Automatic recovery method for narrow-space unmanned rotorcraft of offshore mobile platform

Abstract

The invention discloses an automatic recovery method for a narrow-space unmanned rotorcraft of a maritime mobile platform. The method comprises the following steps: step 1, the maritime mobile platform sends a withdrawal instruction and a body position instruction; step 2, the rotor unmanned aerial vehicle captures a maritime mobile platform by using an extensible FPN algorithm and moves to a position above the platform; step 3, the rotor unmanned aerial vehicle uses an extensible FPN algorithm to identify a recovery area checkerboard, captures checkerboard vector information, and starts to descend to approach the recovery area; 4, the unmanned rotorcraft descends to the position 10 meters away, the unmanned rotorcraft obtains an RTK signal, and a geodetic coordinate system is converted into a relative coordinate system of a recovery system; and step 5, after passing through the functional support column with the height of 2 meters, the rotor unmanned aerial vehicle can slowly land on the vector checkerboard according to the accurate elevation information of the infrared distance meter. The method has the characteristics of no need of personnel participation in the whole process, outstanding safety and accuracy and high recovery efficiency.

Description

Technical Field

[0001] The invention relates to the technical field of unmanned aerial vehicle control, and in particular to an automatic recovery method for a narrow-space rotary-wing unmanned aerial vehicle on an offshore mobile platform. Background Art

[0002] The recovery of traditional rotary-wing UAVs relies on remote control recovery by humans in a loop, or semi-automatic recovery by preset routes. However, in the final stage of recovery, it relies on close-range visual remote control landing. If the recovery of rotary-wing UAVs is carried out in the case of wave formation and extremely narrow landing area on the mobile offshore platform, it is difficult to accurately monitor and adjust the spatial position and attitude of the UAV, and the landing accuracy cannot be fully guaranteed. There is also the possibility of collision with other protruding equipment on the mobile offshore platform, causing a larger secondary accident. Therefore, the traditional rotary-wing UAV recovery control technology has the disadvantages of lack of automation, low precision, and high risk, and cannot meet the requirements of the mobile offshore platform for efficient and automatic recovery of UAVs. Summary of the invention

[0003] The purpose of the present invention is to provide a method for automatically recovering a narrow-space rotary-wing UAV on an offshore mobile platform, comprising the following steps:

[0004] 1) The offshore mobile platform sends a withdrawal command to the rotary-wing UAV, and periodically sends the offshore mobile platform body position command to the rotary-wing UAV;

[0005] 2) After receiving the withdrawal command, the rotary-wing UAV sends a withdrawal command response to the offshore mobile platform, and approaches the position of the offshore mobile platform according to the position command of the offshore mobile platform;

[0006] During the approach process, the rotor UAV uses the binocular camera gimbal to take real-time exterior images, and uses the image recognition model based on the improved YOLO v8 framework to determine whether there is an offshore mobile platform in the exterior image. If so, the horizontal angle and pitch angle of the binocular camera gimbal to the offshore mobile platform are calculated by its own posture and position, and then proceed to step 3);

[0007] 3) The rotor drone retrieves the stored top deck image set and moves to the top of the mobile platform at a speed of v1 based on the horizontal angle, pitch angle and the position command of the offshore mobile platform. At this time, the binocular camera is facing vertically downward;

[0008] 4) The rotor UAV uses the binocular camera gimbal to take real-time images of the image below the rotor UAV, and uses the recovery area recognition model based on the ResNet-50 network to determine whether there is a recovery area of ​​the offshore mobile platform in the image below. If so, the rotor UAV enters the descent phase;

[0009] The rotor drone descends at a speed of v2 to a height of h1 above the recovery area and determines whether to continue descending according to the signal from the take-off and landing indicator light. If it continues to descend, it proceeds to step 5);

[0010] 5) Two RTK reference stations on the functional pillar of the offshore mobile platform divide the area where the recovery area is located and the space above it at a height of h1 into multiple spatial grids;

[0011] The offshore mobile platform sends base station positioning information to the rotary-wing UAV;

[0012] 6) The RTK receiver of the rotary-wing UAV obtains the positioning information of the RTK base station and obtains a fixed solution positioning. The coordinate translation matrix and the coordinate rotation matrix are used to determine the Gaussian projection transformation formula after parameter optimization, and the position information of the offshore mobile platform is converted into the relative position relationship information between the rotary-wing UAV and the recovery area;

[0013] 7) The rotor UAV determines the elevation information between the rotor UAV and the recovery area through an infrared rangefinder, and based on the relative position relationship information between the rotor UAV and the recovery area, lands in the recovery area at a speed v3; v3<v2.

[0014] Furthermore, the offshore mobile platform main body position instruction carries the offshore mobile platform main body position information, including the longitude, latitude, speed and direction of the rotor UAV in the CGCS2000 geodetic coordinate system.

[0015] Furthermore, the image recognition model based on the YOLO v8 framework includes a backbone network, a head network, and a neck network;

[0016] The backbone network adopts the DeformableConvNets v1-3 improved C2f module as the basic component unit;

[0017] The backbone network is used to extract image features;

[0018] The neck network uses an adaptive spatial feature fusion method to fuse the image features extracted by the backbone network;

[0019] The neck network includes an SPPF module, a PAA module and a PAN module combined with Multi-scale context Aggregation;

[0020] The head network is a Dynamic Head detection head, which is used to output recognition results.

[0021] Further, in step 3), the rotor drone hovers on the top of the mobile platform, and the absolute value of the horizontal displacement speed is 1 / 10 of the longest side of the mobile platform.

[0022] Furthermore, the recycling area recognition model based on the ResNet-50 network includes an input module, a multi-branch dilated convolution module, a multi-branch pooling module, a lightweight channel interaction fusion module, a global context module, and an output layer;

[0023] The input module is used to obtain the image below the rotary-wing UAV;

[0024] The multi-branch dilated convolution module and the multi-branch pooling module expand the output receptive field of the ResNet-50 network;

[0025] The lightweight channel interaction fusion module is used to alleviate the problem of sparse receptive field distribution;

[0026] The global context module learns image perspective information and improves multi-scale detection performance.

[0027] The output layer is used to output the recognition result.

[0028] Furthermore, if the signal sent by the take-off and landing indicator light is descending, the rotor UAV continues to descend; if the signal sent by the take-off and landing indicator light is hovering, the rotor UAV hovers at the current height; if the signal sent by the take-off and landing indicator light is ascending, the rotor UAV rises to hover at a height h2 above the vector chessboard, where h2 is greater than h1.

[0029] Furthermore, the recycling area is a vector checkerboard area.

[0030] Furthermore, after the rotor UAV lands smoothly, the propeller shaft motor is turned off and a UAV docking instruction is sent to the offshore mobile platform.

[0031] The technical effect of the present invention is unquestionable. The present invention enables the rotary-wing UAV to efficiently identify and capture characteristic images of the offshore mobile platform and the landing area on the platform. The present invention can ensure the dynamic recovery of the rotary-wing UAV in a small space. The present invention can be widely used in the automatic recovery of rotary-wing UAVs by offshore mobile platforms.

[0032] The present invention solves the technical problem of stably recovering a rotary-wing UAV in a relatively small space when the offshore mobile platform's navigation wave formation causes turbulence. The present invention has the characteristics of no need for human involvement throughout the process, outstanding safety and accuracy, and high recovery efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the automatic recovery framework of the rotary-wing UAV.

[0034] Figure 2 This is a schematic diagram of the automatic recovery process of the rotary-wing UAV.

[0035] Figure 3 Identify and capture test images for offshore mobile platforms. ;

[0036] Figure 4 Schematic diagram of the recovery of the rotary-wing UAV. DETAILED DESCRIPTION

[0037] The present invention is further described below in conjunction with the embodiments, but it should not be understood that the above subject matter of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes are made according to the common technical knowledge and customary means in the art, which should all be included in the protection scope of the present invention.

[0038] Embodiment 1:

[0039] See also Figures 1 to 4 , a method for automatically recovering a narrow-space rotor drone on an offshore mobile platform, comprising the following steps:

[0040] 1) The offshore mobile platform sends a withdrawal command to the rotary-wing UAV, and periodically sends the offshore mobile platform body position command to the rotary-wing UAV;

[0041] 2) After receiving the withdrawal command, the rotary-wing UAV sends a withdrawal command response to the offshore mobile platform, and approaches the position of the offshore mobile platform according to the position command of the offshore mobile platform;

[0042] During the approach process, the rotor UAV uses the binocular camera gimbal to take real-time exterior images, and uses the image recognition model based on the improved YOLO v8 framework to determine whether there is an offshore mobile platform in the exterior image. If so, the horizontal angle and pitch angle of the binocular camera gimbal to the offshore mobile platform (that is, the angle of the gimbal offset relative to the reference point) are calculated through its own posture and position, and then proceed to step 3);

[0043] 3) The rotor drone retrieves the stored top deck image set and moves to the top of the mobile platform at a speed of v1 based on the horizontal angle, pitch angle and the position command of the offshore mobile platform. At this time, the binocular camera is facing vertically downward;

[0044] 4) The rotor UAV uses the binocular camera gimbal to take real-time images of the image below the rotor UAV, and uses the recovery area recognition model based on the ResNet-50 network to determine whether there is a recovery area of ​​the offshore mobile platform in the image below. If so, the rotor UAV enters the descent phase;

[0045] The rotor drone descends at a speed of v2 to a height of h1 above the recovery area and determines whether to continue descending according to the signal from the take-off and landing indicator light. If it continues to descend, it proceeds to step 5);

[0046] 5) Two RTK reference stations on the functional pillar of the offshore mobile platform divide the area where the recovery area is located and the space above it at a height of h1 into multiple spatial grids;

[0047] The offshore mobile platform sends base station positioning information to the rotary-wing UAV;

[0048] 6) The RTK receiver of the rotary-wing UAV obtains the positioning information of the RTK base station and obtains a fixed solution positioning. The coordinate translation matrix and the coordinate rotation matrix are used to determine the Gaussian projection transformation formula after parameter optimization, and the position information of the offshore mobile platform is converted into the relative position relationship information between the rotary-wing UAV and the recovery area;

[0049] 7) The rotor UAV determines the elevation information between the rotor UAV and the recovery area through an infrared rangefinder, and based on the relative position relationship information between the rotor UAV and the recovery area, lands in the recovery area at a speed v3; v3<v2.

[0050] The offshore mobile platform main body position instruction carries the offshore mobile platform main body position information, including the longitude, latitude, speed and direction of the rotorcraft in the CGCS2000 geodetic coordinate system.

[0051] The image recognition model based on the YOLO v8 framework includes a backbone network, a head network, and a neck network;

[0052] The backbone network adopts the DeformableConvNets v1-3 improved C2f module as the basic component unit;

[0053] The backbone network is used to extract image features;

[0054] The neck network uses an adaptive spatial feature fusion method to fuse the image features extracted by the backbone network;

[0055] The neck network includes an SPPF module, a PAA module and a PAN module combined with Multi-scale context Aggregation;

[0056] The head network is a Dynamic Head detection head, which is used to output recognition results.

[0057] In step 3), the rotor drone hovers on the top of the mobile platform, and the absolute value of the horizontal displacement speed is 1 / 10 of the longest side of the mobile platform.

[0058] The recycling area recognition model based on the ResNet-50 network includes an input module, a multi-branch dilated convolution module, a multi-branch pooling module, a lightweight channel interaction fusion module, a global context module, and an output layer;

[0059] The input module is used to obtain the image below the rotary-wing UAV;

[0060] The multi-branch dilated convolution module and the multi-branch pooling module expand the output receptive field of the ResNet-50 network;

[0061] The lightweight channel interaction fusion module is used to alleviate the problem of sparse receptive field distribution;

[0062] The global context module learns image perspective information and improves multi-scale detection performance.

[0063] The output layer is used to output the recognition result.

[0064] If the signal sent by the take-off and landing indicator light is descending, the rotor UAV continues to descend. If the signal sent by the take-off and landing indicator light is hovering, the rotor UAV hovers at the current height. If the signal sent by the take-off and landing indicator light is ascending, the rotor UAV rises to hover at a height of h2 above the vector chessboard, where h2 is greater than h1.

[0065] The recycling area is a vector checkerboard area.

[0066] After the rotor UAV lands smoothly, the propeller shaft motor is turned off and the UAV docking command is sent to the offshore mobile platform.

[0067] Embodiment 2:

[0068] A method for automatically recovering a narrow-space rotor drone on an offshore mobile platform comprises the following steps:

[0069] 1) The offshore mobile platform sends a withdrawal command to the rotary-wing UAV, and periodically sends the offshore mobile platform body position command to the rotary-wing UAV;

[0070] 2) After receiving the withdrawal command, the rotary-wing UAV sends a withdrawal command response to the offshore mobile platform, and approaches the position of the offshore mobile platform according to the position command of the offshore mobile platform;

[0071] During the approach process, the rotor UAV uses the binocular camera gimbal to take real-time exterior images, and uses the image recognition model based on the improved YOLO v8 framework to determine whether there is an offshore mobile platform in the exterior image. If so, the horizontal angle and pitch angle of the binocular camera gimbal to the offshore mobile platform are calculated by its own posture and position, and then proceed to step 3);

[0072] 3) The rotor drone retrieves the stored top deck image set and moves to the top of the mobile platform at a speed of v1 based on the horizontal angle, pitch angle and the position command of the offshore mobile platform. At this time, the binocular camera is facing vertically downward;

[0073] 4) The rotor UAV uses the binocular camera gimbal to take real-time images of the image below the rotor UAV, and uses the recovery area recognition model based on the ResNet-50 network to determine whether there is a recovery area of ​​the offshore mobile platform in the image below. If so, the rotor UAV enters the descent phase;

[0074] The rotor drone descends at a speed of v2 to a height of h1 above the recovery area and determines whether to continue descending according to the signal from the take-off and landing indicator light. If it continues to descend, it proceeds to step 5);

[0075] 5) Two RTK reference stations on the functional pillar of the offshore mobile platform divide the area where the recovery area is located and the space above it at a height of h1 into multiple spatial grids;

[0076] The offshore mobile platform sends base station positioning information to the rotary-wing UAV;

[0077] 6) The RTK receiver of the rotary-wing UAV obtains the positioning information of the RTK base station and obtains a fixed solution positioning. The coordinate translation matrix and the coordinate rotation matrix are used to determine the Gaussian projection transformation formula after parameter optimization, and the position information of the offshore mobile platform is converted into the relative position relationship information between the rotary-wing UAV and the recovery area;

[0078] 7) The rotor UAV determines the elevation information between the rotor UAV and the recovery area through an infrared rangefinder, and based on the relative position relationship information between the rotor UAV and the recovery area, lands in the recovery area at a speed v3; v3<v2.

[0079] Embodiment 3:

[0080] A method for automatically recovering a narrow-space rotor drone on an offshore mobile platform, the technical content of which is the same as that of Example 2. Furthermore, the offshore mobile platform body position instruction carries the offshore mobile platform body position information, including longitude, latitude, speed and direction in the CGCS2000 geodetic coordinate system.

[0081] Embodiment 4:

[0082] A method for automatically recovering a narrow-space rotor drone on an offshore mobile platform, the technical content of which is the same as any one of Embodiments 2-3, and further, an image recognition model based on a YOLO v8 framework includes a backbone network, a head network, and a neck network;

[0083] The backbone network adopts the DeformableConvNets v1-3 improved C2f module as the basic component unit;

[0084] The backbone network is used to extract image features;

[0085] The neck network uses an adaptive spatial feature fusion method to fuse the image features extracted by the backbone network;

[0086] The neck network includes an SPPF module, a PAA module and a PAN module combined with Multi-scale context Aggregation;

[0087] The head network is a Dynamic Head detection head, which is used to output recognition results.

[0088] Embodiment 5:

[0089] A method for automatically recovering a rotor drone in a narrow space on an offshore mobile platform, the technical content of which is the same as any one of Examples 2-4, and further, in step 3), the rotor drone hovers on the top of the mobile platform, and the absolute value of the horizontal displacement speed is 1 / 10 of the longest side of the mobile platform.

[0090] Embodiment 6:

[0091] A method for automatically recovering a narrow-space rotor drone on an offshore mobile platform, the technical content of which is the same as any one of Embodiments 2-5, and further, a recovery area recognition model based on a ResNet-50 network includes an input module, a multi-branch dilated convolution module, a multi-branch pooling module, a lightweight channel interaction fusion module, a global context module, and an output layer;

[0092] The input module is used to obtain the image below the rotary-wing UAV;

[0093] The multi-branch dilated convolution module and the multi-branch pooling module expand the output receptive field of the ResNet-50 network;

[0094] The lightweight channel interaction fusion module is used to alleviate the problem of sparse receptive field distribution;

[0095] The global context module learns image perspective information and improves multi-scale detection performance.

[0096] The output layer is used to output the recognition result.

[0097] Embodiment 7:

[0098] A method for automatically recovering a rotor UAV in a narrow space on an offshore mobile platform, the technical content of which is the same as any one of Examples 2-6, and further, if the signal emitted by the take-off and landing indicator light is descending, the rotor UAV continues to descend; if the signal emitted by the take-off and landing indicator light is hovering, the rotor UAV hovers at the current height; if the signal emitted by the take-off and landing indicator light is ascending, the rotor UAV rises to hover at a height of h2 above the vector chessboard, where h2 is greater than h1.

[0099] Embodiment 8:

[0100] A method for automatically recovering a narrow-space rotor drone on an offshore mobile platform, the technical content of which is the same as any one of Examples 2-7, and further, the recovery area is a vector checkerboard area.

[0101] Embodiment 9:

[0102] A method for automatically recovering a rotor UAV in a narrow space on an offshore mobile platform, the technical content of which is the same as any one of Examples 2-8. Furthermore, after the rotor UAV lands smoothly, the propeller shaft motor is turned off and a UAV docking instruction is sent to the offshore mobile platform.

[0103] Embodiment 10:

[0104] A method for automatically recovering a narrow-space rotor drone on an offshore mobile platform, the steps are as follows:

[0105] Step 1: The offshore mobile platform sends a withdrawal command and a main body position command;

[0106] After the rotary-wing UAV sends the withdrawal command response, it begins to approach the main position of the offshore mobile platform at a speed of 20m / s, with the altitude controlled at 150m above sea level. The main position command adopts the CGCS2000 geodetic coordinate system, including longitude, latitude, speed and direction, and the altitude is 3 meters above sea level, and is sent to the rotary-wing UAV at 1 / second. The rotary-wing UAV calculates the horizontal angle and pitch angle of the binocular camera gimbal to the offshore mobile platform through its own attitude and position.

[0107] Step 2: The rotary-wing UAV uses the scalable FPN algorithm to capture the offshore mobile platform and moves above the platform;

[0108] DeformableConvNets v1-3 is introduced to improve the C2f module, and the features of offshore mobile platforms of different sizes can be extracted more flexibly. Multi-scale context Aggregation is combined with the SPPF module in YOLO v8 to capture long-range dependencies and adaptability, realize the reuse and fusion of shallow to deep features, and reduce the interference of complex background on the detection of offshore mobile platforms. Adaptive Spatial Feature Fusion (ASFF) is used in the YOLO v8 neck network to enhance the detection effect of small-target offshore mobile platforms. The ASFF structure can effectively combine image perspective information and suppress conflicting information between features of different scales, thereby enhancing the feature learning ability of the network. The YOLO v8 head is replaced with the Dynamic Head detection head. By combining the three attention functions of scale, space and task, the performance of the model under the above three improvements is further improved.

[0109] This step helps the rotorcraft UAV to identify the offshore mobile platform in the complex background image. The appropriately improved YOLO v8 algorithm has a strong ability to capture large targets with background interference and long distance. After the rotorcraft UAV identifies the corresponding offshore mobile platform, it calls out the top deck image set and starts to hover on the top of the mobile platform. At this time, the altitude is 80m, and the binocular camera is vertically downward to find the recovery area. The absolute value of the horizontal displacement speed is 1 / 10 of the longest side of the mobile platform.

[0110] Step 3: The rotary-wing UAV uses the scalable FPN algorithm to capture the vector checkerboard of the offshore mobile platform, allowing the UAV to enter the descent phase;

[0111] CA attention is applied to the outputs of the backbone network Stage3 and Stage4, and the feature matrices are average pooled in the W and H directions respectively to obtain three-dimensional feature matrices of C*1*H and C*W*1. The two feature matrices are convolved, concatenated (Contact) and nonlinearly processed, and the output matrix size is (C / r)*1*(H+W). The number of channels is restored by convolution and sigmoid function, and the output is the weight information of each position of the original feature map. By analyzing the receptive field of ResNet-50 and the vector checkerboard, it is found that the actual receptive field of each layer of ResNet-50 output is much smaller than the vector checkerboard receptive field. The present invention designs a multiple branch dilated convolution (MBD) module and a multiple branch pooling (MBP) module to double the original receptive field of the C2 and C4 layers. In view of the sparse distribution of the actual receptive field, a lightweight channel interaction fusion (Channel interaction Fusion) is designed with the C3 layer as the central layer. The sparse receptive field distribution problem is effectively alleviated by the CIF (Convolutional Fiber Interference Integrator) module. In view of the fact that FPN loses channel information when unifying the number of channels through lateral convolution, the present invention uses SubPixel convolution to derive an additional bidirectional path to supplement the channel information. The global context (GC) module is introduced to further learn the image perspective information, thereby improving the multi-scale detection performance.

[0112] This step is conducive to the rapid identification and anchoring of the vector chessboard during the recovery process of the rotor UAV. The scalable FPN algorithm improves the effectiveness of multi-scale object detection performance. The acquisition of vector chessboard image information during the gradual recovery and descent of the rotor UAV is a process of gradual expansion of small targets at close range, and it must be anchored in real time. This method allows the image information to contain both channel information and spatial position information, achieving lightweight models while enabling the rotor UAV to arrange its position in the direction of the vector chessboard. The rotor UAV descends at 4m / s to 10m on the vector chessboard and begins to decelerate. If the take-off and landing indicator light is green, it will continue to descend, if it is yellow, it will hover, and if it is red, it will hover at an altitude of 80m.

[0113] Step 4: The rotary-wing UAV obtains the RTK signal and converts the earth coordinate system into the relative coordinate system of the recovery system;

[0114] The two RTK base stations on the functional pillars divide the space around the vector chessboard and 10 meters above it into a spatial grid with an accuracy of 2 cm. After the rotorcraft UAV RTK receiver obtains the positioning information of the base station, it obtains a fixed solution positioning, and uses the coordinate translation matrix and coordinate rotation matrix to determine the Gaussian projection transformation formula after parameter optimization, which is converted from the CGCS2000 geodetic coordinate system into a reference coordinate system used to describe the relative position relationship between the UAV and the recovery area.

[0115] Step 5: The rotor UAV uses the infrared rangefinder to obtain precise elevation information, allowing the UAV to slowly land on the vector chessboard.

[0116] The rotor UAV descends at a speed of 1m / s at a height of 10m. After passing the top of the 2m-high functional pillar, it relies on an infrared rangefinder with centimeter-level accuracy to obtain high-precision elevation information, which is fed back to the rotor UAV for a smooth landing. After the UAV lands smoothly, the propeller shaft motor is turned off and the UAV docking command is sent to the offshore mobile platform.

Claims

1. A method for automatically recovering a narrow-space rotor drone on an offshore mobile platform, characterized in that: The following steps are involved: 1) The offshore mobile platform sends a withdrawal command to the rotary-wing UAV, and periodically sends the offshore mobile platform body position command to the rotary-wing UAV. 2) After receiving the withdrawal command, the rotor UAV sends a withdrawal command response to the offshore mobile platform, and approaches the position of the offshore mobile platform according to the position command of the offshore mobile platform. During the approach process, the rotor UAV uses the binocular camera gimbal to take real-time exterior images, and uses the image recognition model based on the improved YOLO v8 framework to determine whether there is an offshore mobile platform in the exterior image. If so, the horizontal angle and pitch angle of the binocular camera gimbal to the offshore mobile platform are calculated by its own posture and position, and then proceed to step 3); 3) The rotor drone retrieves the stored top deck image set and moves to the top of the mobile platform at a speed of v1 based on the horizontal angle, pitch angle and the position command of the offshore mobile platform. At this time, the binocular camera is facing vertically downward; 4) The rotor UAV uses the binocular camera gimbal to take real-time images of the image below the rotor UAV, and uses the recovery area recognition model based on the ResNet-50 network to determine whether there is a recovery area of ​​the offshore mobile platform in the image below. If so, the rotor UAV enters the descent phase; The rotor drone descends at a speed of v2 to a height of h1 above the recovery area and determines whether to continue descending according to the signal from the take-off and landing indicator light. If it continues to descend, it proceeds to step 5); 5) Two RTK reference stations on the functional pillar of the offshore mobile platform divide the area where the recovery area is located and the space above it at a height of h1 into multiple spatial grids; The offshore mobile platform sends base station positioning information to the rotary-wing UAV; 6) The RTK receiver of the rotary-wing UAV obtains the positioning information of the RTK base station and obtains a fixed solution positioning. The coordinate translation matrix and the coordinate rotation matrix are used to determine the Gaussian projection transformation formula after parameter optimization, and the position information of the offshore mobile platform is converted into the relative position relationship information between the rotary-wing UAV and the recovery area; 7) The rotor UAV determines the elevation information between the rotor UAV and the recovery area through an infrared rangefinder, and based on the relative position relationship information between the rotor UAV and the recovery area, lands in the recovery area at a speed v3; v3<v2.

2. The method for automatically recovering a narrow-space rotor drone on an offshore mobile platform according to claim 1, characterized in that: The offshore mobile platform main body position instruction carries the offshore mobile platform main body position information, including the longitude, latitude, speed and direction of the rotorcraft in the CGCS2000 geodetic coordinate system.

3. The automatic recovery method of a narrow space rotary wing UAV on an offshore mobile platform according to claim 1 is characterized in that: The image recognition model based on the YOLO v8 framework includes a backbone network, a head network, and a neck network; The backbone network adopts the DeformableConvNets v1-3 improved C2f module as the basic component unit; The backbone network is used to extract image features; The neck network uses an adaptive spatial feature fusion method to fuse the image features extracted by the backbone network; The neck network includes an SPPF module, a PAA module and a PAN module combined with Multi-scale context Aggregation; The head network is a Dynamic Head detection head, which is used to output recognition results.

4. The method for automatically recovering a narrow-space rotor drone on an offshore mobile platform according to claim 1, characterized in that: In step 3), the rotor drone hovers on the top of the mobile platform, and the absolute value of the horizontal displacement speed is 1 / 10 of the longest side of the mobile platform.

5. The method for automatically recovering a narrow-space rotor drone on an offshore mobile platform according to claim 1, characterized in that: The recycling area recognition model based on the ResNet-50 network includes an input module, a multi-branch dilated convolution module, a multi-branch pooling module, a lightweight channel interaction fusion module, a global context module, and an output layer; The input module is used to obtain the image below the rotary-wing UAV; The multi-branch dilated convolution module and the multi-branch pooling module expand the output receptive field of the ResNet-50 network; The lightweight channel interaction fusion module is used to alleviate the problem of sparse receptive field distribution; The global context module learns image perspective information to improve multi-scale detection performance; The output layer is used to output the recognition result.

6. The method for automatically recovering a narrow-space rotary-wing UAV on an offshore mobile platform according to claim 1, characterized in that: If the signal sent by the take-off and landing indicator light is descending, the rotor UAV continues to descend. If the signal sent by the take-off and landing indicator light is hovering, the rotor UAV hovers at the current height. If the signal sent by the take-off and landing indicator light is ascending, the rotor UAV rises to hover at a height of h2 above the vector chessboard, where h2 is greater than h1.

7. The method for automatically recovering a narrow-space rotary-wing UAV on an offshore mobile platform according to claim 1, characterized in that: The recycling area is a vector checkerboard area.

8. The method for automatically recovering a narrow-space rotary-wing UAV on an offshore mobile platform according to claim 1, characterized in that: After the rotor UAV lands smoothly, the propeller shaft motor is turned off and the UAV docking command is sent to the offshore mobile platform.

Images