Unmanned aerial vehicle video positioning method and device fusing artificial intelligence and beidou navigation

Abstract

This application discloses a method and apparatus for UAV video positioning that integrates artificial intelligence and BeiDou navigation, relating to the field of video positioning technology. The method includes: acquiring an initial target image; measuring the straight-line distance between the UAV and the target; setting a spherical region centered on the UAV's current position, establishing an objective function with the goals of minimizing the total flight time of the UAV to all sampling points and optimizing the uniformity of the viewing angles of all sampling points, and solving for multiple sampling points; acquiring a sampled target image at each sampling point; determining the preliminary geodetic coordinates of the target corresponding to each sampling point; calculating the weight of each sampling point based on the sharpness score of the sampled target image and the preliminary geodetic coordinates; and performing a weighted summation of all the preliminary geodetic coordinates according to the weights to obtain the final coordinates of the target. This application can obtain relatively accurate target coordinates in a short time, providing a feasible solution for the need to use UAVs for video positioning when mobile space is limited.

Description

Technical Field

[0001] This application relates to the field of video positioning technology, and in particular to a method for video positioning that integrates artificial intelligence and BeiDou navigation technology. Background Technology

[0002] Video positioning is a technology that identifies the location of a drone or a target in its vicinity by recognizing captured images. Video positioning technology can accurately locate targets when drones cannot rely on satellite positioning or when satellite positioning cannot directly obtain the target's location, and therefore it is being used in various military and civilian applications.

[0003] Currently, most video positioning technologies primarily target situations where drones can hover above a target. This method allows for quick and accurate target location once the drone has established its own position. However, this approach is unsuitable for certain scenarios. For example, when obstructed by buildings, mountains, or power towers, if the target is very close to these obstacles, or even inside them, the drone cannot reach directly above the target. In such cases, the drone needs to use oblique photography to capture images of the target for analysis. Furthermore, if there are other obstacles nearby, such as trees or other buildings, the range of adjustment available for the drone during image acquisition becomes extremely limited. Under these limited and fixed-direction conditions, existing video positioning technologies cannot function effectively. Summary of the Invention

[0004] This application provides a UAV video positioning method and apparatus that integrates artificial intelligence and BeiDou navigation, in order to solve the problem that existing video positioning methods cannot be applied to small-scale and fixed-direction scenarios.

[0005] On the one hand, embodiments of this application provide a UAV video positioning method integrating artificial intelligence and BeiDou navigation, including:

[0006] Initial target images were acquired by a drone using oblique photography.

[0007] Identify the approximate location of the target in the initial target image, and orient the laser rangefinder toward the approximate location to measure the straight-line distance between the UAV and the target;

[0008] Centered on the current position of the drone, a spherical region with a set length as its diameter is set. An objective function is established with the goal of minimizing the total flight time of the drone to all sampling points and optimizing the uniformity of the viewpoints of all sampling points. Solving the objective function yields multiple sampling points on the spherical surface of the spherical region. The uniformity of viewpoints is the standard deviation of the angle between the viewpoints of all sampling points and the minimum angle. The minimum angle is determined based on the straight-line distance.

[0009] The drone sequentially collects images of the target at each sampling point, while simultaneously acquiring the BeiDou navigation and positioning error at each sampling point; when collecting the target images, the drone adjusts the camera's attitude to center the target in the image.

[0010] A coordinate transformation method based on a camera projection model is adopted to determine the preliminary geodetic coordinates of the target corresponding to each sampling point based on the sampled target image;

[0011] Using deep reinforcement learning, the weight of each sampling point is calculated based on the sharpness score of the sampled target image, the preliminary geodetic coordinates, and the BeiDou navigation positioning error.

[0012] The final coordinates of the target are obtained by weighted summation of all preliminary geodetic coordinates according to their weights.

[0013] On the other hand, embodiments of this application also provide a drone video positioning device that integrates artificial intelligence and BeiDou navigation, including:

[0014] A camera used to acquire initial target images via oblique photography when carried by a drone;

[0015] The data processing unit is used to identify the approximate location of the target in the initial target image, and to orient the laser rangefinder toward the approximate location to measure the straight-line distance between the UAV and the target. A spherical region is set with the current position of the UAV as the center and a set length as the diameter. An objective function is established with the goal of minimizing the total flight time of the UAV to all sampling points and optimizing the uniformity of the viewing angles of all sampling points. The objective function is solved to obtain multiple sampling points on the spherical surface of the spherical region. The uniformity of the viewing angles is the standard deviation of the angles between all sampling points from the minimum angle. The minimum angle is determined based on the straight-line distance.

[0016] The camera sequentially acquires images of the target at each sampling point, while simultaneously obtaining the BeiDou navigation positioning error at each sampling point; when acquiring images of the target, the UAV adjusts the camera's attitude to center the target in the image.

[0017] The data processing unit adopts a coordinate transformation method based on the camera projection model to determine the preliminary geodetic coordinates of the target corresponding to each sampling point based on the sampled target image; using a deep reinforcement learning method, it calculates the weight of each sampling point based on the sharpness score of the sampled target image, the preliminary geodetic coordinates, and the BeiDou navigation positioning error; and performs a weighted summation of all the preliminary geodetic coordinates based on the weights to obtain the final coordinates of the target.

[0018] The UAV video positioning method and apparatus integrating artificial intelligence and BeiDou navigation in this application have the following advantages:

[0019] A video localization method was specifically designed for scenarios with limited range. By optimizing the selection of sampling points, image acquisition, and coordinate fusion based on deep reinforcement learning technology, relatively accurate target coordinates can be obtained in a short time, providing a feasible solution for the need to use drones for video localization when mobile space is limited. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating the UAV video positioning method integrating artificial intelligence and BeiDou navigation provided in this application embodiment. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Figure 1 A flowchart illustrating the UAV video positioning method integrating artificial intelligence and BeiDou navigation provided in this application embodiment. This application embodiment provides a UAV video positioning method integrating artificial intelligence and BeiDou navigation, including:

[0024] The S100 uses oblique photography to capture initial target images by a drone.

[0025] For example, before acquiring the initial target image, some preparatory work needs to be done, which mainly includes hardware configuration and core parameter calibration. These preparatory work will be described in detail below.

[0026] Hardware configuration. The UAV is equipped with a Beidou-3 RTK (Real-time Dynamic Carrier Phase Differential Technology) module (planar accuracy ≤2cm, elevation accuracy ≤5cm), a three-axis stabilized gimbal, an industrial camera, an edge computing terminal, and is additionally equipped with a laser rangefinder (range 0.5m~100m, error ±0.1m).

[0027] Core parameter calibration. Camera intrinsic parameter matrix: obtained using Zhang Zhengyou calibration method, describing the mapping relationship between camera pixel coordinates and physical coordinates. K The formula is:

[0028]

[0029] in, f The focal length of the camera (unit: mm), such as 8mm, which is commonly used in industrial cameras. d This refers to the pixel size of the camera (unit: mm / px), such as a 1 / 2.3-inch sensor with a pixel size of approximately 1.12μm (i.e., mm / px). u 0 and v 0 represents the principal pixel coordinates of the camera (unit: pixels), i.e., the position of the center pixel of the image, such as in a 1920×1080 resolution image. u 0 = 960px v 0 = 540px.

[0030] BeiDou-Camera Coordinate Transformation Matrix T U2C The rigid transformation matrix (3×3 rotation matrix + 3×1 translation vector, a total of 12 parameters) is used to convert the UAV's BeiDou coordinate system (CGCS2000 geodetic coordinate system, unit: m) into the camera coordinate system (right-handed system: X-axis forward, Y-axis to the right, Z-axis downward, unit: m), which is obtained by taking pictures of three calibration points with known geodetic coordinates.

[0031] UAV motion parameters: including maximum translation speed and turning time coefficient. v max The maximum translation speed (unit: m / s) for small-range fine-tuning is set to 0.5 m / s in this embodiment (to avoid attitude fluctuations caused by rapid movement). k turn The steering time coefficient (unit: s / °) is set to 0.1s / ° in this embodiment (the larger the steering angle, the longer the time).

[0032] S110 identifies the approximate location of the target in the initial target image and directs the laser rangefinder toward the approximate location to measure the straight-line distance between the UAV and the target.

[0033] For example, the drone hovers at its current location. P init ( X init , Y init , Z init (Real-time geodetic coordinates output by BeiDou navigation, unit: m), the straight-line distance between the UAV and the target is measured using a laser rangefinder. D (Unit: m) If the drone is not equipped with a laser rangefinder, the approximate location of the target can be determined through BeiDou navigation. Qpre ( X pre , Y pre , Z pre ), calculate using the two-point distance formula:

[0034] .

[0035] S120: A spherical region is set with the current position of the UAV as the center and a set length as the diameter. An objective function is established with the goal of minimizing the total flight time of the UAV to all sampling points and optimizing the uniformity of the viewing angle of all sampling points. Solving the objective function yields multiple sampling points on the spherical surface of the spherical region. The uniformity of the viewing angle is the standard deviation of the angle between all sampling points and the minimum angle. The minimum angle is determined based on the straight-line distance.

[0036] For example, the modeling process for the spherical region is as follows: based on the rough location of the target... Q pre The vertex represents the current position of the drone. P init Using the center of the sphere as the reference point and setting the length, this application selects 1m as the diameter to construct the sphere, and all sampling points... P i All are located on the surface of the sphere or inside the sphere (i.e. P i arrive P init The distance is ≤0.5m.

[0037] Any two sampling points P i , P j With the goal Q pre The angle of view formed α i,j (Unit: rad), its minimum value, i.e., the minimum included angle. α min (Dynamic constraint threshold, unit: rad) is determined by the straight-line distance. D The decision is made using the following formula:

[0038]

[0039] Summarized as follows:

[0040]

[0041] in, α min The minimum included angle, also used as the dynamic viewpoint constraint threshold, means that the angle between any two sampling points and the target must be ≥ αmin ; L max This represents the maximum straight-line distance (in meters) between two sampling points within a 1-meter range. Since the range is a 1-meter sphere, the maximum distance is equal to the diameter of the sphere. L max =1m; D The straight-line distance between the UAV and the target (unit: m).

[0042] Furthermore, if D > 57m, then α min <1° (because 2arcsin(1 / (2×57))≈1°), to avoid excessive overlap of viewpoints, forced α min =1° (approximately 0.017 rad); if D < 1m (close-up scene), then α min =45° (approximately 0.785 rad). Since a large angle can be formed within a 1m range, no dynamic adjustment is required.

[0043] The primary objective of the objective function is the total flight time. T total (Unit: s) The shortest time is the secondary objective, and the uniformity of the viewpoint is the best.

[0044] Total flight time T total (Unit: s) Shortest, the formula is:

[0045]

[0046] in, T trans The total translation time (in seconds) of the UAV between sampling points is given by the formula:

[0047]

[0048] Where 5 represents the number of sampling points pre-set in this embodiment of the application. d 0,1 for P init To the first sampling point P A distance of 1 d i-1,i ( i ≥2) is the first i -1 sampling point to the i The distance between each sampling point v max This is the maximum translational speed of the drone (unit: m / s), which is 0.5 m / s.

[0049] Tturn The total turning time (in seconds) of the UAV between sampling points is given by the formula:

[0050]

[0051] in, k turn This is the turning time coefficient (unit: s / °), i.e., 0.1 s / °. For the first i The azimuth angle of each sampling point (unit: °, relative to) P init (horizontal angle) The azimuth difference between adjacent sampling points (unit: °) indicates that the larger the difference, the longer the turning time.

[0052] The optimal uniformity of the viewing angle is achieved by considering the angle between all sampling points and... α min standard deviation s α (Unit: rad) measurement s α The smaller the value, the more uniform the viewing angle. The formula is:

[0053]

[0054] in, C (5,2)=10 represents the number of combinations of selecting 2 out of 5 sampling points. α i,j For the first i , j The angle between each sampling point and the target (unit: rad).

[0055] The process of solving the objective function using a genetic algorithm is as follows:

[0056] Variable encoding: each sampling point P i Using spherical coordinates ( r i , i i , ) Description (relative to Pinit), where, r i For sampling points to P init The radial distance (in meters), ranging from [0,1]. i i The polar angle of the sampling point (unit: rad), which is the angle with the vertical direction (i.e., the Z-axis), in the range [0, π] (0 is directly above, π is directly below). The azimuth angle of the sampling point (unit: rad) is the horizontal angle with due north, ranging from [0, 2π] (increasing clockwise).

[0057] Population initialization: Generate 50 initial chromosomes (each chromosome corresponds to the spherical coordinates of 5 sampling points + flight order), which must satisfy: ① r i ≤0.5m; ②Any α i,j ≥ α min If the conditions are not met, the system will regenerate.

[0058] Fitness function and termination condition:

[0059] Fitness function (comprehensively evaluates the merits of the sampling point scheme; a larger value indicates better performance):

[0060]

[0061] in, Fit The fitness value is 0.6, the time weight coefficient is 0.4, and the view weight coefficient is 0.4. Both of these values ​​were determined through offline verification. F t Let be the time fitness, and be the normalized value of the total flight time. The formula is:

[0062]

[0063] F v For the uniformity of visual field fitness, the formula is:

[0064]

[0065] Termination condition: After 50 iterations, or if the maximum value of fitness changes by ≤0.01 for 10 consecutive generations, output the chromosome with the highest fitness, corresponding to the optimal set of sampling points { P 1, P 2,..., P 5} (Geometric coordinates, unit: m) and flight sequence P init → P 1→...→ P 5.

[0066] S130: The UAV sequentially collects images of the target at each sampling point, while simultaneously acquiring the BeiDou navigation and positioning error at each sampling point; when collecting the target images, the UAV adjusts the camera's attitude to center the target in the target image.

[0067] For example, drones v maxThe sampler moves to the sampling point at a speed of 0.5 m / s. P i After hovering and stabilizing (position fluctuation ≤ 0.1m), record the geodetic coordinates output by BeiDou RTK. X i , Y i , Z i (Unit: m) and BeiDou navigation and positioning error s i (Unit: m).

[0068] Use an IMU (Inertial Measurement Unit) to obtain the drone's body attitude angles. α body,i , β body,i , c body,i (Unit: °), where α body,i The pitch angle of the aircraft (forward pitch is positive, backward pitch is negative). β body,i This is the roll angle of the aircraft (rolling to the right is positive, rolling to the left is negative). c body,i This is the aircraft's heading angle (clockwise is positive, counterclockwise is negative).

[0069] Fine-tune the tilt angle of the camera mount. α gimbal,i (Unit: °, downward is positive), ensure the target is located in the center region of the image, and calculate the actual camera attitude angle (unit: °):

[0070]

[0071]

[0072] Where, Δ α 0, Δ β 0, Δ c 0 represents the fixed attitude offset angle between the gimbal and the device (unit: °), obtained through static calibration on a horizontal ground. α i , β i , c i The final attitude angles of the camera (unit: °) correspond to pitch, roll, and yaw, respectively.

[0073] Capture a single frame of sampled target image and calculate the image sharpness score using the Laplacian operator. C i (Range 0~1), the formula is:

[0074]

[0075] in, M × N To sample the resolution of the target image, For the Laplace operator, I ( x , y ) represents the coordinates in the sampled target image. x , y The pixel value at () To take the absolute value, I ( x , y ) is pixels ( x , y The Laplacian value of the image is used (a larger value indicates a clearer image), and the target is identified using the YOLOv8 model, outputting the pixel coordinates of the target's center point. u i , v i (Unit: px).

[0076] S140 uses a coordinate transformation method based on a camera projection model to determine the preliminary geodetic coordinates of the target corresponding to each sampling point based on the sampled target image.

[0077] For example, a coordinate transformation method based on a camera projection model includes:

[0078] Convert the pixel coordinates of the target in the sampled target image to camera coordinates in the camera coordinate system;

[0079] The camera coordinates are converted to geodetic coordinates in the geodetic coordinate system to obtain preliminary geodetic coordinates.

[0080] Specifically, the process of converting pixel coordinates to camera coordinate vectors is as follows:

[0081] Let the target's coordinates in the camera coordinate system be... Q C,i ( X C,i , Y C,i , Z C,i (Unit: m), where Z C,i The vertical distance from the camera to the target (initially set to...) D (Unit: m), according to the camera intrinsic model, the relationship between pixel coordinates and camera coordinates is:

[0082]

[0083] The inverse kinematics yields the target coordinates in the camera coordinate system (unnormalized):

[0084]

[0085] in, u i and v i The target center point is at the th i The pixel coordinates (unit: px) of the sampled target image at each sampling point. u 0、 v 0、 f and d For camera internal parameters, X C,i , Y C,i and Z C,i For the goal in the i The coordinates of each sampling point in the camera coordinate system (unit: m). Z C,i Initially take D (Initial straight-line distance between the drone and the target).

[0086] The process of transforming the camera coordinate system to the geodetic coordinate system is as follows:

[0087] via camera attitude rotation matrix R i (Depend on α i , β i , c i Construction and transformation matrix T U2C ,Will Q C,i Preliminary coordinates converted to geodetic coordinate system Q i ( X Q,i , Y Q,i , Z Q,i (Unit: m), the formula is:

[0088]

[0089] in, R i This is the camera attitude rotation matrix (3×3), used to convert the camera coordinate system vector into the UAV body coordinate system vector. R α ( αi ( ) is the pitch angle α i The corresponding rotation matrix:

[0090]

[0091] R β ( β i () is the roll angle β i The corresponding rotation matrix:

[0092]

[0093] R γ ( c i ( ) is the heading angle c i The corresponding rotation matrix:

[0094]

[0095] T U2C This is the BeiDou-camera coordinate transformation matrix (3×4 dimensions, rotation + translation, unit: m). X i , Y i , Z i For the first i The BeiDou geodetic coordinates (unit: m) of each sampling point. X Q,i , Y Q,i , Z Q,i For the goal in the i Preliminary geodetic coordinates (unit: m) for each sampling point.

[0096] S150 uses deep reinforcement learning to calculate the weight of each sampling point based on the sharpness score of the sampled target image, preliminary geodetic coordinates, and BeiDou navigation positioning error.

[0097] For example, deep reinforcement learning methods include the following:

[0098] state space S (15-dimensional): Input 5 sampling points Q i ( X Q,i , Y Q,i , ZQ,i (Unit: m) s i (Unit: m) C i There are a total of 5 × 3 = 15 dimensions.

[0099] Action space A (10-dimensional): Output the weights of 5 sampling points w i The distance correction value Δ from the camera to the target Z C,i (Unit: m, range ±2m), a total of 5+5=10 dimensions.

[0100] reward function R (A higher value indicates a better action):

[0101]

[0102] in, To obtain the corrected geodetic coordinates of the target (in meters), Z C,i Replace with Z C,i +Δ Z C,i Recalculate using the formula from step S140. The standard deviation (in meters) of the five corrected coordinates reflects coordinate consistency; a smaller value is better. s i For the first i The BeiDou positioning error (unit: m) at each sampling point comes from the data collection results of step S130. w i For the first i The weight of each sampling point, the larger the weight... s i / w i The smaller the value, the lower the impact of the error. 0.7 and 0.3 are weighting coefficients, which are determined through offline training (prioritizing coordinate consistency).

[0103] Model reasoning: The current state s Input a pre-trained PPO (Proximal Policy Optimization) model (trained offline with 1000 sets of "sampled data - real coordinates", convergence error ≤ 0.05), output the optimal action. a ( w i With Δ Z C,i ).

[0104] S160: Weighted summation of all preliminary geodetic coordinates is performed according to the weights to obtain the final coordinates of the target.

[0105] For example, by w i right We obtain the final coordinates of the target by weighted summation. Q final (Unit: m):

[0106]

[0107] in, w i The sampling point weights are the output of DRL (Deep Reinforcement Learning). The corrected target coordinates (unit: m).

[0108] If the drone is equipped with a laser rangefinder, it measures the actual distance from the camera to the target. Z real (Unit: m), Verification:

[0109]

[0110] If the above formula is satisfied, then repeat the positioning three times consecutively and calculate three times. Q final The standard deviation is taken as follows: if the standard deviation is ≤0.3m, the accuracy is considered to be up to standard, and the output is determined. Q final If the standard is not met, repeat steps S120-160.

[0111] This application also provides a drone video positioning device that integrates artificial intelligence and BeiDou navigation, the device comprising:

[0112] A camera used to acquire initial target images via oblique photography when carried by a drone;

[0113] The data processing unit is used to identify the approximate location of the target in the initial target image, and to orient the laser rangefinder toward the approximate location to measure the straight-line distance between the UAV and the target. A spherical region is set with the current position of the UAV as the center and a set length as the diameter. An objective function is established with the goal of minimizing the total flight time of the UAV to all sampling points and optimizing the uniformity of the viewing angles of all sampling points. The objective function is solved to obtain multiple sampling points on the spherical surface of the spherical region. The uniformity of the viewing angles is the standard deviation of the angles between all sampling points from the minimum angle. The minimum angle is determined based on the straight-line distance.

[0114] The camera sequentially acquires images of the target at each sampling point, while simultaneously obtaining the BeiDou navigation positioning error at each sampling point; when acquiring images of the target, the UAV adjusts the camera's attitude to center the target in the image.

[0115] The data processing unit adopts a coordinate transformation method based on the camera projection model to determine the preliminary geodetic coordinates of the target corresponding to each sampling point based on the sampled target image; using a deep reinforcement learning method, it calculates the weight of each sampling point based on the sharpness score of the sampled target image, the preliminary geodetic coordinates, and the BeiDou navigation positioning error; and performs a weighted summation of all the preliminary geodetic coordinates based on the weights to obtain the final coordinates of the target.

[0116] For example, the data processing unit is deployed in an edge computing terminal.

[0117] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0118] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for positioning a UAV video by fusing artificial intelligence and Beidou navigation, characterized in that, include: Initial target images were acquired by a drone using oblique photography. Identify the approximate location of the target in the initial target image, and orient the laser rangefinder toward the approximate location to measure the straight-line distance between the UAV and the target; Centered on the current position of the UAV, a spherical region with a set length as its diameter is set. An objective function is established with the goal of minimizing the total flight time of the UAV to all sampling points and optimizing the uniformity of the viewing angles of all the sampling points. Solving the objective function yields multiple sampling points on the spherical surface of the spherical region. The uniformity of the viewing angles is the standard deviation of the angles between all the sampling points from the minimum angle. The minimum angle is determined based on the straight-line distance. The UAV sequentially acquires target images at each of the sampling points, and simultaneously obtains the BeiDou navigation positioning error at each sampling point; when acquiring the target images, the UAV adjusts the camera's attitude so that the target is centered in the target image; A coordinate transformation method based on a camera projection model is used to determine the preliminary geodetic coordinates of the target corresponding to each sampling point based on the sampled target image; Using deep reinforcement learning, the weight of each sampling point is calculated based on the sharpness score of the sampled target image, the preliminary geodetic coordinates, and the BeiDou navigation positioning error. The final coordinates of the target are obtained by weighted summation of all the preliminary geodetic coordinates according to the weights. The minimum included angle is expressed as: wherein α min is the minimum angle, D is the straight line distance, L max is the diameter of the sphere region.

2. The UAV video positioning method integrating artificial intelligence and BeiDou navigation according to claim 1, characterized in that, If the straight-line distance D If the angle is greater than 57m, then the minimum included angle is... α min The value is set to 1° if the straight-line distance is... D If the angle is less than 1m, then the minimum included angle is... α min The value is set to 45°.

3. The UAV video positioning method integrating artificial intelligence and BeiDou navigation according to claim 1, characterized in that, The objective function is solved using a genetic algorithm, and the fitness function for solving the objective function is expressed as follows: in, Fit For fitness value, F t For time-fitness, the value is the total flight time of the UAV to all the sampling points. T total The reciprocal of is expressed as: F v The uniformity of viewpoint fitness is expressed as: in, The standard deviation of the angle between the viewing angles of all the sampling points and the minimum angle is given.

4. The UAV video positioning method integrating artificial intelligence and BeiDou navigation according to claim 3, characterized in that, The total flight time T total Represented as: in, T trans The total translation time for the UAV to move between all the sampling points is expressed as: in, d i-1,i For the first i -1 of the sampling points and the first i The straight-line distance between the sampling points v max This represents the maximum translational speed of the drone; T turn The total turning time of the UAV between all the sampling points is expressed as: in, k turn For turning time coefficient, φ i For the first i The azimuth angle of each of the sampling points. To take the absolute value.

5. The UAV video positioning method integrating artificial intelligence and BeiDou navigation according to claim 3, characterized in that, The standard deviation of the deviation Represented as: in, C (5,2) represents the number of combinations of selecting 2 points from the 5 sampling points. α i,j For the first i , j The angle between the sampling points and the viewpoint of the target. α min The minimum included angle is denoted as .

6. The UAV video positioning method integrating artificial intelligence and BeiDou navigation according to claim 1, characterized in that, The clarity score C i Represented as: in, M × N The resolution of the sampled target image. For the Laplace operator, I ( x , y ) is the coordinate in the sampled target image ( x , y The pixel value at () To take the absolute value.

7. The UAV video positioning method integrating artificial intelligence and BeiDou navigation according to claim 1, characterized in that, Coordinate transformation methods based on camera projection models include: Convert the pixel coordinates of the target in the sampled target image to camera coordinates in the camera coordinate system; The camera coordinates are converted into geodetic coordinates in the geodetic coordinate system to obtain the preliminary geodetic coordinates.

8. A UAV video positioning device integrating artificial intelligence and BeiDou navigation, wherein the device applies the method described in any one of claims 1-7, characterized in that, include: A camera used to acquire initial target images via oblique photography when carried by a drone; A data processing unit is used to identify the approximate location of the target in the initial target image, and to orient the laser rangefinder toward the approximate location to measure the straight-line distance between the UAV and the target; to establish a spherical region with the current location of the UAV as the center and a set length as the diameter, and to establish an objective function with the shortest total flight time of the UAV to all sampling points and the optimal uniformity of the viewing angle of all the sampling points as the objectives, and to solve the objective function to obtain multiple sampling points on the spherical surface of the spherical region, wherein the uniformity of the viewing angle is the standard deviation of the angle between all the sampling points and the minimum angle, and the minimum angle is determined based on the straight-line distance; The camera sequentially acquires the target image at each sampling point, and simultaneously obtains the BeiDou navigation positioning error at each sampling point; when acquiring the target image, the UAV adjusts the camera's attitude so that the target is centered in the target image; The data processing unit employs a coordinate transformation method based on a camera projection model to determine the preliminary geodetic coordinates of the target corresponding to each sampling point based on the sampled target image; and uses a deep reinforcement learning method to calculate the weight of each sampling point based on the sharpness score of the sampled target image, the preliminary geodetic coordinates, and the BeiDou navigation positioning error. The final coordinates of the target are obtained by weighted summation of all the preliminary geodetic coordinates according to the weights.

9. The UAV video positioning device integrating artificial intelligence and BeiDou navigation according to claim 8, characterized in that, The data processing unit is deployed on an edge computing terminal.

Images