An autonomous landing method for unmanned aerial vehicles (UAVs) on an inclined platform

Abstract

This invention proposes an autonomous landing method for unmanned aerial vehicles (UAVs) on tilted platforms. The steps are as follows: constructing an autonomous landing system; calculating the center coordinates of the area where the determined landing landmark is located; calculating the center position of the landing landmark and the normal vector of its plane; planning a reference trajectory for UAV landing based on a differential flatness trajectory planning algorithm; and controlling the autonomous landing of the UAV based on a model predictive control algorithm. This invention plans the reference trajectory for UAV landing using reference trajectories calculated from the acceleration reference trajectory and the normal vector of the plane where the landing landmark is located. The normal vector takes into account the tilt angle of the landing platform. The defined optimization problem includes the objective function of the reference trajectory and constraints on UAV velocity, control variables, and the rate of change of control variables. By solving the optimization problem and obtaining the optimal sequence of control variables, the UAV rotor speed is adjusted, reducing the position and attitude errors during autonomous landing and effectively improving landing accuracy.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) technology and relates to an autonomous landing method for UAVs using an inclined platform, which can be applied to logistics distribution, power / bridge facility inspection, disaster emergency rescue and other fields. Background Technology

[0002] As drone applications become increasingly complex, autonomous landing sites now include not only level ground but also sloped terrain and inclined platforms such as sloping roofs. The accuracy of the drone's perception of the landing platform's location and its flight control are key factors influencing the accuracy of autonomous landing. Traditional drone landing location perception methods primarily rely on satellite positioning systems, which can achieve relatively stable landings in level terrain. However, in inclined platform scenarios, satellite positioning systems can only perceive the platform's position, not its normal direction. This prevents the drone from dynamically adjusting its attitude angle to maintain alignment with the landing platform, potentially leading to collisions, rollovers, and other safety issues.

[0003] Unlike satellite positioning systems, the vision-based autonomous landing method for UAVs first acquires patterned markings painted or affixed to the landing platform using an image acquisition device including a monocular or binocular vision camera. These markings reflect the pitch angle information of the landing platform. A processor then uses perspective information from the images to obtain the UAV's pitch angle data relative to the landing platform. The UAV then dynamically adjusts its flight attitude in real time based on this pitch angle data, ensuring the landing gear is parallel to the landing platform. This process is repeated until the UAV lands smoothly. This method can dynamically adjust the UAV's flight attitude in real time based on the landing platform's position and pitch angle, resulting in high landing accuracy. However, the attitude control strategy of this invention only considers the tilt angle of the landing platform, leading to a deviation between the attitude adjustment commands and the actual motion response, which hinders further improvement in landing accuracy. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and propose an autonomous landing method for UAVs on tilted platforms, which aims to improve the autonomous landing accuracy of UAVs on tilted platforms.

[0005] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0006] (1) Constructing an autonomous landing system for unmanned aerial vehicles:

[0007] The system comprises an image acquisition device, a positioning device, and a processor mounted on a multi-rotor UAV, as well as landing landmarks affixed to the landing area to guide the UAV's descent. The image acquisition device includes two monocular cameras positioned to the left and right for acquiring ground images. The positioning device is used to acquire the UAV's own position and attitude information. The processor is used to determine the area, location, and normal vector of the plane containing the landing landmarks based on the ground images sent by the image acquisition device, and simultaneously controls the UAV to land at the landing landmark location according to a reference trajectory planned based on the UAV's position and attitude information sent by the positioning device.

[0008] (2) Calculate the center coordinates of the area where the determined landing landmark is located:

[0009] Based on a pre-trained MobileNetV2 model, and by dividing ground images captured by the drone during flight into multiple image patches, the region where the landing landmark is located in the pixel coordinate system is determined. Then, the center coordinates (x, y, z) of the region where the landing landmark is located in the world coordinate system are calculated. c ,y c ,z c );

[0010] (3) Calculate the center position of the landing landmark and the normal vector of its plane:

[0011] By using the drone to fly towards the center coordinates (x) of the area where the landing landmark is located. c ,y c ,z c The ground images captured during the process are used to identify the landing landmarks, and the position P of the center of the landing landmarks is calculated based on the identification results. aruco and the normal vector of the plane where the landing landmark is located

[0012] (4) Planning the reference trajectory for UAV landing based on the differential flatness trajectory planning algorithm:

[0013] A trajectory planning algorithm based on differential flatness is used, and the location P of the landing landmark center is determined. aruco and the normal vector of the plane where the landing landmark is located Plan the reference trajectory for drone landing;

[0014] (5) Autonomous landing of UAVs based on model predictive control algorithms:

[0015] Based on the model predictive control algorithm, we define an objective function containing the UAV control variables and an optimization problem Q with constraints including UAV speed, control variables, and the rate of change of control variables. We solve this problem to control the position and attitude of the UAV and guide it to land along a planned reference trajectory to the landing landmark.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] This invention plans a reference trajectory for UAV landing by calculating the roll and pitch angles using the acceleration reference trajectory and the normal vector of the plane where the landing landmark is located. The normal vector takes into account the tilt angle of the landing platform. The defined optimization problem includes the objective function of the reference trajectory and constraints on UAV velocity, control variables, and the rate of change of control variables. By solving the optimization problem, the optimal sequence of control variables is used to adjust the UAV rotor speed, enabling the UAV to land along the planned reference trajectory towards the landing landmark. This reduces the position and attitude errors of autonomous landing and effectively improves the accuracy of autonomous landing of the UAV. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the implementation of the present invention.

[0019] Figure 2 This is a simulation comparison of the autonomous landing accuracy of the present invention and existing technologies. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0021] Reference Figure 1 The present invention includes the following steps:

[0022] Step 1) Construct an autonomous landing system for unmanned aerial vehicles:

[0023] The system comprises an image acquisition device, a positioning device, and a processor mounted on a multi-rotor UAV, as well as landing landmarks affixed to the landing area to guide the UAV's descent. The image acquisition device includes two monocular cameras positioned to the left and right for acquiring ground images. The positioning device is used to acquire the UAV's own position and attitude information. The processor is used to determine the area, location, and normal vector of the plane containing the landing landmarks based on the ground images transmitted by the image acquisition device, while simultaneously controlling the UAV to land at the landing landmark location according to a reference trajectory planned based on the UAV's position and attitude information transmitted by the positioning device.

[0024] In this embodiment, two monocular cameras are mounted on the bottom of the drone, facing downwards and arranged horizontally; the landing markers are square.

[0025] Step 2) Calculate the center coordinates of the area where the determined landing landmark is located:

[0026] Based on a pre-trained MobileNetV2 model, and by dividing ground images captured by the drone during flight into multiple image patches, the region where the landing landmark is located in the pixel coordinate system is determined. Then, the center coordinates (x, y, z) of the region where the landing landmark is located in the world coordinate system are calculated. c ,y c ,z c ).

[0027] The trained MobileNetV2 model includes a cascaded initial convolutional module, an inverse residual module, and a classification module. The inverse residual module consists of D cascaded inverse residual units, and the classification module consists of stacked convolutional layers, pooling layers, and fully connected layers. In this embodiment, D = 17. The initial convolutional module extracts features from each image patch and then downsamples them to obtain primary visual features including edges, textures, and colors. The inverse residual module expands the primary visual features and then performs depthwise convolution, compressing the result of the depthwise convolution to achieve feature transformation of the primary visual features and obtain deep semantic features. The classification module performs global average pooling on the deep semantic features and then maps them to fully connected layers to obtain the confidence probability of each image patch. The ground region corresponding to the image patch with the highest confidence probability is taken as the landing landmark region.

[0028] The center coordinates (x) of the area where the landmark is located in the world coordinate system c ,y c ,z c The calculation formula is:

[0029]

[0030] in, Let x be the rotation matrix from the UAV body coordinate system to the world coordinate system. cam ,y cam ,z cam () represents the coordinates of the center of the descent region in the left monocular camera coordinate system. Let x be the translation matrix from the camera coordinate system to the UAV body coordinate system. d ,y d ,z d (x) represents the position coordinates of the UAV in the world coordinate system. lp ,y lp ), (x rp ,y rp ) represent the center coordinates of the area where the landmark is located, which are the pixel coordinates of the images captured by the two monocular cameras. x with f y These are the focal lengths of the monocular camera along the x and y axes, respectively. x ,c y) represents the coordinates of the optical center of the monocular camera in the pixel coordinate system, b represents the distance between the centers of the two monocular cameras, and θ, φ and ψ represent the current pitch angle, roll angle and yaw angle of the UAV, respectively.

[0031] In this embodiment, the initial convolutional module includes one convolutional layer with an input data dimension of 224×224×3 and an output data dimension of 112×112×32; the inverse residual module includes 17 cascaded inverse residual units with an input data dimension of 112×112×32 and an output data dimension of 7×7×1280; each inverse residual unit includes an extended convolutional layer, a depthwise convolutional layer, and a compressed convolutional layer; the classification module has an input data dimension of 7×7×1280 and an output data dimension of 2×1. The unit is m; f x =409, f y =408, (c x ,c y = (318, 221), in pixels.

[0032] Step 3) Calculate the position of the center of the landing landmark and the normal vector of the plane containing the landing landmark:

[0033] (3a) The center coordinates (x) of the area where the UAV flies towards the landing landmark. c ,y c ,z c During the process, ground images captured by two monocular cameras undergo preprocessing including grayscale conversion and Gaussian filtering. The Harris corner detection algorithm is then used to detect candidate corner points in the two preprocessed images, resulting in a set C of corner points in the two ground images. l ={c l1 ,c l2 ,…,c lm ,…,c lM} and C r ={c r1 ,c r2 ,…,c rm ,…,c rM}, where M represents the number of candidate corner points, M≥4, c lm c rm Let m represent the m-th candidate corner point in the preprocessed images from the left and right monocular cameras, respectively.

[0034] (3b) Based on the epipolar constraint relationship between the left and right monocular camera images, C l c in lm With C r c in rmStereo matching verification is performed to obtain the coordinates of the I corner points of the landing landmarks in the preprocessed images of the left and right monocular cameras in the pixel coordinate system, where I = 4, and the coordinates of the i-th corner point of the landing landmarks in the preprocessed images of the left and right monocular cameras are respectively (x... lpi ,y lpi ), (x rpi ,y rpi ), i∈[1,I).

[0035] The epipolar constraint relationship between the left and right monocular camera images refers to the projection point (x, y) of any object in the left camera image. l ,y l ) and its projection point (x) in the right camera image r ,y r The polar constraint equations must be satisfied:

[0036]

[0037] F = E -T [A] × WE -1

[0038] Where F is the fundamental matrix of the two monocular cameras, E is the intrinsic parameter matrix of the monocular cameras, A is the translation matrix between the coordinate systems of the two monocular cameras, W is the rotation matrix between the coordinate systems of the two monocular cameras, and E -T Let [A] be the inverse matrix of the transpose of E. × Let E be the antisymmetric matrix of A. -1 Let E be the inverse matrix; in this embodiment, A = [0.0500] T W is a 3D identity matrix.

[0039] In this embodiment, stereo matching verification refers to calculating the feature descriptor of each candidate corner point using the SIFT algorithm; calculating candidate corner point c lm c rm The cosine similarity between SIFT feature descriptors is used; the pixel coordinates of candidate corner points with cosine similarity > 0.7 are substituted into the epipolar constraint equation; and the candidate corner points that satisfy the epipolar constraint equation are used as the corner points of the landing landmark.

[0040] (3c) The coordinates (x) of the i-th corner point of the landing landmark in the preprocessed images of the left and right monocular cameras in the pixel coordinate system. lpi ,y lpi ), (x rpi ,y rpi Map each point to the world coordinate system and calculate the position P of the i-th corner point of the landmark in the world coordinate system. i :

[0041]

[0042] Among them, (x cami ,y cami ,z cami ) represents the coordinates of the i-th corner point in the left monocular camera coordinate system.

[0043] (3d) By determining the position P of the i-th corner point of the landmark in the world coordinate system. i Calculations were performed to obtain the position P of the landing landmark center. aruco and the normal vector of the plane where the landing landmark is located The calculation formulas are as follows:

[0044]

[0045] Where ∑ represents the summation operation.

[0046] Step 4) Plan the reference trajectory for UAV landing based on the differential flatness trajectory planning algorithm:

[0047] (4a) Based on the fifth-order polynomial interpolation method, and according to the position P of the UAV at t=0 d Speed ​​v d and acceleration a d and the landing landmark center location P aruco Given a preset time T for the reference trajectory, calculate the reference trajectory P(t), v(t), and a(t) for position, velocity, and acceleration that are continuous in time:

[0048]

[0049]

[0050] (4b) Based on the reference trajectory of acceleration a(t) and the normal vector of the plane where the landing landmark is located. Calculate the reference trajectories φ(t) and θ(t) for continuous roll and pitch angles over time:

[0051]

[0052] θ(t)=arctan2(l x ,l z )

[0053]

[0054] Where arctan2(·) is the arctangent function in the four quadrants, arcsin(·) is the arcsine function, and l x , l y and l zThese are the desired acceleration vectors of the UAV after attitude angle compensation. Components along the x, y, and z axes, where g is the acceleration due to gravity, and k... p For attitude compensation gain; in this embodiment, g = 9.8 N / kg; k p =1.2.

[0055] normal vector The directional information reflects the roll and pitch angles around the x and y axes of the plane where the landing landmark is located. This is achieved through the normal vector. The desired acceleration vector of the UAV after attitude angle compensation is calculated from the UAV acceleration reference trajectory a(t). and through Generate reference trajectories φ(t) and θ(t) for the roll and pitch angles of the UAV, enabling the UAV to compensate for attitude errors caused by platform tilt in real time during landing, thereby improving landing accuracy in tilted platform scenarios.

[0056] (4c) The yaw angle ψ of the UAV at t=0 d The reference trajectory ψ(t) serves as the yaw angle for the UAV, ψ(t) = ψ d The reference trajectory for UAV landing, including position, attitude angle, and its first derivative, is calculated using ψ(t).

[0057]

[0058] Where d· represents the differentiation operation. Let φ(t), θ(t), and ψ(t) represent the first derivatives, respectively.

[0059] Step 5) Controlling the UAV to land autonomously based on model predictive control algorithm:

[0060] (5a) Based on the Newton-Euler equations, a nonlinear state-space model of the UAV is constructed using the UAV's state variable X and control variable U:

[0061]

[0062] Where m is the mass of the drone, K x K y With K z J represents the air drag coefficient of the UAV along the x, y, and z axes, respectively. x J y With J z J represents the moments of inertia of the UAV about the x-axis, y-axis, and z-axis of the body coordinate system, respectively. RPLet be the total moment of inertia of the motor rotor and rotor about their axes, and Ω be the difference between the counterclockwise and clockwise rotor speeds. The derivatives of the UAV's state variables X and X' are (x, y, z, v). x ,v y ,v z ,φ,θ,ψ,p,q,r) The control variable U of the UAV is (f, τ) x ,τ y ,τ z ); (x,y,z), (v x ,v y ,v z (φ,θ,ψ) and (p,q,r) represent the position, velocity, attitude angle, and attitude angular velocity of the UAV, respectively; f is the resultant thrust of the propeller on the UAV, (τ) x ,τ y ,τ z ) represents the three-axis torque of the drone;

[0063] In this embodiment, m = 1.45 kg; K x =0.5, K y =0.5, K z =1.2; J x =1.145×10 -2 J y =1.145×10 -2 J z =2.071×10 -2 J RP =1.30×10 -4 The unit is kg·m 2 The four rotors of the drone are arranged clockwise. The rotor on the front right of the nose is rotor 1. The rotor speeds are (n1, n2, n3, n4) and the rotor thrust coefficient is C. T =1.208×10 - 5 N / (rad / s) 2 Rotor moment coefficient C M =1.749×10 -7 N·m / (rad / s) 2 The distance e between the rotor shaft and the center of the fuselage is 0.205m; Ω, f, τ x τ y and τ z The calculation formulas are as follows:

[0064] Ω=n1-n2+n3-n4

[0065]

[0066] (5b) Using the current state variable X and control variable U as the initial state variable X0 and initial control variable U0, the UAV state variables in the prediction time domain N are iteratively predicted based on the UAV's nonlinear state-space model to obtain the UAV state variable sequence X. s ={X1,X2,…,X n ,…,X N}, where N≥10, n∈[1,N]; in this embodiment, N=16.

[0067] (5c) Define the UAV state variable X based on time n. n and the reference state variable X at time n obtained by discretizing the reference trajectory r,n The objective function J contains the UAV velocity constraint S. v Control variable constraints S U and control variable change rate constraint S dU The optimization problem Q; J, S v S U and S dU They are respectively:

[0068]

[0069] Q = [Q] p Q a ] T

[0070]

[0071] Where min represents the minimum value operation, ||·|| p Let X be the p-norm function, |·| be the absolute value operation, and X be the absolute value of the function. k X r,k U k Let X be the UAV state variable, reference state variable, and control variable at time k. N and X r,N Let N be the UAV state variables and the reference state variables at time N; Q is the weight matrix of the state variables, including (x, y, z, v). x ,v y ,v z The weight matrix Q p and The weight matrix Q a R is the weight matrix of the control variables, Q N Let Q be the weight matrix of the control variables at time N. a,0 For Q a The initial weight matrix, α is Q a The adjustment factor, d, is the distance between the UAV and the center of the landing area. maxLet be the historical maximum distance between the UAV and the center of the landing area, R0 be the initial weight matrix of R, and β be the adjustment factor of R0; (df,dτ) x ,dτ y ,dτ z ) represents the rate of change of the control variable; max_v xy max_v represents the maximum speed of the drone in the x and y directions. z The maximum velocity of the drone along the z-axis is given by max_f, and the maximum tensile force acting on the drone is given by max_τ. xy max_τ represents the maximum torque experienced by the drone in the x and y directions. z The maximum torque experienced by the UAV in the z-axis direction is given by max_df, and the maximum rate of change of tension experienced by the UAV is given by max_dτ. xy max_dτ represents the maximum rate of change of torque experienced by the UAV in the x and y directions. z This represents the maximum rate of change of torque experienced by the UAV in the z-axis direction.

[0072] In this embodiment, Q p =diag(100,100,100,10,10,10),Q a,0 =diag(10,10,10,10,10,10), α=1.5, R0=diag(1,1,1,1), β=2; max_v xy =5m / s, max_v z =2m / s, max_f=25N, max_τ xy =2N·m, max_τ z =0.5N·m, max_df=25N / s, max_dτ xy = 2N·m / s, max_dτ z =1 N·m / s, K = 15;

[0073] Define an objective function J containing a reference trajectory and a UAV velocity constraint S. v Control variable constraint S U and the control variable change rate constraint S dUThe value of the optimization problem Q lies in achieving accurate trajectory tracking and stable control in dynamic environments. Traditional control methods have three limitations when dealing with tilted platform landing scenarios: first, they ignore the physical limits of the UAV's actuators, which can easily lead to control command overshoot; second, they lack dynamic constraints, which may cause sudden changes in control quantities that cause oscillations in the UAV control system; and third, single-objective optimization makes it difficult to balance trajectory accuracy and energy efficiency. This invention achieves this by defining an optimization problem Q: the objective function guides the UAV to move along a reference trajectory, ensuring landing accuracy by minimizing position and attitude deviations; velocity constraints prevent stall or overspeed hazards caused by environmental disturbances; control variable constraints ensure that the output of actuators such as motors and rotors is always within safe thresholds; and control variable rate of change constraints suppress control variable jumps and avoid severe fuselage shaking.

[0074] (5d) Solve the optimization problem Q to obtain the optimal control variable sequence U at time K. s ={U1,U2,…,U k ,…,U K}, and through each of its control variables U k The rotor speed of the UAV is adjusted so that it lands along the planned reference trajectory toward the landing landmark, where k∈[1,K];

[0075] In this embodiment, the weighted least squares method is used to solve the optimization problem Q.

[0076] (5e) Determine whether the drone altitude H and threshold h satisfy H≤h. If yes, the drone lands at the landing landmark. Otherwise, proceed to step (5b).

[0077] The technical effects of the present invention will be further explained below with reference to simulation experiments:

[0078] 1. Simulation conditions and content:

[0079] This invention uses Gazebo software on a computer configured with an Intel i5-8300H CPU 2.30GHz*8 processor, GTX1050Ti GPU, 16GB of memory, and Ubuntu 20.04 system. The landing accuracy of this invention is compared with that of existing technologies using the coordinates (3,3,0) as the center of the landing landmark. The results are as follows: Figure 2 As shown.

[0080] 2. Simulation Result Analysis:

[0081] Reference Figure 2 ,in Figure 2 (a) Figure 2 (b) and Figure 2(c) These are comparison charts showing the landing accuracy of the present invention and the prior art, with time as the horizontal axis and the position coordinates of the UAV on the x-axis, y-axis and z-axis as the vertical axes, respectively.

[0082] from Figure 2 (a) It can be seen that in the prior art and the simulation of this invention, the position coordinate of the UAV on the x-axis when it begins to land is 0m, and the position coordinates on the x-axis when the landing is completed are 3.134m and 3.020m respectively, and the landing errors on the x-axis are 0.134m and 0.020m respectively;

[0083] from Figure 2 (b) It can be seen that in the prior art and the simulation of this invention, the position coordinate of the UAV on the y-axis when it begins to land is 0m, and the position coordinates on the y-axis when the landing is completed are 2.906m and 3.025m, respectively, and the landing errors on the y-axis are 0.094m and 0.025m, respectively.

[0084] from Figure 2 (c) It can be seen that the position coordinate of the UAV on the z-axis when it begins to land is 10m, that is, the height of the UAV when it lands is H=10; in the simulation of the prior art and the present invention, the position coordinate of the UAV on the z-axis starts to fall from 0s and stops falling at 29.05s and 26.45s respectively, indicating that the landing is completed.

[0085] In summary, compared with the prior art, the present invention significantly reduces the landing error on the x-axis and y-axis, indicating that the present invention has higher landing accuracy than the prior art.

Claims

1. A method for autonomous landing of an unmanned aerial vehicle (UAV) on an inclined platform, characterized in that, Includes the following steps: (1) Constructing an autonomous landing system for unmanned aerial vehicles: The system comprises an image acquisition device, a positioning device, and a processor mounted on a multi-rotor UAV, as well as landing landmarks affixed to the landing area to guide the UAV's descent. The image acquisition device includes two monocular cameras positioned to the left and right for acquiring ground images. The positioning device is used to acquire the UAV's own position and attitude information. The processor is used to determine the area, location, and normal vector of the plane containing the landing landmarks based on the ground images sent by the image acquisition device, and simultaneously controls the UAV to land at the landing landmark location according to a reference trajectory planned based on the UAV's position and attitude information sent by the positioning device. (2) Calculate the center coordinates of the area where the determined landing landmark is located: Based on a pre-trained MobileNetV2 model, and by dividing ground images captured by the drone during flight into multiple image patches, the region where the landing landmark is located in the pixel coordinate system is determined. Then, the center coordinates (x, y, z) of the region where the landing landmark is located in the world coordinate system are calculated. c ,y c ,z c ); (3) Calculate the center position of the landing landmark and the normal vector of its plane: By using the drone to fly towards the center coordinates (x) of the area where the landing landmark is located. c ,y c ,z c The ground images captured during the process are used to identify the landing landmarks, and the position P of the center of the landing landmarks is calculated based on the identification results. aruco and the normal vector of the plane where the landing landmark is located (4) Planning the reference trajectory for UAV landing based on the differential flatness trajectory planning algorithm: A trajectory planning algorithm based on differential flatness is used, and the location P of the landing landmark center is determined. aruco and the normal vector of the plane where the landing landmark is located Plan the reference trajectory for drone landing; (5) Autonomous landing of UAVs based on model predictive control algorithms: Based on the model predictive control algorithm, we define an objective function containing the UAV control variables and an optimization problem Q with constraints including UAV speed, control variables, and the rate of change of control variables. We solve this problem to control the position and attitude of the UAV and guide it to land along a planned reference trajectory to the landing landmark.

2. The method according to claim 1, characterized in that, The trained MobileNetV2 model mentioned in step (2) includes a cascaded initial convolutional module, an inverse residual module, and a classification module. The inverse residual module includes D cascaded inverse residual units, and the classification module includes stacked convolutional layers, pooling layers, and fully connected layers, where D≥10.

3. The method according to claim 2, characterized in that, The steps for determining the region where the pixel coordinate system drops the landmark in step (2) are as follows: (2a) The initial convolution module extracts features from each image block and then downsamples them to obtain primary visual features including edges, textures, and colors; (2b) The inverse residual module expands the primary visual features and then performs deep convolution, and compresses the result of the deep convolution to realize the feature transformation of the primary visual features and obtain deep semantic features. (2c) The classification module performs global average pooling on the deep semantic features and then maps them to a fully connected layer to obtain the confidence probability of each image patch. The ground area corresponding to the image patch with the highest confidence probability is taken as the landing landmark area.

4. The method according to claim 3, characterized in that, The center coordinates (x, y) of the area where the world coordinate system landmark is located, as described in step (2), are as follows: c ,y c ,z c The calculation formula is: in, Let x be the rotation matrix from the UAV body coordinate system to the world coordinate system. cam ,y cam ,z cam () represents the coordinates of the center of the descent region in the left monocular camera coordinate system. Let x be the translation matrix from the camera coordinate system to the UAV body coordinate system. d ,y d ,z d () represents the position coordinates of the UAV in the world coordinate system.

5. The method according to claim 4, characterized in that, The steps for identifying the landing landmarks described in step (3) are as follows: (3a) The ground images captured by the two monocular cameras are preprocessed, including grayscale conversion and Gaussian filtering. The Harris corner detection algorithm is then used to detect candidate corner points in the two preprocessed images, resulting in a set C of corner points in the two ground images. l ={c l1 ,c l2 ,…,c lm ,…,c lM } and C r ={c r1 ,c r2 ,…,c rm ,…,c rM }, where M represents the number of candidate corner points, M≥4, c lm c rm Let m represent the m-th candidate corner point in the preprocessed images of the left and right monocular cameras, respectively. (3b) Based on the epipolar constraint relationship between the left and right monocular camera images, C l c in lm With C r c in rm Stereo matching verification is performed to obtain the coordinates of the I corner points of the landing landmarks in the preprocessed images of the left and right monocular cameras in the pixel coordinate system, where I = 4, and the coordinates of the i-th corner point of the landing landmarks in the preprocessed images of the left and right monocular cameras are respectively (x... lpi ,y lpi ), (x rpi ,y rpi ), i∈[1,I]; (3c) The coordinates (x) of the i-th corner point of the landing landmark in the preprocessed images of the left and right monocular cameras in the pixel coordinate system. lpi ,y lpi ), (x rpi ,y rpi Map each point to the world coordinate system and calculate the position P of the i-th corner point of the landmark in the world coordinate system. i : Among them, (x cami ,y cami ,z cami ) represents the coordinates of the i-th corner point in the left monocular camera coordinate system.

6. The method according to claim 5, characterized in that, The epipolar constraint relationship between the left and right monocular camera images mentioned in step (3b) refers to the projection point (x) of any object in the left camera image. l ,y l ) and its projection point (x) in the right camera image r ,y r Satisfies the polar constraint equations: Where F is the fundamental matrix of the two monocular cameras.

7. The method according to claim 5, characterized in that, The location P of the landing landmark center mentioned in step (3) aruco and the normal vector of the plane where the landing landmark is located The calculation formulas are as follows: Where ∑ represents the summation operation.

8. The method according to claim 7, characterized in that, The trajectory planning algorithm based on differential flatness described in step (4) plans the reference trajectory for UAV landing. The implementation steps are as follows: (4a) Based on the fifth-order polynomial interpolation method, and according to the position P of the UAV at t=0 d Speed ​​v d and acceleration a d and the landing landmark center location P aruco Given a preset time T for the reference trajectory, calculate the reference trajectory P(t), v(t), and a(t) for position, velocity, and acceleration that are continuous over time: (4b) Based on the reference trajectory of acceleration a(t) and the normal vector of the plane where the landing landmark is located. Calculate the reference trajectories φ(t) and θ(t) for continuous roll and pitch angles over time: θ(t)=arctan2(l x ,l z ) Where arctan2(·) is the arctangent function in the four quadrants, arcsin(·) is the arcsine function, and l x , l y and l z These are the desired acceleration vectors of the UAV after attitude angle compensation. Components along the x, y, and z axes, where g is the acceleration due to gravity, and k... p For attitude compensation gain, 1≤k p ≤2; (4c) The yaw angle ψ of the UAV at t=0 d The reference trajectory ψ(t) serves as the yaw angle for the UAV, ψ(t) = ψ d The reference trajectory for UAV landing, including position, attitude angle, and its first derivative, is calculated using ψ(t). Where d· represents the differentiation operation. Let φ(t), θ(t), and ψ(t) represent the first derivatives, respectively.

9. The method according to claim 8, characterized in that, The steps for controlling the UAV to land autonomously based on the model predictive control algorithm described in step (5) are as follows: (5a) Based on the Newton-Euler equations, a nonlinear state-space model of the UAV is constructed using the UAV's state variable X and control variable U: Where sin, cos, tan, and sec are the sine, cosine, tangent, and secant functions, respectively; m is the mass of the UAV; and K... x K y With K z J represents the air drag coefficient of the UAV along the x, y, and z axes, respectively. x J y With J z J represents the moments of inertia of the UAV about the x-axis, y-axis, and z-axis of the body coordinate system, respectively. RP Let be the total moment of inertia of the motor rotor and rotor about their axes, and Ω be the difference between the counterclockwise and clockwise rotor speeds. The derivatives of the UAV's state variables X and X' are (x, y, z, v). x ,v y ,v z ,φ,θ,ψ,p,q,r) The control variable U of the UAV is (f, τ) x ,τ y ,τ z ); (x,y,z), (v x ,v y ,v z (φ,θ,ψ) and (p,q,r) represent the position, velocity, attitude angle, and attitude angular velocity of the UAV, respectively; f is the resultant thrust of the propeller on the UAV, (τ) x ,τ y ,τ z ) represents the three-axis torque of the drone; (5b) Using the current state variable X and control variable U as the initial state variable X0 and initial control variable U0, the UAV state variables in the prediction time domain N are iteratively predicted based on the UAV's nonlinear state-space model to obtain the UAV state variable sequence X. s ={X1,X2,…,X n ,…,X N }, where N≥10, n∈[1,N]; (5c) Define the UAV state variable X based on time n. n and the reference state variable X at time n obtained by discretizing the reference trajectory r,n The objective function J contains the UAV velocity constraint S. v Control variable constraints S U and control variable change rate constraint S dU The optimization problem Q; (5d) Solve the optimization problem Q to obtain the optimal control variable sequence U at time K. s ={U1,U2,…,U k ,…,U K }, and through each of its control variables U k The rotor speed of the UAV is adjusted so that it lands along the planned reference trajectory toward the landing landmark, where K = N-1, k ∈ [1, K]; (5e) Determine whether the drone altitude H and threshold h satisfy H≤h. If yes, the drone lands at the landing landmark. Otherwise, proceed to step (5b).

10. The method according to claim 9, characterized in that, The objective function J and the UAV velocity constraint S mentioned in step (5c) v Control variable constraints S U and control variable change rate constraint S dU Their expressions are as follows: Q=[Q p Q a ] T Where min represents the minimum value operation, ||·|| p Let X be the p-norm function, |·| be the absolute value operation, and X be the absolute value of the function. k X r,k U k Let X be the UAV state variable, reference state variable, and control variable at time k. N and X r,N Let N be the UAV state variables and the reference state variables at time N; Q is the weight matrix of the state variables, including (x, y, z, v). x ,v y ,v z The weight matrix Q p and The weight matrix Q a R is the weight matrix of the control variables, Q N Let Q be the weight matrix of the control variables at time N. a,0 For Q a The initial weight matrix, α is Q a The adjustment factor, d, is the distance between the UAV and the center of the landing area. max Let be the historical maximum distance between the UAV and the center of the landing area, R0 be the initial weight matrix of R, and β be the adjustment factor of R0; (df,dτ) x ,dτ y ,dτ z ) represents the rate of change of the control variable; max_v xy max_v represents the maximum speed of the drone in the x and y directions. z The maximum velocity of the drone along the z-axis is given by max_f, and the maximum tensile force acting on the drone is given by max_τ. xy max_τ represents the maximum torque experienced by the drone in the x and y directions. z The maximum torque experienced by the UAV in the z-axis direction is given by max_df, and the maximum rate of change of tension experienced by the UAV is given by max_dτ. xy max_dτ represents the maximum rate of change of torque experienced by the UAV in the x and y directions. z This represents the maximum rate of change of torque experienced by the UAV in the z-axis direction.

Images