Unmanned aerial vehicle based on visual guidance and control method thereof

Abstract

The invention relates to an unmanned aerial vehicle based on visual guidance and a control method thereof. The unmanned aerial vehicle comprises an unmanned aerial vehicle body, a mechanical arm assembly connected with the unmanned aerial vehicle body, a mechanical arm fixing platform and a control mainboard. The mechanical arm assembly comprises a rotary positioning module, a horizontal guide module, an axial telescopic module, a rotary cutting module and a binocular camera module. The mechanical arm assembly is installed below the unmanned aerial vehicle body and connected through the mechanical arm fixing platform. The rotation positioning module comprises a servo motor, a planetary reducer and a coupler. The control mainboard comprises a control module, a positioning module, a path planning module, a communication module and a deep learning module. Compared with the prior art, the corn tassel removing device has the advantages that omissive corn tassels are accurately identified and positioned by adopting a visual technology, meanwhile, the mechanical arm device is simple in structure and large in working space, the error-tolerant rate and the stability are improved, the tassels can be accurately removed, the seed corn tassel removing efficiency is improved, and then the seed quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles and mechanical arms, and in particular to a vision-guided unmanned aerial vehicle and a control method thereof. Background Art

[0002] Corn production has surpassed rice and wheat to become China's largest grain crop, which is of great significance in ensuring food security. Hybrid breeding is an effective way to increase corn yield and improve grain quality. Corn is one of the earliest crops to use hybrid vigor, and purity is an important indicator for evaluating the quality of corn hybrid seeds. Maize female parent emasculation is the most critical and difficult link to ensure seed purity in the hybrid corn seed production process.

[0003] At present, there are two main methods of detasseling for seed corn in my country: manual detasseling and mechanical detasseling. Manual detasseling is the traditional method of detasseling and is also the main method of detasseling in my country. This method requires a lot of labor, has low operating efficiency and it is difficult to ensure the quality of detasseling. Mechanized detasseling has high operating efficiency and is less affected by natural conditions and human factors. However, due to the diversity of corn plant morphology and the uneven distribution of male ears, mechanical detasseling may result in single missed or wrong detasseling, requiring multiple missed detasseling.

[0004] In order to ensure the quality of corn seed production, manual surveys are required to find the missed tassels and remove them in time. This method will consume a lot of manpower and material resources, and may result in missing the window period for emasculation, causing the missed female tassels to shed pollen and contaminate the seed production field, affecting the quality of corn seeds over a wide range. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art that the mechanized removal device cannot achieve precise positioning and efficient removal when dealing with a small number of missed male ears, and to provide a vision-guided drone and its control method to achieve accurate, rapid and automated removal of missed male ears scattered in the farmland.

[0006] The present invention provides a vision-guided drone, comprising: a drone body, a mechanical arm assembly connected to the drone body, a mechanical arm fixing platform and a control mainboard;

[0007] The mechanical arm assembly includes: a rotary positioning module, a horizontal guidance module, an axial telescopic module, a rotary cutting module and a binocular camera module; the mechanical arm assembly is installed below the UAV body and connected through a mechanical arm fixing platform;

[0008] The rotary positioning module includes: a servo motor, a planetary reducer and a coupling, wherein the coupling is installed on the driving shaft at the end of the planetary reducer;

[0009] The horizontal guide module includes: a stepping motor, a spiral transmission slide rail, a horizontal connecting plate and a slider, and the middle part of the spiral transmission slide rail is connected to the rotary positioning module through the horizontal connecting plate;

[0010] The axial telescopic module is vertically fixed on the slider at the bottom of the horizontal guide module, and the axial telescopic module includes: a push rod motor, a first connecting member, a motor mounting seat, a second connecting member and a cross-shear telescopic unit; the rotary cutting module is fixed to the end of the axial telescopic module through the second connecting member, and the rotary cutting module includes: a cutting motor, a shield and a blade;

[0011] The binocular camera module comprises: a camera frame, a binocular camera, a camera base and a camera swivel, wherein the binocular camera is mounted on the camera swivel; one end of the stepping motor is connected to the binocular camera module through the camera frame;

[0012] The control mainboard includes: a control module, a positioning module, a path planning module, a communication module and a deep learning module, and the control mainboard is fixed on the robot arm fixed platform.

[0013] Furthermore, the robot arm fixing platform includes a mounting crossbeam and a platform fixing part; a DC power supply is used to power the control main board, binocular camera, servo motor, stepper motor, push rod motor and cutting motor.

[0014] Furthermore, the drone body includes a drone body, propellers, RTK and a brushless motor. The eight propellers are installed at the end of the drone's arm and driven by four brushless motors to provide power for the drone.

[0015] Furthermore, the cross-scissor type telescopic unit includes: a first-stage connecting rod, a second-stage connecting rod, a third-stage connecting rod, a fourth-stage connecting rod, a buffer pin, a rotating shaft connecting piece, a bracket and a rotating shaft fixing piece, the first-stage connecting rod, the second-stage connecting rod, the third-stage connecting rod and the fourth-stage connecting rod are all overlapping double layers and are reinforced and connected by the bracket, and are hinged by the buffer pin; the push rod motor is fixed to the motor mounting seat and vertically fixed to the first connecting piece, and the end of the push rod motor drive shaft is connected to the rotating shaft fixing piece; the end of the first-stage connecting rod is hinged to the rotating shaft connecting pieces on both sides of the motor mounting seat, and the fourth-stage connecting rod is hinged to the second connecting piece through the rotating shaft connecting piece.

[0016] The present invention also provides a control method of a UAV based on visual guidance, comprising the following steps:

[0017] S1: The robotic arm assembly returns to its original position, and the drone flies to the target and hovers;

[0018] S2: A real-time image of the target object below is taken through a binocular camera (82); after the deep learning module (35) recognizes the target object, pixel coordinate information of the target in the image is obtained, and then the distance of the target from the set coordinate origin is obtained through a coordinate conversion module, and the drone adjusts the hovering height according to the information; the coordinate conversion formula is:

[0019]

[0020] Among them, Z c is the coordinate of the target in the Z-axis direction in the camera coordinate system, (u, v) is the coordinate of the target in the pixel coordinate system, (u0, v0) is the midpoint of the pixel, f is the focal length of the binocular camera, R is the rotation matrix, T is the offset matrix, (X w ,Y w ,Z w ) are the coordinates of the target in the world coordinate system.

[0021] S3: After the drone obtains the three-dimensional information of the target, it calculates the error between the blade end and the target, drives the mechanical arm to operate through the driving module, and starts the rotary cutting module equipped at the end before the blade position reaches the target position, and removes the tassel with the blade;

[0022] S4: After the emasculation task is completed, the binocular camera takes a photo to record the removal status of the current task point, which is uploaded for confirmation by the staff; the robotic arm assembly returns to its original position;

[0023] S5: The drone climbs, switches to flight mode and flies to the next tassel missing point, and repeats S1 to S5.

[0024] Furthermore, in S1, the robotic arm assembly starts self-checking before the drone takes off to confirm that each module of the robotic arm assembly is in an initial state; the drone central control loads the latitude and longitude coordinates of the missing tassel and plans the path of each working point; then, the drone flies along the path to the sky above the target tassel.

[0025] Furthermore, in S2, the target image acquired by the binocular camera is segmented into target tassels through the deep learning module, and the pixel coordinates of the target in the image are converted into the world coordinate system coordinates of the target relative to the coordinate origin of the robot arm through the coordinate conversion module.

[0026] Furthermore, in S3, the control mainboard calculates the motion parameters of each module of the robotic arm assembly according to the three-dimensional coordinates of the target point and is driven by the motor; the start timing of the cutting tool loaded at the end of the robotic arm is determined by the axial telescopic module, and the cutting motor is started when the telescopic length reaches two-thirds of the distance of the instruction to cut the tassel. At the same time, the binocular camera recognizes the target in real time, and the control mainboard calculates the distance between the blade and the target. If the difference is less than the preset error threshold, the current blade position is determined to be the cutting point; in S4, the control mainboard returns the robotic arm assembly to its original position after completing the cutting instruction, and the binocular camera shoots the lower area during the movement and is detected by the deep learning module. If the missing tassel is detected again and the distance between the target and the blade is greater than the preset error threshold, it is determined that the current task point excision task is not completed; return to S2 and S3, and re-excise the target tassel.

[0027] Furthermore, the coordinate information of the target to be detasseled is loaded before the task starts, and the positioning module and the path planning module divide the task groups and plan the path based on the take-off point of the UAV, the terrain of the plot, and the target distribution status; after arriving above the target point, the binocular camera module is used to collect the image data of the male tassel, and the three-dimensional point cloud information of the male tassel is calculated by the deep learning module, and then the obtained point cloud information is filtered to extract the precise position information of the male tassel; the control module removes the male tassel through the mechanical arm assembly according to the visual recognition result, and continues to execute the next detasseling task according to the planned path.

[0028] Furthermore, the deep learning module constructs a tassel data set and trains a convolutional neural network (CNN) model to achieve accurate segmentation and fitting judgment of corn tassel targets. The training process is as follows: an initial network structure model is established, and the weights of each layer of the model are randomly initialized. The tassel images in the training set are preprocessed and input into the initial model. The input images are feature extracted through convolution operations. The parameters are updated through back propagation and iterative training. The parameters in the neural network are continuously optimized and updated to obtain the optimal model. The last training weights are continuously loaded for transfer learning during use, and the model is continuously optimized to maintain a high generalization ability.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] (1) The visual guidance technology adopted in the present invention can realize the precise removal of male tassels by real-time acquisition and processing of image data by the UAV-mounted robotic arm; the binocular camera can obtain information such as the position and shape of the target, and improve the accuracy of the target position through the filtering algorithm; and the deep learning technology can iteratively optimize the recognition of targets in different operating environments, thereby improving the success rate and rapidity of detasseling.

[0031] (2) The robotic arm device designed in the present invention has a relatively simple structure. The visual sensor is arranged at the end of the horizontal guide rail and follows the rotation module to adjust the viewing angle, thereby achieving visual coverage within a larger range of space below the UAV. At the same time, the coordination of the three degrees of freedom is simpler than the control operation of other UAV robotic arms. The axial telescopic module greatly expands the working space below the UAV, improves the fault tolerance and stability, and is more suitable for the UAV to operate on the target below.

[0032] (3) The visual control method adopted by the present invention can realize the accurate identification and positioning of the missing corn tassels, and can automatically and accurately remove the tassels by combining with the visually guided robotic arm device, thereby improving the efficiency of emasculation of seed corn and thus improving the seed quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the overall structure of a vision-guided drone of the present invention;

[0034] Figure 2 This is a schematic diagram of the structure of the UAV precision emasculation mechanical arm of the present invention;

[0035] Figure 3 It is a structural schematic diagram of a binocular camera module of the present invention;

[0036] Figure 4 It is a structural schematic diagram of the axial telescopic module of the present invention in a contracted state;

[0037] Figure 5 It is a structural schematic diagram of the axial telescopic module of the present invention in an extended state;

[0038] Figure 6 This is a module architecture diagram of the robot arm control mainboard of the present invention;

[0039] Figure 7 is a flow chart of the control method of the present invention;

[0040] Figure 8 A flowchart of the high-altitude operation of precise detasseling by a mechanical arm provided by the present invention;

[0041] Figure numerals: 1. UAV body; 11. UAV fuselage; 12. Propeller blades; 13. RTK; 14. Brushless motor; 2. Robot arm fixing platform; 21. Mounting beam; 22. Platform fixing parts; 3. Control main board; 31. Control module; 32. Positioning module; 33. Path planning module; 34. Communication module; 35. Deep learning module; 4. Rotary positioning module; 41. Servo motor; 42. Planetary reducer; 43. Coupling; 5. Horizontal guide module; 51. Stepper motor; 52. Spiral transmission slide rail; 53. Horizontal connecting plate; 54. Slider; 6. Axial telescopic module; 61. Push rod motor; 62. First connecting piece; 63. Motor mounting seat; 64. Second connecting piece; 65. Cross-shear telescopic unit; 651. First-stage connecting rod; 652. Second-stage connecting rod; 653. Third-stage connecting rod; 654. Fourth-stage connecting rod; 655. Buffer pin; 656. Rotating shaft connecting piece; 657. Bracket; 658. Rotating shaft fixing piece; 7. Rotating cutting module; 71. Cutting motor; 72. Protective cover; 73. Blade; 8. Binocular camera module; 81. Camera stand; 82. Binocular camera; 83. Camera base; 84. Camera swivel. DETAILED DESCRIPTION

[0042] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. Component models, material names, connection structures, control methods, algorithms and other features not clearly described in this technical solution are all considered to be common technical features disclosed in the prior art.

[0043] Example 1

[0044] This embodiment provides a vision-guided drone, such as Figure 1 , 2 , 3, 4, 5, 6, comprising: a drone body 1, a mechanical arm assembly connected to the drone body 1, a mechanical arm fixing platform 2 and a control mainboard 3;

[0045] The mechanical arm assembly includes: a rotary positioning module 4, a horizontal guiding module 5, an axial telescopic module 6, a rotary cutting module 7 and a binocular camera module 8; the mechanical arm assembly is installed below the drone body 1 and connected through the mechanical arm fixing platform 2;

[0046] The rotary positioning module 4 includes: a servo motor 41, a planetary reducer 42 and a coupling 43, wherein the coupling 43 is mounted on a driving shaft at the end of the planetary reducer 42;

[0047] The horizontal guide module 5 includes: a stepping motor 51, a spiral transmission slide rail 52, a horizontal connecting plate 53 and a slider 54. The middle part of the spiral transmission slide rail 52 is connected to the rotary positioning module 4 through the horizontal connecting plate 53.

[0048] The axial telescopic module 6 is vertically fixed on the slider 54 at the bottom of the horizontal guide module 5, and the axial telescopic module 6 includes: a push rod motor 61, a first connecting member 62, a motor mounting seat 63, a second connecting member 64 and a cross-shear telescopic unit 65; the rotary cutting module 7 is fixed to the end of the axial telescopic module 6 through the second connecting member 64, and the rotary cutting module 7 includes: a cutting motor 71, a shield 72 and a blade 73;

[0049] The binocular camera module 8 includes: a camera frame 81, a binocular camera 82, a camera base 83 and a camera swivel 84, wherein the binocular camera 82 is mounted on the camera swivel 84; one end of the stepper motor 51 is connected to the binocular camera module 8 through the camera frame 81; the model of the binocular camera 82 is ZED 2, and the rotation adjustment member can be adjusted in the range of 0° to 90° to adapt to the recognition of male spikes at different angles. The advantages of the ZED2 camera are high-precision depth perception and built-in IMU sensor, providing a 110° wide viewing angle for rapid target search.

[0050] The control mainboard 3 includes: a control module 31, a positioning module 32, a path planning module 33, a communication module 34 and a deep learning module 35, and the control mainboard 3 is fixed on the mechanical arm fixed platform 2. The control module 31 adopts the NVIDIA edge computing processor of model JetsonXavier NX, the positioning module 32 adopts GPS to obtain the current position of the drone, the communication module 34 is used for remote control, the path planning module 33 and the positioning module 32 are combined to plan the emasculation task path, and the deep learning module 35 extracts the collected image information and calculates the target position information.

[0051] In a specific embodiment, the robot arm fixing platform 2 includes a mounting beam 21 and a platform fixing member 22; a DC power supply is used to power the control main board 3, the binocular camera 82, the servo motor 41, the stepper motor 51, the push rod motor 61 and the cutting motor 71.

[0052] In a specific embodiment, the drone body 1 includes a drone body 11, blades 12, RTK 13 and a brushless motor 14. Eight blades 12 are installed at the end of the drone's arm and driven by four brushless motors 14 to provide power for the drone.

[0053] In a specific embodiment, the cross-scissor type telescopic unit 65 includes: a first-stage connecting rod 651, a second-stage connecting rod 652, a third-stage connecting rod 653, a fourth-stage connecting rod 654, a buffer pin 655, a rotating shaft connecting piece 656, a bracket 657 and a rotating shaft fixing piece 658. The first-stage connecting rod 651, the second-stage connecting rod 652, the third-stage connecting rod 653 and the fourth-stage connecting rod 654 are all overlapping double layers and are reinforced and connected by the bracket 657, and are hinged by the buffer pin 655; the push rod motor 61 is fixed to the motor mounting seat 63 and vertically fixed to the first connecting piece 62, and the end of the driving shaft of the push rod motor 61 is connected to the rotating shaft fixing piece 658; the end of the first-stage connecting rod 651 is hinged to the rotating shaft connecting pieces 656 on both sides of the motor mounting seat 63, and the fourth-stage connecting rod 654 is hinged to the second connecting piece 64 through the rotating shaft connecting piece 656.

[0054] This embodiment also provides a control method of a drone based on visual guidance, such as Figure 7 As shown, the following steps are included:

[0055] S1: The robotic arm assembly returns to its original position, and the drone flies to the target and hovers;

[0056] S2: Take a real-time picture of the target below through the binocular camera 82; after the deep learning module 35 recognizes the target, obtain the pixel coordinate information of the target in the image, and then obtain the distance of the target from the set coordinate origin through the coordinate conversion module, and the drone adjusts the hovering height according to the information; the coordinate conversion formula is:

[0057]

[0058] Among them, Z c is the coordinate of the target in the Z-axis direction in the camera coordinate system, (u, v) is the coordinate of the target in the pixel coordinate system, (u0, v0) is the midpoint of the pixel, f is the focal length of the binocular camera, R is the rotation matrix, T is the offset matrix, (X w ,Y w ,Z w ) are the coordinates of the target in the world coordinate system.

[0059] S3: After the drone obtains the three-dimensional information of the target, it calculates the error between the end of the blade 73 and the target, drives the mechanical arm to operate through the driving module, and starts the rotary cutting module 7 equipped at the end before the end blade 73 reaches the target position, and removes the tassel through the blade 73;

[0060] S4: After the emasculation task is completed, the binocular camera takes a photo to record the removal status of the current task point, which is uploaded for confirmation by the staff; the robotic arm assembly returns to its original position;

[0061] S5: The drone climbs, switches to flight mode and flies to the next tassel missing point, and repeats S1 to S5.

[0062] In a specific implementation, in S1, the robotic arm assembly starts self-checking before the drone takes off to confirm that each module of the robotic arm assembly is in an initial state; the drone central control loads the latitude and longitude coordinates of the missing tassel and plans the path of each working point; then, the drone flies along the path to the sky above the target tassel.

[0063] In a specific implementation, in S2, the target image acquired by the binocular camera 82 is segmented into target tassels through the deep learning module 35, and the pixel coordinates of the target in the image are converted into the world coordinate system coordinates of the target relative to the coordinate origin of the robot arm through the coordinate conversion module.

[0064] In a specific implementation, in S3, the control mainboard 3 calculates the motion parameters of each module of the mechanical arm assembly according to the three-dimensional coordinates of the target point and is driven by a motor; the start timing of the cutting tool loaded at the end of the mechanical arm is judged by the axial telescopic module 6, and the cutting motor is started when the telescopic length reaches two-thirds of the distance of the instruction to cut the tassel. At the same time, the binocular camera 82 recognizes the target in real time, and the control mainboard 3 calculates the distance between the blade 73 and the target. If the difference is less than the preset error threshold, the current position of the blade 73 is determined to be the cutting point; in S4, the control mainboard 3 performs a homing operation on the mechanical arm assembly after completing the cutting instruction, and the binocular camera 82 shoots the lower area during the movement and is detected by the deep learning module 35. If the missing tassel is detected again and the distance between the target and the blade 73 is greater than the preset error threshold, it is determined that the current task point excision task is not completed; return to S2 and S3, and re-excise the target tassel.

[0065] In a specific implementation, the coordinate information of the target to be detasseled is loaded before the task starts, and the positioning module 32 and the path planning module 33 divide the task groups and plan the path based on the take-off point of the UAV, the terrain of the plot, and the target distribution conditions; after arriving above the target point, the binocular camera module 8 is used to collect the image data of the male tassel, and the three-dimensional point cloud information of the male tassel is calculated by the deep learning module 34, and then the obtained point cloud information is filtered to extract the precise position information of the male tassel; the control module 31 removes the male tassel through the robotic arm assembly according to the visual recognition result, and continues to execute the next detasseling task according to the planned path.

[0066] In a specific embodiment, the deep learning module 34 constructs a tassel data set and trains a convolutional neural network (CNN) model to achieve accurate segmentation and fitting judgment of corn tassel targets. The training process is as follows: an initial network structure model is established, and the weights of each layer of the model are randomly initialized. The tassel images in the training set are preprocessed and input into the initial model. The input images are feature extracted through convolution operations. The parameters are updated through back propagation and iterative training. The parameters in the neural network are continuously optimized and updated to obtain the optimal model. The last training weights are continuously loaded for transfer learning during use, and the model is continuously optimized to maintain a high generalization ability.

[0067] like Figure 8 As shown, the mechanical arm performs precise emasculation at high altitude, positioning: the drone flies to the sky above the corn plant with tassels according to the positioning information, and hovers within a certain height from the tassels. At this time, the hovering height and positioning information of the drone can ensure that the located tassels are located in the working space of the emasculation device. Identification: the binocular camera 82 takes images in the field of view in real time, and the corn tassels are identified through the target recognition algorithm. There are two situations: first, the captured picture does not identify the target after the recognition algorithm, and the three-dimensional information cannot be output. In this case, the dynamic recognition stage is entered, that is, the servo motor is driven to rotate, and the binocular camera 82 continuously takes pictures along the circular space for identification (the rotation angle does not exceed 180 °), and enters the second situation within the range of one rotation until the tassels are identified. Second, the binocular camera 82 recognizes the tassel when it starts working, and outputs the three-dimensional coordinate information of the tassel. At this time, the posture information of the tassel and the adjacent leaves is divided into two situations: 1. No obstruction (the relative area of ​​the tassel and the leaf can be considered in the analysis of this situation) 2. Occlusion (the leaf partially obstructs the tassel) The dynamic recognition stage is performed by rotation until the obstructed part reaches an acceptable range. Feed (consider reducing leaf damage): After the recognition stage, it is assumed that the tassel image recognized by the binocular camera 82 is unobstructed or less obstructed. At this time, the relative position of the leaf and the tassel is divided into two situations: 1. The leaf grows along the diameter direction of the workspace, the tool is retracted to the specified height, and then the horizontal guide tool is first moved away from a certain distance, and the rotation module rotates a certain angle so that the tool passes through the gap of the corn leaf along the linear speed direction to complete the cutting operation. 2. The leaf is at a certain angle to the diameter direction, the tool is retracted to the specified height, and then the rotation module rotates a certain angle first, and the horizontal guide tool feeds and cuts along the radial direction. Upload: After completing the emasculation task, the binocular camera 82 takes a photo to record the removal status of the current task point, which is uploaded for confirmation by the staff. The tool retracts to the initial position, the guide module moves to the center point of the robotic arm, and the robotic arm returns to its position to enter the next workflow.

[0068] Components not described in detail in this embodiment are all existing components that can be purchased through public channels.

[0069] The above description of the embodiments is to facilitate the understanding and use of the invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A vision-guided drone, characterized in that: include: An unmanned aerial vehicle body (1), a mechanical arm assembly connected to the unmanned aerial vehicle body (1), a mechanical arm fixing platform (2), and a control mainboard (3); The mechanical arm assembly comprises: a rotary positioning module (4), a horizontal guiding module (5), an axial telescopic module (6), a rotary cutting module (7) and a binocular camera module (8); the mechanical arm assembly is installed below the drone body (1) and is connected via a mechanical arm fixing platform (2); The rotary positioning module (4) comprises: a servo motor (41), a planetary reducer (42) and a coupling (43), wherein the coupling (43) is mounted on a drive shaft at the end of the planetary reducer (42); The horizontal guide module (5) comprises: a stepping motor (51), a spiral transmission slide rail (52), a horizontal connecting plate (53) and a sliding block (54); the middle portion of the spiral transmission slide rail (52) is connected to the rotary positioning module (4) via the horizontal connecting plate (53); The axial telescopic module (6) is vertically fixed on a slider (54) at the bottom of the horizontal guide module (5), and the axial telescopic module (6) comprises: a push rod motor (61), a first connecting member (62), a motor mounting seat (63), a second connecting member (64) and a shearing telescopic unit (65); the rotary cutting module (7) is fixed to the end of the axial telescopic module (6) through the second connecting member (64), and the rotary cutting module (7) comprises: a cutting motor (71), a shield (72) and a blade (73); The binocular camera module (8) comprises: a camera frame (81), a binocular camera (82), a camera base (83) and a camera rotating seat (84), wherein the binocular camera (82) is mounted on the camera rotating seat (84); one end of the stepping motor (51) is connected to the binocular camera module (8) via the camera frame (81); The control main board (3) comprises: a control module (31), a positioning module (32), a path planning module (33), a communication module (34) and a deep learning module (35); the control main board (3) is fixed on a mechanical arm fixed platform (2).

2. A vision-guided drone according to claim 1, characterized in that: The mechanical arm fixing platform (2) comprises a mounting crossbeam (21) and a platform fixing member (22).

3. The vision-guided drone according to claim 1, characterized in that: The drone body (1) comprises a drone body (11), propeller blades (12), an RTK (13) and a brushless motor (14).

4. The vision-guided drone according to claim 1, characterized in that: The cross-shear telescopic unit (65) comprises: a first-stage connecting rod (651), a second-stage connecting rod (652), a third-stage connecting rod (653), a fourth-stage connecting rod (654), a buffer pin (655), a rotating shaft connecting member (656), a bracket (657) and a rotating shaft fixing member (658); the first-stage connecting rod (651), the second-stage connecting rod (652), the third-stage connecting rod (653) and the fourth-stage connecting rod (654) are all overlapping double layers and are connected by the bracket (657). ) are reinforced and connected, and are hinged by a buffer pin shaft (655); the push rod motor (61) is fixed to the motor mounting seat (63) and vertically fixed to the first connecting member (62), and the end of the driving shaft of the push rod motor (61) is connected to the rotating shaft fixing member (658); the end of the first-stage connecting rod (651) is hinged to the rotating shaft connecting members (656) on both sides of the motor mounting seat (63), and the fourth-stage connecting rod (654) is hinged to the second connecting member (64) through the rotating shaft connecting member (656).

5. A control method for a UAV based on visual guidance as claimed in any one of claims 1 to 4, characterized in that: The following steps are involved: S1: The robotic arm assembly returns to its original position, and the drone flies to the target and hovers; S2: taking a real-time picture of the target object below through a binocular camera (82); after the deep learning module (35) recognizes the target object, obtaining pixel coordinate information of the target in the image, and then obtaining the distance of the target from the set coordinate origin through a coordinate conversion module, and the drone adjusts the hovering height according to the information; S3: After the drone obtains the three-dimensional information of the target, it calculates the error between the end of the blade (73) and the target, drives the mechanical arm to operate through the driving module, and starts the rotary cutting module (7) equipped at the end before the end blade (73) reaches the target position, and removes the tassel through the blade (73); S4: After the emasculation task is completed, the binocular camera takes a photo to record the removal status of the current task point, which is uploaded for confirmation by the staff; the robotic arm assembly returns to its original position; S5: The drone climbs, switches to flight mode and flies to the next tassel missing point, and repeats S1 to S5.

6. The control method of a UAV based on visual guidance according to claim 5, characterized in that: In S1, the robotic arm assembly starts self-checking before the drone takes off to confirm that each module of the robotic arm assembly is in the initial state; The drone's central control loads the latitude and longitude coordinates of the missing tassel and plans the path for each work point; then, the drone flies along the path to the target tassel.

7. The control method of a UAV based on vision guidance according to claim 5, characterized in that: In S2, the target image acquired by the binocular camera (82) is segmented into target tassels through the deep learning module (35), and the pixel coordinates of the target in the image are converted into the world coordinate system coordinates of the target relative to the coordinate origin of the robot arm through the coordinate conversion module.

8. The control method of a UAV based on vision guidance according to claim 5, characterized in that: In S3, the control main board (3) calculates the motion parameters of each module of the mechanical arm assembly according to the three-dimensional coordinates of the target point and drives the module by the motor; the start timing of the cutting tool loaded at the end of the mechanical arm is determined by the axial telescopic module (6), and the cutting motor is started when the telescopic length reaches two-thirds of the distance of the instruction to cut the male spike; at the same time, the binocular camera (82) recognizes the target in real time, and the control main board (3) calculates the distance between the blade (73) and the target. If the difference is less than a preset error threshold, the current position of the blade (73) is determined to be the cutting point; in S4, the control main board (3) performs a return operation on the mechanical arm assembly after completing the cutting instruction, and the binocular camera (82) photographs the lower area during the movement and is detected by the deep learning module (35). If the missing male spike is detected again and the distance between the target and the blade (73) is greater than the preset error threshold, it is determined that the current task point excision task is not completed; Return to S2 and S3 to cut off the target tassel again.

9. The control method of a UAV based on vision guidance according to claim 5, characterized in that: Before the task starts, the coordinate information of the target to be detasseled is loaded, and the positioning module (32) and the path planning module (33) divide the task groups and plan the path based on the take-off point of the UAV, the terrain of the plot, and the target distribution status; after arriving above the target point, the binocular camera module (8) is used to collect image data of the male tassel, and the three-dimensional point cloud information of the male tassel is calculated by the deep learning module (34), and then the obtained point cloud information is filtered to extract the precise position information of the male tassel; the control module (31) removes the male tassel through the mechanical arm component according to the visual recognition result, and continues to execute the next detasseling task according to the planned path.

10. The control method of a UAV based on vision guidance according to claim 5, characterized in that: The deep learning module (34) constructs a tassel data set and performs convolutional neural network (CNN) model training to achieve accurate segmentation and fitting judgment of corn tassel targets. The training process is as follows: an initial network structure model is established, and the weights of each layer of the model are randomly initialized. The tassel images in the training set are preprocessed and input into the initial model. The input images are subjected to feature extraction through convolution operations. The parameters are updated through back propagation and iterative training is performed. The parameters in the neural network are continuously optimized and updated to obtain the optimal model. During use, the last training weights are continuously loaded for transfer learning, and the model is continuously optimized to maintain a high generalization ability.

Images