An aerial robotic system and method for grabbing drones

Abstract

The application provides an aerial robot system and method for capturing a drone. The system provided by the application comprises a target detection module for identifying a real-time position of the drone according to a real-time image; a pose control module for generating a real-time attitude control quantity by PID hierarchical control based on the real-time position of the drone, and controlling the aerial robot to fly; a capturing module for controlling a capturing mechanism to capture the drone in response to a signal indicating that the aerial robot reaches a capturing position; and the pose control module for predicting the mass of the drone according to a current signal detected by the capturing mechanism after the capturing module captures the drone, calculating a mass center offset based on the mass of the drone and the mass of the aerial robot, calculating a feedforward quantity of the PID hierarchical control based on the mass center offset, summing the real-time attitude control quantity and the feedforward quantity as a corrected real-time attitude control quantity, and controlling the aerial robot to fly. The aerial robot system and method for capturing a drone provided by the application are used for accurately and safely capturing a drone.

Description

Technical Field

[0001] This application relates to the field of drone capture technology, and more particularly to an aerial robot system and method for capturing drones. Background Technology

[0002] With the rapid development and widespread application of drone technology, illegal intrusion incidents involving low-altitude micro-drones have occurred frequently, posing a serious threat to social security, public order, and personal privacy. Disorderly drone flights not only disrupt civil aviation operations but may also be used for illegal activities. Furthermore, drones can be used for cyberattacks, becoming tools for hackers to infiltrate corporate networks and critical infrastructure. These incidents not only highlight the multifaceted harms of illegal drone intrusion but also expose the current shortcomings in drone regulation and defense.

[0003] Current prevention and control measures against unauthorized drone intrusion mainly include passive and active defense. Passive defense uses radar, photoelectric detectors, acoustic sensors, and signal monitoring equipment to detect intruding drones and provide early warnings. However, it relies on early warning and coordinated response mechanisms, making it unable to respond promptly to rapid emergencies and susceptible to environmental factors. Especially when precise location and capture are required, passive defense systems struggle to provide real-time and accurate target location information. While active defense can directly intervene in or destroy intruding drones through electromagnetic interference, laser weapons, and net-trapping systems, existing active defense methods generally have functional defects: electromagnetic interference and laser weapons cannot achieve complete capture of drones; while net-trapping systems can achieve physical contact, they suffer from drawbacks such as the inability to autonomously detect and short operating range, making it impossible to accurately, completely, and thoroughly capture intruding drones. Summary of the Invention

[0004] In view of this, this application provides an aerial robot system and method for capturing drones, so as to capture drones accurately, stably and safely.

[0005] Specifically, this application is implemented through the following technical solution:

[0006] The first aspect of this application provides an aerial robot system for capturing drones, the system comprising:

[0007] The target detection module is used to identify the real-time position of the drone based on real-time images;

[0008] The pose control module is used to generate real-time attitude control quantities based on the real-time position and PID hierarchical control of the UAV, thereby controlling the flight of the aerial robot.

[0009] The grasping module is used to control the grasping mechanism to grasp the drone in response to a signal indicating that the grasping position has been reached;

[0010] The pose control module is further configured to, after the grasping module grasps the drone, predict the mass of the drone based on the current signal detected by the grasping mechanism, calculate the center of gravity offset based on the mass of the drone and the mass of the aerial robot, calculate the feedforward quantity of PID hierarchical control based on the center of gravity offset, and sum it with the real-time attitude control quantity as the corrected real-time attitude control quantity to control the flight of the aerial robot; wherein, calculating the center of gravity offset based on the mass of the drone and the self-weight of the aerial robot includes:

[0011] Construct the position vector of the grasping point of the grasping drone relative to the initial centroid of the aerial robot;

[0012] The mass percentage of the drone is calculated based on the mass of the drone and the weight of the aerial robot.

[0013] The centroid offset is calculated based on the product of the mass percentage and the position vector.

[0014] A second aspect of this application provides a method for capturing a drone, the method comprising:

[0015] Acquire real-time images of the drone, calculate a disparity map based on the real-time images, and calculate a depth value based on the disparity map;

[0016] The real-time image is identified using the YOLOv5 algorithm to obtain the two-dimensional bounding box of the UAV;

[0017] The three-dimensional bounding box is calculated by combining the depth value and the two-dimensional bounding box, and the position of the UAV is determined based on the three-dimensional bounding box;

[0018] Adjust the flight direction and attitude of the aerial robot according to the location, and control the aerial robot to grab the drone;

[0019] After the drone is captured, the mass of the drone is predicted based on the current signal detected by the capturing mechanism, and the centroid offset is calculated based on the mass and the mass of the aerial robot.

[0020] The feedforward quantity of the PID hierarchical control is calculated based on the centroid offset, and summed with the real-time attitude control quantity to obtain the corrected real-time attitude control quantity, thereby controlling the flight of the aerial robot.

[0021] The aerial robot system and method for capturing unmanned aerial vehicles (UAVs) provided in this application integrate target detection, pose control and other methods to achieve accurate control of the aerial robot's pose. During the capture process, the change in the center of mass caused by the mass of the captured UAV is evaluated in real time to achieve accurate control of the aerial robot's pose. This realizes pose control in response to changes in mass estimation and collaborative control of the capture module, thereby achieving precise capture and stable control of the UAV. First, the mass of the UAV is inverted in real time by the current signal of the grasping mechanism, and the position vector of the grasping point is constructed simultaneously. The centroid offset is calculated based on the product of the mass ratio and the position vector, realizing dynamic modeling of the entire process from "mass perception - offset quantification - compensation generation". The centroid offset is converted into feedforward torque and integrated with the real-time attitude control output of PID hierarchical control to form a composite control mode of "predictive compensation + feedback adjustment", which reduces the attitude response delay of the aerial robot after grasping. Second, the collaborative mechanism of centroid offset quantization and pose control solves the pain point of "load change - attitude instability" in traditional grasping. When the mass of the UAV changes and causes the centroid to shift, the system dynamically adjusts the magnitude and direction of the feedforward torque by updating the mass ratio parameter in real time. In addition, the grasping module and the pose control module are deeply coupled: the mass estimation process is triggered by the current change detection at the moment of grasping, and the centroid offset calculation result drives the pose control module to adjust the feedforward parameters synchronously, forming a closed-loop response chain of "grasping - mass assessment - attitude compensation". This architecture enables closed-loop control throughout the entire process, from target localization and flight control to attitude correction after capture. It can effectively solve problems such as large positioning errors and flight instability caused by centroid shift in traditional capture systems, thereby improving the accuracy and safety of UAV capture and significantly increasing the success rate and safety of UAV capture in complex environments. Attached Figure Description

[0022] Figure 1 A schematic diagram of the structure of an aerial robot system for capturing drones provided in this application;

[0023] Figure 2 A flowchart of an embodiment of the method for capturing a drone provided in this application. Detailed Implementation

[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0025] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0026] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0027] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0028] Figure 1 This is a schematic diagram of the structure of an aerial robot system for capturing drones provided in this application, according to a first embodiment. Please refer to... Figure 1 The system provided in this embodiment may include:

[0029] The target detection module is used to identify the real-time position of the drone based on real-time images;

[0030] The pose control module is used to generate real-time attitude control quantities based on the real-time position and PID hierarchical control of the UAV, thereby controlling the flight of the aerial robot.

[0031] The grasping module is used to control the grasping mechanism to grasp the drone in response to a signal indicating that the grasping position has been reached;

[0032] The pose control module is further configured to, after the grasping module grasps the drone, predict the mass of the drone based on the current signal detected by the grasping mechanism, calculate the center of gravity offset based on the mass of the drone and the mass of the aerial robot, calculate the feedforward quantity of PID hierarchical control based on the center of gravity offset, and sum it with the real-time attitude control quantity as the corrected real-time attitude control quantity to control the flight of the aerial robot; wherein, calculating the center of gravity offset based on the mass of the drone and the self-weight of the aerial robot includes:

[0033] Construct the position vector of the grasping point of the grasping drone relative to the initial centroid of the aerial robot;

[0034] The mass percentage of the drone is calculated based on the mass of the drone and the weight of the aerial robot.

[0035] The centroid offset is calculated based on the product of the mass percentage and the position vector.

[0036] Specifically, real-time images refer to images captured by the aerial robot using its onboard binocular cameras. These real-time images include left-view and right-view images. The steps for identifying the drone's real-time position based on these images include:

[0037] (1) Generate a disparity map between the real-time image from the left perspective and the real-time image from the right perspective based on a semi-global matching algorithm;

[0038] Specifically, the semi-global matching algorithm is a block-matching-based stereo matching algorithm that optimizes disparity calculation through multi-path cost aggregation. First, using camera calibration parameters, the real-time images from the left and right views are corrected to the same plane, ensuring that corresponding points are on the same horizontal line. This reduces the matching search from two dimensions to one dimension, decreasing computational complexity. For each pixel in the left-view real-time image, the grayscale difference between it and pixels at different disparities in the same row of the right-view real-time image is calculated to obtain the initial matching cost. Further, the initial matching cost is optimized using a cost function, which can be expressed by the following formula:

[0039] ;

[0040] in, The coordinates of the pixel in the real-time image from the left perspective Pixel value at;

[0041] The coordinates of the pixel in the real-time image from the right perspective Pixel value at;

[0042] This represents the disparity value of a pixel in the real-time images from the left and right perspectives.

[0043] The weights for the smoothing terms control the degree of smoothing in the disparity map;

[0044] This is the squared difference of the disparity between neighboring pixels.

[0045] For each pixel, the disparity value with the lowest cost is selected to obtain the disparity map.

[0046] (2) Traverse each pixel in the disparity map and calculate the disparity value of each pixel;

[0047] (3) Calculate the depth value of the pixel based on the disparity value, and determine the position of each pixel based on the depth value.

[0048] Specifically, the disparity value of each pixel in the disparity map represents the horizontal disparity of that point in the left and right images. The larger the value, the closer the object is to the camera. The disparity value of each pixel can be calculated using a stereo matching algorithm. After obtaining the disparity value, the depth value of the pixel is calculated using the following formula:

[0049] ;

[0050] Where B is the baseline of the binocular camera;

[0051] This represents the disparity value.

[0052] f is the focal length of the binocular camera.

[0053] This yields the depth value of each pixel, which is then combined with the optical center coordinates of the binocular camera to calculate the three-dimensional world coordinates of each pixel in the disparity map.

[0054] (4) For any one of the real-time images in the left-view real-time image and the right-view real-time image, identify the two-dimensional bounding box of the UAV in the real-time image based on the YOLOV5 algorithm;

[0055] Specifically, YOLOv5 is a lightweight object detection algorithm that extracts features at different scales and applies convolutional neural networks to identify objects in images. For any real-time image, whether it is a left-view or right-view real-time image, the YOLOv5 algorithm can identify the two-dimensional bounding box of the object in the real-time image.

[0056] (5) Combine the depth value and the two-dimensional bounding box to generate a three-dimensional bounding box;

[0057] (6) Determine the position of the UAV based on the three-dimensional bounding box.

[0058] Specifically, the two-dimensional bounding box of the target is set as follows: And obtain the depth value of the region from the disparity map. and Then the 3D bounding box coordinates of the target can be calculated using the following formula:

[0059] ;

[0060] ;

[0061] ;

[0062] in, The coordinates of the center of the two-dimensional bounding box;

[0063] ( () represents the optical center coordinates of the binocular camera;

[0064] The focal length of the camera;

[0065] The depth of the target.

[0066] The three-dimensional coordinates of the UAV in the three-dimensional coordinate system can be determined based on the three-dimensional bounding box of the real-time image, and then the position of the UAV can be obtained based on the three-dimensional coordinates.

[0067] Furthermore, after obtaining the real-time position of the drone, the pose control module is used to control the flight of the aerial robot, specifically including:

[0068] (1) Determine the target position based on the real-time position of the UAV, and calculate the position deviation based on the target position and the current position of the aerial robot;

[0069] Specifically, the target position of the aerial robot is determined based on the real-time position of the drone. The target position is the position that the aerial robot needs to reach to grab the drone. The position deviation is the spatial distance and direction between the current position of the aerial robot and the target position, which is the input basis of the outer loop PID control. It is calculated from the difference between the current position of the aerial robot and the target position.

[0070] Furthermore, an EKF state estimation method based on binocular vision is used to estimate the position, velocity, and acceleration of the target UAV. The target motion state is defined as a 9-dimensional vector containing position, velocity, and acceleration. The constant acceleration model (CA model) is used to recursively deduce the state:

[0071] ;

[0072] in This is the state vector from the previous time step;

[0073] This is the state transition matrix;

[0074] This is process noise;

[0075] Furthermore, the state transition matrix is:

[0076] ;

[0077] The sampling time interval for binocular vision;

[0078] The process noise covariance matrix;

[0079] It is a 3x3 identity matrix;

[0080] It is a 3x3 zero matrix.

[0081] Binocular vision directly provides the observation of the target's location: ;

[0082] in, For observation vectors;

[0083] It is a state vector;

[0084] To reduce observation noise, EKF is used to transform discrete position observations from binocular vision into continuous motion state (velocity, acceleration) estimations, avoiding the dependence of traditional methods on IMU.

[0085] Considering that pure visual EKF suffers from state estimation divergence due to missing observations during target occlusion or rapid maneuvers, a temporal error compensation mechanism (LSTM) is employed to reduce the target detection error in pure visual detection.

[0086] Construct the LSTM input sequence as the historical state estimation residual: ;

[0087] in, To estimate the residual, at time k, the error between the observed value predicted by the current EKF state and the actual observed value is quantified and input into the LSTM network as an input feature;

[0088] Let be the observation vector, which is the value actually measured by a purely visual target detection system (such as a camera) at time k;

[0089] The observation matrix is ​​the matrix that maps the system's state space to the observation space. It defines how to extract the quantities that should theoretically be observed from the state vector.

[0090] Let be the EKF state estimation vector, which is the target state estimate predicted by pure visual EKF at time k.

[0091] The mapping relationship of the LSTM network is as follows: The estimated residual sequence from the past n+1 time steps (from kn to k) is input into the LSTM neural network. The LSTM learns the temporal patterns contained in these historical residual sequences (such as target initiation of maneuver, error accumulation due to visual occlusion, etc.) and outputs a state correction. .

[0092] The collaborative output is: ;

[0093] The final target state estimate after LSTM error compensation is obtained by adding the target state estimate from EKF and the state correction learned by LSTM.

[0094] The final collaboratively estimated state vector is the target state estimate output by the system after LSTM compensation.

[0095] This is the EKF state estimation vector;

[0096] This is the state correction vector output by the LSTM.

[0097] Generate residual sequences of target abrupt motions (such as sharp turns or instantaneous accelerations) in a simulation environment.

[0098] The loss function is defined as:

[0099] ;

[0100] in, To estimate the final collaborative state vector;

[0101] The observation matrix;

[0102] These are actual observed values;

[0103] T represents the time step.

[0104] The loss function is calculated over T time steps, resulting in the final collaboratively estimated state. The predicted observations ( ) and actual observed values The goal of the loss function is to minimize the sum of mean squared errors (MSEs) between the two sides, and to minimize the amount of correction learned by the LSTM. This makes the final The predicted observations should be as close as possible to the actual observed values ​​Z. k .

[0105] Through the above mechanisms, online estimation of velocity / acceleration is achieved in a pure vision system, eliminating the dependence on IMU; and robustness in complex motion scenes is improved by compensating for the intermittent lack of visual observations through LSTM.

[0106] (2) Based on the position deviation combined with the outer loop PID control algorithm, output the desired speed of the aerial robot;

[0107] Specifically, taking positional deviation as input and outputting the desired velocity, can be calculated using the following formula:

[0108] ;

[0109] in, For the desired speed;

[0110] (ΔX, ΔY, ΔZ) represents the positional deviation;

[0111] This is the proportionality coefficient;

[0112] The integral coefficient;

[0113] is the differential coefficient.

[0114] (3) Calculate the desired attitude angle based on the desired velocity;

[0115] Specifically, the desired attitude angles include yaw angle, pitch angle, and roll angle, which can be calculated using the following formula:

[0116] ;

[0117] ;

[0118] ;

[0119] in, Yaw angle;

[0120] The pitch angle;

[0121] For roll angle;

[0122] ( , , () represents the desired speed;

[0123] k is the steering compensation coefficient.

[0124] (4) Based on the attitude deviation between the desired attitude angle and the current attitude angle, and combined with the inner loop PID control algorithm, output the real-time attitude control quantity of the aerial robot;

[0125] Specifically, the attitude deviation is calculated based on the difference between the desired attitude angle and the current attitude angle. This attitude deviation is then used as input to calculate the real-time attitude control variable, which can be calculated using the following formula:

[0126] ;

[0127] in, This is the proportionality coefficient;

[0128] This is for attitude deviation;

[0129] is the integral coefficient.

[0130] (5) Control the motor speed according to the real-time attitude control quantity, and control the flight of the aerial robot.

[0131] Specifically, the torque control quantity is determined based on the real-time attitude control quantity, and the motor speed is calculated based on the torque control quantity, thereby controlling the flight of the aerial robot.

[0132] Furthermore, after controlling the aerial robot to fly to the target location, the grasping mechanism is controlled to grasp the drone, specifically including:

[0133] (1) Calculate the initial torque of the gripping mechanism based on the preset initial drive current;

[0134] Specifically, the initial drive current is a pre-set current used to drive the gripping mechanism. The magnitude of the initial drive current is determined based on empirical values, and in this embodiment, it is not limited. Furthermore, the initial torque is calculated based on the robotic arm dynamics and the initial drive current using the following formula:

[0135] ;

[0136] in, The torque constant of the motor;

[0137] This is the initial drive current.

[0138] (2) Control the aerial robot to fly to the grasping position, and control the grasping mechanism to grasp the UAV at the initial grasping angle with the initial torque;

[0139] Specifically, after the aerial robot reaches the grasping position, an initial grasping angle is set according to the geometric characteristics of the drone to ensure that the gripper of the grasping mechanism fits against the surface of the drone. For example, in one embodiment, the geometric characteristics of the drone are detected to be cylindrical, and the initial angle is set to 60° to grasp the drone with the initial torque and the initial angle.

[0140] As an optional embodiment, the aerial robot's binocular vision system captures real-time images of the drone, identifies the drone boundaries in the real-time images, locates the fusion point of the left and right eye images based on the identified drone boundaries, and stitches the left and right eye images together based on the fusion point to obtain a complete image of the drone. Based on the drone boundary information in the complete image, the initial torque and initial angle of the grasping mechanism are calculated. Specifically, the drone boundary information in the complete image does not need to be detected again; instead, it is fused from the drone boundaries in the left and right eye images during stitching. Specifically, the image pixels of the fusion point are aligned, the focal points of the fusion point and other boundaries in the real-time image are located, and the focal points of the same boundaries in the left and right eye images are overlapped.

[0141] (3) Detect the driving current of the gripping mechanism and the pressure information of the contact surface between the gripping mechanism and the UAV;

[0142] Specifically, the driving current of the gripping mechanism can be directly determined by the input current of the gripping mechanism drive motor, and the pressure information of the contact surface between the gripping mechanism and the UAV can be detected by the pressure sensor arranged on the contact surface of the gripper of the gripping mechanism.

[0143] (4) The torque of the gripping mechanism is corrected according to the changing trend of the driving current and the changing information of the pressure information.

[0144] Specifically, by measuring the changes in drive current and pressure information of the gripping mechanism within a preset time after gripping the drone, the trend of drive current change and pressure information change are obtained. It should be noted that the preset time is set according to actual needs, and is not limited in this embodiment. For example, in one embodiment, the preset time is 2 seconds. 2 seconds after the gripping mechanism has gripped the drone, the drive current of the gripping mechanism is measured, and the trend of drive current change is calculated based on the difference between this drive current and the initial drive current. Similarly, 2 seconds after the gripping mechanism has gripped the drone, the pressure information of the gripping mechanism's claws is measured, and the change of pressure information is calculated based on the difference between this pressure information and the pressure information before 2 seconds.

[0145] Furthermore, the steps for adjusting the torque of the gripping mechanism based on the changing trend and information include:

[0146] 4.1 Determine whether the grasping mechanism has grasped the drone based on the changing trend of the driving current;

[0147] Specifically, it is understandable that when the gripping mechanism grips the drone based on the initial drive current, if the gripping mechanism successfully grips the drone, the change in the initial drive current will not be significant. If the gripping mechanism fails to grip the drone, the gripper will perform a gripping action even though it has not successfully gripped the drone. Therefore, after the gripper closes, the change in the drive current relative to the initial drive current will be significant.

[0148] Therefore, a current change threshold can be set based on empirical values. If the trend of the driving current change is less than or equal to the current change threshold, it is determined that the gripping mechanism has grabbed the drone. If the trend of the driving current change is greater than the current change threshold, it is determined that the gripping mechanism has not grabbed the drone.

[0149] Furthermore, after determining that the gripping mechanism has failed to grasp the drone, the torque and gripping angle of the gripping mechanism are updated. Based on the updated torque and gripping angle, the mechanism attempts to grasp the drone again. The process then returns to the steps of calculating the trend of the drive current and the change in pressure information. It should be noted that if the number of times the gripping mechanism has failed to grasp the drone exceeds a preset number, the gripping operation stops and the system returns to a standby point. The preset number of times is set according to actual needs; in this embodiment, it is not limited.

[0150] 4.2 After the grasping mechanism grasps the unmanned aircraft, it determines whether the unmanned aircraft will slide down based on the change in pressure information;

[0151] Specifically, after the grasping mechanism grasps the drone, it is necessary to judge the stability of the grasping mechanism in grasping the drone. It can be understood that if the drone does not slide down in the grasping mechanism, then the pressure information changes of the pressure sensors at various positions in the grasping mechanism should be the same. If the drone slides down in the grasping mechanism, then the pressure information changes from top to bottom in the grasping mechanism will be different. Therefore, the standard deviation of the pressure information or the pressure change rate can be used to judge whether the drone has slid down in the grasping mechanism. If the drone does not slide down, the posture control module will compensate the center of gravity of the aerial robot according to the weight of the drone and then control the aerial robot to grasp the drone and return to the base station.

[0152] 4.3 After determining that the drone is descending, calculate the sliding speed based on the change information, calculate the torque correction value based on the sliding speed, and correct the torque of the gripping mechanism based on the torque correction value.

[0153] Specifically, after determining that the drone is descending, the sliding speed can be calculated based on the rate of change of pressure information. The sliding speed can be calculated using the following formula:

[0154] ;

[0155] in, This is the pressure-velocity conversion factor;

[0156] This represents the rate of change of pressure.

[0157] Furthermore, after obtaining the sliding speed, the sliding speed can be used as the input to the PID algorithm to calculate the torque. Alternatively, the sliding speed can be input into a pre-trained speed-torque model, and the model can output the torque corresponding to the sliding speed. This torque can then be used as a torque correction value to control the grasping mechanism to grasp the drone.

[0158] Furthermore, after the drone is captured, the steps for predicting the drone's mass based on the current signal detected by the capturing mechanism include:

[0159] (1) Calculate the output torque of the control motor of the gripping mechanism based on the current signal and the torque-current correlation of the gripping mechanism motor;

[0160] (2) Input the output torque and the length of the mechanical arm of the gripping mechanism into the static equilibrium equation to obtain the mass of the UAV.

[0161] Specifically, based on the above description, the current value of the current signal detected by the gripping mechanism is input into the torque calculation formula to obtain the output torque of the control motor. Furthermore, the static equilibrium equation is expressed using the following formula:

[0162] ;

[0163] By combining the torque calculation formula, we can obtain the formula for calculating the mass of the drone:

[0164] ;

[0165] in, For output torque;

[0166] It is the acceleration due to gravity;

[0167] The length of the robotic arm;

[0168] For driving current;

[0169] This is the motor torque constant.

[0170] Optionally, the steps for calculating the centroid offset based on the mass of the UAV and the weight of the aerial robot include:

[0171] (1) Construct the position vector of the grasping point of the grasping drone relative to the initial centroid of the aerial robot;

[0172] Specifically, after the aerial robot grasps the drone, the displacement of the aerial robot's center of mass is determined by both the drone's mass and the position of the grasping point. The position vector reflects the spatial distribution of the drone's center of mass relative to the aerial robot's center of mass and is the geometric basis for calculating the displacement of the center of mass. The position vector can be constructed using the following formula:

[0173] ;

[0174] in, These are the coordinates of the capture point on the three coordinate axes;

[0175] The X-axis represents the direction of movement for the aerial robot.

[0176] The Y-axis is perpendicular to the X-axis and points to the right side of the aerial robot;

[0177] The Z-axis represents the direction of gravity.

[0178] (2) Calculate the mass percentage of the UAV based on the mass of the UAV and the self-weight of the aerial robot;

[0179] Specifically, the mass percentage can be calculated using the following formula:

[0180] ;

[0181] in, For the quality of the drone;

[0182] This refers to the weight of the aerial robot.

[0183] It should be noted that the weight of the aerial robot is determined based on its own parameters.

[0184] (3) Calculate the centroid offset based on the product of the mass ratio and the position vector.

[0185] Specifically, the centroid offset can be calculated using the following formula:

[0186] ;

[0187] Where A represents the percentage of mass;

[0188] This is a position vector.

[0189] By detecting the current signal of the grasping structure at the moment of grasping the drone, the mass of the drone is predicted, and the center of mass offset is further calculated. This allows the aerial robot to adjust according to the center of mass offset in the early stage of the grasping phase, compensating for the impact of the center of mass offset caused by sudden load changes on the flight attitude. This significantly reduces the attitude response delay of the aerial robot after grasping the drone and greatly reduces the position tracking error.

[0190] Furthermore, the position vector of the grasping point directly determines the spatial relative relationship between the grasping point and the initial centroid of the aerial robot, avoiding the offset estimation deviation caused by ignoring the difference in grasping position in traditional methods. The mass proportion clarifies the weight of the UAV's mass in the total mass of the system, enabling the calculation of the centroid offset to closely match the actual load distribution. The centroid offset obtained by multiplying the two can truly reflect the impact of load changes on the system's centroid, reducing errors. From a real-time perspective, it does not rely on complex multi-sensor data fusion or long-term excitation trajectories. It can quickly complete the calculation by inverting the mass using the current signal of the grasping mechanism and combining it with the preset initial centroid position. It can output the centroid offset result instantly after the grasping action is completed, solving the problem of lag in centroid estimation in existing technologies.

[0191] Furthermore, after the drone grasps the target object, its overall center of mass shifts. Based on the state of the grasping mechanism at the previous moment, the grasping state is identified, and the grasping phase is divided into the contact moment phase, the glide contact phase, and the stable grasping phase. The center of mass shift of each phase is calculated, and the feedforward control quantity is adjusted according to the center of mass shift of each phase.

[0192] Specifically, after the drone grasps the target object, the overall center of mass will shift. The center of mass shift is estimated in real time and continuously updated based on the state detection cycle of the grasping mechanism to ensure that every state change is promptly fed back into the center of mass calculation. The grasping state is identified by combining the force sensor data at the front end of the claw with changes in the driving current, and is specifically divided into three stages: the instantaneous contact stage is identified by the force sensor suddenly detecting a contact force and the driving current showing a momentary fluctuation without continuous change; during the gliding contact period, the pressure value detected by the force sensor shows a continuous changing trend, and the pressure distribution gradually shifts from the front end of the grasp to the rear end; the stable grasping stage is identified by the pressure value detected by the force sensor remaining stable without significant fluctuations and the driving current remaining within a constant range. For all three stages, the object's center of mass is estimated in real time based on data collected by force sensors, making the calculated center of mass more consistent with actual working conditions: at the moment of contact, the initial center of mass offset is quickly calculated to provide basic compensation for attitude control; during the glide contact, the center of mass change is dynamically tracked and the feedforward control is adjusted in real time to counteract the attitude interference caused by the glide; in the stable grasping stage, the center of mass position is accurately locked and a stable feedforward control is output to ensure that the aerial robot maintains a stable flight attitude. Through phased real-time center of mass estimation and feedforward control adjustment, the accuracy and stability of flight control after grasping are further improved.

[0193] Optionally, the steps for calculating the feedforward quantity for PID hierarchical control based on the centroid offset include:

[0194] (1) Calculate the first gravity vector of the UAV based on the product of the mass of the UAV and the gravitational acceleration;

[0195] (2) Calculate the second gravity vector of the aerial robot based on the product of its own weight and gravitational acceleration;

[0196] Specifically, the first and second gravity vectors can be calculated using the following formulas:

[0197] ;

[0198] ;

[0199] in, This is the first gravity vector;

[0200] This is the second gravitational vector;

[0201] For the quality of the drone;

[0202] The weight of the aerial robot;

[0203] This is the acceleration due to gravity.

[0204] (3) Calculate the feedforward force vector based on the sum of the first gravity vector and the second gravity vector;

[0205] Specifically, the feedforward force vector can be calculated using the following formula:

[0206] ;

[0207] in, This is the first gravity vector;

[0208] This is the second gravity vector.

[0209] (4) Calculate the feedforward torque based on the cross product of the centroid offset and the feedforward force vector.

[0210] Specifically, the feedforward torque can be calculated using the following formula:

[0211] ;

[0212] in, This is the feedforward force vector;

[0213] This represents the centroid offset.

[0214] Furthermore, when operating in complex environments, this system may encounter abnormal situations such as sensor failure, communication interruption, and capture failure. To ensure the stability and security of the system, a multi-level redundancy design, intelligent communication switching, and emergency braking strategy are proposed to achieve high-reliability operation of the UAV in extreme environments.

[0215] Furthermore, the aerial robot is equipped with two inertial measurement units (three-axis accelerometers + gyroscopes) and dual GPS devices. The two inertial measurement units are fixed at symmetrical positions at the center of the body and rigidly connected to the core controller to ensure consistency in measuring the body's movement. The accelerometer output values ​​of the two inertial measurement units are compared to calculate the absolute value of the X, Y, and Z axis acceleration deviations. If the deviation between the two acceleration measurement values ​​is detected to be greater than 0.2g (2% of the Earth's gravitational acceleration of 9.81m / s²), the robot automatically switches to the backup inertial measurement unit module to ensure the accuracy of attitude calculation.

[0216] Furthermore, dual GPS devices are deployed on the unobstructed area above the aerial robot. Differential positioning using dual GPS devices improves positioning accuracy and ensures an error of ≤10cm in open environments. When the GPS signal error exceeds 1m or the GNSS signal quality deteriorates (e.g., in urban canyons or indoor environments), the system automatically activates visual SLAM (RGB-D sensor + VIO inertial odometry) for GPS-free autonomous positioning, ensuring uninterrupted mission operation.

[0217] Furthermore, to ensure real-time data transmission between the aerial robot and the ground station, this system adopts a Wi-Fi + LoRa dual-link communication strategy to ensure the reliability of the communication link.

[0218] The Wi-Fi solution utilizes a high-speed, short-range configuration with a 2.4GHz frequency and a 20MHz bandwidth, providing high-speed data transmission for low-latency control and video transmission. This solution is suitable for aerial robot formations and close-range tasks (<300m), and is prioritized when the signal strength is >-70dBm. When the distance between the aerial robot and the ground station is greater, 433MHz LoRa technology is used, which supports long-range (maximum 3km) data transmission and is suitable for beyond-line-of-sight tasks. If the Wi-Fi signal strength drops below -90dBm (in weak signal areas or during long-distance operations), the system automatically switches to the LoRa communication link to maintain basic control information transmission.

[0219] The aerial robot system for capturing drones provided in this embodiment achieves precise and safe drone capture through multi-module collaboration and intelligent algorithms. The target detection module utilizes binocular vision and the YOLOv5 algorithm, combined with semi-global matching to generate a disparity map and calculate depth values, constructing a 3D bounding box to achieve high-precision real-time positioning of the drone, providing an accurate coordinate basis for capture. The pose control module employs PID hierarchical control; the outer loop generates the desired velocity based on position deviation, while the inner loop outputs control quantities based on attitude deviation. Combined with centroid offset feedforward compensation, this improves the flight attitude control response speed, reduces position tracking errors, and ensures the aerial robot flies stably to the capture position. The capture module uses a dual feedback mechanism of current and pressure, combining initial torque with an adaptive correction strategy to achieve stable capture of drones of different masses. Simultaneously, the system dynamically predicts the drone's mass through current signals, calculates the centroid offset, and generates a feedforward torque, which is fused with real-time attitude control quantities to effectively compensate for the impact of load changes on flight attitude after capture, improving the system's anti-interference capability. In addition, the system also features safety designs such as sensor redundancy and multimodal communication to ensure reliable operation in complex environments and achieve precise control of the UAV throughout the entire process from detection, positioning, capture to stable flight.

[0220] Corresponding to the aforementioned embodiment of an aerial robot system for capturing drones, this application also provides an embodiment of a method for capturing drones.

[0221] Figure 2 This is a flowchart illustrating an embodiment of the method for capturing a drone provided in this application. Please refer to... Figure 2 The method provided in this embodiment includes:

[0222] S101. Acquire real-time images of the UAV, calculate a disparity map based on the real-time images, and calculate a depth value based on the disparity map;

[0223] S102. Use the YOLOv5 algorithm to identify the real-time image and obtain the two-dimensional bounding box of the UAV;

[0224] S103. Calculate the three-dimensional bounding box by combining the depth value and the two-dimensional bounding box, and determine the position of the UAV based on the three-dimensional bounding box;

[0225] S104. Adjust the flight direction and attitude of the aerial robot according to the position, and control the aerial robot to grab the drone;

[0226] S105. After capturing the drone, predict the mass of the drone based on the current signal detected by the capturing mechanism, and calculate the centroid offset based on the mass and the mass of the aerial robot.

[0227] S106. Calculate the feedforward quantity of the PID hierarchical control based on the centroid offset, sum it with the real-time attitude control quantity to obtain the corrected real-time attitude control quantity, and control the aerial robot to fly.

[0228] The method in this embodiment can be used to execute Figure 1 The steps of the device embodiment shown are similar in principle and process, and will not be repeated here.

[0229] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An aerial robot system for capturing unmanned aerial vehicles, characterized in that, The system includes: The target detection module is used to identify the real-time position of the drone based on real-time images; The pose control module is used to generate real-time attitude control quantities based on the real-time position and PID hierarchical control of the UAV, thereby controlling the flight of the aerial robot. A grasping module, configured to control a grasping mechanism to grasp the drone in response to a signal indicating arrival at the grasping position; wherein, controlling the grasping mechanism to grasp the drone includes: The initial torque of the gripping mechanism is calculated based on the preset initial drive current; Control the aerial robot to fly to the grasping position, and use the initial torque to control the grasping mechanism to grasp the drone at the initial grasping angle; The driving current of the gripping mechanism and the pressure information between the gripping mechanism and the contact surface of the drone are detected. The torque of the gripping mechanism is corrected based on the changing trend of the driving current and the changing pressure information; wherein, the correction of the torque of the gripping mechanism based on the changing trend of the driving current and the changing pressure information includes: The gripping mechanism determines whether it has captured the drone based on the changing trend of the driving current. After the grasping mechanism grasps the unmanned aircraft, it determines whether the unmanned aircraft will descend based on the change in pressure information. After determining that the drone is descending, the sliding speed is calculated based on the change information, the torque correction value is calculated based on the sliding speed, and the torque of the gripping mechanism is corrected based on the torque correction value. The pose control module is further configured to, after the grasping module grasps the drone, predict the mass of the drone based on the current signal detected by the grasping mechanism, calculate the center of gravity offset based on the mass of the drone and the mass of the aerial robot, calculate the feedforward quantity of PID hierarchical control based on the center of gravity offset, and sum it with the real-time attitude control quantity as the corrected real-time attitude control quantity to control the flight of the aerial robot; wherein, calculating the center of gravity offset based on the mass of the drone and the self-weight of the aerial robot includes: Construct the position vector of the grasping point of the grasping drone relative to the initial centroid of the aerial robot; The mass percentage of the drone is calculated based on the mass of the drone and the weight of the aerial robot. The centroid offset is calculated based on the product of the mass percentage and the position vector.

2. The system according to claim 1, characterized in that, The step of predicting the mass of the drone based on the current signal detected by the grasping mechanism includes: The output torque of the gripping mechanism control motor is calculated based on the current signal and the torque-current correlation of the gripping mechanism motor. The mass of the UAV is obtained by inputting the output torque and the length of the robotic arm of the gripping mechanism into the static equilibrium equation.

3. The system according to claim 1, characterized in that, The calculation of the feedforward quantity for PID hierarchical control based on the centroid offset includes: The first gravity vector of the drone is calculated based on the product of the drone's mass and gravitational acceleration. The second gravity vector of the aerial robot is calculated based on the product of its own weight and gravitational acceleration. Calculate the feedforward force vector based on the sum of the first and second gravity vectors; The feedforward torque is calculated based on the cross product of the centroid offset and the feedforward force vector.

4. The system according to claim 1, characterized in that, The real-time images include left-view real-time images and right-view real-time images; identifying the real-time position of the UAV based on the real-time images includes: A disparity map between the real-time images from the left and right perspectives is generated based on a semi-global matching algorithm. Iterate through each pixel in the disparity map and calculate the disparity value for each pixel; The depth value of the pixel is calculated based on the disparity value, and the position of each pixel is determined based on the depth value.

5. The system according to claim 4, characterized in that, The method of identifying the real-time location of the drone based on real-time images also includes: For any one of the real-time images from the left-view and the right-view, the two-dimensional bounding box of the UAV in the real-time image is identified based on the YOLOv5 algorithm. A three-dimensional bounding box is generated by combining the depth value and the two-dimensional bounding box; The location of the UAV is determined based on the three-dimensional bounding box.

6. The system according to claim 1, characterized in that, Controlling the flight of the aerial robot includes: The target position is determined based on the real-time position of the UAV, and the position deviation is calculated based on the target position and the current position of the aerial robot. The desired speed of the aerial robot is output based on the position deviation combined with the outer loop PID control algorithm. Calculate the desired attitude angle based on the desired velocity; The real-time attitude control quantity of the aerial robot is output based on the attitude deviation between the desired attitude angle and the current attitude angle, combined with the inner-loop PID control algorithm. The motor speed is controlled based on the real-time attitude control parameters, thereby controlling the flight of the aerial robot.

7. A method for capturing a drone, characterized in that, The method is implemented based on the system according to any one of claims 1-6, and the method includes: Acquire real-time images of the drone, calculate a disparity map based on the real-time images, and calculate a depth value based on the disparity map; The real-time image is identified using the YOLOv5 algorithm to obtain the two-dimensional bounding box of the UAV; The three-dimensional bounding box is calculated by combining the depth value and the two-dimensional bounding box, and the position of the UAV is determined based on the three-dimensional bounding box; Adjust the flight direction and attitude of the aerial robot according to the location, and control the aerial robot to grab the drone; After the drone is captured, the mass of the drone is predicted based on the current signal detected by the capturing mechanism, and the centroid offset is calculated based on the mass and the mass of the aerial robot. The feedforward quantity of the PID hierarchical control is calculated based on the centroid offset, and summed with the real-time attitude control quantity to obtain the corrected real-time attitude control quantity, thereby controlling the flight of the aerial robot.

Images