Unmanned aerial vehicle impact interception method and impact unmanned aerial vehicle based on monocular vision

Abstract

The present application relates to a kind of unmanned aerial vehicle impact interception methods based on monocular vision, the azimuth measurement information and angle measurement information of target unmanned aerial vehicle are calculated by target detection algorithm based on monocular vision;Azimuth measurement information and angle measurement information are used, through orthogonal projection matrix and algebraic transformation, non-linear measurement is converted into pseudo-linear form, then pseudo-linear adaptive Kalman filtering algorithm is carried out to estimate the position of target unmanned aerial vehicle, finally, unmanned aerial vehicle maneuvering strategy is used to control impact unmanned aerial vehicle to realize the impact interception of target unmanned aerial vehicle.The method of the present application is obviously improved in system noise robustness, target state estimation accuracy and pursuit efficiency.The present application also relates to impact unmanned aerial vehicle guided by the above impact interception method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle control, in particular to a method for intercepting unmanned aerial vehicles by collision based on monocular vision and a collision unmanned aerial vehicle. BACKGROUND

[0002] An unmanned aerial vehicle (UAV) is a pilotless aircraft that is controlled by radio remote control equipment and self-provided program control device. The booming UAV industry, mature technology and greatly reduced production cost have made UAVs deeply integrated into people's lives. However, the wide application of UAVs has also brought many safety hazards. Illegal "black flight" occurs frequently, and UAVs intrude into airports and air defense identification zones, seriously threatening aviation safety. After modification, the load capacity of UAVs can be converted into offensive weapons, endangering public safety. Therefore, it has become a social problem to be solved to implement effective UAV countermeasures in low-altitude airspace.

[0003] The current mainstream UAV countermeasures include net gun / net bomb launching, laser attack and counter UAV collision. The net gun / net bomb launching captures the propeller of a UAV by launching a winding net to make the UAV lose power and force-land. The laser attack uses a high-power laser beam to irradiate the photoelectric sensor, battery or fuselage of a UAV to make it paralyzed or crash. The counter UAV collision deploys a counter UAV with autonomous navigation capability, which is not affected by weather conditions and radio signals and can realize over-the-horizon operation.

[0004] Compared with the net gun / net bomb and laser attack, the counter UAV collision technology has unique advantages in adaptability to complex environments and cost-effectiveness. The net gun needs to rely on manual or auxiliary aiming systems, and when multiple targets invade simultaneously, there are problems of slow response speed and untimely ammunition replenishment. Although the laser attack can achieve long-distance precision attack, the equipment cost is high, and the effectiveness is greatly reduced in extreme weather (such as sandstorm and heavy rain). However, the traditional UAV collision method brings many challenges due to complex sensors. To achieve precise attack, multiple high-precision sensors are often needed to cooperate, which not only greatly increases the cost but also reduces the reliability of the equipment. SUMMARY

[0005] The purpose of the present application is to provide an efficient unmanned aerial vehicle collision interception method based on monocular vision and a collision unmanned aerial vehicle guided by the method.

[0006] The present application adopts the following technical solutions:

[0007] The application discloses a monocular vision-based unmanned aerial vehicle collision interception method, which comprises the following steps: calculating the azimuth measurement information and the angle measurement information of a target unmanned aerial vehicle through a monocular vision-based target detection algorithm; converting the nonlinear measurement into a pseudo-linear form through an orthogonal projection matrix and an algebraic transformation by using the azimuth measurement information and the angle measurement information, and then performing a pseudo-linear adaptive Kalman filtering algorithm to estimate the position of the target unmanned aerial vehicle; and finally, adopting an unmanned aerial vehicle maneuvering strategy to control a collision unmanned aerial vehicle to realize the collision interception of the target unmanned aerial vehicle.

[0008] Further, the monocular vision-based target detection algorithm calculates the azimuth measurement information and the angle measurement information of the target unmanned aerial vehicle, specifically as follows: imaging the target unmanned aerial vehicle by using a monocular camera, performing 2D target extraction on the imaging obtained by the monocular camera by using a tiny-YOLOv4 to obtain a bounding box of the target unmanned aerial vehicle, and calculating the direction measurement information and the angle measurement information by using the bounding box information.

[0009] Further, the method for calculating the direction measurement information and the angle measurement information by using the bounding box information is as follows:

[0010] the pixel coordinates of the center point of the bounding box , the internal parameter matrix of the camera and the rotation matrix are used to calculate a three-dimensional direction vector g:

[0011]

[0012] wherein, is the pixel coordinates of the center point of the bounding box.

[0013] the size of the bounding box is converted into the angle measurement information :

[0014]

[0015] wherein, , are the distances from the center of the camera to the left and right midpoints of the bounding box respectively; and respectively represent the focal length and the single-pixel size of the camera; , is the distance between the center of the bounding box and the center of the image, and respectively represent the width and the height of the entire image in pixels, is the size of the bounding box.

[0016] Further, the azimuth measurement information g and the angle measurement information are used to convert the nonlinear measurement into a pseudo-linear form through an orthogonal projection matrix and an algebraic transformation to obtain a state transition matrix and measurement matrix .

[0017] state transition matrix :

[0018]

[0019] measurement matrix :

[0020] .

[0021] Further, the pseudo-linear Kalman filtering algorithm comprises:

[0022] (1) Prediction stage

[0023]

[0024]

[0025] wherein, is the predicted prior state at the current time, is the state estimation at the previous time, is the input matrix, is the input.

[0026] combining the state covariance at the previous time , the state transition matrix and the process noise covariance , the prior state covariance is updated.

[0027] (2) State update

[0028] using the prior state combining the Kalman gain , the residual , the posterior state estimation is obtained, which fuses the prediction and observation information.

[0029]

[0030] the final estimated target UAV position information is obtained from the posterior state estimation obtained in the state update stage.

[0031] Further, the Kalman gain is:

[0032]

[0033]

[0034] wherein, is an observation noise covariance; is a forgetting factor; is a previous time observation noise covariance estimate.

[0035] the residual is:

[0036]

[0037] wherein, is an observation value.

[0038] Further, the unmanned aerial vehicle maneuvering strategy is a reinforcement learning guidance decision method based on a proximal policy optimization; an input of the unmanned aerial vehicle maneuvering strategy is a target state obtained by a pseudo-linear Kalman filter target position estimator and a state of the counter-unmanned aerial vehicle, and an output of the unmanned aerial vehicle maneuvering strategy is an acceleration of an axis direction of the counter-unmanned aerial vehicle.

[0039] A reward function of the unmanned aerial vehicle maneuvering strategy is:

[0040]

[0041] wherein, is a distance reward; is a direction reward; is a collision reward.

[0042] A collision unmanned aerial vehicle is guided by using the unmanned aerial vehicle collision interception method based on monocular vision.

[0043] Specifically, the collision unmanned aerial vehicle comprises a frame and arms arranged on side portions of the frame, and free ends of the arms are respectively provided with motors and propellers connected with rotating shafts of the motors.

[0044] An on-board computer support is arranged on an upper portion of the frame, and a battery fixing support is arranged on a lower portion of the frame; an on-board computer, a flight controller and an electronic speed controller are fixedly arranged on the on-board computer support, and a monocular gimbal camera is arranged at a top end of the on-board computer support; and a battery is fixedly arranged on the battery fixing support.

[0045] The monocular gimbal camera is connected with the on-board computer through a USB, the on-board computer is in communication connection with the flight controller and sends a flight control signal, the flight controller controls the electronic speed controller through PWM, and the electronic speed controller controls the motor; and the battery supplies power to the electronic speed controller, the flight controller and the on-board computer.

[0046] A streamlined shell is arranged outside the on-board computer support, a gimbal shell is arranged outside the monocular gimbal camera, and a battery shell is arranged outside the battery fixing support.

[0047] Specifically, the gimbal shell and the streamlined shell below the gimbal shell jointly form a bullet-shaped shell structure.

[0048] The application has the beneficial effect that the application proposes a bearing-angle target position estimation method based on visual measurement, the bearing-angle information is modeled by an image bounding box to improve the target observability, and the dependence on the horizontal movement of the camera in the traditional method is broken through.

[0049] The application also discloses a bullet-shaped shell used for the unmanned aerial vehicle impact interception method, which is used for high-speed impact of the black flying unmanned aerial vehicle and completes interception. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 It is a system framework schematic diagram of the method of the application.

[0051] Figure 2 It is a schematic diagram of the size conversion of the bounding box into angle measurement information.

[0052] Figure 3 It is a comparison diagram of the tracking effects of the non-adaptive and adaptive pseudo-linear Kalman filters.

[0053] Figure 4 It is a comparison diagram of the performance of the adaptive and non-adaptive pseudo-linear Kalman filter algorithms.

[0054] Figure 5 It is a structure schematic diagram of the impact unmanned aerial vehicle.

[0055] Figure 6 It is a structure schematic diagram of the impact unmanned aerial vehicle with a shell.

[0056] Figure 5 And Figure 6 In the figure, 1 is a rack, 2 is an arm, 3 is an on-board computer support, 4 is a battery fixing support, 5 is an on-board computer, 6 is a flight controller, 7 is an electronic speed controller, 8 is a single-lens gimbal camera, 9 is a battery, 10 is a streamlined shell, 11 is a gimbal shell, and 12 is a battery shell.

[0057] Figure 7 It is a system framework schematic diagram of the impact unmanned aerial vehicle. DETAILED DESCRIPTION

[0058] The technical solutions of the application are described in detail below with specific embodiments and drawings.

[0059] As shown in the figure, a method for estimating the position of a UAV and intercepting a collision based on monocular vision information, calculates the azimuth angle of the target UAV through a monocular vision-based target detection algorithm, then estimates the position of the target UAV using a pseudo-linear Kalman filter algorithm, and controls the collision UAV to achieve the interception of the target UAV using a UAV maneuvering strategy. Figure 1

[0060] I. Vision-based target detection algorithm

[0061] A monocular camera is used to image the target UAV, tiny-YOLOv4 is used to extract 2D targets from the imaging obtained by the monocular camera, the bounding box of the target UAV is obtained, and the composite measurement information is calculated using the bounding box information. The composite measurement information includes direction measurement information g and angle measurement information .

[0062] The center pixel coordinates of the bounding box are used , the three-dimensional direction vector g is calculated through the camera intrinsic matrix and the rotation matrix :

[0063]

[0064] wherein, is the pixel coordinates of the center point of the bounding box.

[0065] The bounding box size is converted into angle measurement information, i.e. the included angle (see Figure 2 ):

[0066]

[0067] wherein,

[0068]

[0069]

[0070] and are the distances from the camera center to the left and right midpoints of the bounding box, respectively, wherein and represent the focal length and the size of a single pixel of the camera, respectively, , is the distance between the center of the bounding box and the center of the image, and represent the width and height of the entire image in pixels, respectively, is the size of the bounding box. ​

[0071] Through the above monocular camera imaging, target detection and calculation using the boundary box pixel coordinates and size, the key measurement quantity for describing the relative relationship between the target and the tracker is obtained.

[0072] Further analysis from the perspective of target motion state, the azimuth vector and the included angle are nonlinear functions of the target state , , is the three-dimensional position of the target, is the three-dimensional position of the tracker, is the physical size of the target.

[0073] II. Pseudo-linear Kalman filtering algorithm

[0074] Using the above azimuth measurement information and the angle measurement information , if the Kalman filtering method is directly used to estimate the position of the target unmanned aerial vehicle, large errors will be introduced, especially in the target close range or high maneuvering scene, which is easy to cause filter divergence. The present application converts the nonlinear measurement into a "pseudo-linear" form through the orthogonal projection matrix and algebraic transformation, and carries out pseudo-linear Kalman filtering.

[0075] Augmented state vector definition: a ten-dimensional state vector

[0076]

[0077] Among them, , are the velocity and acceleration of the target respectively.

[0078] State transition equation: adopt the discrete time double integrator model

[0079]

[0080] Among them, the state transition matrix , is the sampling time, is the process noise and , , , .

[0081] Pseudo-linear measurement conversion: through the orthogonal projection matrix , the nonlinear measurement of the azimuth vector and the included angle , are the measurement noises of the direction vector and the included angle respectively)

[0082]

[0083]

[0084] Transforming to pseudo-linear equations:

[0085]

[0086] .

[0087] Pseudo-linear measurement equation is represented as:

[0088]

[0089] where,

[0090]

[0091]

[0092] .

[0093] The state transition matrix and the measurement matrix are obtained.

[0094] (1) Prediction stage

[0095]

[0096]

[0097] To predict the current time prior state, is the previous time state estimate, is the input matrix, is the input. Combined with the previous time state covariance , state transition matrix and process noise covariance , the prior state covariance is updated, reflecting the prediction uncertainty.

[0098] The residual is calculated by subtracting the observation value from the observation value based on the prior state prediction, obtaining the residual , measuring the deviation between prediction and actual observation:

[0099] .

[0100] The forgetting factor (Balance new and old information weights), combined with the previous time observation noise covariance estimation Update observation noise covariance online with residual information Let the filter adapt to the observation noise changes:

[0101] .

[0102] Calculate the Kalman gain, weighing the state prediction uncertainty and observation noise uncertainty , calculate the Kalman gain , determine the contribution of observation information to state update:

[0103] .

[0104] (2) State update

[0105] Use the prior state combined with the Kalman gain , residual , get the posterior state estimate , fusion prediction and observation information. Among them,

[0106] .

[0107] Update the posterior state covariance , reflecting the uncertainty of the state estimate after fusion observation, identity matrix:

[0108] .

[0109] After the pseudo-linear Kalman filter algorithm, the final estimated target position information , target speed , target acceleration , target size can be obtained from the posterior state estimate obtained in the state update stage.

[0110] Three, the maneuvering strategy of the UAV

[0111] After obtaining the position information of the target, autonomous guidance impact on the target is realized by using reinforcement learning decision. The input of the reinforcement learning guidance decision module is the target state (position , speed ) obtained by the pseudo-linear Kalman filter target position estimator and the countermeasure UAV's own state (position , speed ), and the output is the flight control command of the countermeasure UAV (the countermeasure UAV , , axial acceleration ), and finally forces the counter-UAV to approach and impact the target along the optimal trajectory through continuous decision-making.

[0112] The structure of the reinforcement learning guidance decision module is shown in Fig. 2. Figure 1 The state space is defined as:

[0113]

[0114] where, is the three-dimensional relative position of the counter-UAV and the target, obtained by subtracting the target position from the counter-UAV position . is the three-dimensional relative velocity of the counter-UAV and the target, obtained by subtracting the target velocity from the counter-UAV velocity .

[0115] The output of the guidance decision module is the axial acceleration , , axial acceleration .

[0116] To guide the counter-UAV to quickly and accurately impact the target while avoiding off-target or pre-impact instability, the reward function is designed as:

[0117]

[0118] where, is the distance reward, encouraging the counter-UAV to continuously approach the target, inversely proportional to the relative distance; is the direction reward, encouraging the counter-UAV to fly towards the target, based on the relative bearing angle (the angle between the counter-UAV's heading direction and the line connecting the counter-UAV and the target), where the closer to 1 (the more consistent the direction), the greater the reward, is a positive coefficient; is the impact reward, giving a positive reward if successful impact.

[0119] In terms of reinforcement learning methods, the proximal policy optimization (PPO) algorithm is chosen. It includes a policy network (Actor) and a value network (Critic). The policy network inputs the state and outputs the probability distribution of the action , where, is the policy network parameter, is the mean function, is the standard deviation function. The value network input state , the output state value , where is the value network parameter.

[0120] The algorithm update process is as follows:

[0121] First, use the current policy to interact with the environment and collect data

[0122]

[0123] Discount report is calculated by the following formula:

[0124]

[0125] where, is the discount factor, , is the terminal time step.

[0126] The generalized advantage estimation is calculated by the following formula:

[0127]

[0128] where, is the time difference residual, is the generalized advantage estimation parameter, which balances the bias and variance of the advantage estimation.

[0129] Update the policy network parameters by maximizing the Clip objective function , limit the difference between the new and old policies, the update formula is:

[0130]

[0131] where, is the probability ratio of the new and old policies, is the clipping parameter, is the old policy parameter before updating.

[0132] Update the value network parameters by minimizing the mean square error loss:

[0133] .

[0134] Through the above design, the policy network and value network of the PPO algorithm work together to form a closed loop of "state awareness - action decision - reward feedback", enabling the counter unmanned aerial vehicle to have the autonomous guidance capability of self-adaptive target motion.

[0135] Four, simulation experiment

[0136] To compare the performance of adaptive pseudo-linear Kalman filter and non-adaptive pseudo-linear Kalman filter in dynamic target pursuit scenario, and verify the robustness of adaptive filter to system noise, the accuracy of target state estimation and the improvement of pursuit efficiency. A single pursuit experiment is designed.

[0137] (1) Experimental method

[0138] The target moves at a constant speed v t = [2m / s, 2m / s]T, and the pursuer controls the motion using the proportional navigation guidance law (PNG) with a navigation gain of 1, with a constant speed of 3 m / s, and only adjusts the direction to approach the target. The simulation is terminated when the pursuit is successful; the initial estimation conditions of the two filtering methods (adaptive Kalman filter, non-adaptive Kalman filter) are consistent, including the initial estimated values of target position, velocity, size and noise covariance parameters (as shown in Table 1 and Table 2), and the trajectory estimation error and convergence are used as indicators to compare the performance of the two methods in the pursuit of escaping target scenario.

[0139] (2) Parameter setting

[0140] Table 1 Basic simulation parameters

[0141]

[0142] Table 2 Initial parameters of filter

[0143] .

[0144] (3) Results and discussion

[0145] Figure 3 The results show that in the case of pseudo-linear Kalman filter (PLKF) ( Figure 3 left), the tracker shows obvious deviation and oscillation behavior, which is a direct result of the filter's inability to adapt to the statistical fluctuations of monocular noise. In contrast, the trajectory of AKF guidance ( Figure 3 right) closely follows the true value of the target position in a smooth and stable manner, highlighting the stabilizing effect of adaptive covariance adjustment.

[0146] AKF is superior to non-adaptive PLKF in terms of position error, target size, estimation error, velocity estimation error and cumulative error (see Figure 4 ). During the entire pursuit process, the position error under AKF remains bounded and low, while the PLKF error accumulates rapidly, eventually leading to divergence. The velocity and target size estimates under AKF also converge stably to the true value, while the PLKF results fluctuate greatly. The cumulative error curve further confirms that the adaptive strategy effectively suppresses long-term error propagation.

[0147] In the three key dimensions of capture efficiency, estimation accuracy and state stability, the adaptive Kalman filter (AKF) is significantly better than the pseudo-linear Kalman filter (PLKF). First, AKF reduces the average estimation error by 43.7% and shortens the capture time by 19.8%, indicating that its adaptive mechanism effectively improves the overall performance of the filtering algorithm. Second, AKF shows stronger robustness by dynamically adjusting the noise covariance matrix, significantly enhancing the ability to adapt to observation noise fluctuations and target motion uncertainty, making it more suitable for scenarios where noise parameters are unknown or time-varying in real environments. Finally, AKF has good engineering practicability - its algorithm complexity is comparable to that of non-adaptive PLKF, only adding residual calculation and parameter adjustment steps in the basic process, without the need for additional hardware resources, making it easy to implement and deploy in practical systems.

[0148] Five, the design of the impact drone

[0149] An impact drone comprises a 5MM carbon fiber frame 1 and a machine arm 2 arranged on the side of the frame 1, and the number of machine arms 2 can be four, six or eight. The free end of each machine arm 2 is provided with a motor and a propeller connected with the rotating shaft of the motor.

[0150] An on-board computer support 3 is arranged on the upper part of the frame 1, and a battery fixing support 4 is arranged on the lower part of the frame 1. An on-board computer 5, a flight controller 6 and an electronic speed controller 7 are fixedly arranged on the on-board computer support 3, and a monocular gimbal camera 8 is arranged at the top end of the on-board computer support 3. A battery 9 is fixedly arranged on the battery fixing support 4.

[0151] As shown in Figures 4-6 , the monocular gimbal camera 8 is connected with the on-board computer 5 through USB, the on-board computer 5 is in communication connection with the flight controller 6 and sends a flight control signal, the flight controller 6 controls the electronic speed controller 7 through PWM, and the electronic speed controller 7 controls the motor, thereby realizing the flight control of the impact drone. The battery 9 supplies power to the electronic speed controller 7, the flight controller 6 and the on-board computer 5 respectively.

[0152] A streamlined shell 10 is sleeved outside the on-board computer support 3, and a gimbal shell 11 is sleeved outside the monocular gimbal camera 8, and the gimbal shell 11 and the streamlined shell 10 below it jointly form a bullet-shaped shell structure, which on the one hand reduces the resistance in the flight process of the drone, and on the other hand the bullet-shaped shell structure is more conducive to realizing the impact and damage to the target drone and improving the interception success rate. A battery shell 12 is sleeved outside the battery fixing support 4.

[0153] In use, the monocular holder camera 8 is used to locate the target, capture the visual features of the black flying unmanned aerial vehicle in the search area, identify and lock the target. The airborne computer 5 combines the image information collected by the camera and uses the unmanned aerial vehicle impact interception method based on monocular vision proposed by the application to determine the relative position and motion trajectory of the black flying unmanned aerial vehicle, continuously track the target and correct the flight parameters. The airborne computer 5 plans the optimal interception path according to the target motion state, the position of the airborne computer 5 and the environment. The electronic speed regulator 7 accurately adjusts the motor speed according to the path planning to drive the unmanned aerial vehicle to fly along the planned path at high speed and high mobility. When approaching the target, the flight controller 6 adjusts the flight attitude according to the real-time position deviation, uses the kinetic energy interception mechanism to implement autonomous high-speed impact on the black flying unmanned aerial vehicle by physically intercepting the shell of the unmanned aerial vehicle, and completes the interception.

Claims

1. A method for intercepting drone collisions based on monocular vision, characterized in that, The azimuth and angle measurement information of the target UAV are calculated by a target detection algorithm based on monocular vision. Using the azimuth and angle measurement information, the nonlinear measurement is converted into a pseudo-linear form through orthogonal projection matrix and algebraic transformation. Then, the position of the target UAV is estimated by a pseudo-linear adaptive Kalman filter algorithm. Finally, the UAV maneuvering strategy is used to control the collision UAV to achieve collision interception of the target UAV. The target detection algorithm based on monocular vision calculates the orientation and angle measurement information of the target UAV as follows: the target UAV is imaged using a monocular camera, and 2D target extraction is performed on the image obtained by the monocular camera using tiny-YOLOv4 to obtain the bounding box of the target UAV. The orientation and angle measurement information are then calculated using the bounding box information. The method for calculating direction and angle measurement information using bounding box information is as follows: Using the pixel coordinates of the bounding box center point Through camera intrinsic parameters With rotation matrix Calculate the three-dimensional direction vector g: in, These are the two-dimensional pixel coordinates of the center point of the bounding box; Convert bounding box dimensions into angle measurements. : in, , These are the distances from the center of the camera to the midpoints of the left and right sides of the bounding box, respectively. and These represent the camera's focal length and the size of a single pixel, respectively. , It is the distance between the center of the bounding box and the center of the image. and The width and height of the entire image are represented in pixels. This refers to the bounding box size.

2. The UAV collision interception method based on monocular vision according to claim 1, characterized in that, Using azimuth measurement information g and angle measurement information By using orthogonal projection matrices and algebraic transformations, nonlinear measurements are converted into pseudolinear forms, resulting in the state transition matrix. and measurement matrix ; State transition matrix : Measurement Matrix : .

3. The UAV collision interception method based on monocular vision according to claim 2, characterized in that, The pseudo-linear adaptive Kalman filter algorithm includes: (1) Prediction stage in, To predict the prior state at the current moment, For the state estimation of the previous moment, For the input matrix, For input; Combined with the state covariance of the previous time step State transition matrix Covariance of process noise Update the prior state covariance ; (2) Status update Using prior states Combined Kalman gain residual , thus obtaining the posterior state estimate It integrates prediction and observation information; Final estimated target drone location information Posterior state estimate obtained from the state update phase Obtained from [the source].

4. The UAV collision interception method based on monocular vision according to claim 3, characterized in that, Kalman gain for: in, To observe the noise covariance; Forgetting factor; This is an estimate of the covariance of the observation noise from the previous time step. The residual for: in, These are the observed values.

5. The UAV collision interception method based on monocular vision according to claim 4, characterized in that, The UAV maneuvering strategy is a reinforcement learning-based guidance and decision-making method based on near-end strategy optimization; The reward function is: in, As a distance reward; As a directional reward; As a reward for impact.

6. An impact drone, characterized in that, It uses the monocular vision-based UAV collision interception method described in any one of claims 1 to 5 for guidance.

7. The impact drone according to claim 6, characterized in that, It includes a frame (1) and a boom (2) located on the side of the frame (1). The free ends of the boom are equipped with motors and propellers connected to the rotating shafts of the motors. An airborne computer bracket (3) is provided on the upper part of the frame (1), and a battery mounting bracket (4) is provided on the lower part of the frame (1); an airborne computer (5), a flight controller (6) and an electronic speed controller (7) are fixedly installed on the airborne computer bracket (3), and a monocular gimbal camera (8) is installed at the top of the airborne computer bracket (3); a battery (9) is fixedly installed on the battery mounting bracket (4). The monocular gimbal camera (8) is connected to the onboard computer (5) via USB. The onboard computer (5) communicates with the flight controller (6) and sends flight control signals. The flight controller (6) controls the electronic speed controller (7) via PWM. The electronic speed controller (7) then controls the motor. The battery (9) supplies power to the electronic speed controller (7), the flight controller (6), and the onboard computer (5) respectively. The airborne computer bracket (3) is covered with a streamlined shell (10), the monocular gimbal camera (8) is covered with a gimbal shell (11), and the battery mounting bracket (4) is covered with a battery shell (12).

8. The impact drone according to claim 7, characterized in that, The gimbal housing (11) and its lower streamlined housing (10) together form a bullet-shaped housing structure.

Images