An autonomous landing gear docking method based on multi-modal perception

Abstract

The application discloses an autonomous docking method for landing gears based on multi-modal perception and belongs to the field of automatic navigation and docking control; the method comprises the following steps: in the long-distance coarse alignment stage, a dual-path visual sensor is used to acquire helicopter images and perform center closed-loop control on the boundary box, and after centering, the fine alignment dominated by the radar centroid direction is switched to and the AGV is controlled to move straight to approach; in the short-distance central axis alignment stage, the landing gear point cloud is subjected to multiple spatial filtering, the central axis is extracted through serialized RANSAC double-line fitting and normal vector consistency correction, the heading alignment is performed by using the direction vector accumulation, the lateral deviation is measured at the far end of the skid and convergence is controlled; in the high-precision docking stage of the belly, a depth camera is used to collect the skid side marker image, the actual depth is estimated by considering the pixel size reverse calculation formula of the inclination correction, and a multi-axis decoupling sequential control strategy is adopted to complete the high-precision docking; and the autonomous docking work with high precision and high safety is realized.

Description

Technical Field

[0001] This invention belongs to the field of automated navigation and docking control of aircraft, and specifically relates to an autonomous docking method for landing gear based on multimodal perception. Background Technology

[0002] In ground support, maintenance, and mobile transport tasks for helicopters and other aircraft, the use of Automated Guided Vehicles (AGVs) to autonomously drive under the helicopter and precisely dock with the landing gear skid structure is a key technology for achieving unmanned and intelligent ground support. However, due to the complex helicopter parking environment and the significant differences in the observability of landing gear skid characteristics at different distances, existing docking navigation solutions face numerous technical bottlenecks.

[0003] Perception failure over long distances: At long distances (e.g., more than 10 meters), the landing gear and skid point cloud is extremely sparse, and traditional lidar cannot provide effective alignment of the centerline. Conventional depth cameras also cannot work in the far field due to noise and resolution limitations.

[0004] Point cloud clutter and centerline jump: During the close-range radar alignment phase, the surrounding ground clutter, AGV body reflections, and the sparse characteristics of the non-repeating scanning lidar can easily cause the landing gear skid profile fitting failure, resulting in severe control heading "jumps" or false convergence, which can easily lead to collision accidents.

[0005] Depth ambiguity and occlusion failure in visual positioning: After entering directly beneath the aircraft's belly, radar signals are lost due to the confined space and the physical self-occlusion of the skids, necessitating extremely high-precision (millimeter-level) docking with visual reference markers. However, the traditional solvePnP algorithm suffers from insurmountable "depth ambiguity" when processing small-sized planar markers due to camera calibration distortion or marker surface tilt, resulting in significant depth-direction positioning errors. Furthermore, sudden changes in illumination, marker contamination, or occlusion can easily lead to instantaneous loss of positioning signals, and a reliable safety backoff mechanism is lacking.

[0006] Therefore, there is an urgent need for a multimodal landing gear autonomous docking solution that can integrate visual semantics, laser geometric features and high-precision image decoupling control, and has a cascaded safety back-off mechanism. Summary of the Invention

[0007] To address the above shortcomings, this invention provides a multimodal sensing-based autonomous landing gear docking method, comprising the following steps:

[0008] Step 1, Long-distance coarse alignment stage: Long-distance alignment is performed using a vision sensor and a 3D LiDAR installed at the front end of the AGV; the long-distance coarse alignment stage includes four sequentially connected sub-states: search sub-state, alignment sub-state, fine alignment sub-state, and straight-line sub-state; the AGV is controlled to switch sequentially between the four sub-states to perform heading closure control, and the AGV is controlled to approach in a straight line to the preset close-range switching distance;

[0009] Step 2, Close-range centerline alignment stage: After the AGV enters the close-range switching distance, the landing gear skid structure is scanned using a 3D LiDAR and multiple spatial filtering of the point cloud is performed. The geometric parameters of the two skid lines are solved by the RANSAC bilinear fitting algorithm and the centerline heading is extracted. The AGV is controlled to perform heading alignment control and lateral deviation convergence control based on the centerline heading, and obstacle detection and collision avoidance are performed during the journey.

[0010] Step 3, High-precision docking stage under the fuselage: When the AGV drives under the fuselage of the helicopter and the radar signal is lost, the system switches to precision alignment control dominated by the depth vision sensor and the visual reference mark. The three-dimensional relative position of the visual reference mark is calculated based on the pixel size back-calculation algorithm with tilt correction, and the AGV chassis is driven to move by a multi-axis decoupled sequential control strategy to complete the high-precision docking. When the preset abnormal judgment conditions are met, cascaded backtracking is performed.

[0011] Furthermore, step one specifically includes:

[0012] In the search sub-state, the AGV is controlled to rotate at a constant angular velocity to search for the target. When any vision sensor detects the helicopter pixel bounding box, it automatically switches to the alignment sub-state.

[0013] In the alignment state, the heading of the AGV is controlled by a proportional-integral-derivative (PID) controller using the center offset of the pixel bounding box of the dual-channel vision sensor; once the center is precisely aligned, the system automatically switches to the precise alignment state.

[0014] In the precise alignment state, the direction of the centroid of the radar point cloud exponential moving average after median filtering and noise reduction is used to guide the alignment, and the PID controller automatically switches between high and low gain levels according to the azimuth deviation to approach the course; when the alignment is completed, it automatically switches to the straight-line state.

[0015] In the straight-line sub-state, the AGV is controlled to move in an open-loop straight path to approach the short-distance switching distance.

[0016] Furthermore, the exponential moving average centroid direction is updated in real time only during the precise alignment state, the straight alignment state, and the close centerline alignment stage, and remains locked during the search and alignment states to suppress interference from distant sparse noise points on the heading estimation.

[0017] Furthermore, in step two, performing multiple spatial filtering on the point cloud data specifically includes:

[0018] A height-direction interval filter is used to filter out noise outside the area at the bottom of the landing gear skids;

[0019] Set an exclusion rectangle for the AGV body to filter out interference points reflected by the AGV body itself;

[0020] A cone-shaped filtering region is defined around the current exponential moving average heading angle to filter out environmental clutter outside the cone angle range, thereby obtaining a clean set of skid feature points.

[0021] Furthermore, in step two, solving for the geometric parameters of the two skid lines and the heading alignment control specifically includes:

[0022] The serialized RANSAC bilinear fitting algorithm with a set number of iterations was used to fit the left and right sliding lines respectively.

[0023] Calculate the product of the normal vectors of the slopes of the two sliding lines. If the product is less than zero, then perform an angle flip correction on the second fitted line.

[0024] Using the centerline of the two skid lines as the target heading, heading alignment control is performed using the direction vector accumulation criterion.

[0025] Furthermore, in step two, the forward obstacle detection and detour collision avoidance specifically includes:

[0026] The density of laser reflection points is obtained by using a polar coordinate sector density peak algorithm. When the detected reflection point density exceeds the set global average density threshold, it is determined that an airborne pod structure has been detected.

[0027] If the airborne pod structure is located directly in front of the AGV, the AGV is controlled to first move laterally for a set time to avoid the projection area, and then travel forward to a preset distance to circle around to the rear of the helicopter.

[0028] Furthermore, in step three, the pixel size inverse estimation algorithm based on tilt angle correction specifically includes:

[0029] When the visual reference marker and the optical axis of the depth camera are tilted at an angle, the actual depth distance is calculated using the following corrected depth formula. :

[0030]

[0031] in, This represents the actual pixel focal length value of the camera. The actual physical side length marked for visual reference. The pixel dimensions and side lengths in the image are used as visual reference markers. The angle between the normal vector and the camera's optical axis;

[0032] The included angle The estimation is based on the ratio of the upper and lower edge pixel lengths of the visual reference marked image, and the estimation formula is as follows:

[0033]

[0034] in, These are the pixel lengths of the upper and lower edges of the visual reference markers in the image, respectively.

[0035] Furthermore, in step three, the multi-axis decoupling sequence control strategy specifically includes:

[0036] The control is decoupled and executed in the order of yaw angle alignment control, longitudinal deviation correction control, lateral deviation correction control, and three-axis residual monitoring control, and at any given time, a non-zero control speed command is output along only a single degree of freedom.

[0037] In the three-axis residual monitoring and control, if the drift in any dimension exceeds the set threshold, the corresponding correction control step will be automatically re-entered for re-alignment. When all three-axis residuals are continuously and stably stable for a preset number of frames, the docking is determined to be complete.

[0038] Furthermore, in step three, the cascading rollback specifically includes:

[0039] When the visual reference marker is continuously lost for a period exceeding a first preset time threshold, or when the total execution time of the high-precision docking phase of the fuselage exceeds a second preset time threshold, the system triggers a cascading rollback:

[0040] Control the AGV to immediately apply emergency braking and record the current pose estimate;

[0041] Control the AGV to reverse and move straight out from under the helicopter's belly, returning to the straight-moving waiting-to-enter state of step two;

[0042] Control the AGV to resume open-loop straight movement and attempt to enter step three again.

[0043] The present invention also discloses a system for the above-mentioned multimodal perception-based autonomous landing gear docking method, comprising: an automated guided vehicle (AGV) chassis, a dual-channel vision sensor and a three-dimensional lidar mounted on the front end of the AGV, a depth vision sensor mounted on the top of the AGV, and an airborne mission computer and flight control system.

[0044] The airborne mission computer integrates:

[0045] An autonomous navigation module for controlling the AGV to sequentially switch between four sub-states: search, alignment, fine alignment, and straight travel in order to perform long-distance coarse alignment.

[0046] Direction measurement module used for filtering landing gear point clouds, skid bilinear fitting, and centerline heading extraction;

[0047] A precision alignment control module is used to calculate the pose of the visual reference markers acquired by the depth vision sensor and to perform multi-axis decoupling control.

[0048] An exception tolerance module used to trigger cascading rollback when exception conditions are met;

[0049] The airborne mission computer is configured to send chassis motion control commands to the flight control system at each stage to drive the AGV chassis to complete autonomous docking.

[0050] Compared with the prior art, the present invention has the following advantages:

[0051] 1. Wide-range high dynamic perception: The multimodal perception stage achieves seamless step-by-step cascading in space and time, which not only ensures rapid approach over long distances, but also ensures high-precision (millimeter-level) alignment in the narrow space under the fuselage.

[0052] 2. Excellent noise and shake resistance: By designing an EMA centroid direction locking mechanism, LiDAR multi-cone / height filtering, serialized RANSAC normal vector correction, and a lateral deviation remote measurement strategy, the amplification effect of angle jitter is eliminated, making the AGV's heading alignment and lateral correction extremely stable.

[0053] 3. Eliminate visual depth blur: The pixel size back-calculation formula with tilt angle cosine correction is used to replace the traditional solvePnP iteration, eliminating up to 55% of the depth direction estimation bias when the marker tilt is 30° to 50°.

[0054] 4. Comprehensive safety and fault tolerance: A cascaded anomaly backoff control is designed for the loss of visual reference markers and non-convergence due to timeout. Through the logical closed loop of "emergency stop - exit - re-shuttle - limited retry", the absolute safety of the expensive helicopter landing gear and airborne equipment is guaranteed. Attached Figure Description

[0055] Figure 1 This is a flowchart of the autonomous docking method in this invention.

[0056] Figure 2 This is the user interface for the autonomous navigation module in this invention.

[0057] Figure 3 This is the operation interface for the high-precision docking stage in this invention.

[0058] Figure 4 This is the user interface for the skid detection process in this invention.

[0059] Figure 5 This is the user interface for the centerline fitting and alignment process in this invention.

[0060] Figure 6 This is the horizontally aligned operation interface in this invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example 1

[0063] This embodiment provides an autonomous landing gear docking system based on multimodal perception. The overall hardware and software control architecture is as follows:

[0064] The system's hardware specifically includes:

[0065] AGV chassis: A four-steering-wheel independent steering chassis, capable of high-precision lateral translation and zero-radius rotation without changing the heading;

[0066] Airborne mission computer: As the central control and computing core, it adopts a computing platform equipped with high computing power chips to run software algorithms such as zero-sample helicopter detection model, point cloud processing algorithm, QR code feature recognition and positioning, and multi-axis decoupling control;

[0067] Flight control system: As the underlying motion controller of the AGV chassis motors, it receives high-frequency control commands (such as the desired velocity vector) from the onboard mission computer via a data bus. ]);

[0068] Front vision sensor: It adopts a dual 640×480 pixel OAK-1-AF wide-angle camera installed at the front of the vehicle body, with a symmetrical installation angle of ±20° on both sides;

[0069] Front-mounted 3D LiDAR: Employs the Livox Mid-360 LiDAR, which boasts centimeter-level ranging accuracy and is aligned with the vehicle's central axis.

[0070] Top depth camera group: Two D455 depth cameras are symmetrically installed on the left and right sides of the top of the AGV body, with an installation elevation angle of 30°, to cover the visual reference mark (AprilTag36h10) on the side of the helicopter skid after entering the belly of the AGV.

[0071] The software module architecture includes a software layer running in the onboard mission computer, specifically including:

[0072] Autonomous navigation module: responsible for efficient asynchronous inference of YOLOv8s-World model, outputting the pixel bounding box of helicopter, and performing EMA smoothing and two-level gain PID heading control on lidar point cloud;

[0073] Orientation Measurement Module: Integrates point cloud multiple filters and serialized RANSAC bilinear fitter for measuring lateral deviation at the far end of the skid and running polar coordinate sector density peak calculation method to detect collision avoidance;

[0074] The precision alignment control module integrates an AprilTag detector, a pixel size back-calculation operator, and a rigorous six-axis decoupled state machine (YAW-LON-LAT-CHECK).

[0075] Anomaly tolerance module: Used to monitor the loss status of visual reference markers and docking time, and trigger cascading rollback when an anomaly occurs.

[0076] Example 2

[0077] This embodiment is based on the autonomous docking system of Embodiment 1, and implements complete closed-loop control of the helicopter landing gear docking method. The specific execution steps and sub-state transition details are as follows:

[0078] Step 1: Coarse Alignment at Long Distance (Phase 1)

[0079] The AGV approaches the helicopter from 15m away, and the autonomous navigation module in the onboard mission computer controls the AGV chassis to sequentially switch between four sub-states:

[0080] S1-1, SEARCH sub-state:

[0081] Control the AGV chassis to rotate counterclockwise at a constant angular velocity. For example... Figure 2As shown, the autonomous navigation module alternately acquires and caches images from the left and right cameras at an asynchronous rate of 3Hz. When either camera detects a valid helicopter bounding box, the system automatically switches to ALIGN state.

[0082] S1-2, ALIGN (Alignment) sub-state:

[0083] The geometric center offset of the dual-camera bounding box on the image plane is used as the error input, and the AGV chassis rotation is controlled in a closed loop using a PID control algorithm. For single-camera detection scenarios, the target is quickly pulled into the overlapping field of view of the dual cameras by rotating at a constant speed with a large deviation. When the image satisfies the condition of a 20-pixel window and is centered for 5 consecutive frames, the alignment is considered complete, and the process proceeds to the next state.

[0084] S1-3, FINE_ALIGN (Fine Alignment) sub-state:

[0085] The main control is handed over to the 3D LiDAR. The orientation measurement module performs median filtering on the LiDAR point cloud to calculate the radar reflection centroid, and smooths the centroid orientation using the EMA (Exponential Moving Average) algorithm. A two-stage gain PID controller adjusts the azimuth deviation accordingly. Automatic gain switching: when At that time, set high gain Drive the chassis to approach rapidly; when Automatically switch to low gain. Prevent chassis overshoot and oscillation. During this period, ground clutter and helicopter fuselage interference are suppressed through corridor filtering and a dynamic minimum distance threshold;

[0086] S1-4, DRIVE (Straight-Ahead) Sub-state:

[0087] The AGV chassis locks in the alignment direction and approaches the helicopter in an open loop at a speed of 0.5m / s until the distance to the helicopter is reduced to 3.5m. At this point, the system automatically switches to the subsequent Phase 2 close-range phase.

[0088] During this period, the EMA centroid direction is updated in real time only during FINE_ALIGN, DRIVE, and Phase2, and is completely locked during SEARCH and ALIGN, thus avoiding contamination of laser heading estimation by distant sparse noise.

[0089] Step 2, Close-range centerline alignment stage (Phase 2):

[0090] When the distance enters within 3.5m, the sled point cloud features are obvious. Perform the following steps:

[0091] S2-1, Point Cloud Multi-Spatial Filtering:

[0092] The orientation measurement module performs multiple filters on the point cloud of the skid bottom acquired by the radar:

[0093] 1) Utilize Height range filtering is performed within the axis interval [-0.4m, -0.2m].

[0094] 2) Set the vehicle body exclusion rectangle to filter out noise from the vehicle body itself;

[0095] 3) Set up a ±60° cone-shaped filtering area around the current EMA direction to finally obtain a pure set of left and right skid feature points;

[0096] S2-2, Serialization RANSAC bilinear fitting and alignment (ALIGN):

[0097] like Figure 4 As shown, the left and right sled lines are fitted using the serialized RANSAC algorithm iterated 80 times. If the consistent product of the normal vectors of the two sled lines is calculated... Then the second fitted line is flipped at an angle to eliminate the abrupt change in the calculation of the central axis;

[0098] Then, using the fitted skid centerline direction as the target, heading alignment control is performed using the direction vector accumulation criterion to eliminate false convergence caused by reverse noise points, such as... Figure 5 As shown. After EMA smoothing, stability for 8 consecutive frames is used as the locking criterion;

[0099] S2-3, Remote Lateral Deviation Control (LOCKED) and Collision Avoidance:

[0100] After the heading is locked, the system switches to the core sub-state of close alignment (LOCKED).

[0101] The orientation measurement module measures lateral deviation at the landing gear skid landing point (far end) rather than the AGV chassis physical origin (near end). A simple moving average (SMA, scale factor 0.15) is used to smooth the lateral deviation, and a PID controller with a 0.02m dead zone limit is utilized. After 25 consecutive frames of detection, the lateral deviation was controlled within 0.03m.

[0102] During this period, the orientation measurement module runs the polar coordinate sector density peak algorithm. When it detects the presence of a photoelectric pod feature (reflection point density reaching 1.5 times the global average density threshold) directly in front of the AGV, it automatically triggers AVOID_FRONT (collision avoidance and detour): controlling the chassis to move laterally for 5 seconds to bypass the pod projection area, then moving forward to the rear of the helicopter, and automatically flipping the detour direction to prevent dead loops.

[0103] S2-4, Enter:

[0104] After centering and convergence, the system uses open-loop straight-line control to latch the current heading and distance parameters and maneuver under the fuselage. A dual protection threshold of 1.0m hard stop distance and at least 20 points within the RANSAC fitting area is set to prevent collisions.

[0105] Step 3: High-precision docking of the fuselage (Phase 3)

[0106] When the AGV enters directly beneath the helicopter's belly, the radar signal is lost due to self-blocking. The system then uses an anomaly-tolerant module to check normal operating conditions: if the visual reference marker is unobstructed, the ambient illuminance is not less than 50 lux, and the marker is within the 0.1–5.0 m detection range of the D455 depth camera (e.g., ...). Figure 3 As shown in the figure, the system automatically switches to precision alignment control dominated by visual reference markers.

[0107] S3-1, Depth Back-Calculation:

[0108] To eliminate depth blur in the solvePnP algorithm when the marker surface is tilted, the precision alignment control module uses a pixel size inverse calculation formula with tilt correction to solve for the actual distance:

[0109]

[0110] in, , The actual physical side length marked for visual reference. The pixel dimensions and side lengths in the image are used as visual reference markers. Let be the angle between the normal vector and the camera's optical axis. These represent the pixel lengths of the upper and lower edges of the visual reference markers in the image, and the actual pixel focal length. The actual value of the camera calibration matrix [0,0], thereby directly eliminating the 15% to 55% depth ranging error caused by the 30° to 50° tilt angle, and transforming the relative coordinates to the chassis coordinate system through the URDF extrinsic rotation matrix;

[0111] S3-2, Multi-axis decoupling control:

[0112] The precision alignment control module performs strict six-axis decoupled sequential control, namely: yaw angle alignment (YAW, which uses the double marker line to estimate the heading deviation and control it to converge to within 0.008 rad). Longitudinal alignment (LON, controlling longitudinal deviation to converge within 3mm) Lateral alignment (LAT) is performed to control lateral deviation to within 3mm, such as... Figure 6 (As shown) Three-axis residual monitoring (CHECK). In any sub-state, a non-zero velocity command is issued only along a single degree of freedom at the same time. In CHECK state, if residual drift exceeds the limit, the system automatically re-enters the corresponding correction state. When all three-axis residuals are stable for 10 consecutive frames, the docking is considered complete.

[0113] S3-3, Cascaded rollback control:

[0114] During the execution of step three, the fault tolerance module continuously monitors the locking status of the visual reference markers.

[0115] If the visual reference marker is continuously lost for more than 1 second, or if the total alignment execution time in step three exceeds 60 seconds and still fails to converge, the system determines that an anomaly has occurred and immediately triggers a cascading rollback:

[0116] P1. The onboard mission computer sends an emergency braking command to the flight control system to immediately stop the AGV and records the current position and attitude estimate.

[0117] P2. The local planning module controls the chassis to reverse and move straight out from under the helicopter belly, returning to the open-loop straight-moving waiting-to-enter state in step two (Phase2ENTER), and controls the AGV chassis to execute the straight-moving shuttle again, attempting to enter under the belly for visual precision alignment once more.

[0118] P3. If the system fails to successfully converge and complete the docking after two consecutive retries, the system will perform a safe shutdown outside the fuselage, mark the current task as abnormal, and send an alarm signal to the ground control center, waiting for manual intervention.

[0119] It should be noted that the structure described in this invention can be implemented in many different forms and is not limited to the embodiments described. Any equivalent transformations made by those skilled in the art based on the description and drawings of this invention, or direct or indirect applications in other related technical fields, such as the loading and unloading of other items, are included within the protection scope of this invention.

Claims

1. A landing gear autonomous docking method based on multimodal sensing, characterized in that, Includes the following steps: Step 1, Long-distance coarse alignment stage: Long-distance alignment is performed using a vision sensor and a 3D LiDAR installed at the front end of the AGV; the long-distance coarse alignment stage includes four sequentially connected sub-states: search sub-state, alignment sub-state, fine alignment sub-state, and straight-line sub-state; the AGV is controlled to switch sequentially between the four sub-states to perform heading closure control, and the AGV is controlled to approach in a straight line to the preset close-range switching distance; Step 2, Close-range centerline alignment stage: After the AGV enters the close-range switching distance, the landing gear skid structure is scanned using a 3D LiDAR and multiple spatial filtering of the point cloud is performed. The geometric parameters of the two skid lines are solved by the RANSAC bilinear fitting algorithm and the centerline heading is extracted. The AGV is controlled to perform heading alignment control and lateral deviation convergence control based on the centerline heading, and obstacle detection and collision avoidance are performed during the journey. Step 3, High-precision docking stage under the fuselage: When the AGV drives under the fuselage of the helicopter and the radar signal is lost, the system switches to precision alignment control dominated by the depth camera and visual reference markers; the three-dimensional relative position of the visual reference markers is calculated based on the pixel size back-calculation algorithm with tilt correction, and the AGV chassis is driven to move by a multi-axis decoupled sequential control strategy to complete the high-precision docking, and cascaded backtracking is executed when the preset abnormal judgment conditions are met.

2. The landing gear autonomous docking method based on multimodal sensing as described in claim 1, characterized in that: Step one specifically includes: In the search sub-state, the AGV is controlled to rotate at a constant angular velocity to search for the target. When any vision sensor detects the helicopter pixel bounding box, it automatically switches to the alignment sub-state. In the alignment state, the heading of the AGV is controlled by a proportional-integral-derivative (PID) controller using the center offset of the pixel bounding box of the dual-channel vision sensor; once the center is precisely aligned, the system automatically switches to the precise alignment state. In the precise alignment state, the direction of the centroid of the radar point cloud exponential moving average after median filtering and noise reduction is used to guide the alignment, and the PID controller automatically switches between high and low gain levels according to the azimuth deviation to approach the course; when the alignment is completed, it automatically switches to the straight-line state. In the straight-line sub-state, the AGV is controlled to move in an open-loop straight path to approach the short-distance switching distance.

3. The landing gear autonomous docking method based on multimodal sensing as described in claim 2, characterized in that: The exponential moving average centroid direction is updated in real time only during the precise alignment state, the straight alignment state, and the close centerline alignment stage, and remains locked during the search and alignment states to suppress interference from distant sparse noise points on the heading estimation.

4. The landing gear autonomous docking method based on multimodal sensing as described in claim 1, characterized in that: Step two, specifically the multi-spatial filtering of the point cloud data, includes: A height-direction interval filter is used to filter out noise outside the area at the bottom of the landing gear skids; Set an exclusion rectangle for the AGV body to filter out interference points reflected by the AGV body itself; A cone-shaped filtering region is defined around the current exponential moving average heading angle to filter out environmental clutter outside the cone angle range, thereby obtaining a clean set of skid feature points.

5. The landing gear autonomous docking method based on multimodal sensing as described in claim 1, characterized in that: Step two, specifically, involves solving for the geometric parameters of the two skid lines and the heading alignment control: The serialized RANSAC bilinear fitting algorithm with a set number of iterations was used to fit the left and right sliding lines respectively. Calculate the product of the normal vectors of the slopes of the two sliding lines. If the product is less than zero, then perform an angle flip correction on the second fitted line. Using the centerline of the two skid lines as the target heading, heading alignment control is performed using the direction vector accumulation criterion.

6. The landing gear autonomous docking method based on multimodal sensing as described in claim 1, characterized in that: Step two specifically includes: obstacle detection and collision avoidance. The density of laser reflection points is obtained by using a polar coordinate sector density peak algorithm. When the detected reflection point density exceeds the set global average density threshold, it is determined that an airborne pod structure has been detected. If the airborne pod structure is located directly in front of the AGV, the AGV is controlled to first move laterally for a set time to avoid the projection area, and then travel forward to a preset distance to circle around to the rear of the helicopter.

7. The landing gear autonomous docking method based on multimodal sensing as described in claim 1, characterized in that: In step three, the pixel size inverse estimation algorithm based on tilt angle correction specifically includes: When the visual reference marker and the optical axis of the depth camera are tilted at an angle, the actual depth distance is calculated using the following corrected depth formula. : in, This represents the actual pixel focal length value of the camera. The actual physical side length marked for visual reference. The pixel dimensions and side lengths in the image are used as visual reference markers. The angle between the normal vector and the camera's optical axis; The included angle The estimation is based on the ratio of the upper and lower edge pixel lengths of the visual reference marked image, and the estimation formula is as follows: in, These are the pixel lengths of the upper and lower edges of the visual reference markers in the image, respectively.

8. The landing gear autonomous docking method based on multimodal sensing as described in claim 1, characterized in that: In step three, the multi-axis decoupling sequence control strategy specifically includes: The control is decoupled and executed in the order of yaw angle alignment control, longitudinal deviation correction control, lateral deviation correction control, and three-axis residual monitoring control, and at any given time, a non-zero control speed command is output along only a single degree of freedom. In the three-axis residual monitoring and control, if the drift in any dimension exceeds the set threshold, the corresponding correction control step will be automatically re-entered for re-alignment. When all three-axis residuals are continuously and stably stable for a preset number of frames, the docking is determined to be complete.

9. The landing gear autonomous docking method based on multimodal sensing as described in claim 1, characterized in that: Step three, specifically, includes: When the visual reference marker is continuously lost for a period exceeding a first preset time threshold, or when the total execution time of the high-precision docking phase of the fuselage exceeds a second preset time threshold, the system triggers a cascading rollback: Control the AGV to immediately apply emergency braking and record the current pose estimate; Control the AGV to reverse and move straight out from under the helicopter's belly, returning to the straight-moving waiting-to-enter state of step two; Control the AGV to resume open-loop straight movement and attempt to enter step three again.

Images