A trajectory prediction-based unmanned aerial vehicle capturing system and capturing method

Abstract

The application discloses a kind of based on trajectory prediction's unmanned plane capture system and capture method, based on trajectory prediction's unmanned plane capture system includes: capture component, adjusting component and visual component, capture component includes cylinder and at least one buffer, the inside of cylinder is hollow, and one end is open, buffer is built-in in cylinder, and with the capture hole of adjustable size, adjusting component includes first driving part, first driving part is connected to cylinder, and can drive cylinder freely move in space, visual component includes vision module, vision module is connected to cylinder, for calibration unmanned plane, vision module is electrically connected with first driving part, for driving cylinder freely move in space, so that unmanned plane enters into capture hole.The present application can effectively solve the problem that the relative motion between unmanned plane and recovery platform has high uncertainty and nonlinear characteristics under the complex sea conditions of continuous ship sway, so as to cause the high failure rate of capture.

Description

Technical Field

[0001] This invention relates to the field of drone recovery technology, specifically to a drone capture system and method based on trajectory prediction. Background Technology

[0002] Traditional methods for recovering fixed-wing drones include net-crash recovery, hook recovery, and parachute recovery. These methods are all passive recovery methods, where the recovery device is fixed in place and the drone relies on itself to fly precisely into the recovery window.

[0003] For example, Chinese invention patent CN109502042B, entitled "A Drone Capture Device," includes a turntable, a base mounted on the turntable, two support rods symmetrically arranged from the base, two vertically arranged pendulum rods at the top of the support rods that can rotate around a rotation axis mounted on the support rods, a bracket perpendicular to the support rods between the two vertically arranged pendulum rods, two vertically arranged first magnetorheological dampers hinged to the ends of the brackets and hinged to the pendulum rods, a barrier net suspended at one end of the pendulum rod away from the support rods, and the other end hinged to the first magnetorheological dampers; the base is equipped with a telescopic device with a rotating bottom and a connecting block at the top, and a pull rod with a strip-shaped through hole in the bottom of the support rod hinged to each end of the connecting block. This device enables rotatable capture and recovery of drones, effectively absorbing the kinetic energy of high-speed drones during recovery and preventing damage to the drones during capture.

[0004] However, in complex sea conditions where ships are constantly rocking, the relative motion between the drone and the recovery platform has high uncertainty and nonlinear characteristics. Relying solely on the drone's autonomous navigation makes it difficult to achieve precise docking, resulting in a high recovery failure rate. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned technical deficiencies and propose a UAV capture system and method based on trajectory prediction, which solves the technical problem of high capture failure rate caused by the high uncertainty and nonlinearity of the relative motion between the UAV and the recovery platform under complex sea conditions of continuous ship swaying.

[0006] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a drone capture system based on trajectory prediction, comprising: A capture assembly includes a cylinder and at least one buffer, the cylinder being hollow inside and open at one end, the buffer being built into the cylinder and having a capture hole of adjustable size; The adjustment assembly includes a first driving member connected to the cylinder and capable of driving the cylinder to move freely within space; and The vision component includes a vision module connected to the cylinder for calibrating the drone. The vision module is electrically connected to the first drive unit for driving the cylinder to move freely in space so that the drone enters the capture hole.

[0007] In some embodiments, the number of buffer elements is multiple, and the multiple buffer elements are arranged sequentially at intervals along the axial direction of the cylinder and are all connected to the inner wall of the cylinder.

[0008] In some embodiments, the cylinder is trumpet-shaped, and the inner diameter of the cylinder gradually increases along the direction near its open end, and the outer diameter of the plurality of buffer members is set to match the inner diameter of the cylinder.

[0009] In some embodiments, the buffer includes a fixed disk, a plurality of movable blocks, and a plurality of first driving units. The fixed disk is coaxially arranged with the cylinder and connected to the inner wall of the cylinder. The fixed disk has a through hole. The plurality of movable blocks are arranged opposite to the through hole and abut against each other in sequence to form the capture hole. The plurality of movable blocks can slide relative to the fixed disk. The first driving units are arranged one-to-one with the movable blocks. The first driving units are connected to the fixed disk and the movable blocks and are used to drive the movable blocks to slide relative to the next adjacent movable block to adjust the size of the capture hole.

[0010] In some embodiments, the vision module includes a connector and a binocular camera. The binocular camera is detachably connected to the side wall of the cylinder via the connector and is positioned relative to the cylinder near its opening end. The binocular camera is used to acquire intrinsic parameter matrices, distortion coefficients, and stereo correction parameters.

[0011] In some embodiments, the trajectory prediction-based UAV capture system further includes an embedded computing platform, which is electrically connected to the first drive unit, the first drive section, and the binocular camera.

[0012] In some embodiments, the embedded computing platform includes a feature extraction module, a trajectory prediction module, a motion control module, and a grasping switching module. The feature extraction module is electrically connected to the binocular camera. The trajectory prediction module is connected to the feature extraction module, the motion control module, and the grasping switching module. The motion control module is electrically connected to the first driving component, and the grasping switching module is electrically connected to the first driving component.

[0013] In some embodiments, the adjustment assembly further includes a clamping member having a fixed end and a clamping end, the fixed end of the clamping member being connected to the first driving member, and the clamping end of the clamping member being detachably connected to the cylinder.

[0014] In some embodiments, the trajectory prediction-based UAV capture system further includes a data processing unit, which is electrically connected to the binocular camera, the first drive unit, the first drive component, the feature extraction module, the trajectory prediction module, the motion control module, and the grasping switching module.

[0015] Secondly, the present invention also provides a method for capturing unmanned aerial vehicles (UAVs), which utilizes the trajectory prediction-based UAV capture system described in any of the above claims, and the specific steps are as follows: Binocular camera is used for binocular positioning to obtain intrinsic parameter matrix, distortion coefficients and stereo correction parameters; The drone's nose is detected and key points are located in the binocular images, and the bounding box of the nose and the pixel coordinates of the preset key points are output. Epipolar matching is performed on the same key point extracted from the binocular images, and the three-dimensional coordinates of the point in the camera coordinate system are obtained by triangulation. Using the nose position and velocity sequences of the past N frames as input, a multi-layer LSTM network is constructed, and the network outputs the predicted nose position sequence in the future time τ. The predicted future position of the nose cone is used as the desired alignment point for the opening end of the cylinder. Based on the DH parameter model of the first driving component, establish the transformation matrix; The process involves: acquiring images with binoculars, extracting feature points using deep learning, calculating the current 3D position / velocity, predicting the future trajectory using LSTM, solving for the desired end pose, calculating joint angles using inverse kinematics, and driving the first actuator. The data processing unit records and corrects tracking errors and adjusts MPC parameters online.

[0016] Compared with existing technologies, the beneficial effects of the trajectory prediction-based UAV capture system and method provided by this invention include: at least one buffer is provided inside the cylinder, the buffer having an adjustable capture hole for capturing the UAV; a first drive unit can drive the cylinder to move freely in space; a vision module is connected to the cylinder for calibrating the UAV, and the vision module is electrically connected to the first drive unit to drive the cylinder to move freely in space so that the UAV enters the capture hole. Compared with existing technologies, by calibrating the position of the UAV through the vision module and coordinating with the first drive unit to drive the cylinder to move freely in space so that the UAV enters the capture hole, and with at least one buffer forming the capture hole, dynamic capture of the UAV can be achieved. This reduces the impact of the high uncertainty and nonlinearity of the relative motion between the UAV and the recovery platform on UAV capture, making it suitable for use in complex sea conditions where ships are constantly swaying. It does not only rely on autonomous navigation of the UAV, but also achieves precise docking of the UAV, solving the technical problem of high capture failure rate caused by the high uncertainty and nonlinearity of the relative motion between the UAV and the recovery platform in complex sea conditions where ships are constantly swaying. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a drone capture system based on trajectory prediction provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the connection between the clamping member and the capturing assembly according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a capture component provided in an embodiment of the present invention; Figure 4 This is a schematic diagram showing the electrical connection between a data processing unit and an embedded computing platform, a binocular camera, a first driving element, and a first driving unit according to an embodiment of the present invention; Figure 5 This is a flowchart of an LSTM-based UAV trajectory prediction method provided in an embodiment of the present invention; Figure 6 This is a closed-loop control flowchart of the motion and visual tracking of the first driving component provided in an embodiment of the present invention; Figure 7 This is a flowchart of the near-field detection and grasping switching logic provided in an embodiment of the present invention; Figure 8 This is a flowchart of system failure protection and anomaly handling provided in an embodiment of the present invention.

[0018] Explanation of reference numerals in the attached figures: Capture component 100; cylinder 110; buffer 120; capture hole 121; fixed disk 122; movable block 123; first drive unit 124; adjustment component 200; first drive unit 210; clamping component 220; vision component 300; vision module 310; connector 311; binocular camera 312; embedded computing platform 400; feature extraction module 410; trajectory prediction module 420; motion control module 430; grasping switching module 440; data processing unit 500. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] To address the high capture failure rate caused by the highly uncertain and nonlinear relative motion between the UAV and the recovery platform under complex sea conditions with continuous ship swaying, this application provides a UAV capture system and method based on trajectory prediction. The system uses a vision module 310 to pinpoint the UAV's position and, in conjunction with a guiding first drive component 210, drives the cylinder 110 to move freely in space, allowing the UAV to enter the capture hole 121. Simultaneously, the capture hole 121 has at least one buffer component 120, enabling dynamic capture of the UAV. This reduces the impact of the highly uncertain and nonlinear relative motion between the UAV and the recovery platform on UAV capture. It is suitable for use under complex sea conditions with continuous ship swaying and, beyond relying solely on autonomous navigation, enables precise docking of the UAV.

[0021] Please see Figure 1 , Figure 2 This is a schematic diagram of a trajectory prediction-based drone capture system and method according to an embodiment of this application. The trajectory prediction-based drone capture system includes a capture component 100, an adjustment component 200, and a vision component 300. The capture component 100 includes a cylinder 110 and at least one buffer 120. The cylinder 110 is hollow inside and open at one end. The buffer 120 is built into the cylinder 110 and has an adjustable capture hole 121. The adjustment component 200 includes a first drive component 210 connected to the cylinder 110 and capable of driving the cylinder 110 to move freely in space. The vision component 300 includes a vision module 310 connected to the cylinder 110 for calibrating the drone. The vision module 310 is electrically connected to the first drive component 210 and is used to drive the cylinder 110 to move freely in space so that the drone enters the capture hole 121.

[0022] In this device, compared to existing technologies, the position of the UAV is determined by the vision module 310, and the first driving component 210 is used to guide the cylinder 110 to move freely in space, allowing the UAV to enter the capture hole 121. At the same time, the capture hole 121 has at least one buffer component 120, which enables dynamic capture of the UAV. This reduces the impact of the high uncertainty and nonlinearity of the relative motion between the UAV and the recovery platform on UAV capture. It is suitable for use in complex sea conditions where the ship is constantly swaying. It does not only rely on the UAV's autonomous navigation, but also enables precise docking of the UAV. It can solve the technical problem of high capture failure rate caused by the high uncertainty and nonlinearity of the relative motion between the UAV and the recovery platform in complex sea conditions where the ship is constantly swaying.

[0023] Furthermore, the cylinder 110 here is a common and readily available stainless steel cylinder structure, which is a conventional setting known to those skilled in the art, and will not be described in detail here.

[0024] In some of these embodiments, such as Figures 1 to 3 As shown, there are multiple buffer elements 120, which are arranged sequentially at intervals along the axial direction of the cylinder 110 and are all connected to the inner wall of the cylinder 110.

[0025] By setting multiple buffers 120, the success rate of drone capture can be improved.

[0026] Furthermore, multiple buffers 120 arranged at intervals can be used to achieve multiple buffering and interception.

[0027] In some of these embodiments, such as Figure 3 As shown, the cylinder 110 is trumpet-shaped, and the inner diameter of the cylinder 110 gradually increases along the direction close to its opening end, and the outer diameter of the multiple buffer members 120 is set to match the inner diameter of the cylinder 110.

[0028] By setting up a trumpet-shaped cylinder 110, the drone can be effectively decelerated, thus improving the success rate of capture.

[0029] Furthermore, the dimensions of the multiple buffer components 120 are matched with the inner diameter of the cylinder 110, which can improve the success rate of drone capture.

[0030] In this embodiment, as Figure 2As shown, the buffer 120 includes a fixed disk 122, multiple movable blocks 123, and multiple first drive units 124. The fixed disk 122 is coaxially arranged with the cylinder 110 and connected to the inner wall of the cylinder 110. The fixed disk 122 has a through hole. The multiple movable blocks 123 are arranged opposite to the through hole and abut against each other in sequence to form a capture hole 121. The multiple movable blocks 123 can slide relative to the fixed disk 122. The first drive units 124 are arranged one-to-one with the movable blocks 123. The first drive units 124 are connected to the fixed disk 122 and the movable blocks 123 and are used to drive the movable blocks 123 to slide relative to the next adjacent movable block 123 to adjust the size of the capture hole 121.

[0031] Multiple movable blocks 123 that can slide relative to the fixed disk 122 are arranged inside the fixed disk 122. The movable blocks 123 abut against each other and surround each other to form a capture hole 121 that can be adjusted in size.

[0032] Furthermore, the number of active blocks 123 is M, and the number of M is greater than or equal to 6, preferably 8 to 12.

[0033] In some embodiments, the movable block 123 slides radially along the fixed disk 122.

[0034] In one embodiment, the active block 123 is a flexible structure.

[0035] By incorporating a flexible movable block 123, the impact force when the recovery device comes into contact with the drone can be reduced.

[0036] Furthermore, the active block 123 here is a common and readily available device on the market, and will not be described in detail here.

[0037] In some embodiments, each movable block 123 is also equipped with an independent airbag mechanism. The airbag is made of a high-strength airtight composite material, which has good flexibility and fatigue resistance.

[0038] Furthermore, all independent airbag mechanisms are evenly distributed circumferentially, and all airbags on the same cross section together form the gripping cavity at that cross section; the airbag arrays with different cross sections along the axial direction together form a three-dimensional deformable gripping interface, and the inner surface of the airbag is provided with anti-slip texture or high friction coating to increase the static friction force with the surface of the machine head after gripping and prevent slippage.

[0039] In this embodiment, as Figure 2 As shown, the vision module 310 includes a connector 311 and a binocular camera 312. The binocular camera 312 is detachably connected to the side wall of the cylinder 110 via the connector 311 and is positioned relative to the cylinder 110 near its opening end. The binocular camera 312 is used to acquire intrinsic parameter matrices, distortion coefficients and stereo correction parameters.

[0040] The binocular camera 312 is detachably connected to the cylindrical body 110 via the connector 311, which can protect the binocular camera 312.

[0041] Furthermore, the binocular camera 312 here is a common and readily available industrial camera on the market, such as the Basler Ensenso N / S series and the Orbbec Gemini 330 series. This is a conventional setup known to those skilled in the art, and will not be described in detail here.

[0042] In some of these embodiments, such as Figure 4 As shown, the trajectory prediction-based UAV capture system also includes an embedded computing platform 400, which is electrically connected to the first drive unit 210, the first drive unit 124, and the binocular camera 312.

[0043] The embedded computing platform 400 is electrically connected to the binocular camera 311, the first drive unit 210 and the first drive unit 124 to improve the drone's capture effect.

[0044] In some of these embodiments, such as Figure 4 As shown, the embedded computing platform 400 includes a feature extraction module 410, a trajectory prediction module 420, a motion control module 430, and a grasping switching module 440. The feature extraction module 410 is electrically connected to the binocular camera 312. The trajectory prediction module 420 is connected to the feature extraction module 410, the motion control module 430, and the grasping switching module 440. The motion control module 430 is electrically connected to the first driving unit 210, and the grasping switching module 440 is electrically connected to the first driving unit 124.

[0045] The automatic capture of the UAV can be achieved by using the feature extraction module 410, trajectory prediction module 420, motion control module 430 and grasping switching module 440.

[0046] Furthermore, the feature extraction module 410, trajectory prediction module 420, motion control module 430, and grasping switching module 440 are deployed in the upper unit to form an embedded computing platform 400.

[0047] In some embodiments, the feature extraction module 410 can perform deep learning feature extraction.

[0048] In some embodiments, the trajectory prediction module 420 is based on LSTM trajectory prediction.

[0049] In some embodiments, the motion control module 430 can achieve automatic control of the first drive element 210.

[0050] In some embodiments, the grasping switching module 440 can realize near-field detection and grasping switching of the UAV.

[0051] In this embodiment, as Figure 2 As shown, the adjustment assembly 200 also includes a clamping member 220, which has a fixed end and a clamping end. The fixed end of the clamping member 220 is connected to the first driving member 210, and the clamping end of the clamping member 220 is detachably connected to the cylinder 110.

[0052] The clamping member 220 can be used to clamp and release the first driving member 210 from the cylinder 110.

[0053] Furthermore, the clamping element 220 can improve the stability of the connection of the cylinder 110.

[0054] In some embodiments, the clamping member 220 includes two clamping claws and a first rotation drive motor. The two clamping claws are arranged opposite each other, and the first rotation drive motor is connected to one end of each of the two clamping claws. It is used to drive the other ends of the two clamping claws to move closer to each other or further away from each other, so that the two clamping claws clamp or release the cylinder 110. The clamping claws are connected to the first rotation drive motor. By operating the first rotation drive motor, the first rotation drive motor can drive the other ends of the two clamping claws to move closer to each other or further away from each other, so that the two clamping claws clamp or release the cylinder 110.

[0055] In some embodiments, the two gripping claws are connected by gear transmission to achieve synchronous adjustment, which will not be described in detail here.

[0056] In some embodiments, the first driving member 210 includes at least two swing arms, at least two second rotation drive motors, a third rotation drive motor, and a fourth rotation drive motor. The swing arms are arranged sequentially. The output end of each second rotation drive motor is connected to the head end of each swing arm, driving each swing arm to rotate around its own axis of rotation in a vertical plane to adjust the tilt angle of each swing arm. The fixed end of the third rotation drive motor is fixedly connected to the end of the last swing arm, and its output end is connected to the fixed end of the first rotation drive motor, driving the first rotation drive motor to rotate to adjust the elevation angle of the cylinder 110. The output end of the fourth rotation drive motor is fixedly connected to the fixed end of the first second rotation drive motor. The first rotary drive motor is connected to the second rotary drive motor to rotate in the horizontal plane to adjust the orientation of the cylinder 110. When the second rotary drive motor is started, its output end rotates, which can drive the corresponding swing arm to rotate around its own rotation axis in the vertical plane. This can adjust the tilt angle of each swing arm, thereby adjusting the height of the cylinder 110. When the third rotary drive motor is started, its output end rotates, driving the first rotary drive motor to rotate, which can adjust the elevation angle of the cylinder 110. When the fourth rotary drive motor is started, its output end rotates, driving the second rotary drive motor to rotate in the horizontal plane, which can adjust the orientation of the cylinder 110, allowing the drone to smoothly enter the cylinder 110.

[0057] In one embodiment, such as Figure 4 As shown, the UAV capture system based on trajectory prediction also includes a data processing unit 500, which is electrically connected to the binocular camera 312, the first drive unit 124, the first drive component 210, the feature extraction module 410, the trajectory prediction module 420, the motion control module 430, and the grasping switching module 440.

[0058] The data processing unit 500 can achieve automated control.

[0059] Furthermore, the data processing unit 500 here is responsible for the collaboration and data synchronization between multiple modules.

[0060] This embodiment also provides a drone capture method, which utilizes a drone capture system based on trajectory prediction. The specific steps are as follows: Binocular camera 312 was used for binocular positioning to obtain intrinsic parameter matrix, distortion coefficients and stereo correction parameters; The drone's nose is detected and key points are located in the binocular images, and the bounding box of the nose and the pixel coordinates of the preset key points are output. Epipolar matching is performed on the same key point extracted from the binocular images, and the three-dimensional coordinates of the point in the camera coordinate system are obtained by triangulation. Using the nose position and velocity sequences of the past N frames as input, a multi-layer LSTM network is constructed, and the network outputs the predicted nose position sequence in the future time τ. The predicted future position of the nose is used as the desired alignment point of the opening end of the cylinder 110. Based on the DH parameter model of the first driving component 210, establish the transformation matrix; Periodic loop execution: binocular image acquisition, deep learning to extract feature points, calculation of current 3D position / velocity, LSTM to predict future trajectory, calculation of desired end pose, inverse kinematics to calculate joint angles, and first actuator 210 driving; The data processing unit 500 records and corrects tracking errors and adjusts MPC parameters online.

[0061] In addition, in some implementations, a binocular camera 312 is used for binocular positioning to obtain the intrinsic parameter matrix, distortion coefficients and stereo correction parameters. The left and right images are acquired synchronously using a hardware triggering method with a frame rate of no less than 60 frames / second. The acquired images are transmitted to the computing platform in real time and stamped with a unified timestamp, laying the foundation for subsequent time series analysis.

[0062] In addition, in some implementations, an improved lightweight convolutional neural network (such as YOLOv8n or MobileNet-SSD) is used to detect and locate key points of the drone's head in the left and right images. The network outputs the head bounding box and the pixel coordinates of preset key points (propeller center, nose tip, and points on both sides of the fuselage axis).

[0063] Furthermore, to improve the robustness of feature point extraction, an attention mechanism is introduced to suppress background interference such as wave reflection and lighting changes. The network is pre-trained on synthetic datasets containing different sea conditions and aircraft models, as well as actual aerial photography datasets, and can maintain high detection accuracy in complex environments.

[0064] Furthermore, in some implementations, based on the principle of binocular stereo matching, epipolar matching is performed on the same key point extracted from the left and right images, and the three-dimensional coordinates (x, y, z) of that point in the camera coordinate system are obtained through triangulation. c ,y c ,z c ).

[0065] Furthermore, considering that the camera is fixed to the cylinder 110, and the cylinder 110 moves with the first driving component 210, it is necessary to transform the head coordinates to the world coordinate system or the base coordinate system of the first driving component 210. The transformation relationship is obtained through robot hand-eye calibration. At the same time, the position data of multiple consecutive frames are differentially analyzed to calculate the three-dimensional velocity vector (v) of the head at the current moment. x ,v y ,v z ).

[0066] Furthermore, to reduce noise, a first-order low-pass filter is used to smooth the position and velocity.

[0067] Furthermore, in some implementations, an LSTM-based method for predicting the trajectory of a fixed-wing UAV is employed. Specifically, the nose position and velocity sequences of the past N frames (N = 10-20) are used as input to construct a multi-layer LSTM network (2-3 hidden layers, 64-128 units per layer). The network outputs the predicted nose position sequence for the next time interval τ (τ = 0.2-0.5 seconds). During training, a large amount of real or simulated UAV approach trajectory data is used for supervised learning, and the loss function is the mean square error between the predicted and actual positions.

[0068] Furthermore, a rolling window strategy is adopted for prediction: each time a new frame of data is received, the historical sequence is updated and the future trajectory is re-predicted to ensure the real-time performance and accuracy of the prediction.

[0069] In addition, in some implementations, the predicted future position of the machine head (with the sampling step size consistent with the control cycle of the first drive unit 210) is used as the expected alignment point of the cylinder 110 opening.

[0070] Furthermore, considering that the cylinder 110 is a hollow cylinder, its central axis should be aligned with the nose approach direction. Therefore, it is also necessary to calculate the desired attitude of the cylinder 110: determine the cylinder opening normal vector based on the tangent direction of the predicted trajectory, so that the angle between the axis of the cylinder 110 and the incoming flow direction of the UAV is minimized; finally, obtain the six-degree-of-freedom desired pose (x) of the end of the first drive component 210. d ,y d ,z d ,R d ,P d ,Y d ).

[0071] In addition, in some implementations, a transformation matrix from the base coordinate system to the end flange is established based on the DH parameter model of the first driving component 210; the inverse kinematics is solved using analytical or numerical iteration methods to obtain the target angles (θ1~θ6) of each joint.

[0072] Furthermore, to ensure smooth motion, fifth-order polynomial interpolation is used to plan the joint trajectory, and amplitude is limited according to the maximum joint velocity and acceleration constraints; the control period is set to 10-20 milliseconds (50-100Hz).

[0073] Furthermore, traditional feedback control has inherent delays, including image acquisition (15-30ms), deep learning inference (5-10ms), communication transmission (1-2ms), and joint servo response (5-10ms), with a total delay of up to 30-50ms.

[0074] To compensate for this delay, this application adopts a model predictive control framework: the dynamic model of the first driving component 210 is used as the prediction model, and the future trajectory predicted by LSTM is used as a reference. The control input sequence is optimized in the rolling time domain to minimize the deviation between the actual trajectory of the first driving component 210 and the predicted trajectory. This MPC controller can predict and "advance" the movement in advance, thereby offsetting the effect of system delay.

[0075] Furthermore, in some implementations, in a ship swaying environment, the base of the first drive component 210 experiences low-frequency swaying. To accurately sense the motion of the UAV relative to inertial space, an inertial measurement unit (IMU) is installed on the base of the first drive component 210 to measure the three-axis angular velocity and linear acceleration of the base. This measurement value is then fused with the feedback from the joint encoder of the first drive component 210, and the actual pose of the end of the cylinder 110 in the inertial coordinate system is estimated through an extended Kalman filter (EKF). This pose is then compared with the predicted trajectory of the UAV to eliminate the influence of the base motion on the tracking control.

[0076] In some implementations, within each control cycle, the system sequentially performs: binocular image acquisition, deep learning to extract feature points, calculation of current 3D position / velocity, LSTM prediction of future trajectory, calculation of desired end pose, inverse kinematics to calculate joint angles, and joint servo drive.

[0077] Furthermore, the closed loop operates at a frequency above 60Hz to ensure that the end of the cylinder 110 is continuously aligned with the predicted impact point, while the data processing unit 500 records and corrects tracking errors for online adjustment of MPC parameters.

[0078] In addition, in some implementations, when the near-field sensor at the opening of the cylinder 110 detects that the head has entered a predetermined range (e.g., 0.1 meters from the opening), the system immediately switches from "active tracking mode" to "grabbing and clamping mode". In this mode, the first drive unit 210 stops active movement (or switches to low-impedance follow-up), and the airbag or mechanical iris mechanism inside the cylinder 110 inflates instantly to clamp the head and complete the capture. This switching signal has the highest priority to ensure that the capture opportunity is not missed.

[0079] Furthermore, if visual tracking is lost or the uncertainty of LSTM prediction is too high (such as the prediction variance exceeding the threshold), the system automatically switches to safety mode: the first drive unit 210 stops moving and is placed in a preset standby position, while issuing a warning signal.

[0080] Furthermore, all state information is recorded for subsequent algorithm optimization.

[0081] To better understand this application, the following is combined with... Figures 1 to 4 The technical solution of this application is described in detail below: At least one buffer 120 is provided inside the cylinder 110. The buffer 120 has an adjustable capture hole 121 for capturing the UAV. A first drive unit 210 can drive the cylinder 110 to move freely in space. A vision module 310 is connected to the cylinder 110 for calibrating the UAV, and the vision module 310 is electrically connected to the first drive unit 210 to drive the cylinder 110 to move freely in space so that the UAV enters the capture hole 121. Compared with the prior art, by using the vision module 310 to calibrate the position of the UAV and cooperating with the first drive unit 210 to drive the cylinder 110 to move freely in space so that the UAV enters the capture hole 121, and the capture hole 121 has at least one buffer 120, dynamic capture of the UAV can be achieved. This reduces the impact of the high uncertainty and nonlinear characteristics of the relative motion between the UAV and the recovery platform on UAV capture. It is suitable for use in complex sea conditions where the ship is constantly swaying. It does not only rely on the UAV's autonomous navigation, but also achieves precise docking of the UAV.

[0082] This application utilizes an LSTM deep learning model to achieve accurate prediction of UAV trajectories, effectively compensate for system latency, and significantly improve the capture success rate.

[0083] Furthermore, traditional visual tracking methods only utilize the current frame position for robotic arm control. Due to delays in image acquisition, processing, communication, and actuators, the robotic arm always "chases" the drone, especially during high-speed approaches, where it is easy to miss the capture window. This application uses an LSTM model to learn the temporal motion pattern of the drone during its approach, predicting its position 0.2 to 0.5 seconds in advance. This allows the robotic arm to "predict" the drone's direction, achieving advanced movement and significantly improving the recovery success rate in complex sea conditions.

[0084] This application constructs an integrated intelligent control chain consisting of visual perception, trajectory prediction, motion planning, and closed-loop control, enabling quasi-static docking in dynamic environments.

[0085] Furthermore, binocular vision, deep learning feature extraction, LSTM prediction, MPC compensation, and robotic arm servo control are deeply integrated to form an end-to-end active acquisition closed loop. By fusing and estimating the base motion caused by ship swaying with visual data through IMU, the impact of environmental disturbances on tracking accuracy is effectively eliminated.

[0086] In sea states 4-5, the recovery canister can still maintain centimeter-level alignment with the UAV's approach path, transforming the traditional "dynamic docking" problem into a stable state close to "static docking," greatly reducing the dependence on the UAV's flight control precision.

[0087] This application introduces model predictive control (MPC) and state estimation techniques to simultaneously compensate for system delays and platform disturbances, thereby improving control robustness.

[0088] Furthermore, existing active tracking methods for robotic arms mostly employ simple PID feedback control, which is difficult to cope with nonlinear time-varying delays and external disturbances. By adopting the MPC framework and using look-ahead prediction trajectory as a reference, the control input is optimized, enabling the robotic arm to exhibit "predictive" motion characteristics and effectively suppressing overshoot and oscillations caused by delays. At the same time, the Kalman filter state estimation that integrates IMU data and joint feedback improves the accuracy of base motion compensation.

[0089] This application uses a lightweight deep learning model to extract nose feature points in real time, balancing detection accuracy and inference speed to meet real-time requirements.

[0090] Furthermore, the deep learning feature extraction network used is pruned and quantized while maintaining a high detection rate for key points on the nose. Compared with traditional methods based on manually designed features (such as HOG+ SVM), the deep learning method has stronger robustness to complex scenes such as changes in lighting, partial occlusion, and reflections from the sea surface, avoiding tracking failures caused by feature loss.

[0091] This application employs a modular algorithm design to support flexible deployment and hardware upgrades, adapting to robotic arms and drone models of different tonnages.

[0092] Furthermore, the vision, prediction, and control modules are all designed with standardized input / output interfaces, which can be easily replaced with more advanced algorithms (such as Transformer predictors and impedance controllers). The prediction time domain and control frequency can also be adjusted according to the load capacity of the robotic arm. For drones of different sizes and speeds, only the trajectory samples in the training dataset need to be updated to adapt, demonstrating good versatility and scalability.

[0093] This application establishes a near-field detection and tracking-grabbing smooth switching mechanism to ensure the reliability and smoothness of the capture moment.

[0094] Furthermore, at the critical stage when the drone is about to make contact with the recovery container, the system triggers a switch from active tracking mode to grasping mode through a near-field sensor. At this time, the impedance control of the robotic arm makes the recovery container have a certain degree of compliance, avoiding rigid collisions; at the same time, the airbag inflates in time to wrap around the nose of the drone. This switching strategy not only ensures the capture accuracy, but also protects the structure of the drone.

[0095] This application, through the aforementioned system and method, can solve the technical problem in the prior art where the relative motion between the UAV and the recovery platform has high uncertainty and nonlinear characteristics under complex sea conditions with continuous ship swaying, resulting in a high capture failure rate.

[0096] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A drone capture system based on trajectory prediction, characterized in that, include: A capture assembly includes a cylinder and at least one buffer, the cylinder being hollow inside and open at one end, the buffer being built into the cylinder and having a capture hole of adjustable size; The adjustment component includes a first driving member, which is connected to the cylinder and can drive the cylinder to move freely in space. as well as The vision component includes a vision module connected to the cylinder for calibrating the drone. The vision module is electrically connected to the first drive unit for driving the cylinder to move freely in space so that the drone enters the capture hole.

2. The UAV capture system based on trajectory prediction according to claim 1, characterized in that, The number of buffer components is multiple, and the multiple buffer components are arranged sequentially at intervals along the axial direction of the cylinder and are all connected to the inner wall of the cylinder.

3. The UAV capture system based on trajectory prediction according to claim 2, characterized in that, The cylinder is trumpet-shaped, and the inner diameter of the cylinder gradually increases along the direction close to its opening end, and the outer diameter of the plurality of buffer members is set to match the inner diameter of the cylinder.

4. The UAV capture system based on trajectory prediction according to claim 2, characterized in that, The buffer includes a fixed disk, multiple movable blocks, and multiple first driving units. The fixed disk is coaxially arranged with the cylinder and connected to the inner wall of the cylinder. The fixed disk has a through hole. The multiple movable blocks are arranged opposite to the through hole and abut against each other in sequence to form the capture hole. The multiple movable blocks can slide relative to the fixed disk. The first driving units are arranged one-to-one with the movable blocks and are connected to the fixed disk and the movable blocks. The first driving units are used to drive the movable blocks to slide relative to the next adjacent movable block to adjust the size of the capture hole.

5. The UAV capture system based on trajectory prediction according to claim 4, characterized in that, The vision module includes a connector and a binocular camera. The binocular camera is detachably connected to the side wall of the cylinder via the connector and is positioned relative to the cylinder and close to its opening end. The binocular camera is used to acquire intrinsic parameter matrix, distortion coefficients and stereo correction parameters.

6. The UAV capture system based on trajectory prediction according to claim 5, characterized in that, The trajectory prediction-based UAV capture system also includes an embedded computing platform, which is electrically connected to the first driving component, the first driving unit, and the binocular camera.

7. The UAV capture system based on trajectory prediction according to claim 6, characterized in that, The embedded computing platform includes a feature extraction module, a trajectory prediction module, a motion control module, and a grasping switching module. The feature extraction module is electrically connected to the binocular camera. The trajectory prediction module is connected to the feature extraction module, the motion control module, and the grasping switching module. The motion control module is electrically connected to the first driving component, and the grasping switching module is electrically connected to the first driving component.

8. The UAV capture system based on trajectory prediction according to claim 6, characterized in that, The adjustment assembly further includes a clamping member having a fixed end and a clamping end. The fixed end of the clamping member is connected to the first driving member, and the clamping end of the clamping member is detachably connected to the cylinder.

9. The UAV capture system based on trajectory prediction according to claim 7, characterized in that, The trajectory prediction-based UAV capture system also includes a data processing unit, which is electrically connected to the binocular camera, the first drive unit, the first drive component, the feature extraction module, the trajectory prediction module, the motion control module, and the capture switching module.

10. A method for capturing unmanned aerial vehicles (UAVs), characterized in that, The specific steps of using the trajectory prediction-based UAV capture system as described in any one of claims 1-9 are as follows: Binocular camera is used for binocular positioning to obtain intrinsic parameter matrix, distortion coefficients and stereo correction parameters; The drone's nose is detected and key points are located in the binocular images, and the bounding box of the nose and the pixel coordinates of the preset key points are output. Epipolar matching is performed on the same key point extracted from the binocular images, and the three-dimensional coordinates of the point in the camera coordinate system are obtained by triangulation. Using the nose position and velocity sequences of the past N frames as input, a multi-layer LSTM network is constructed, and the network outputs the predicted nose position sequence in the future time τ. The predicted future position of the nose cone is used as the desired alignment point for the opening end of the cylinder. Based on the DH parameter model of the first driving component, establish the transformation matrix; The process involves: acquiring images with binoculars, extracting feature points using deep learning, calculating the current 3D position / velocity, predicting the future trajectory using LSTM, solving for the desired end pose, calculating joint angles using inverse kinematics, and driving the first actuator. The data processing unit records and corrects tracking errors and adjusts MPC parameters online.

Images