Method for capturing unmanned aerial vehicle by using unmanned aerial vehicle

Abstract

The invention relates to the technical field of unmanned aerial vehicle capturing, and discloses a method for capturing an unmanned aerial vehicle by using an unmanned aerial vehicle, which comprises the following steps: S1, detecting an invader by an early warning system, triggering and initializing Kalman filter parameters, and obtaining a prediction state and a prediction error covariance matrix P; s3, if a target is detected, the target state is an observation value, the Kalman filter inputs the observation value into an updating equation to obtain a Kalman gain K, the state is updated, and the estimation covariance matrix P is updated; aiming at double dynamic conditions of intercepting and harassing the unmanned aerial vehicle, Kalman filtering and optical flow method speed measurement, accurate planning and dynamic adjustment of an interception path are realized, the positioning accuracy is greatly improved, the real-time speed of a target is judged by using an optical flow method under the double dynamic conditions, autonomous identification and tracking of the enemy unmanned aerial vehicle are realized, and the emission capture rate is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of drone capture technology, specifically a method for capturing drones using drones. Background Technology

[0002] To protect no-fly zones, it is necessary to capture and intercept drones within these zones. Net guns, with their wide capture range and high fault tolerance, are widely used for drone capture.

[0003] The net gun's firing area is located on the ground, and it is only possible to capture drones when they are flying at low altitudes. When the drone's flight altitude is higher than the net gun's firing distance, long-range capture tools such as infrared and microwave are required.

[0004] Compared to other tools, nets are simple in structure, easy to maintain, and have low operating costs, and can even be reused. If a net gun is mounted on a drone, using image recognition for rapid positioning and locking, and the drone approaches the intruder to shorten the distance to the net gun, then the intruder is captured by the net gun, this method solves the shortcomings of ground-based capture with net guns, increases the applicability of net guns, and reduces operating costs.

[0005] Since the states of aerial targets and drones are dynamic, there are situations where camera shake and targets are out of sight, which may cause targets to be unable to be continuously tracked. Furthermore, in complex environments, it cannot be guaranteed that targets will always be captured. Summary of the Invention

[0006] The purpose of this invention is to provide a method for capturing drones using drones, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for capturing a drone using a drone includes:

[0009] S1: The early warning system detects an intruder. The drone carrying the capture device takes off and hovers. YOLO performs target detection, detects the target, triggers and initializes the Kalman filter parameters, and the Kalman filter performs prediction to obtain the predicted state. And the prediction error covariance matrix P;

[0010] S2: Target is continuously detected with a confidence level greater than or equal to 0.7: The target state is the observed value. The Kalman filter uses the observations as input to the update equation to obtain the Kalman gain K and update the state. Update the estimated covariance matrix P. If the confidence level is less than 0.7, start the prediction process using Kalman filtering. The target state is the predicted value.

[0011] S3: If no target is detected, input the predicted parameters obtained in S1 into the Kalman filter update process, and output the posterior estimate and target state. By fusing target state and depth information, the following conclusions are drawn. The system then enters the tracking and control logic module. Based on the target information, it adjusts the UAV's position to perform tracking.

[0012] S4: When there are insufficient optical flow feature points, the prediction of the Kalman filter is relied upon; when there are sufficient feature points, the results of optical flow and Kalman filtering are fused.

[0013] S5: Based on the error between the target center coordinates and the image center, corresponding control commands are generated, such as left / right movement and up / down movement. The drone's current attitude information is obtained using its own attitude sensor. Based on the target's position in the image and the drone's attitude information, the control algorithm calculates the attitude angle that the drone needs to adjust to achieve aiming of the net gun. The PD algorithm is used to dynamically adjust the drone's flight speed according to the error magnitude. Data from the DJI Magic Calculation's built-in inertial measurement unit is introduced, combined with depth information from the Celtic laser rangefinder, to compensate for the impact of drone attitude jitter on aiming in real time.

[0014] S6: When the target is identified and located in the center of the image, it is assumed that the net gun has been aimed at the target. At the same time, the target distance d < 20m, the firing conditions are met, and the gun can be fired to capture the target.

[0015] As a further aspect of the present invention: through target recognition, the predicted bounding box of the target object in each frame of the camera can be obtained in real time. The predicted bounding box contains four parameters. , , , and With the center coordinates, and Let S be the width and height, denoted as the region of interest (ROI). The predicted bounding box of the target object is scaled, denoted as S. The scaling factor is θ, expressed as:

[0016]

[0017] Based on the depth information provided by the camera, filter out Take the pixel information other than those in the image. The average value of the medium-depth information is used as the relative position of the target.

[0018] As a further aspect of this invention: The object's state vector xk is established using the camera's relative position and velocity. It is assumed that the target state changes with a constant velocity, and all measurements and noise follow a Gaussian distribution. The recursive equation for estimating the system under discrete conditions is:

[0019]

[0020] F is the state transition matrix; B is the input control matrix; This is the noise matrix; H represents the observed values; H is the measurement matrix. To measure noise, the system can be further divided into two steps: prediction and update. The prediction equation is:

[0021]

[0022] P is the covariance matrix for calculating the predicted state value and the prediction error between the predicted and actual values; Q is the system prediction noise covariance matrix, representing the reliability of the entire system, and is generally initialized to a very small value. When the UAV fails to capture the target object, the Kalman filter is activated to predict the target's current position using the data from the previous frame. The update equation is:

[0023]

[0024] K is the optimal Kalman gain. H is the measurement matrix; H is the transition matrix; R is the observation noise covariance matrix. The motion states of the Kalman filter have a recursive relationship, and the state information can be updated in each iteration.

[0025] As a further aspect of the present invention: after the video stream transmitted back by the gimbal is processed, feature points within the target UAV region are selected in two adjacent frames. Using optical flow calculation methods, the displacement of the feature points between the two frames is tracked. By analyzing the displacement changes of the feature points and combining them with the time interval, the two-dimensional velocity component of the target UAV on the image plane is calculated. At the same time, by combining the depth information provided by the KJIT laser ranging radar and the system, the two-dimensional velocity on the image plane is converted into the actual motion velocity in three-dimensional space.

[0026] As a further step of this invention: using the coordinates of the image pixel center point and the intrinsic parameter matrix of the Zenmuse x3 gimbal, the relative position of the target and my drone is estimated. The formula for the pixel center coordinates is:

[0027]

[0028] Using the intrinsic parameter matrix of Zenmuse x3, the pixel coordinates are transformed using the following formula:

[0029]

[0030] in( , ) are the main pixel coordinates, ( , () represents the focal length in pixels.

[0031] As a further improvement of the present invention: after the pixel coordinates are transformed:

[0032] (Horizontal aiming threshold, 2°), the drone shifts to the right; (Horizontal aiming threshold, 2°), the drone shifts to the left;

[0033] (Horizontal aiming threshold, 1.5°), the drone descends; (Horizontal aiming threshold, 1.5°), drone climb.

[0034] As a further aspect of this invention: the PD algorithm is used to dynamically adjust the drone's flight speed based on the magnitude of the error.

[0035]

[0036] Where e is the current error and ∆e is the error rate of change, data from the DJI Inertial Measurement Unit (IMU) is incorporated into the system, combined with depth information d from the Celtic laser rangefinder, to compensate for the impact of drone attitude jitter on aiming in real time. The error correction formula is as follows:

[0037]

[0038] in, This is the corrected azimuth angle.

[0039] As a further aspect of the present invention: when the target is identified and located in the central region of the image (the error is within an acceptable range, i.e., satisfying the formula), It is assumed that the net gun has been aimed at the target, and the target distance d < 20m, which meets the firing conditions.

[0040] As a further aspect of the present invention: an array-type integrated capture device for anti-drone carry, comprising a central transmitter and a ring transmitter, wherein the central transmitter is controlled to launch in the manner described above, and the compensation launch unit of the ring transmitter obtains the two-dimensional velocity component direction of the drone on the image plane according to the optical flow calculation method, adjusts the direction of the line connecting the center of the compensation launch unit and the central transmitter, the direction of the line corresponding to the two-dimensional velocity component direction, and the compensation launch unit launches synchronously with the central transmitter.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] Artificial intelligence training is conducted using the YOLOv11 algorithm. Through learning from massive amounts of data, it accurately identifies various types of drones. For the dual dynamic conditions of intercepting and harassing drones, Kalman filtering and optical flow velocity measurement are used to achieve precise interception path planning and dynamic adjustment, which greatly improves positioning accuracy. At the same time, a KJIT laser rangefinder radar is added. Combined with depth information, optical flow is used to determine the real-time speed of the target under dual dynamic conditions, enabling autonomous identification and tracking of enemy drones. Single-shot or dual-shot launch modes are adopted to greatly improve the launch acquisition rate. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the state prediction and update process using Kalman filtering in a method for capturing drones.

[0045] Figure 2 A schematic diagram of the Labelimg annotation tool interface;

[0046] Figure 3 A diagram illustrating the process of adding training code to a Python script for training;

[0047] Figure 4 A schematic diagram illustrating the effect of cruising around a ground target in an urban environment;

[0048] Figure 5 A graph showing the evaluation metrics during the training process;

[0049] Figure 6 This is a schematic diagram of a dual-launch system. Detailed Implementation

[0050] Please see Figure 1-6 In this embodiment:

[0051] Unmanned aerial vehicle (UAV) target detection methods:

[0052] The YOLOv11 computer vision model was adopted, based on the PyTorch deep learning framework, using the labelimg annotation software, and the artificial intelligence was trained using an open-source drone dataset.

[0053] First, the dataset required for training was collected, covering drone images from various scenes including skies, cities, and mountains, under different lighting conditions such as sunny days, cloudy days, and nighttime, and from different angles. Labelimg annotation tools were used, such as... Figure 2 As shown, drones in the images are labeled using the YOLO format, and the labeled dataset is then divided into training, validation, and test sets.

[0054] Add training code to the Python script to perform training, such as... Figure 3 As shown. After training, the model is evaluated to understand its performance. The trained model is then used to perform inference on new images or videos to achieve autonomous recognition of the drone.

[0055] We tested the drone's recognition and tracking performance separately, such as Figure 4 As shown, the cruise effect around ground targets in an urban environment was tested.

[0056] After locking onto an enemy harassing drone, the next step is to adjust its speed and direction to intercept or track the intruder. Therefore, the core task is to determine the location of the drone and the intruder, estimate the speed of the intruder, and then design the direction and speed for the drone to track.

[0057] Target recognition can obtain the predicted bounding box of the target object in each frame of the camera in real time. The predicted bounding box contains four parameters. , , , and With the center coordinates, and Let S be the width and height, denoted as the region of interest (ROI). The predicted bounding box of the target object is scaled, denoted as S. The scaling factor is θ, expressed as:

[0058]

[0059] Based on the depth information provided by the camera, filter out Take the pixel information other than those in the image. The average value of the medium-depth information is used as the relative position of the target.

[0060] While the YOLOv11 algorithm boasts good speed and accuracy, the dynamic nature of both aerial targets and UAVs, coupled with camera shake and situations where targets are out of sight, can prevent continuous tracking. Furthermore, it cannot guarantee that targets will always be captured in complex environments. Referring to commonly used auxiliary techniques, we employed a Kalman filter as an auxiliary technique to address short-term occlusion, thereby improving the system's tracking performance.

[0061] The Kalman filter is a recursive algorithm. First, the object's state vector xk is established using the camera's relative position and velocity, representing the x, y, and z coordinates in the camera's coordinated coordinate system. It is assumed that the target state changes with a constant velocity, and all measurements and noise follow a Gaussian distribution. The recursive equation for estimating the system under discrete conditions is:

[0062] (1) (2)

[0063] In equations (1) and (2): F is the state transition matrix; B is the input control matrix; This is the noise matrix; H represents the observed values; H is the measurement matrix. To measure noise, the system can be further divided into two steps: prediction and update. Figure 1 This describes the state prediction and update process using Kalman filtering.

[0064] The prediction equation is

[0065] (3)

[0066] P is the covariance matrix for calculating the predicted state value and the prediction error between the predicted and actual values; Q is the system prediction noise covariance matrix, representing the reliability of the entire system, and is generally initialized to a very small value. When the UAV fails to capture the target object, the Kalman filter is activated to predict the target's current position using the data from the previous frame. The update equation is:

[0067] (4)

[0068] In equation (4): K is the optimal Kalman gain. H is the measurement matrix; H is the transition matrix; R is the observation noise covariance matrix. The motion states of the Kalman filter have a recursive relationship, and the state information can be updated in each iteration. In summary, the entire vision algorithm implementation process is as follows:

[0069] S1: The early warning system detects an intruder. The drone carrying the capture device takes off and hovers. YOLO performs target detection, detects the target, triggers and initializes the Kalman filter parameters, and the Kalman filter performs prediction to obtain the predicted state. And the prediction error covariance matrix P;

[0070] S2: Target is continuously detected with a confidence level greater than or equal to 0.7: The target state is the observed value. The Kalman filter uses the observations as input to the update equation to obtain the Kalman gain K and update the state. Update the estimated covariance matrix P. If the confidence level is less than 0.7, start the prediction process using Kalman filtering. The target state is the predicted value.

[0071] S3: If no target is detected, input the predicted parameters obtained in S1 into the Kalman filter update process, and output the posterior estimate and target state. By fusing target state and depth information, the following conclusions are drawn. The system then enters the tracking and control logic module. Based on the target information, it adjusts the UAV's position to perform tracking.

[0072] S4: To address complex situations such as brief target loss and changes in lighting, we introduced a combination of optical flow and Kalman filters to further improve the accuracy and real-time performance of enemy UAV speed estimation. The processing mechanism is as follows:

[0073] When there are insufficient optical flow feature points, the prediction relies on the Kalman filter;

[0074] When there are enough feature points, the results of optical flow and Kalman filtering are fused.

[0075] Simultaneously, it outputs the fused velocity estimate, improving stability and accuracy;

[0076] Optical flow is a commonly used motion analysis method in computer vision, which calculates the velocity of an object based on the pixel motion of an object in an image sequence. Its core principle is to calculate the motion vector of the object on the image plane by analyzing the displacement changes of pixels in two consecutive frames of images, and then combine the intrinsic and extrinsic parameters of the camera and the depth information of the scene to deduce the actual velocity of the object.

[0077] First, the video stream transmitted back by the gimbal is processed. In two adjacent frames, feature points (such as corner points or other pixels with obvious characteristics) within the target UAV region are selected. Classical optical flow calculation methods, such as the Lucas-Kanade algorithm, are used to track the displacement of these feature points between the two frames. By analyzing the displacement changes of these feature points and combining them with the time interval (determined by the video frame rate), the two-dimensional velocity component of the target UAV on the image plane can be calculated. Simultaneously, by combining the depth information provided by the KJT laser ranging radar and the system, the two-dimensional velocity on the image plane is converted into the actual motion velocity in three-dimensional space.

[0078] S5: Based on the error between the target center coordinates and the image center, corresponding control commands are generated, such as left / right movement and up / down movement. The drone's current attitude information is obtained using its own attitude sensor. Based on the target's position in the image and the drone's attitude information, the control algorithm calculates the attitude angle that the drone needs to adjust to achieve aiming of the net gun. The PD algorithm is used to dynamically adjust the drone's flight speed according to the error magnitude. Data from the DJI Magic Calculation's built-in inertial measurement unit is introduced, combined with depth information from the Celtic laser rangefinder, to compensate for the impact of drone attitude jitter on aiming in real time.

[0079] S6: When the target is identified and located in the center of the image, it is assumed that the net gun has been aimed at the target. At the same time, the target distance d < 20m, the firing conditions are met, and the gun can be fired to capture the target.

[0080] Using the coordinates of the center point of the image pixels and the intrinsic parameter matrix of the Zenmuse x3 gimbal, the relative position of the target and my drone is estimated.

[0081] Pixel center coordinate formula:

[0082] (5)

[0083] The pixel coordinates are transformed using the intrinsic parameter matrix of Zenmuse x3:

[0084] (6)

[0085] Where (μ0, ν0) are the principal pixel coordinates, (f x f y () represents the focal length in pixels.

[0086] Using DJIOSDK-Python to connect to the drone, corresponding control commands are generated based on the error between the target's center coordinates and the image center, such as left / right movement and up / down movement. The drone's own attitude sensors (such as gyroscopes and accelerometers) are used to obtain the drone's current attitude information. Based on the target's position in the image and the drone's attitude information, a control algorithm calculates the required attitude angle adjustment for the drone to achieve aiming with the net gun.

[0087] (Horizontal aiming threshold, 2°), the drone shifts to the right; (Horizontal aiming threshold, 2°), the drone shifts to the left.

[0088] (Horizontal aiming threshold, 1.5°), the drone descends; (Horizontal aiming threshold, 1.5°), drone climb.

[0089] The PD algorithm is used to dynamically adjust the drone's flight speed based on the magnitude of the error.

[0090] (7)

[0091] Where e is the current error and ∆e is the rate of change of error.

[0092] By incorporating data from the DJI Inertial Measurement Unit (IMU) built into the DJI Calculator and combining it with depth information 'd' from the Celtic laser rangefinder, the system compensates in real time for the impact of UAV attitude jitter (such as pitch and yaw angles) on aiming. The error correction formula is as follows:

[0093] (8)

[0094] in, This is the corrected azimuth angle.

[0095] Please see Figure 5 , Figure 5 It includes various evaluation metrics during the training process, such as loss function, precision, recall, and mAP, allowing a clear view of how the model's performance changes during training. The first three columns show that the loss value gradually decreases with increasing training iterations, stabilizing after 50 iterations.

[0096] The last two columns of graphs show that the recall rate is improving, indicating that the detection range is continuously expanding, fluctuating around 0.9. Therefore, based on the first two graphs, we can say that the precision of this training is 0.9, and the recall rate is 0.9.

[0097] Experiments show that fusion velocity measurement can ensure good recognition and tracking performance even in situations such as temporary target loss and changes in lighting.

[0098] Through 150 capture experiments in different scenarios, as shown in the table below, the system exhibits the following characteristics:

[0099] Test conditions Successful capture count Average interception distance (m) Capture time (s) Net winding success rate Static target (hovering) 48 / 50 8.2 ± 1.5 4.3 ± 0.8 98% Dynamic target (10m / s) 81 / 90 11.7 ± 2.3 6.1 ± 1.2 92% Complex lighting (nighttime) 27 / 30 9.5 ± 1.8 5.0 ± 1.0 95%

[0100] Experimental results show that the overall capture success rate is 90%, with a slightly lower success rate (81 / 90) in dynamic target scenarios, mainly affected by target speed and tracking delay. The optimal effective range is 5-15m; too close a distance may result in insufficient net deployment, while too far a distance is significantly affected by wind speed. At night, through Kalman filter and YOLOv11 low-light optimization, the success rate reaches 90%, with no significant difference from daytime performance.

[0101] The reasons are as follows: Although the accuracy has been improved after algorithm processing, the flight trajectory of the drone fluctuates. The drone speeding up causes it to leave the capture range. In addition, the flight trajectory of the capture net has deviations due to the influence of wind resistance, humidity and other factors. The capture net has a large area, fixed kinetic energy, high wind resistance and short flight distance. Therefore, in order to make up for the tolerance, the following adjustments are made.

[0102] The array-type integrated capture device carried by the anti-drone includes a central transmitter and a ring transmitter. The central transmitter controls the launch according to the above. The compensation launch unit of the ring transmitter obtains the two-dimensional velocity component direction of the drone on the image plane based on the optical flow calculation method, adjusts the direction of the line connecting the center of the compensation launch unit and the central transmitter, and the direction of the line connecting the center corresponds to the direction of the two-dimensional velocity component. The compensation launch unit launches synchronously with the central transmitter.

[0103] Two units are launched at a time. The compensation launch unit obtains the two-dimensional velocity component direction of the UAV on the image plane based on the optical flow calculation method. The direction of the line connecting the center of the compensation launch unit and the center launcher is adjusted so that the compensation launch unit can play the role of compensating for the capture area.

[0104] Please see Figure 6 The eccentric disk drives the central transmitter to switch to a new launch tube, and the outer ring disk drives multiple compensated launch units to switch, ensuring that there are at least two 360° adjustable launch tubes on the radial line of the outer ring disk.

[0105] Indoor testing

[0106] A simulated launch test was conducted indoors using footage captured by a drone, and the following test results were obtained.

[0107] Static target detection error: ≤2 pixels (corresponding to actual distance ≤0.5m, test distance 50m);

[0108] Tracking error of dynamic target (uniform speed of 10m / s): ≤5 pixels (after Kalman filter compensation).

[0109] Tracking delay:

[0110] Image processing latency: 120ms (including YOLOv11 inference + optical flow calculation);

[0111] Control response latency: 80ms (servo motor action + UAV attitude adjustment), total closed-loop latency ≤200ms.

[0112] Hardware compatibility:

[0113] The synchronization error between the lidar and the gimbal is ≤5ms, which meets the requirements for aligning depth information with image frames.

[0114] The impact of recoil from the net-catching gun on the attitude of the UAV: ​​pitch angle fluctuation ≤ 1.2°, which can be compensated in real time by the IMU.

[0115] Outdoor testing

[0116] Objective: To verify the system reliability under complex environments (the influence of other external objects).

[0117] Location: Bushes, buildings, and other terrain features.

[0118] Test scenarios and data:

[0119] Our group uses a fixed target for testing.

[0120] Deployment method: Deploy one drone in clusters within a 20m radius of the simulated protected object;

[0121] Test conditions: Simulate an enemy intruder (DJI Mini 3, speed 15m / s) attacking from random directions;

[0122] result:

[0123] Average warning time: 8.2 seconds (from radar detection to drone takeoff);

[0124] Interception success rate: 9 out of 10.

[0125] Additional notes: Sufficient early warning information is also needed when enemy drones attack. Building upon existing systems, this can be combined with drone detection stations, early warning radars, and low-altitude gap-filling radars. Simultaneously, the software can be upgraded to enable drones to achieve a higher degree of autonomy, freeing up operators and saving manpower.

[0126] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for capturing a drone using a drone, characterized in that: include: S1: The early warning system detects an intruder. The drone carrying the capture device takes off and hovers. YOLO performs target detection, detects the target, triggers and initializes the Kalman filter parameters, and the Kalman filter performs prediction to obtain the predicted state. And the prediction error covariance matrix P; S2: Target is continuously detected with a confidence level greater than or equal to 0.7: The target state is the observed value. The Kalman filter uses the observations as input to the update equation to obtain the Kalman gain K and update the state. Update the estimated covariance matrix P. If the confidence level is less than 0.7, start the prediction process using Kalman filtering. The target state is the predicted value. S3: If no target is detected, input the predicted parameters obtained in S1 into the Kalman filter update process, and output the posterior estimate and target state. By fusing target state and depth information, the following conclusions are drawn. The system then enters the tracking and control logic module. Based on the target information, it adjusts the UAV's position to perform tracking. S4: When there are insufficient optical flow feature points, the prediction of the Kalman filter is relied upon; when there are sufficient feature points, the results of optical flow and Kalman filtering are fused. S5: Based on the error between the target center coordinates and the image center, corresponding control commands are generated, such as left / right movement and up / down movement. The drone's current attitude information is obtained using its own attitude sensor. Based on the target's position in the image and the drone's attitude information, the control algorithm calculates the attitude angle that the drone needs to adjust to achieve aiming of the net gun. The PD algorithm is used to dynamically adjust the drone's flight speed according to the error magnitude. Data from the DJI Magic Calculation's built-in inertial measurement unit is introduced, combined with depth information from the Celtic laser rangefinder, to compensate for the impact of drone attitude jitter on aiming in real time. S6: When the target is identified and located in the center of the image, it is assumed that the net gun has been aimed at the target. At the same time, the target distance d < 20m, the firing conditions are met, and the gun can be fired to capture the target.

2. The method for capturing a drone using a drone according to claim 1, characterized in that: Target recognition can obtain the predicted bounding box of the target object in each frame of the camera in real time. The predicted bounding box contains four parameters. , , , and With the center coordinates, and Let S be the width and height, denoted as the region of interest (ROI). The predicted bounding box of the target object is scaled, denoted as S. The scaling factor is θ, expressed as: Based on the depth information provided by the camera, filter out Take the pixel information other than those in the image. The average value of the medium-depth information is used as the relative position of the target.

3. The method for capturing a drone using a drone according to claim 1, characterized in that: The object's state vector xk is established using the camera's relative position and velocity. It is assumed that the target state changes with a constant velocity, and all measurements and noise follow a Gaussian distribution. The recursive equation for estimating the system under discrete conditions is: F is the state transition matrix; B is the input control matrix; This is the noise matrix; H represents the observed values; H is the measurement matrix. To measure noise, the system can be further divided into two steps: prediction and update. The prediction equation is: P is the covariance matrix for calculating the predicted state value and the prediction error between the predicted and actual values; Q is the system prediction noise covariance matrix, representing the reliability of the entire system, and is generally initialized to a very small value. When the UAV fails to capture the target object, the Kalman filter is activated to predict the target's current position using the data from the previous frame. The update equation is: K is the optimal Kalman gain. H is the measurement matrix; H is the transition matrix; R is the observation noise covariance matrix. The motion states of the Kalman filter have a recursive relationship, and the state information can be updated in each iteration.

4. A method for capturing a drone using a drone according to claim 1, characterized in that: After the video stream transmitted back by the gimbal is processed, feature points within the target UAV area are selected in two adjacent frames. Using optical flow calculation methods, the displacement of the feature points between the two frames is tracked. By analyzing the displacement changes of the feature points and combining them with the time interval, the two-dimensional velocity component of the target UAV on the image plane is calculated. At the same time, combined with the depth information provided by the KJIT laser ranging radar and the system, the two-dimensional velocity on the image plane is converted into the actual motion velocity in three-dimensional space.

5. A method for capturing a drone using a drone according to claim 1, characterized in that: Using the image pixel center coordinates and the intrinsic parameter matrix of the Zenmuse x3 gimbal, the relative position of the target and my drone is estimated. The formula for pixel center coordinates is: Using the intrinsic parameter matrix of Zenmuse x3, the pixel coordinates are transformed using the following formula: in( , ) are the main pixel coordinates, ( , () represents the focal length in pixels.

6. A method for capturing a drone using a drone according to claim 5, characterized in that: After pixel coordinate transformation: (Horizontal aiming threshold, 2°), the drone shifts to the right; (Horizontal aiming threshold, 2°), the drone shifts to the left; (Horizontal aiming threshold, 1.5°), drone descends; #imgpt34# (Horizontal aiming threshold, 1.5°), drone ascends.

7. A method for capturing a drone using a drone according to claim 6, characterized in that: The PD algorithm is used to dynamically adjust the drone's flight speed based on the magnitude of the error. #imgpt35# Where e is the current error and ∆e is the error rate of change, data from the DJI Inertial Measurement Unit (IMU) is incorporated into the system, combined with depth information d from the Celtic laser rangefinder, to compensate for the impact of drone attitude jitter on aiming in real time. The error correction formula is as follows: #imgpt36# Where #imgpt37# represents the corrected azimuth angle.

8. A method for capturing a drone using a drone according to claim 1, characterized in that: When the target is identified and located in the center of the image (within an acceptable range of error, i.e., satisfying the formula #imgpt38#), it is considered that the net gun has been aimed at the target, and the target distance d < 20m, thus satisfying the firing conditions.

9. A method for capturing a drone using a drone according to claim 4, characterized in that: An array-type integrated capture device for anti-drone carry includes a central transmitter and a ring transmitter. The central transmitter is controlled to launch in accordance with the method of claims 1-8. The compensation launch unit of the ring transmitter obtains the two-dimensional velocity component direction of the drone on the image plane according to the optical flow calculation method, adjusts the direction of the line connecting the center of the compensation launch unit and the central transmitter, and the direction of the line connecting the center corresponds to the direction of the two-dimensional velocity component. The compensation launch unit launches synchronously with the central transmitter.

Images