that moves the inspection robot is set, and control access is locked within this cycle. This control lock ensures that the chassis will not frequently change its motion state due to momentary disturbances within a set time, thus providing a stable physical inertial environment for imaging.

[0065] In this embodiment, it should be noted that in S41, visual degradation parameters are calculated based on the length of the synthetic motion blur vector corresponding to the center point of the license plate candidate region and the average recognition confidence. Specifically, the calculation expression used to quantify the current perception state is as follows: This computational process establishes a comprehensive evaluation mechanism that spans physical sensor data and deep learning algorithm outputs.

[0066] Wherein, the numerator of the expression The Euclidean distance formula (L2 norm) was used to calculate the vector length components of the center point of the license plate candidate region in the horizontal and vertical directions. and The length of the synthesized motion blur vector, which objectively reflects the physical pixel offset that occurs at the moment of exposure. A constant compensation term is added. The mathematical significance lies in preventing the numerator from approaching zero and masking other visual degradation factors when the robot is stationary or moving at a constant speed on a smooth surface, causing the physical fuzziness vector to be extremely small (or even zero); in terms of physical logic, A baseline offset was established, representing the fundamental amount of inherent image degradation caused by non-motion factors such as uneven illumination, sensor dark current noise, or lens smudges.

[0067] In practical applications, the above constant compensation term The specific value is not a subjective assumption, but is obtained through statistical analysis of historical static images of the site. Specifically, when the inspection robot is stationary and its motors are off, 1000 baseline license plate images are collected under different lighting conditions and environmental noise. The inherent blur pixel equivalent of these static images is calculated using the Laplacian variance or edge gradient algorithm. After removing extreme values, the average value is taken as the constant compensation term. For example, the statistical result shows that the average inherent blur of this batch of static images is 1.02 pixels. To facilitate engineering calculation efficiency, the system rounds the result to the nearest integer. It is calibrated to 1.0 pixel, which is used as the true baseline degradation caused by non-motor factors.

[0068] Furthermore, the denominator of the expression introduces the average recognition confidence level. This creates an inverse adjustment mechanism: when environmental interference makes it difficult for the neural network to extract character features, The value will decrease, causing the entire score to amplify non-linearly. Based on data from the application scenario, when the robot is in a routine inspection state and physical vibration is controllable, assuming... , constant compensation term And the average recognition confidence level First, calculate the molecule: the length of the synthetic motion fuzzy vector is... , plus The corrected blur amount was then obtained as follows: Divide it by Visual degradation parameters were calculated. At this point, the value is below the set safety threshold, and the chassis maintains normal speed. When the robot enters harsh working conditions, such as severely bumpy roads causing a significant increase in the lateral physical offset of the center point, the robot will maintain its normal speed. Furthermore, the deterioration of lighting conditions caused a significant drop in the confidence level of the neural network. At that time, the calculation is re-executed: the length of the synthesized motion blur vector increases to Correcting the fuzzy quantity becomes Divide it by the low confidence level. Afterwards, visual degradation parameters .because If the value exceeds the set safety threshold, a single scalar signal will immediately trigger a deceleration and initiate a hysteresis lock command to the downstream servo control module.

[0069] This computational logic solves the problem of "high-confidence misidentification" that easily occurs when relying solely on the confidence level of deep learning models. By combining the algorithm's probability judgment with the objective physical data measured by the inertial measurement unit through a division operation, it ensures that high-speed operation is maintained only under the dual conditions of a stable physical environment and reliable visual algorithm output, thereby providing a quantifiable closed-loop judgment basis for the smooth control of the chassis.

[0070] The set thresholds (safety threshold and recovery threshold) used were calculated based on curve fitting of a limited number of historical dynamic inspection data. Specifically, 5000 dynamic license plate recognition records were collected beforehand on various bumpy road sections. A scatter plot was plotted with the visual degradation parameter as the x-axis and the OCR effective recognition accuracy as the y-axis, and then polynomial fitting was performed. Analysis of the fitted curve characteristics revealed that when the visual degradation parameter exceeded 8.1 pixels, the recognition accuracy dropped sharply to below 85%; while when the parameter was below 4.8 pixels, the accuracy remained stable above 98%. Based on this data distribution pattern, the system set the safety threshold down to 8 pixels and the recovery threshold up to 5 pixels, thus ensuring recognition accuracy while allowing for a reasonable physical hysteresis range.

[0071] In S42, a decision is made based on the calculated visual degradation parameters, and a deceleration command is triggered and a hysteresis holding cycle with control lock is initiated when necessary. When the visual degradation parameter is 2.498 pixels, which is lower than the set safety threshold of 8 pixels, the robot maintains a speed of 2.0 m / s for inspection. If environmental interference causes the parameter to increase to 12.906 pixels and exceed the set threshold, a deceleration command is immediately triggered, reducing the robot's movement speed to the target ratio of 1.5 m / s. At the same time, a hysteresis holding cycle lasting 5 seconds is initiated, during which the maximum speed limit of the drive motor is set, and any acceleration command is refused. Only when the 5-second cycle ends and the visual degradation parameter is lower than the recovery threshold of 5 pixels for 1 second is the restriction lifted and the speed restored to 2.0 m / s. In addition, the target running speed ratio and the time parameters related to the hysteresis holding cycle in the above servo control actions are determined through chassis dynamics testing and vibration damping feedback. In a pre-built bumpy test site, the damping response curve of the chassis from receiving the deceleration command to the complete dissipation of the high-frequency mechanical vibration is recorded. Test data shows that reducing the initial velocity from 2.0 m / s to 75% (i.e., 1.5 m / s) significantly reduces the amplitude without excessively affecting the efficiency of garage inspections. Simultaneously, the average time for the chassis shock absorbers to absorb aftershocks is approximately 4.2 seconds. To ensure sufficient environmental stability redundancy, the system sets the hysteresis hold period to 5 seconds and the continuous observation period before exiting to 1 second, ensuring that the physical vibrations of the chassis motor completely subside before resuming the initial inspection speed.

[0072] This technology, which introduces a time buffer and speed lock, avoids the problem of high-frequency acceleration and deceleration of the chassis motor caused by relying solely on single-frame image quality fluctuations. While ensuring the stable improvement of the perception environment, it also maintains the operating life of the underlying hardware and the smoothness of the chassis driving.

[0073] This invention also provides a license plate recognition system for a parking lot patrol robot based on dynamic visual compensation. The system is used to implement a license plate recognition method for a parking lot patrol robot based on dynamic visual compensation. The system includes: The synchronous sensing module is used to acquire motion posture data sequence, predict the transient rest point of vibration direction reversal based on the motion posture data sequence, and trigger the image acquisition command for the imaging component to acquire the license plate image when the transient rest point is predicted. The kernel parameter generation module is used to acquire angular velocity data during the imaging process and calculate the motion blur vector of the coordinate points in the image within a single exposure time of the imaging component based on the imaging intrinsic parameters of the imaging component, so as to generate a spatial variation degradation kernel. The joint recognition module is used to extract the license plate candidate region in the license plate image, perform non-uniform deblurring on the license plate candidate region using the spatial variation degradation kernel, extract the restored texture features and the undeblurred topological features, and fuse the texture features and topological features to output the license plate number and the average recognition confidence. The servo control module is used to calculate the visual degradation parameters based on the length of the synthetic motion blur vector corresponding to the center point of the license plate candidate area and the average recognition confidence. When the visual degradation parameters are greater than the set threshold, a deceleration command is triggered and a hysteresis hold cycle with control lock is started.

[0074] In one specific implementation, the synchronous sensing module is further configured to: acquire the synthetic acceleration magnitude sequence in the motion attitude data sequence; perform a first-order backward difference operation on the synthetic acceleration magnitude sequence to calculate the difference between the current sampling time and the previous sampling time; when the difference value of the previous sampling time is detected to be negative and the difference value of the current sampling time is non-negative, determine the current sampling time as a transient stationary point and trigger an image acquisition command.

[0075] In one specific implementation, the servo control module is provided with a degradation calculation unit, which is used to: obtain the vector length components of the center point of the license plate candidate area in the horizontal and vertical directions, and calculate the synthetic motion blur vector length; add a constant compensation term used to establish the reference offset to the synthetic motion blur vector length to obtain the corrected blur amount; divide the corrected blur amount by the average recognition confidence to obtain the visual degradation parameters.

[0076] To further clarify the operating mechanism and physical quantification process of the technical solution of this invention, the following analysis will be conducted in detail on the underlying derivation logic of the license plate recognition method and system for parking lot patrol robots based on dynamic visual compensation, using a scenario containing specific parameters and data.

[0077] Inside a large underground parking garage, a parking garage patrol robot is... The robot performs automatic inspection tasks along the lane at a preset inspection speed. Due to long-term use, the underground parking lot floor has localized unevenness deviations and small cracks, causing high-frequency vibrations in the robot's chassis during operation. At this time, the robot's internal synchronous sensing module... The system acquires motion attitude data sequences in real time at a high frequency to obtain discrete physical information including triaxial acceleration and angular velocity. To capture high-quality raw images, the system enters the upper-level S1 and begins predicting the transient stationary point of vibration direction reversal based on the motion attitude data sequence. At this time, the system monitors the values ​​of the synthetic acceleration modulus at three consecutive sampling times as follows: , as well as By performing a first-order backward difference operation, the following is calculated: The difference value at the current sampling time is According to the judgment logic, because and The system recognizes that the chassis acceleration modulus changes from decreasing to increasing, successfully capturing the physical vibration trough, i.e. the transient stationary point, and triggers an image acquisition command to the camera within this millisecond window to obtain an image containing the target license plate.

[0078] After acquiring the license plate image, the system immediately enters the upper-level S2, where the kernel parameter generation module performs pixel-level spatial variation degradation kernel generation on the image frame. At this point, the imaging intrinsic parameters are known: the equivalent focal length of the camera. Single exposure time The instantaneous angular velocity data fed back by the inertial measurement unit, after mapping, has the following components in the imaging coordinate system: , and Using a specific pixel coordinate point on the license plate image. For example, this point represents the edge of the Chinese character region to the left of the license plate number. Using the differential optical flow model formula in the instruction manual, the motion blur vector length component of this pixel is calculated. and First calculate : After substituting the specific values, ; ; plus .final .

[0079] Next, the vertical component of the pixel is calculated. : Substitute the values: ; ;minus .final These two components This characterizes the objective physical fuzzy vector experienced at that point, reflecting the effect of the robot's turning angular velocity. The dominant effect is the horizontal pixel trailing phenomenon. The system performs this non-uniform calculation on each coordinate point within the license plate candidate area, and the generated spatial variation degradation kernel can realistically depict the differentiated motion degradation of different characters on the license plate.

[0080] After completing the fuzzy parameter modeling, the system executes the upper-level S3. The joint recognition module uses the generated spatial variation degradation kernel array to assign corresponding restoration operators to pixels at different locations within the license plate area, performing non-uniform deblurring processing. For the above coordinates... For nearby textures, the restoration operator will prioritize targeting approximately [texture type]. The lateral offset is compensated by deconvolution. The deblurred image is then input into a dual-branch feature extraction network. The first branch extracts the restored texture features, significantly improving the sharpness of the character stroke edges; the second branch extracts topological features from the unblurred license plate image to ensure the stability of the overall geometric structure of the license plate. The system then uses a channel attention mechanism to calculate the response weights of the two types of features and fuses them to generate a joint feature map, which is then input into the sequence recognition decoder. Finally, the OCR system outputs the license plate number as "XA88888" and provides the average recognition confidence score for this frame. (Right now (confidence probability).

[0081] Finally, the system enters the upper-level S4 for performance evaluation and feedback control. The servo control module retrieves the motion fuzzy vector length component at the center point of the license plate candidate region, assuming the center point calculation result is... , At this point, a constant compensation term is set. The system calculates the visual degradation parameters for the current state based on the calculation model in the manual. The calculation process is as follows: First, calculate the length of the synthesized motion blur vector. Next, the visual degradation parameters are calculated: Because at this time Far below the set safety threshold This indicates that although there is some physical vibration, the recognition results after algorithm compensation are still highly reliable.

[0082] If the robot enters an area with rapidly changing lighting, causing increased image noise, the average recognition confidence level will drop significantly. At this point, the visual degradation parameters will become If the physical vibration intensifies further... Increase to Then the synthesized physical fuzzy quantity reaches Calculation .because If the visual degradation parameters exceed the set threshold, the system immediately triggers a deceleration command, reducing the robot's operating speed from... Reduce to the target ratio (e.g.) ), and start continuous The system maintains a hysteresis hold period. During this period, the system locks the maximum speed limit of the drive motor that propels the inspection robot, prohibiting any acceleration. Only when this... The observation period ended, and the system detected continuous Visual degradation parameters within Stable decline to the recovery threshold (e.g.) Only when the speed lock is engaged below a certain threshold will the robot release the speed lock and resume the initial speed inspection, thus achieving a robust closed loop of perception and motion in complex dynamic environments.

[0083] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0084] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0085] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A license plate recognition method for a parking lot patrol robot based on dynamic visual compensation, characterized in that the method... include: Acquire motion posture data sequences; Acquire the synthetic acceleration magnitude sequence from the motion attitude data sequence; perform a first-order backward difference operation on the synthetic acceleration magnitude sequence to calculate the difference between the current sampling time and the previous sampling time; when the difference value of the previous sampling time is detected to be negative and the difference value of the current sampling time is non-negative, determine the current sampling time as a transient stationary point and trigger an image acquisition command for the imaging component to acquire the license plate image; Acquire angular velocity data during the imaging process; acquire the equivalent focal length of the imaging component and the single exposure time; The horizontal and vertical coordinates of the coordinate points within the image relative to the optical center of the image are determined. Based on the horizontal and vertical coordinates, equivalent focal length, and single exposure time, combined with the instantaneous angular velocity components of the angular velocity data in the three axes of the spatial coordinate system, the vector length components of the coordinate points in the image plane in the horizontal and vertical directions are calculated according to the projection mapping relationship. Based on the horizontal and vertical vector length components corresponding to each pixel coordinate point, restoration operators of corresponding scales are assigned to pixels in different coordinate positions within the license plate candidate area, forming a spatial variation degradation kernel. Extract the license plate candidate region from the license plate image, and perform non-uniform deblurring on the license plate candidate region using the spatial variation degradation kernel to extract the restored texture features and the unblurred topological features. By fusing texture features and topological features, the license plate number and average recognition confidence are output. The visual degradation parameters are calculated based on the length of the synthetic motion blur vector corresponding to the center point of the license plate candidate area and the average recognition confidence. When the visual degradation parameters are greater than the set threshold, a deceleration command is triggered and a hysteresis holding cycle with control lock is started, including: when the visual degradation parameters are greater than the set threshold, a deceleration command is triggered to reduce the current moving speed of the inspection robot to the target running speed. During the initial hysteresis hold period, the maximum speed limit of the drive motor that drives the inspection robot is set, and acceleration commands exceeding the maximum speed limit are rejected. When the hysteresis hold period ends and the visual degradation parameters are lower than the recovery threshold within a continuously set time period, the maximum speed limit is released and the initial speed command before the deceleration command is restored.

2. The license plate recognition method for parking lot patrol robots based on dynamic visual compensation according to claim 1, characterized in that, The process of extracting the license plate candidate region from the license plate image, performing non-uniform deblurring on the license plate candidate region using a spatial variation degradation kernel, and extracting the restored texture features and the unblurred topological features includes: The license plate candidate region after non-uniform deblurring is input into the first feature extraction branch to extract texture features that represent the character boundary recovery state. The unblurred license plate candidate region is input into the second feature extraction branch to extract topological features that represent the overall semantics of the license plate.

3. The license plate recognition method for parking lot patrol robots based on dynamic visual compensation according to claim 2, characterized in that, The fusion of texture features and topological features outputs the license plate number and average recognition confidence, including: The response weights of texture features and topological features in each channel dimension are calculated using a channel attention mechanism. Based on response weights, texture features and topological features are weighted and fused to generate a joint feature map; The joint feature map is input into the sequence recognition decoder, which outputs the license plate number and the average recognition confidence.

4. The license plate recognition method for parking lot patrol robots based on dynamic visual compensation according to claim 1, characterized in that, The calculation of visual degradation parameters based on the length of the synthetic motion blur vector corresponding to the center point of the license plate candidate region and the average recognition confidence includes: Obtain the vector length components of the center point of the license plate candidate region in the horizontal and vertical directions, and calculate the length of the synthetic motion blur vector; The corrected ambiguity is obtained by adding a constant compensation term used to establish the reference offset to the length of the synthesized motion ambiguity vector. Divide the amount of blurred correction by the average recognition confidence to obtain the visual degradation parameters.

5. A license plate recognition system for a parking lot patrol robot based on dynamic visual compensation, characterized in that, The system is used to implement the license plate recognition method for parking lot patrol robots based on dynamic visual compensation as described in any one of claims 1 to 4, and the system includes: The synchronous sensing module is used to acquire motion posture data sequence, predict the transient rest point of vibration direction reversal based on the motion posture data sequence, and trigger the image acquisition command for the imaging component to acquire the license plate image when the transient rest point is predicted. The synchronous sensing module is also used to: acquire the synthetic acceleration magnitude sequence in the motion posture data sequence; perform a first-order backward difference operation on the synthetic acceleration magnitude sequence to calculate the difference between the current sampling time and the previous sampling time; when the difference value of the previous sampling time is detected to be negative and the difference value of the current sampling time is non-negative, determine the current sampling time as a transient stationary point and trigger an image acquisition command. The kernel parameter generation module is used to acquire angular velocity data during the imaging process and calculate the motion blur vector of the coordinate points in the image within a single exposure time of the imaging component based on the imaging intrinsic parameters of the imaging component, so as to generate a spatial variation degradation kernel. The joint recognition module is used to extract the license plate candidate region in the license plate image, perform non-uniform deblurring on the license plate candidate region using the spatial variation degradation kernel, extract the restored texture features and the undeblurred topological features, and fuse the texture features and topological features to output the license plate number and the average recognition confidence. The servo control module is used to calculate the visual degradation parameters based on the length of the synthetic motion blur vector corresponding to the center point of the license plate candidate area and the average recognition confidence. When the visual degradation parameters are greater than the set threshold, a deceleration command is triggered and a hysteresis hold cycle with control lock is started.

6. The license plate recognition system for a parking lot patrol robot based on dynamic visual compensation according to claim 5, characterized in that, The servo control module includes a degradation calculation unit, which is used for: Obtain the vector length components of the center point of the license plate candidate region in the horizontal and vertical directions, and calculate the length of the synthetic motion blur vector; The corrected ambiguity is obtained by adding a constant compensation term used to establish the reference offset to the length of the synthesized motion ambiguity vector; Divide the amount of blurred correction by the average recognition confidence to obtain the visual degradation parameters.

Images