Unmanned aerial vehicle dynamic platform autonomous landing control method and system based on depth vision perception and interactive multi-model prediction

Abstract

The application relates to a kind of unmanned aerial vehicle dynamic platform autonomous landing control method and system based on depth vision perception and interactive multi-model prediction, belong to unmanned aerial vehicle autonomous navigation and computer vision field.The method and system pass through vision perception module and utilize YOLO11 algorithm to carry out real-time target detection and continuous tracking to mobile landing platform;Trajectory prediction module introduces interactive multi-model filter, and the motion trajectory of strong maneuvering platform is adaptively estimated and accurately predicted;Flight control module inputs predicted trajectory into model predictive control framework, and calculates the optimal flight control instruction of unmanned aerial vehicle at present;Landing execution module adjusts rotor speed according to control instruction, and completes attitude adjustment and safe landing.The application realizes the cooperative work of vision recognition and adaptive motion prediction, significantly overcomes the tracking error caused by system delay, and improves the landing success rate and autonomous control level of unmanned aerial vehicle in complex dynamic environment.

Description

Technical Field

[0001] This invention belongs to the field of autonomous navigation and computer vision for unmanned aerial vehicles (UAVs), and relates to an autonomous landing control method and system for a dynamic UAV platform based on depth visual perception and interactive multi-model prediction. Background Technology

[0002] With the rapid development of the low-altitude economy and the continuous advancement of UAV autonomous control technology, UAVs are increasingly being used in logistics delivery, emergency rescue, agricultural plant protection, and maritime operations. In future fully automated operation scenarios, UAVs will often need to frequently land on moving vehicles, ships, and other dynamic landing platforms to complete material resupply or data retrieval. However, existing UAV landing control systems still have significant shortcomings in terms of visual perception robustness and trajectory tracking accuracy when facing complex maneuvering targets, making it difficult to meet the actual needs of safe and precise landing in highly dynamic environments.

[0003] Currently, mainstream UAV dynamic landing schemes mainly rely on traditional visual markers (such as ArUco codes) and basic filtering algorithms, which have significant limitations. On the one hand, recognition relying solely on traditional visual markers is easily affected by environmental factors such as sudden changes in lighting, complex backgrounds, or partial target occlusion, leading to the easy loss and difficulty in re-capturing visual features, resulting in weak perceptual robustness of the system. On the other hand, in the target trajectory prediction and control stage, most existing systems employ a single Kalman filter and a classic feedback control strategy. This motion prediction model exhibits severe prediction lag when facing strong maneuvers such as acceleration, deceleration, or sharp turns of the landing platform. This not only fails to effectively compensate for the system delay caused by visual sensing in acquiring pose but also easily leads to deviations in the UAV's flight trajectory, and may even result in landing failure and collisions.

[0004] Utilizing deep convolutional neural networks for target detection and correlation tracking can extract more discriminative image features, significantly improving the stability of target recognition in complex backgrounds. Meanwhile, predictive control theory has significant advantages in handling multiple dynamic constraints and nonlinear systems. However, research on the closed-loop integration of visual perception and multi-mode adaptive prediction in scenarios where UAVs land on highly maneuverable dynamic platforms is still in its infancy.

[0005] Most existing technical solutions adopt a modular, independent design approach, failing to construct a truly adaptive prediction and control mechanism to cope with strong maneuverability. For example, some systems only improve the accuracy of visual detection but still use basic trajectory prediction models, which cannot adaptively match the sudden maneuvering modes of a dynamic landing platform, leading to prediction failure. Other systems focus on improving control algorithms but fail to consider trajectory reconstruction and prediction compensation when visual perception is briefly lost. The lack of effective linkage between the perception, prediction, and control modules results in weak adaptability of existing algorithms in real-world dynamic scenarios. They often only achieve basic landings under low-speed or uniform-speed conditions, making it difficult to achieve high-precision, robust, and accurate landings under complex platform maneuvering conditions.

[0006] Therefore, it is necessary to propose a dynamic landing system for UAVs that integrates visual detection, trajectory prediction and optimized control to improve the landing accuracy and system stability of UAVs in complex dynamic environments. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide an autonomous landing control method and system for a UAV dynamic platform based on deep visual perception and interactive multi-model prediction. By constructing a closed-loop control framework that integrates deep learning target detection, multi-target tracking and motion prediction, and using an interactive multi-model filter to adaptively estimate and predict the strong maneuver trajectory of the dynamic landing platform, the UAV can still accurately and robustly estimate the motion state of the landing platform in high-dynamic and strong interference scenarios, supporting it to complete a safe, stable and high-precision autonomous landing, so as to achieve the goal of reliable recovery and application on the dynamic landing platform.

[0008] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, an autonomous landing control method for a dynamic UAV platform based on depth visual perception and interactive multi-model prediction is provided, the method including: S1. Extract the bounding box information of the dynamic landing platform from the continuous image frames acquired by the UAV, and then combine the appearance re-identification features with the motion state to associate the previous and next frames with the multi-target tracking algorithm. Based on the camera intrinsic parameters and UAV attitude, back-project the bounding box pixel coordinates to the body coordinate system to obtain the three-dimensional position estimate of the dynamic landing platform and output the continuous observation state sequence of the dynamic landing platform. S2. Based on the acquired continuous observation state sequence, the trajectory of the dynamic landing platform is predicted. In this process, through an interactive multi-model filtering mechanism, the matching probability weights of multiple dynamic models reflecting the motion characteristics of the dynamic landing platform and the current observation state sequence are calculated in real time, and the local estimation results of each model are weighted and fused. S3. Using the predicted trajectory information of the acquired dynamic landing platform as a reference desired trajectory, the final control command sequence is obtained by rolling optimization within the prediction time series through the model predictive control algorithm, combined with the nonlinear dynamic model of the UAV and its safety constraints. S4. Adjust the rotor speed of the UAV according to the control command sequence, and monitor the relative height and relative speed between the UAV and the dynamic landing platform in real time. When both are below the preset safety threshold and remain below the set time, trigger the landing lock command.

[0009] Furthermore, in the dynamic landing platform bounding box information extraction process of step S1, the YOLO11 network is used to extract image features and detect targets. Multi-scale feature maps of the image are obtained through the backbone feature extraction module, and bounding box regression prediction is performed in the detection head. The formula for predicting the coordinates of the bounding box center is as follows:

[0010] in, and For the predicted center coordinates, and This is the offset of the network output. and The coordinates of the top left corner of the grid. This represents the Sigmoid activation function.

[0011] Furthermore, in the multi-target tracking algorithm and coordinate transformation process of step S1, a weighted cascaded matching mechanism of Mahalanobis distance and cosine distance is used for target association. The distance metric formula is as follows:

[0012] in, To comprehensively consider the associated costs, The Mahalanobis distance is used to characterize the state of motion. The cosine distance is used to characterize the similarity of appearance features. These are weighting coefficients; After considering the related costs After confirming the target bounding box, set the center pixel coordinates of the target bounding box. By combining the camera's focal length, pixel size, and the drone's current flight altitude, a pinhole camera model is used for back projection to obtain the three-dimensional position of the landing platform in the camera coordinate system. Then, through rotation and translation transformation using the extrinsic parameter matrix, a three-dimensional observation vector is constructed to the body coordinate system.

[0013] Furthermore, in step S2, during adaptive state estimation using the interactive multi-model filtering mechanism, in each filtering iteration, the mixture probability is first calculated based on the Markov transition probability matrix, and its formula is:

[0014] in, For the model To the model The mixed probability of the transition, These are the preset model state transition probabilities. Model of the previous time step The probability, The total number of models; Then, each filter is updated in parallel, and state estimation is fused. The fusion formula is as follows:

[0015] in, express The target state estimation vector after time-mapping fusion Indicates in Time of the first The model probability of a dynamic model Indicates the time of the first time. The local state estimation vector output by each dynamic model.

[0016] Furthermore, in step S2, the dynamic model reflecting the motion characteristics of the dynamic landing platform includes at least a uniform velocity model, a uniform acceleration model, and a coordinated turning model, covering the typical maneuvering modes of the dynamic landing platform.

[0017] Furthermore, in step S3, the optimal control command sequence for the UAV is generated by solving a constrained optimization problem, the objective cost function of which is expressed as follows:

[0018] in, Let cost function be To predict the time domain, To control the time domain, and They represent in Predicting the future at any moment The system state vector and reference state vector of the step, Indicates the control increment sequence. Here is the state error weight matrix. To control the incremental weight matrix; The control input vector is defined as:

[0019] in, for The first time predicted Step control input vector, For the total thrust of the drone, These are the expected roll angle and pitch angle reference values ​​for the inner attitude ring, respectively. The optimization solution process is constrained by the physical constraints of the UAV, including the maximum attitude angle constraint, maximum lift constraint, and control variable rate of change constraint, among which:

[0020]

[0021] in, These represent the lower and upper bounds of each component of the control input vector, respectively. These are the lower and upper bounds of each component of the control increment within adjacent control cycles, respectively.

[0022] On the other hand, a system is also provided for implementing the aforementioned autonomous landing control method for a dynamic UAV platform based on depth visual perception and interactive multi-model prediction. This system includes: The visual perception module is used to acquire images from the airborne camera and output the observation status of the dynamic landing platform in real time. The trajectory prediction module, connected to the visual perception module, is used to process the observation state, adaptively estimate the platform maneuver mode, and output future trajectory prediction information. The flight control module, connected to the trajectory prediction module, is used to perform rolling optimization calculations based on future trajectory prediction information and generate UAV control commands. The landing execution module, connected to the flight control module, is used to execute the UAV control commands to adjust the rotor speed, thereby achieving attitude adjustment and safe landing of the UAV.

[0023] Furthermore, the visual perception module includes: The target detection unit uses the YOLO11 network model to extract spatial features from the image and outputs the target bounding box and confidence score. The target association unit uses the StrongSORT algorithm to extract target appearance re-identification features and combines the algorithm's built-in Kalman filter to smooth and match target trajectories between frames.

[0024] Furthermore, the model predictive control algorithm in the flight control module integrates the nonlinear kinematics and dynamics model of the quadcopter UAV; when performing rolling optimization, its constraints include the maximum attitude angle constraint, maximum lift constraint, and control variable rate of change constraint of the UAV.

[0025] Furthermore, it is equipped with a landing determination unit, which is used to monitor the relative altitude and relative speed between the drone and the landing platform in real time. When the relative altitude and relative speed are both less than their respective preset safety thresholds and remain so for a set time, the final landing lock command is triggered.

[0026] The beneficial effects of this invention are as follows: (1) This invention can effectively improve the problem of trajectory tracking lag and easy failure when UAVs land on complex maneuvering platforms. By constructing a closed-loop architecture of "visual perception-multi-model prediction-rolling optimization control", the system enables the UAV to actively predict the future movement trend of the platform instead of simply passively following the current observation position. This overcomes the system delay caused by the processing time of the visual sensor, so that the UAV can still maintain smooth and accurate tracking when facing the platform's rapid acceleration, deceleration or continuous turning, which greatly improves the success rate and safety of landing in high dynamic environments.

[0027] (2) This invention possesses strong robustness and anti-interference capabilities at the visual perception level. Compared to traditional visual marking methods that are susceptible to illumination and occlusion, this invention employs a YOLO11 network structure combined with the StrongSORT target tracking algorithm. Relying on dual measurements of appearance and motion features, even under adverse conditions such as brief occlusion of the landing platform, drastic attitude changes, or complex backgrounds, the system can still stably maintain continuous target locking, providing a stable and reliable observation data source for subsequent prediction and control.

[0028] (3) This invention introduces an interactive multi-model mechanism, enabling the system to adaptively estimate the strong maneuvering behavior of a dynamic landing platform. Traditional single-filter algorithms often produce divergence or large errors when the target undergoes sudden maneuvering changes. However, this invention utilizes parallel computation of multiple motion models and probability transfer of Markov chains to automatically evaluate and match the platform's current motion mode. This adaptive soft-switching mechanism effectively improves the prediction accuracy for complex nonlinear trajectories.

[0029] (4) Regarding the control strategy, this invention employs a model predictive control algorithm, which ensures the global optimality of control commands while satisfying multiple flight safety constraints. When generating flight control commands, this algorithm directly incorporates physical limitations such as the UAV's maximum attitude angle, lift limit, and rate of change of control variables into the rolling optimization solution process. This not only ensures that attitude commands are always within the flight safety constraints, but also effectively avoids severe UAV shaking caused by sudden changes in control commands, achieving a smooth approach and final landing.

[0030] (5) This invention has good engineering application value. The fully autonomous landing method does not require the installation of complex communication base stations or high-precision differential GPS equipment on the dynamic landing platform. It can complete the fully autonomous closed loop solely with the onboard computing power and visual sensors of the UAV itself. This provides effective technical support for UAVs in complex dynamic operation scenarios where prior trajectories cannot be provided, such as supplying materials to ships at sea, autonomously landing on top of moving vehicles, and emergency rescue.

[0031] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0032] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is an overall flowchart of the autonomous landing control method for a dynamic UAV platform according to an embodiment of the present invention; Figure 2 This is an overall architecture diagram of the autonomous landing control system for a UAV dynamic platform according to an embodiment of the present invention; Figure 3 This is a diagram illustrating the visual perception and target association architecture according to an embodiment of the present invention; Figure 4 This is a flowchart of the interactive multi-model prediction process according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the rolling optimization principle of model predictive control according to an embodiment of the present invention. Detailed Implementation

[0033] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0034] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0035] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0036] Please see Figures 1-5 This invention relates to an autonomous landing control method and system for a dynamic UAV platform based on depth visual perception and interactive multi-model prediction.

[0037] Example 1 This embodiment provides an autonomous landing method for a dynamic UAV platform based on depth visual perception and interactive multi-model prediction. For example... Figure 1 As shown, the specific implementation process includes the following steps: Step 1-1: As the UAV approaches the dynamic landing platform, it acquires a continuous sequence of images of the mission scene using its onboard high-definition camera. The image sequence is input into a pre-trained deep learning network to extract platform features and output an initial two-dimensional bounding box. Subsequently, combined with a multi-target tracking algorithm, the appearance re-identification features are correlated with the motion state to eliminate false detections. Furthermore, using the camera intrinsic parameter matrix and the UAV's current attitude, the center pixel coordinates of the bounding box are back-projected to the body coordinate system through a gimbal camera model to obtain a three-dimensional position estimate of the platform, outputting a continuous and stable sequence of observations of the landing platform.

[0038] Steps 1-2: After coordinate system transformation, the acquired continuous observation state sequence is input into the trajectory prediction module. This module internally constructs multiple dynamic models reflecting the typical motion characteristics of the platform. Through an interactive multi-model (IMM) filtering mechanism, the probability weights of each model matching the current observation data are calculated in real time, and the local estimation results are weighted and fused to adaptively output accurate trajectory prediction information for the platform over a future period, effectively compensating for the system delay caused by visual sensing.

[0039] Steps 1-3: The future trajectory prediction information output by the IMM filter is used as the reference desired trajectory and fed into the model predictive control framework of the flight control module. The MPC algorithm combines the nonlinear dynamic model of the quadrotor UAV with safety constraints such as attitude angle and rate of change of control variables to perform rolling optimization within the set prediction time domain, and calculates the optimal control command sequence.

[0040] Steps 1-4: The landing execution module adjusts the rotor speed according to the flight control command to complete the attitude adjustment and safe landing of the UAV; and the landing determination unit monitors the relative altitude and relative speed between the UAV and the landing platform in real time. When both are below the preset safety threshold and remain below it for a set time, the landing lock command is triggered.

[0041] In step 1-1 of this embodiment, the multi-target tracking algorithm uses a dual measurement mechanism of appearance features and motion state to associate data, so as to update the target trajectory and suppress detection loss caused by visual occlusion.

[0042] In steps 1-2 of this embodiment, the final state estimation fusion formula of the cross-mode multi-model filter is:

[0043] in, express The target state estimation vector after time-mapping fusion Indicates in Time of the first The model probability of a motion model. Indicates the time of the first time. The local state estimation vector output by each motion model.

[0044] In steps 1-3 of this embodiment, the model predictive control framework generates control commands by solving for the optimal solution of the objective cost function, the expression of which is:

[0045] in, Let cost function be To predict the time domain, To control the time domain, and They represent in Predicting the future at any moment The system state vector and reference state vector of the step, Indicates the control increment sequence. Here is the state error weight matrix. To control the incremental weight matrix.

[0046] Example 2 This embodiment provides a dynamic landing control system for unmanned aerial vehicles (UAVs) to implement the method, such as... Figure 2 As shown, the system includes: The visual perception module is used to acquire images from the airborne camera and output the observation status of the dynamic landing platform in real time.

[0047] The trajectory prediction module, connected to the visual perception module, is used to process the observation status, adaptively estimate the platform's maneuvering mode, and output future trajectory prediction information.

[0048] The flight control module, connected to the trajectory prediction module, is used to perform rolling optimization calculations based on future trajectory prediction information to generate UAV control commands.

[0049] The landing execution module, connected to the flight control module, is used to execute UAV control commands to adjust the rotor speed, thereby enabling the UAV to adjust its attitude and land safely.

[0050] The visual perception module includes: The target detection unit uses the YOLO11 network model to extract spatial features from the image and outputs the target bounding box and confidence score.

[0051] The target association unit uses the StrongSORT algorithm to extract target appearance re-identification features and combines the algorithm's built-in Kalman filter to smooth and match target trajectories between frames.

[0052] The interactive multi-model filter in the trajectory prediction module includes at least a constant velocity (CV) model, a constant acceleration (CA) model, and a coordinated turn (CT) model to cover the typical maneuvering patterns of a dynamic landing platform.

[0053] The model predictive control algorithm in the flight control module integrates the nonlinear kinematics and dynamics model of the quadcopter UAV.

[0054] When the flight control module performs rolling optimization, the constraints include the maximum attitude angle constraint, the maximum lift constraint, and the control variable change rate constraint of the UAV.

[0055] The system is also equipped with a landing determination unit, which is used to monitor the relative altitude and relative speed between the drone and the landing platform in real time. When the relative altitude and relative speed are both less than their respective preset safety thresholds and remain so for a set time, the final landing lock command is triggered.

[0056] Example 3 This embodiment provides a more detailed description of the method in Embodiment 1 and the system in Embodiment 2, including the target detection and association process of the visual perception module, such as... Figure 3As shown, the network receives raw image data, extracts features through a convolutional structure to generate multi-scale feature maps, and outputs target bounding boxes via a detection head regression. Subsequently, by extracting feature re-identification vectors and combining them with a distance metric and a Kalman filter, stable tracking of the same target across frames is achieved. The specific steps are as follows: The YOLO11 network is used for image feature extraction and object detection. The network obtains multi-scale feature maps of the image through a backbone feature extraction module and performs bounding box regression prediction in the detection head. The formula for predicting the coordinates of the bounding box center is:

[0057] in, and For the predicted center coordinates, and This is the offset of the network output. and The coordinates of the top left corner of the grid. This represents the Sigmoid activation function. The specific YOLO11 object detection hyperparameter configurations are shown in Table 1. Table 1

[0058] Target association employs the StrongSORT algorithm. The bounding boxes output by YOLO11 are used as input to extract the appearance re-identification (ReID) feature vectors of the targets. Inter-frame data association uses a weighted concatenated matching mechanism of Mahalanobis distance and cosine distance, with the distance metric formula as follows:

[0059] in, To comprehensively consider the associated costs, The Mahalanobis distance is used to characterize the state of motion. The cosine distance is used to characterize the similarity of appearance features. These are the weighting coefficients. The specific StrongSORT target tracking associated hyperparameter configurations are shown in Table 2: Table 2

[0060] After considering the related costs After confirming the target bounding box, set its center pixel coordinates. By combining the camera's focal length, pixel size, and the drone's current flight altitude, a pinhole camera model is used for back projection to obtain the three-dimensional position of the landing platform in the camera coordinate system. Then, through rotation and translation transformation using the extrinsic parameter matrix, a three-dimensional observation vector is constructed, which serves as the input for the subsequent interactive multi-model filter.

[0061] Comprehensive related costs By using distance metrics across multiple candidate bounding boxes in the current frame, it is determined which bounding box corresponds to the historically tracked target; the identity has already been determined through comprehensive association cost. Once confirmed, the pixel coordinates of the target bounding box are used, combined with camera intrinsics, pose, and height, to perform pure geometric calculations to obtain the 3D position. In other words, the bounding box used for back projection is precisely the result of comprehensive correlation cost calculations. The target box after association confirmation.

[0062] Example 4 In this embodiment, to achieve accurate prediction of the strong maneuver trajectory of a dynamic landing platform, the present invention employs an interactive multi-model filter for adaptive state estimation. For example... Figure 4 As shown, the input observation sequence first calculates the mixture probability and performs state interaction through the Markov transition matrix, and then feeds it into multiple parallel filters such as uniform speed, uniform acceleration, and coordinated turning for state update. Finally, the weights of each model are dynamically adjusted according to the likelihood probability matched by the model, and the weighted fusion outputs the optimal estimate of the current state and the predicted trajectory for the future.

[0063] In each filtering iteration, the mixing probability is first calculated based on the Markov transition probability matrix, and the formula is as follows:

[0064] in, For the model To the model The mixed probability of the transition, These are the preset model state transition probabilities. Model of the previous time step The probability, This represents the total number of models.

[0065] Then, each filter is updated in parallel, and state estimation is fused. The fusion formula is as follows:

[0066] in, express The target state estimation vector after time-mapping fusion Indicates in Time of the first The model probability of a motion model. Indicates the time of the first time. The local state estimation vectors output by each motion model. The specific IMM prediction hyperparameters are shown in Table 3. Table 3

[0067] Example 5 In this embodiment, the predicted desired trajectory information is input to the flight control module. The model predictive control algorithm generates the optimal control command sequence for the UAV by solving a constrained optimization problem. For example... Figure 5 As shown, at each sampling moment, the control system uses the current state as a starting point and calculates the state evolution over a future period using the UAV prediction model. By minimizing the error cost function between the desired trajectory and the predicted trajectory, a control sequence is solved, and the first step of this sequence is used as the actual control quantity applied to the UAV to achieve closed-loop rolling optimization control. The objective cost function expression is:

[0068] in, Let cost function be To predict the time domain, To control the time domain, and They represent in Predicting the future at any moment The system state vector and reference state vector of the step, Indicates the control increment sequence. Here is the state error weight matrix. To control the incremental weight matrix.

[0069] The control input vector is defined as: U(k+i|k) = [T, φ_ref, θ_ref]

[0070] Where T is the total thrust (total lift) of the UAV, and φ_ref and θ_ref are the expected roll angle and pitch angle reference values ​​of the inner attitude loop, respectively.

[0071] Detailed explanations of the three variables: ① U(k+i|k): The control input vector predicted at time k for the (k+i)th step, i.e., the vector composed of [total thrust T, desired roll angle φ_ref, desired pitch angle θ_ref]; ② U_min, U_max: control the lower and upper bounds of each component of the input vector, where the upper and lower bounds of the T component correspond to the "maximum lift constraint", and the upper and lower bounds of the φ_ref and θ_ref components correspond to the "maximum attitude angle constraint". ③ ΔU_min, ΔU_max: The lower and upper bounds of each component of the control increment within adjacent control cycles, corresponding to the "control quantity change rate constraint".

[0072] Therefore, the formula (U_min≤U(k+i|k)≤U_max) simultaneously covers both the maximum lift constraint and the maximum attitude angle constraint, and the formula (ΔU_min≤ΔU(k+i|k)≤ΔU_max) corresponds to the control variable change rate constraint. Together, the two formulas fully embody the three constraints mentioned in the invention, without the need for further splitting or supplementing of the formulas.

[0073] Meanwhile, the optimization solution process is constrained by the physical limitations of the UAV:

[0074]

[0075] The specific model predictive control hyperparameters are shown in Table 4: Table 4

[0076] The pseudocode for the UAV visual perception and multi-model prediction landing process in this invention is as follows: image = read_camera(), uav_state = get_uav_state() bboxes = YOLO11_detect(image, conf_thresh=0.65) targets = StrongSORT_update(bboxes, image, ReID_dim=512) if is_empty(targets): cmd = EXECUTE_SAFETY_HOVER(uav_state) else: z_k = extract_observation(targets) probs = calc_markov_mixing(transition_matrix) state_CV = Kalman_Filter_CV(z_k, probs) state_CA = Kalman_Filter_CA(z_k, probs) state_CT = Kalman_Filter_CT(z_k, probs) pred_state = IMM_fusion(state_CV, state_CA, state_CT) ref_traj = generate_ref_traj(pred_state, N_p=20) opt_seq = MPC_optimize(uav_state, ref_traj, Q, R, constraints) cmd = extract_first_action(opt_seq) if calc_relative_error(uav_state, z_k) <safe_thresh:cmd = EXECUTE_LANDING_LOCK() cmd = safety_filter_smoothing(cmd) send_to_flight_controller(cmd) Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An autonomous landing control method for a dynamic UAV platform based on depth visual perception and interactive multi-model prediction, characterized in that: The method includes: S1. Extract the bounding box information of the dynamic landing platform from the continuous image frames acquired by the UAV, and then combine the appearance re-identification features with the motion state to associate the previous and next frames with the multi-target tracking algorithm. Based on the camera intrinsic parameters and UAV attitude, back-project the bounding box pixel coordinates to the body coordinate system to obtain the three-dimensional position estimate of the dynamic landing platform and output the continuous observation state sequence of the dynamic landing platform. S2. Based on the acquired continuous observation state sequence, the trajectory of the dynamic landing platform is predicted. In this process, through an interactive multi-model filtering mechanism, the matching probability weights of multiple dynamic models reflecting the motion characteristics of the dynamic landing platform and the current observation state sequence are calculated in real time, and the local estimation results of each model are weighted and fused. S3. Using the predicted trajectory information of the acquired dynamic landing platform as a reference desired trajectory, the final control command sequence is obtained by rolling optimization within the prediction time series through the model predictive control algorithm, combined with the nonlinear dynamic model of the UAV and its safety constraints. S4. Adjust the rotor speed of the UAV according to the control command sequence, and monitor the relative height and relative speed between the UAV and the dynamic landing platform in real time. When both are below the preset safety threshold and remain below the set time, trigger the landing lock command.

2. The autonomous landing control method for a UAV dynamic platform based on depth visual perception and interactive multi-model prediction according to claim 1, characterized in that: In step S1, during the extraction of the bounding box information of the dynamic landing platform, a YOLO11 network is used to extract image features and detect targets. Multi-scale feature maps of the image are obtained through the backbone feature extraction module, and bounding box regression prediction is performed in the detection head. The formula for predicting the coordinates of the bounding box center is as follows: in, and For the predicted center coordinates, and This is the offset of the network output. and The coordinates of the top left corner of the grid. This represents the Sigmoid activation function.

3. The autonomous landing control method for a UAV dynamic platform based on depth visual perception and interactive multi-model prediction according to claim 2, characterized in that: In the multi-target tracking algorithm and coordinate transformation process in step S1, a weighted concatenated matching mechanism of Mahalanobis distance and cosine distance is used for target association. The distance metric formula is as follows: in, To comprehensively consider the associated costs, The Mahalanobis distance is used to characterize the state of motion. The cosine distance is used to characterize the similarity of appearance features. These are weighting coefficients; After considering the related costs After confirming the target bounding box, set the center pixel coordinates of the target bounding box. By combining the camera's focal length, pixel size, and the drone's current flight altitude, a pinhole camera model is used for back projection to obtain the three-dimensional position of the landing platform in the camera coordinate system. Then, through rotation and translation transformation using the extrinsic parameter matrix, a three-dimensional observation vector is constructed to the body coordinate system.

4. The autonomous landing control method for a UAV dynamic platform based on depth visual perception and interactive multi-model prediction according to claim 1, characterized in that: In step S2, during adaptive state estimation using the interactive multi-model filtering mechanism, the mixture probability is first calculated based on the Markov transition probability matrix in each filtering iteration. The formula is as follows: in, For the model To the model The mixed probability of the transition, These are the preset model state transition probabilities. Model of the previous time step The probability, The total number of models; Then, each filter is updated in parallel, and state estimation is fused. The fusion formula is as follows: in, express The target state estimation vector after time-mapping fusion Indicates in Time of the first The model probability of a dynamic model Indicates the time of the first time. The local state estimation vector output by each dynamic model.

5. The autonomous landing control method for a UAV dynamic platform based on depth visual perception and interactive multi-model prediction according to claim 4, characterized in that: In step S2, the dynamic model reflecting the motion characteristics of the dynamic landing platform includes at least a uniform velocity model, a uniform acceleration model, and a coordinated turning model, covering the typical maneuvering modes of the dynamic landing platform.

6. The autonomous landing control method for a UAV dynamic platform based on depth visual perception and interactive multi-model prediction according to claim 1, characterized in that: In step S3, the optimal control command sequence for the UAV is generated by solving a constrained optimization problem, the objective cost function of which is expressed as follows: in, Let cost function be To predict the time domain, To control the time domain, and They represent in Predicting the future at any moment The system state vector and reference state vector of the step, Indicates the control increment sequence. Here is the state error weight matrix. To control the incremental weight matrix; The control input vector is defined as: in, for The first time predicted Step control input vector, For the total thrust of the drone, These are the expected roll angle and pitch angle reference values ​​for the inner attitude ring, respectively. The optimization solution process is constrained by the physical constraints of the UAV, including the maximum attitude angle constraint, maximum lift constraint, and control variable rate of change constraint, among which: in, These represent the lower and upper bounds of each component of the control input vector, respectively. These are the lower and upper bounds of each component of the control increment within adjacent control cycles, respectively.

7. A system for executing the autonomous landing control method for a dynamic UAV platform based on depth vision perception and interactive multi-model prediction as described in any one of claims 1-6, characterized in that: The system includes: The visual perception module is used to acquire images from the airborne camera and output the observation status of the dynamic landing platform in real time. The trajectory prediction module, connected to the visual perception module, is used to process the observation state, adaptively estimate the platform maneuver mode, and output future trajectory prediction information. The flight control module, connected to the trajectory prediction module, is used to perform rolling optimization calculations based on future trajectory prediction information and generate UAV control commands. The landing execution module, connected to the flight control module, is used to execute the UAV control commands to adjust the rotor speed, thereby achieving attitude adjustment and safe landing of the UAV.

8. The autonomous landing control system for a UAV dynamic platform based on depth vision perception and interactive multi-model prediction according to claim 7, characterized in that: The visual perception module includes: The target detection unit uses the YOLO11 network model to extract spatial features from the image and outputs the target bounding box and confidence score. The target association unit uses the StrongSORT algorithm to extract target appearance re-identification features and combines the algorithm's built-in Kalman filter to smooth and match target trajectories between frames.

9. The autonomous landing control system for a UAV dynamic platform based on depth vision perception and interactive multi-model prediction according to claim 7, characterized in that: The model predictive control algorithm in the flight control module integrates the nonlinear kinematics and dynamics model of the quadcopter UAV. When performing rolling optimization, the constraints include the maximum attitude angle constraint, the maximum lift constraint, and the control variable change rate constraint of the UAV.

10. The autonomous landing control system for a UAV dynamic platform based on depth vision perception and interactive multi-model prediction according to claim 7, characterized in that: It is also equipped with a landing determination unit, which is used to monitor the relative altitude and relative speed between the drone and the landing platform in real time. When the relative altitude and relative speed are both less than their respective preset safety thresholds and remain so for a set time, the final landing lock command is triggered.

Images