License Plate Recognition Method and System Based on Adaptive Aerial Photography Space Modeling of UAV

Abstract

The present invention relates to the technical field of data recognition, and particularly to a license plate recognition method and system based on adaptive shooting space modeling of an unmanned aerial vehicle. The method comprises the following steps: S10, constructing an adaptive shooting space model; S20, the unmanned aerial vehicle carrying a camera to collect license plate images within the range of the adaptive shooting space model; S30, recognizing the license plate number based on the collected license plate images. The system includes an adaptive shooting space model construction device for constructing an adaptive shooting space model; an unmanned aerial vehicle for carrying a camera to collect license plate images within the range of the adaptive shooting space model; and an image recognition device for recognizing the license plate number based on the collected license plate images. By restricting the unmanned aerial vehicle camera within a specific shooting space model range, the present invention can ensure that the quality of the collected license plate images is high, thereby greatly improving the license plate recognition accuracy and also improving the license plate recognition efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of data recognition, and in particular to a license plate recognition method and system based on unmanned aerial vehicle adaptive shooting space modeling. Background Art

[0002] Although people's awareness of traffic safety is gradually increasing, the phenomenon of randomly parking vehicles still exists. In response to the phenomenon of illegal parking, the current detection method is for traffic police to take photos on the spot to collect evidence, which is inefficient and has also caused the problem of "guerrilla warfare" between car owners and traffic police.

[0003] With the rapid development of drone technology, drone inspections have been widely used in many fields. In view of the low efficiency, high cost, and narrow monitoring area of ​​traditional inspections, which mainly rely on traffic police patrols and fixed camera capture, some scholars have proposed the idea of ​​using drones to inspect and take photos of illegally parked vehicles. For example, Wang Chen et al. proposed a method for urban road parking inspection based on drone images, which improves the all-weather inspection capability through illumination enhancement and deblurring algorithms, and improves YOLOv5 to improve detection accuracy. However, if the shooting distance is too far or the angle is inappropriate, the license plate will be blurred, deformed, or missing details, which will lead to low license plate recognition accuracy. Therefore, the existing methods still have certain limitations in adaptability and stability. Summary of the invention

[0004] The object of the present invention is to provide a license plate recognition method and system based on unmanned aerial vehicle adaptive shooting space modeling, so as to improve the efficiency and accuracy of license plate recognition.

[0005] In order to achieve the above-mentioned object of the invention, the embodiment of the present invention provides the following technical solutions:

[0006] A license plate recognition method based on drone adaptive shooting space modeling includes the following steps:

[0007] S10, constructing an adaptive shooting space model;

[0008] S20, the camera carried by the drone collects license plate images within the range of the adaptive shooting space model;

[0009] S30, recognizing the license plate number based on the collected license plate image.

[0010] Compared with the traditional method of using drones to collect license plate images at random locations, the above solution can greatly improve the quality of license plate images by limiting the drones to carry cameras to collect license plate images within a specific spatial range, thereby improving the accuracy of license plate recognition. Moreover, due to the high quality of license plate images, there is no need to collect them multiple times, thereby improving the efficiency of license plate image collection.

[0011] The S10 comprises the following steps:

[0012] S101. With the length direction of the road as the x-axis, the width direction of the road as the y-axis, and the direction perpendicular to the road plane as the z-axis, construct an ellipsoid model as the adaptive shooting space model;

[0013] The ellipsoid model is , where a is the horizontal distance limit between the UAV and the license plate; b is the lateral offset limit of the UAV; c is the shooting height limit of the UAV; (x i , y i , z i ) is the UAV position coordinate point, (x0, y0, z0) is the coordinate of the ellipsoid center point, z i ≥z0, and x1 and x2 are respectively the nearest horizontal shooting distance and the farthest horizontal shooting distance of the UAV from the license plate.

[0014] Through comparative analysis of multiple models, the ellipsoid model is the most suitable as the shooting space model. That is, in the above solution, by defining the ellipsoid model as the adaptive shooting space model of the UAV, the quality of the collected license plate images can be reliably guaranteed.

[0015] After the S101, there is also a step S102, using the convex hull method to fit the convex hull area with a license plate recognition accuracy higher than the accuracy threshold, and using the minimum volume enclosing ellipsoid method to fit the ellipsoid model.

[0016] In the above solution, fitting the ellipsoid model based on the license plate recognition accuracy and using the minimum volume enclosing ellipsoid method for fitting can narrow the range of the ellipsoid model as much as possible, and then reduce the occlusion of the camera by obstacles as much as possible, improving the shooting efficiency.

[0017] After the S102, there is also a step S103, correcting the ellipsoid model based on the obstacles;

[0018] In the S20, the UAV carries a camera to collect license plate images within the range of the corrected adaptive shooting space model.

[0019] There will inevitably be obstacles such as trees in the vehicle parking area. In the above solution, after initially formulating the shooting space model, correcting the model based on the obstacles, and then arranging the UAV within the corrected space range, this can eliminate the interference of the obstacles and further improve the quality of the license plate images and the shooting efficiency.

[0020] In the S103, the process of correcting the ellipsoid model based on the obstacles includes:

[0021] Define the ray R(t) from the UAV position U to the license plate position P as: where t represents the distance that the ray extends from the UAV, and the direction vector Expressed as: ;

[0022] Detect whether there is an obstacle intersecting with the ray R(t). If not, do not correct the ellipsoid model. If so, calculate the occlusion area of the obstacle, and define the occlusion area of the obstacle as the set of intersection positions of the ray R(t) and the obstacle Q i intersecting, and if there is , then the calculation formula for the corrected shooting space model is: , O st represents the occlusion area of the obstacle.

[0023] a = 6.23, b = 5.15, c = 6.24, and the coordinates of the center point are (6.69, 0, 0.05).

[0024] For the camera used in the experiment, the ellipsoid model defined in the above solution is better, with the smallest spatial range on the basis of ensuring the shooting quality.

[0025] In S30, the YOLOv5-LPRNet model is used to recognize the license plate image.

[0026] The YOLO series algorithms have fast detection speed and high accuracy. The LPRNet license plate recognition algorithm has a highly lightweight network structure, which can meet the requirements of limited hardware resources while ensuring high recognition accuracy. Using YOLOv5-LPRNet for license plate detection and recognition tasks in the above solution can improve the license plate recognition accuracy, occupy less resources, and have high efficiency.

[0027] A license plate recognition system based on adaptive shooting space modeling of an unmanned aerial vehicle (UAV), comprising:

[0028] An adaptive shooting space model construction device for constructing an adaptive shooting space model;

[0029] A UAV for carrying a camera to collect license plate images within the range of the adaptive shooting space model;

[0030] An image recognition device for recognizing the license plate number based on the collected license plate images.

[0031] The adaptive shooting space model is , where a is the horizontal distance limit between the UAV and the license plate; b is the lateral offset limit of the UAV; c is the shooting height limit of the UAV; (x i , y i , z i ) are the coordinate points of the UAV position, (x0, y0, z0) are the coordinates of the center point of the ellipsoid, and z i≥z0, where x1 and x2 are respectively the closest horizontal shooting distance and the farthest horizontal shooting distance of the drone from the license plate.

[0032] The adaptive shooting space model construction device is also used to correct the adaptive shooting space model based on obstacles. Specifically, the drone is used to carry a camera to collect license plate images within the range of the corrected adaptive shooting space model.

[0033] Compared with the prior art, in view of the problem of low license plate recognition accuracy caused by factors such as license plate tilt, blur, inconsistent scale, and obstacle occlusion during the recognition process of ground illegally parked vehicles, the present invention constructs an adaptive ellipsoidal shooting space model considering the relative distance between the drone and the license plate, enabling the drone to flexibly adjust its position within the range of this shooting space to optimize the shooting angle and imaging effect, resulting in high-quality license plate images being collected, thereby greatly improving the license plate recognition accuracy and also enhancing the license plate recognition efficiency.

[0034] Other technical advantages of the present invention are described correspondingly in the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0036] Figure 1 It is a flowchart of a license plate recognition method based on drone adaptive shooting space modeling in the embodiment.

[0037] Figure 2 It is a flowchart of constructing a shooting space model in the embodiment.

[0038] Figure 3 It is a schematic diagram of an ellipsoid model.

[0039] Figure 4 It is a grid schematic diagram.

[0040] Figure 5 It is a three-dimensional surface graph of license plate recognition accuracy.

[0041] Figure 6 It is a two-dimensional surface graph of license plate recognition accuracy at different heights.

[0042] Figure 7 It is a schematic diagram of a three-dimensional convex hull.

[0043] Figure 8 It is a schematic diagram of the minimum volume enclosing ellipsoid.

[0044] Figure 9 It is a schematic diagram of an optimized ellipsoid. Specific implementation manners

[0045] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Usually, the components of the embodiments of the present invention described and shown in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the present invention to be protected, but only represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present invention.

[0046] Please refer to Figure 1 , the license plate recognition method based on the adaptive shooting space modeling of an unmanned aerial vehicle provided in this embodiment includes the following steps:

[0047] S10. Construct an adaptive shooting space model.

[0048] S20. The unmanned aerial vehicle carries a camera to collect license plate images within the range of the adaptive shooting space model.

[0049] S30. Recognize the license plate number based on the collected license plate images.

[0050] At a relatively short distance, the shooting height and left - right offset of the unmanned aerial vehicle are small, and the license plate image is large and clear. As the horizontal distance increases, the shooting range gradually expands, but the proportion of the license plate in the image gradually decreases. When the distance continues to increase, the image details may not be sufficient to ensure high - precision recognition. Therefore, after exceeding a certain critical distance, in order to ensure that the license plate image has enough pixels, the shooting height and left - right offset will tend to contract, so as to meet the accuracy requirements of license plate recognition. That is to say, compared with randomly controlling the flight height and distance of the unmanned aerial vehicle, restricting the unmanned aerial vehicle within a specific space range will help improve the quality of the license plate images collected by the camera carried by the unmanned aerial vehicle, and then improve the accuracy of license plate recognition. This specific space range is the shooting space, which refers to constructing a three - dimensional space for license plate image collection with high recognition accuracy according to the position of the illegally parked vehicle.

[0051] Please refer to Figure 2 , the above step S10 in this embodiment includes the following steps:

[0052] S101. Construct an ellipsoid as the adaptive shooting space model.

[0053] Although a cuboid can provide a large degree of freedom for the drone to take pictures, it lacks effective constraints on the shooting angle, especially when shooting at the top or edge of the cuboid. The hemisphere provides a large degree of freedom in the vertical direction, which helps to adjust the height of the drone. However, the constraints in the horizontal direction are weak, especially in the front-back and left-right directions, which is likely to cause image scale distortion. The cone performs well in terms of the constraints on the shooting angle, but its boundary is relatively single, and the adjustment in the front-back and left-right directions is not flexible enough, making it difficult to provide balanced image quality at all shooting positions. Based on the above analysis, the ellipsoid is the most suitable choice. The three-dimensional space of the ellipsoid can provide sufficient freedom for shooting, and its three axes can effectively control the horizontal distance, lateral offset, and height between the drone and the license plate. At the same time, the ellipsoid has a shrinking trend at the center, which helps to optimize the shooting angle and image quality. Therefore, the ellipsoid is used as the adaptive shooting space model in this embodiment.

[0054] Set the midpoint on one side of the road as the global coordinate system for describing the entire road environment, which is used to determine the positions of vehicles and obstacles. Take the length direction of the road as the x-axis, the width direction of the road as the y-axis, and the direction perpendicular to the road plane as the z-axis.

[0055] Take the center of the license plate as the local coordinate system to define the position of the drone relative to the license plate. In actual shooting, the drone takes pictures at a certain height from the license plate, so only the part where z≥z0 is taken. Assume that the closest horizontal shooting distance of the drone from the license plate is x1, and this distance is determined by the camera focus distance; the farthest horizontal distance is x2, and this distance is determined by the number of pixels required for license plate recognition and the proportion of the license plate target in the image (if the proportion is too small, the target may not be detected). The ellipsoid is truncated at x1 and x2 respectively. The schematic diagram of the shooting space of the ellipsoid is as Figure 3 shown. The rectangular body on the left in the figure represents the vehicle, and the black rectangular area in the rectangular body represents the license plate.

[0056] Assume that the center point of the ellipsoid is O, then the expression of the shooting space G is:

[0057] ;

[0058] In the formula, a is the radius of the horizontal axis, corresponding to the horizontal distance limit between the drone and the license plate; b is the radius of the lateral axis, corresponding to the lateral offset limit of the drone; c is the radius of the vertical axis, corresponding to the shooting height limit of the drone; (x i ,y i ,z i ) is the coordinate point of the drone position. (x0,y0,z0) is the coordinate of the center point of the ellipsoid.

[0059] S102. Fit the convex hull area with a license plate recognition accuracy higher than the accuracy threshold by using the convex hull method, and fit the ellipsoid model by using the minimum volume enclosing ellipsoid method.

[0060] For the evidence collection requirements of illegally parked vehicles, the recognition algorithm not only needs to have high-precision detection capabilities, but also must have a fast processing speed and real-time performance. In this embodiment, YOLOv5-LPRNet is used for license plate detection and recognition tasks. The license plate recognition accuracy can be expressed as: , where P r represents the license plate character recognition accuracy; NC A represents the total number of characters on the license plate; C t represents the number of correctly recognized characters. Let the accuracy threshold be R th , and the set of coordinate points with a recognition accuracy greater than the threshold is screened as M, .

[0061] In this embodiment, the three-dimensional convex hull method is used to construct the smallest convex hull body containing the entire high-recognition-accuracy (i.e., the recognition accuracy is higher than the set value, and the recognition accuracy refers to the ratio of the correctly recognized license plates) point set, effectively determining the outer boundary of the point set, and extracting the distribution of the high-recognition-accuracy point set in space, providing a geometric boundary for subsequent ellipsoid fitting.

[0062] For example, let the set of shooting coordinate points with a recognition accuracy of 100% be screened as , and the convex hull calculation formula is:

[0063] ;

[0064] where represents a convex combination, represents the weight coefficient, satisfying the conditions of being non-negative and having a sum of one.

[0065] When fitting the ellipsoid, it is required to contain the convex hull boundary points without introducing other point sets, so the minimum volume enclosing ellipsoids (MVEE) method is selected for fitting. MVEE is the smallest ellipsoid containing the convex body, and the calculation formula is:

[0066] ; ;

[0067] In the formula: represents the ellipsoid, which is represented by the 3×3 positive definite matrix Q and the center . Among them, the eigenvectors and eigenvalues of Q are the semi-axis length directions and semi-major axis lengths of the ellipsoid, and the semi-major axis direction and semi-major axis length refer to the directions and vectors of the long, medium, and short axes of the ellipsoid respectively. represents the pair of Transpose. d is the scaling factor, and d / 1 determines the degree of ellipsoid scaling.

[0068] To verify the effectiveness of the ellipsoid shooting space model, an actual dataset was collected and experimental analysis was carried out. The goal of the experiment was to examine whether the shooting space defined by the ellipsoid model could describe the relationship between the position coordinates of the UAV and the license plate recognition accuracy in a real environment, and to ensure that the recognition accuracy of the license plate was optimized within the range defined by the shooting space model.

[0069] Since the shooting space without obstacle influence was established first, the data collection was completed in an open-air parking lot scene. In this scene, assuming there were no interference factors such as weather and light, only the spatial relationship between the UAV position and the vehicle parking was considered to verify the universality and effectiveness of the shooting space model. During the shooting, the height, horizontal distance from the license plate, and lateral distance of the UAV were controlled, and the UAV camera was adaptive to always keep the license plate at the imaging center.

[0070] Table 1: Data collection equipment

[0071]

[0072] The parameters of the data collection equipment are shown in Table 1. The focusing distance of the main camera of DJI MAVIC 3 is from 1 meter to infinity, so the starting horizontal shooting distance is 1 meter. As Figure 4 shown, a 0.5m×1m grid was divided every one meter (denoted by d in the figure) directly in front of the license plate, and multiple (for example, 7) images were collected for each grid. The grid corresponding to one x was called one group. The recorded data for each group LPDS was: , then the "camera-license plate" distance dataset can be expressed as: , where n is the value of the horizontal distance x.

[0073] This experiment was based on the Windows environment, using the PyTorch framework, combining the CCPD public dataset and the license plate images collected by the UAV, and pre-trained the YOLOv5-LPRnet model. Finally, the YOLOv5-LPRnet model achieved a recognition accuracy of 99.3%. The pre-trained model was used to perform license plate localization and character recognition on the CCLP dataset (license plate images collected at various angles and heights), and the recognition accuracy of each shooting coordinate point was counted.

[0074] The UAV coordinate points and the corresponding recognition accuracy values were mapped into a three-dimensional surface graph and a two-dimensional plan graph, as shown in Figure 5 and Figure 6 shown respectively. Figure 5Shows the distribution of license plate recognition accuracy in space. The results indicate that the recognition accuracy reaches 100% in certain areas, but as the horizontal distance increases (e.g., after x = 11 meters), the recognition accuracy drops significantly and the 100% accuracy area shrinks significantly. Figure 6 Shows the distribution of license plate recognition accuracy at different heights (z = 0.55, 2.55, 4.55, 6.55). From the comparison at different heights, it can be seen that as the height increases, the high-precision recognition area begins to shrink.

[0075] From Figure 5 and 6 it is obtained that the points with 100% recognition accuracy show a certain regional distribution. The distribution of data points with 100% recognition accuracy on the x-axis is from 1 meter to 12 meters, and the y and z axes also expand first and then approach the central axis. Calculate the convex hull based on the coordinate point set with 100% recognition accuracy to obtain a three-dimensional convex hull, as Figure 7 shown. The convex hull boundary can be used as the boundary of the actual shooting space model in theory. Use the MVEE algorithm to fit the boundary points of the convex hull to obtain an approximate ellipsoid of the convex hull, as Figure 8 shown.

[0076] Use the coverage rate as an index to measure the effect of the fitted ellipsoid approaching the three-dimensional convex hull. It represents the ratio of half of the volume of the ellipsoid covering the three-dimensional convex hull volume v s to the three-dimensional convex hull volume v g , denoted as , . Divide the three-dimensional convex hull into n triangular pyramids, and denote their volumes as v g , , a i is the base area of the triangular pyramid, u i is the height of the triangular pyramid, and the covering volume v s is calculated using the Monte Carlo sampling method.

[0077] For the camera used in Table 1, the center point coordinates of the initially fitted ellipsoid are x0 = 6.69, y0 = 0, z0 = 0.05, and the a, b, and c axes are 6.23, 6.15, and 6.74 respectively, and the volume is 540.85. However, the ellipsoid model does not completely coincide with the outer boundary of the convex hull, especially there is a certain gap in the c-axis direction, indicating that the fitting result does not fully capture the spatial characteristics of the scattered points and there is still invalid space.

[0078] To improve the fitting accuracy, the semi-axis parameters of the ellipsoid are further optimized. Keep the center point and the major semi-axis (a-axis) of the ellipsoid unchanged, gradually reduce the lengths of the minor semi-axis (b-axis) and the c-axis by 0.5 meters each time, and judge the optimization effect by observing the redundant space between the ellipsoid model and the convex hull and calculating the coverage rate until the redundant space between the ellipsoid model and the convex hull approaches 0.

[0079] The adjusted parameters are a = 6.23, b = 5.15, c = 6.24, and the volume of the ellipsoid is 419.31. After optimization, the ellipsoid is truncated at x = 1 and x = 12. The half-ellipsoid after truncation closely adheres to the three-dimensional convex hull, reducing the introduction of invalid space, as Figure 9 shown, improving the spatial adaptability and accuracy of the license plate recognition system in practical applications.

[0080] There may be trees or other objects around the road where the vehicle is parked. Therefore, after the drone enters the shooting space, there may be obstacles at certain positions that block the shooting line of sight. Therefore, after constructing the shooting space model, it is necessary to correct the shooting space model based on the obstacles. That is, in the above step S10, step S103 may further be included to correct the fitted ellipsoidal shooting space model based on the obstacles.

[0081] In this embodiment, the Ray Casting method is used to detect whether this ray intersects with any obstacles. If there is an intersection point, the line of sight is blocked; otherwise, the line of sight is unobstructed. The ray R(t) pointing from the drone position U to the license plate position P can be expressed as: , where the direction vector is expressed as: , and t represents the distance that the ray extends from the drone.

[0082] Use the Slab Method intersection detection algorithm to check whether the ray R(t) intersects with the obstacle O i intersects. For the x / y / z axes, calculate the intersection points of the ray and the obstacle on this axis , , and the calculation formula is:

[0083] ;

[0084] In the formula, , respectively represent the minimum boundary and the maximum boundary of the obstacle on this axis; represents the coordinate of the ray starting point on this axis; represents the component of the direction vector on this axis.

[0085] If , it means that the component of the direction vector of the ray on a certain axis is negative, that is, the moving direction on this axis is opposite to the positive direction of the coordinate axis. The originally calculated time to enter the obstacle may be greater than the time to leave, which does not conform to the actual situation because the entry time should be earlier than the departure time. Therefore, exchange these two values to ensure that the entry time is always less than or equal to the departure time. Comprehensively calculate the time t enter to enter the obstacle and the time t exit, the calculation formula is: , then there is: . That is, by calculating the entry and exit times, it is determined whether the obstacle blocks the ray.

[0086] Define the occlusion area of the obstacle as the set of intersection positions of the ray R(t) and the obstacle O i intersecting, there is , then the calculation formula of the corrected shooting space model is: .

[0087] That is to say, if the obstacle blocks the ray, then according to the time when the ray enters and leaves each boundary of the obstacle, the intersection points of the ray at each boundary of the obstacle are calculated, and the relatively occluded space of the obstacle is calculated based on the obtained intersection points. Then, by removing this occluded space, the corrected shootable space model can be obtained. Taking the x-axis boundary as an example, the coordinates of the obstacle entering and leaving the boundary can be expressed as:

[0088] ;

[0089] ;

[0090] According to the above formula, the intersection points of each boundary can be obtained, and connecting these intersection points can obtain the occluded space.

[0091] This embodiment also provides a license plate recognition system based on UAV adaptive shooting space modeling, including an adaptive shooting space model construction device, a UAV equipped with a camera, and an image recognition device.

[0092] The adaptive shooting space model construction device is used to construct an adaptive shooting space model. The adaptive shooting space model construction device is a software system, integrated in a computer to implement, and executes the foregoing steps S101-S103.

[0093] The UAV is used to carry the camera to collect license plate images within the range of the adaptive shooting space model. In actual applications, after the UAV is arranged within the range of the adaptive shooting space model, it is also necessary to detect whether there are obstacles blocking in this space. If there are obstacles blocking, the adaptive shooting space model construction device will also correct the adaptive shooting space model, and then arrange the UAV within the range of the corrected adaptive shooting space model for license plate image collection, so as to avoid affecting license plate recognition due to unclear license plate images caused by obstacles.

[0094] The image recognition device is used to recognize the license plate number based on the collected license plate images. The image recognition device is a software system, integrated in a computer to implement, and executes the training steps and recognition steps of the YOLOv5-LPRNet model. Of course, the image recognition device can also use other models to recognize license plates.

[0095] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A license plate recognition method based on UAV adaptive shooting space modeling, characterized in that: The following steps are involved: S10, constructing an adaptive shooting space model; S10 includes the following steps: S101, constructing an ellipsoid model as the adaptive shooting space model with the length direction of the road as the x-axis, the width direction of the road as the y-axis, and the direction perpendicular to the road plane as the z-axis; The ellipsoid model is , a is the horizontal distance limit between the drone and the license plate; b is the lateral offset limit of the drone; c is the shooting height limit of the drone; (x i ,y i ,z i ) is the coordinate point of the drone position, (x0, y0, z0) is the coordinate point of the ellipsoid center, z i ≥z0, x1 and x2 are the shortest and farthest horizontal shooting distances of the drone from the license plate respectively; After S101, the method further includes step S102, using a convex hull method to fit a convex hull area where the license plate recognition accuracy is higher than an accuracy threshold, and using a minimum volume enclosing ellipsoid method to fit the ellipsoid model; After S102, the method further includes step S103 of correcting the ellipsoid model based on obstacles; S20, the camera carried by the drone collects the license plate image within the range of the modified adaptive shooting space model; S30, recognizing the license plate number based on the collected license plate image.

2. The license plate recognition method based on UAV adaptive shooting space modeling according to claim 1 is characterized in that: In S103, the process of correcting the ellipsoid model based on the obstacle includes: The ray R(t) pointing from the drone position U to the license plate position P is defined as: , where t represents the distance the ray extends from the drone and the direction vector It is expressed as: ; Detect whether there is an obstacle intersecting with the ray R(t). If not, do not modify the ellipsoid model. If so, calculate the occlusion area of ​​the obstacle and define the occlusion area of ​​the obstacle as the intersection of the ray R(t) and the obstacle Q. i The set of intersection positions has , then the calculation formula of the modified shooting space model is: .

3. The license plate recognition method based on UAV adaptive shooting space modeling according to claim 1 is characterized in that: a=6.23, b=5.15, c=6.24, and the coordinates of the center point are (6.69,0,0.05).

4. The license plate recognition method based on UAV adaptive shooting space modeling according to claim 1 is characterized in that: In S30, the YOLOv5-LPRNet model is used to recognize the license plate image.

5. A license plate recognition system based on UAV adaptive shooting space modeling, characterized in that: include: The adaptive shooting space model construction device is used to construct an adaptive shooting space model, use the convex hull method to fit a convex hull area with a license plate recognition accuracy higher than the accuracy threshold, use the minimum volume enclosing ellipsoid method to fit the adaptive shooting space model, and modify the adaptive shooting space model based on obstacles; the adaptive shooting space model is , a is the horizontal distance limit between the drone and the license plate; b is the lateral offset limit of the drone; c is the shooting height limit of the drone; (x i ,y i ,z i ) is the coordinate point of the drone position, (x0, y0, z0) is the coordinate point of the ellipsoid center, z i ≥z0, x1 and x2 are the shortest and farthest horizontal shooting distances of the drone from the license plate respectively; A drone, used to carry a camera to collect license plate images within the modified adaptive shooting space model; The image recognition device is used to recognize the license plate number based on the collected license plate image.

Images