Unmanned aerial vehicle autonomous landing system and method suitable for unattended scene

Abstract

The invention discloses an unmanned aerial vehicle autonomous landing system and method suitable for an unattended scene, and belongs to the field of unmanned aerial vehicles, and the method comprises the steps: carrying out the staged landmark recognition, which comprises high-altitude global recognition and low-altitude mark tracking; performing dynamic pose estimation, including high-altitude angular point matching and low-altitude mark analysis; environment adaptive flight control, including multi-sensor data fusion and near-earth landing control; and intelligent emergency processing, including abnormity monitoring and dynamic decision making. According to the method, the accuracy of landmark recognition and the precision of pose estimation are improved by adopting a staged landmark recognition and pose estimation method, and the requirement for height change in the autonomous landing process of the unmanned aerial vehicle and the requirement for landing of the unmanned aerial vehicle in an unattended scene are met. The reliability and safety of vertical landing and autonomous landing of the unmanned aerial vehicle are improved. According to the invention, various emergencies can be dealt with in real time, and property losses such as unmanned aerial vehicle damage and cargo damage caused by a landing problem are reduced.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) technology, specifically relating to an autonomous landing system and method for UAVs in unattended scenarios. Background Technology

[0002] In recent years, drone technology has developed rapidly and has been widely used in many fields such as aerial photography, logistics, surveying and mapping, agriculture, and emergency rescue. Its market size continues to expand, and the requirements for the autonomous operation capabilities of drones are also increasing.

[0003] Autonomous landing of drones is a crucial step for their safe return after completing their missions. This is especially true in unmanned airport environments, where the lack of human intervention places higher demands on the accuracy, reliability, and stability of autonomous landing.

[0004] Existing autonomous drone landing technologies partially rely on GPS for navigation and landing. However, GPS accuracy drops significantly in complex terrain or under signal interference, making it difficult to meet the landing requirements of unmanned airports. Furthermore, some landmark-based autonomous drone landing technologies exhibit inconsistent accuracy in landmark recognition and pose estimation as the drone's altitude changes. At higher altitudes, it is difficult to clearly identify landmarks for accurate pose estimation; at lower altitudes, the identification of specific markings on landmarks is easily affected by environmental factors, and there is a lack of effective guidance methods for autonomous drone landing at very close to the ground, impacting the safety and accuracy of autonomous drone landing. Summary of the Invention

[0005] The purpose of this invention is to provide an autonomous landing system and method for unmanned aerial vehicles (UAVs) in unattended scenarios, providing key technical support for autonomous operation of UAVs in unattended scenarios.

[0006] The technical solution adopted by this invention to solve the technical problem is as follows: This invention provides an autonomous landing method for unmanned aerial vehicles (UAVs) in unattended scenarios, which mainly includes the following steps: Step S1: Phased landmark identification, including high-altitude global identification and low-altitude marker tracking; Step S2: Dynamic pose estimation, including high-altitude corner point matching and low-altitude marker resolution; Step S3: Environmental adaptive flight control, including multi-sensor data fusion and near-ground landing control; Step S4: Intelligent emergency response, including anomaly monitoring and dynamic decision-making.

[0007] Furthermore, the specific steps of the high-altitude global recognition are as follows: for the entire landmark with a height of more than 10 meters, high-resolution imaging and feature pyramid algorithm are used to extract the contour and enhance the edge, and extract the key feature information of the landmark.

[0008] Furthermore, the specific steps of the low-altitude marker tracking are as follows: for ArUco markers with an altitude of less than 10 meters, dynamic tracking is performed using a deep learning-based target detection model.

[0009] Furthermore, the specific steps of the high-altitude corner point matching are as follows: for the extracted key feature information of the landmark, mismatches are eliminated by SIFT feature point extraction and RANSAC algorithm, and the six-degree-of-freedom pose parameters are calculated based on the perspective transformation relationship between the three-dimensional coordinates of the landmark corner point and the imaging coordinates of the UAV.

[0010] Furthermore, the specific steps of the low-altitude marker parsing are as follows: the three-dimensional pose of the UAV is directly obtained using the encoded information of the ArUco marker, and Kalman filtering is performed to fuse the data with the IMU inertial measurement data.

[0011] Furthermore, the specific steps of the multi-sensor data fusion are as follows: integrating data from visual sensors, millimeter-wave radar, and barometers to construct a dynamic weighted fusion model and correcting altitude measurement errors in real time.

[0012] Furthermore, the specific steps of the near-ground landing control are as follows: when the UAV's flight altitude is below 20 centimeters, the vertical descent speed is controlled based on the PID parameter self-tuning algorithm, and a soft landing is achieved by combining the preset parameters of the buffer mechanism.

[0013] Furthermore, the specific steps of the anomaly monitoring are as follows: real-time monitoring of landmark recognition confidence and pose estimation variance index through a sliding window algorithm, and setting multi-level thresholds to trigger emergency response.

[0014] Furthermore, the specific steps of the dynamic decision-making are as follows: construct a finite state machine model, and automatically execute strategies such as hovering and waiting, adjusting the drone's flight trajectory to avoid obstacles, or emergency return to home in the event of landmark recognition failure, drone equipment failure, or sudden obstacles, so as to ensure the safety of the drone.

[0015] This invention provides an autonomous landing method for unmanned aerial vehicles (UAVs) in unattended scenarios, which mainly includes the following modules: The landmark recognition module is used to extract the contours and enhance the edges of landmarks when the drone is flying at an altitude of more than 10 meters, and to extract key feature information of landmarks; and to recognize ArUco marks on landmarks when the drone is flying at an altitude of less than 10 meters. The pose estimation module is used to estimate the pose of the UAV based on the key feature information of the landmarks extracted by the landmark recognition module when the UAV is flying at an altitude of more than 10 meters; and to estimate the pose of the UAV directly based on the identified ArUco markers when the UAV is flying at an altitude of less than 10 meters. The flight control module is used to plan the flight trajectory of the UAV based on the pose information obtained by the pose estimation module when the UAV's flight altitude is above 10 meters and below 10 meters, respectively, and guide the UAV to fly smoothly toward the landmark; and to control the UAV to make a vertical landing when the UAV's flight altitude is less than about 20 centimeters and the ArUco marker cannot be detected. The environmental perception module is used to monitor environmental factors in real time during the autonomous landing process of the UAV, and to transmit the monitored environmental information to the landmark recognition module and the flight control module. The emergency response module is used to monitor various emergencies during the autonomous landing process of the drone in real time, and to activate corresponding emergency measures when an emergency is detected.

[0016] The beneficial effects of this invention are: 1. Improved the accuracy of landmark recognition and the precision of pose estimation; This invention employs a phased landmark recognition and pose estimation method, including high-altitude global recognition and low-altitude marker tracking, as well as high-altitude corner point matching and low-altitude marker parsing. By recognizing the entire landmark at a height above 10 meters and recognizing ArUco markers at a height below 10 meters, it is not overly affected by complex environmental factors, thus improving the accuracy of landmark recognition and pose estimation at different altitudes. This meets the needs of altitude changes during UAV autonomous landing and the landing requirements of UAVs in unattended scenarios.

[0017] 2. Improved the safety and accuracy of vertical landing of drones; This invention ensures the safety and accuracy of vertical landing of a drone when its flight altitude is less than approximately 20 centimeters and the ArUco marker cannot be detected, through a vertical landing control strategy.

[0018] 3. Improved the reliability and safety of autonomous landing of drones; This invention integrates an environmental perception module and an emergency response module, enabling the entire autonomous landing system to have high autonomy and adaptability in complex unattended environments. It can respond to various emergencies in real time, improve the safety and reliability of autonomous drone landing, reduce property losses such as drone damage and cargo damage caused by landing problems, and protect the interests of the public when using drone-related services.

[0019] 4. This invention lowers the barrier to drone operation, enabling more people to easily utilize drone technology and improve their lives and work efficiency. At the same time, this invention will help promote the standardization and normalization of the drone industry, facilitate effective supervision of drone flight activities, and ensure the safety of public airspace. Attached Figure Description

[0020] Figure 1 This invention provides a structural block diagram of an autonomous landing system for unmanned aerial vehicles (UAVs) suitable for unattended scenarios.

[0021] Figure 2 The present invention provides a detailed flowchart of an autonomous landing method for unmanned drones in unattended scenarios. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings.

[0023] In a first aspect, the present invention provides an autonomous landing system for unmanned aerial vehicles (UAVs) suitable for unattended scenarios.

[0024] See Figure 1 As shown, the present invention provides an autonomous landing system for unmanned aerial vehicles (UAVs) suitable for unattended scenarios, which mainly includes the following functional modules: a landmark recognition module, a pose estimation module, a flight control module, an environmental perception module, and an emergency response module; the specific functions and roles of each module are as follows: 1) Landmark recognition module; a. In the first stage, when the drone flies at an altitude of more than 10 meters, the outline of the landmark is extracted and the edge is enhanced to extract the key feature information of the landmark, such as the corner information of the landmark.

[0025] b. In the second phase, when the drone flies at an altitude of less than 10 meters, it can accurately identify ArUco markings on landmarks without being overly affected by complex environmental factors.

[0026] 2) Pose estimation module; a. In the first stage, when the drone flies at an altitude of more than 10 meters, pose estimation is performed based on the landmark corner information extracted by the landmark recognition module to calculate the position and attitude of the drone relative to the landmark.

[0027] b. In the second stage, when the UAV is flying at an altitude of less than 10 meters, pose estimation is performed directly based on the ArUco markers identified by the landmark recognition module to obtain more accurate pose data of the UAV relative to the landmarks.

[0028] 3) Flight control module; a. Based on the pose information obtained from the pose estimation module, the flight trajectory of the UAV is planned when the UAV's flight altitude is above 10 meters and below 10 meters, respectively, to guide the UAV to fly smoothly toward the landmark.

[0029] b. When the drone is flying at an altitude of less than approximately 20 centimeters and the ArUco marker cannot be detected, control the drone to make a vertical descent, ensuring a smooth and safe landing process.

[0030] 4) Environmental perception module; a. Real-time monitoring of environmental factors such as light intensity, weather conditions, and ground obstacles during the autonomous landing process of the drone.

[0031] b. The monitored environmental information is transmitted to the landmark recognition module and the flight control module so that these modules can adjust the landmark recognition strategy and flight control parameters according to the environmental conditions.

[0032] 5) Emergency response module; a. Real-time monitoring of various emergencies during the autonomous landing process of the drone, such as landmark recognition failure, drone equipment failure, and sudden obstacles.

[0033] b. Upon detecting an emergency, immediately activate the corresponding emergency measures, such as controlling the drone to hover, adjusting the drone's flight path to avoid obstacles, or making an emergency return to base, to ensure the safety of the drone.

[0034] This invention integrates an environmental perception module and an emergency response module, enabling the entire UAV autonomous landing system to have high autonomy and adaptability in unattended scenarios, and to respond to various emergencies in real time, thereby improving the reliability and safety of UAV autonomous landing.

[0035] Secondly, the present invention provides an autonomous landing method for unmanned aerial vehicles (UAVs) in unattended scenarios. This method is implemented using an autonomous landing system for unattended UAVs provided in the first aspect of the present invention.

[0036] See Figure 2 As shown, the present invention provides an autonomous landing method for unmanned aerial vehicles (UAVs) in unattended scenarios, and its specific implementation process is as follows: Step S1: Phased landmark recognition, the specific implementation process of which is as follows: S1.1: High-altitude full-domain recognition; The first stage is high-altitude full-domain recognition, which mainly uses high-resolution imaging and feature pyramid algorithms. By extracting the contours and enhancing the edges of the entire landmark with a height of more than 10 meters, the key feature information of the landmark, such as the corner information of the landmark, is extracted to achieve accurate capture of the complete shape of the landmark.

[0037] The specific implementation steps of the high-resolution imaging and feature pyramid algorithm are as follows: I. High-resolution imaging; High-resolution imaging is the foundation for all subsequent image processing and analysis. Its goal is to acquire the clearest and most information-rich raw images of the Earth's surface at the physical level. The specific implementation steps are as follows: 1. Sensor selection and platform deployment; Sensor: A high-resolution optical sensor (such as CCD or CMOS) with a large array and small pixel size is selected. High resolution means higher spatial resolution (e.g., Ground Sampling Distance, GSD, reaching the centimeter level), which allows landmark details (such as brick seams and window edges) at heights of 10 meters or more to be clearly recorded.

[0038] Platform Deployment: The sensors are mounted on stable platforms such as drones, high-altitude aerial photography aircraft, or satellites. The platform needs to be equipped with a high-performance image stabilization system to reduce the impact of flight shake on image quality.

[0039] 2. Imaging parameter optimization; Lighting conditions selection: Prioritize shooting in well-lit, even weather conditions (such as midday on a clear day) to avoid excessively long shadows obscuring landmark features.

[0040] Exposure control: Use automatic exposure or preset exposure parameters to ensure that the overall brightness of the image is moderate, and avoid overexposure of bright areas or underexposure of dark areas, so as to lose texture information.

[0041] Focal length and focus: Use a telephoto lens and set it to infinity focus to ensure a sharp image across the entire shooting area.

[0042] 3. Perform image preprocessing before the data enters the algorithm; Radiometric calibration and atmospheric correction: Correcting radiometric distortion caused by atmospheric scattering, absorption, and inconsistencies in sensor response, restoring the true reflectivity information of landmarks.

[0043] Coarse geometric correction: Based on sensor parameters and platform attitude (POS data: position, attitude), perform preliminary geometric distortion correction on the image.

[0044] Image denoising and enhancement: Non-local means denoising or deep learning-based methods are employed to suppress image noise. Simultaneously, techniques such as adaptive histogram equalization may be used to initially enhance the contrast between landmarks and background, preparing for subsequent feature extraction.

[0045] II. Feature Pyramid Algorithm; The feature pyramid algorithm is the core of solving the scale variation problem. In high-altitude recognition, the same landmark may appear at different sizes in an image, and different parts of the landmark itself (such as the roof as a whole and its corners) also have different scale characteristics. The feature pyramid achieves robust recognition of landmarks of different sizes by extracting and fusing features at multiple scales. Its specific implementation steps are as follows: 1. Backbone network feature extraction; The preprocessed high-resolution image is input into a pre-trained convolutional neural network (CNN) as the backbone network, such as ResNet, VGG, or the more efficient CSPDarknet. As the number of network layers increases, multi-scale feature maps are obtained at different depths, as follows: Shallow feature maps (high resolution, low semantics) come from the first few layers of the network. They are high-resolution and contain rich details (such as edges, textures, and corners), crucial for accurately locating landmark outlines and corners. Deep feature maps (low resolution, high semantics) come from the last few layers of the network. After multiple downsampling steps, they have low resolution, but each pixel has a large receptive field, containing abstract semantic information, which helps in understanding the overall category and shape of the landmark.

[0046] 2. Construct a feature pyramid with top-down paths and lateral connections. This is the core idea of ​​Feature Pyramid Network (FPN), which aims to propagate and enhance strong semantic features from high levels to all scales.

[0047] Top-down path: Start with the deepest, most semantically strongest feature map, and upsample it (such as bilinear interpolation or transposed convolution) to make its size match the feature map of the previous layer.

[0048] Lateral connection: The upsampled feature map is fused with shallow feature maps of the same scale from the backbone network (usually by element-wise addition or channel concatenation). This step is crucial; it essentially "refines" the semantics of the deep features with the details of the shallow features.

[0049] Repeat the above process to build a series of feature maps that integrate high semantics and high detail from top to bottom, forming the final pyramid.

[0050] 3. Multi-scale feature fusion and prediction; At each level (i.e., each scale) of the constructed feature pyramid, a small prediction head is independently attached. The prediction head is usually composed of several convolutional layers, thereby achieving division of labor and cooperation, as follows: The pyramid's top layer (low resolution) is responsible for detecting larger landmarks in the image. Because of its large field of view, it can see the entire landmark. The pyramid's bottom layer (high resolution) is responsible for detecting smaller landmarks or fine parts of landmarks (such as corners) in the image, because it preserves rich spatial details. The middle layer of the pyramid is responsible for detecting medium-sized landmarks. This structure ensures that regardless of the size of the landmark in the image, there is always a feature layer of appropriate scale that can accurately identify and locate it.

[0051] 4. Implementation of contour extraction and edge enhancement; The high-resolution, high-detail feature maps output from the bottom layer of the feature pyramid are naturally suitable for contour extraction tasks. Specifically, these feature maps can be input into a specific segmentation head (such as a U-Net-style decoder) or edge detection head, and through supervised learning (using labeled landmark contours as ground truth), binary contour maps of the landmarks are directly output. Edge enhancement is implicitly implemented in this process: the CNN learns to strengthen edge-sensitive filters through training, and during the feature fusion stage, shallow edge information is guided and enhanced by deep semantic information, making the final contour prediction more continuous, accurate, and robust to noise.

[0052] 5. Extraction of key feature information such as corner points; Corner points, as a strong feature, can be obtained in two ways: direct regression and indirect calculation from the contour. Direct regression involves adding a branch to the prediction head to directly regress the heatmap or coordinates of the corner points. Indirect calculation from the contour involves using traditional corner detection algorithms (such as Harris corner detection) or deep learning-based methods after obtaining a precise contour to find the point with the greatest curvature on the contour line, which is the corner point of the landmark.

[0053] S1.2: Low-altitude marker tracking; The second stage, low-altitude marker tracking, is primarily achieved through a deep learning-based object detection model. This stage dynamically tracks ArUco markers with a height below 10 meters, introducing an attention mechanism to enhance marker region features and resist interference from lighting changes and partial occlusion. To achieve fast and robust tracking of low-altitude ArUco markers, this stage recommends using a single-stage object detection model, such as YOLO (YouOnly Look Once) or its variants (e.g., YOLOv5, YOLOv7, YOLOX), as the deep learning-based object detection model. This type of model achieves a good balance between speed and accuracy, making it very suitable for real-time dynamic tracking tasks. Furthermore, to cope with lighting changes and occlusion, mechanisms such as Convolutional Block Attention (CBAM) or Coordinate Attention are integrated into the model backbone. The specific implementation steps are as follows: Step 1: Data preparation and augmentation; 1. Data Acquisition and Annotation: Acquire video streams and images containing ArUco markers in various real-world low-altitude scenarios (indoor, outdoor, and different time periods). Use annotation tools (such as LabelImg) to annotate the bounding boxes of ArUco markers in the images, labeling them as ArUco; the annotation file format is PASCAL VOC or COCO.

[0054] 2. Data augmentation strategies: (1) Simulate illumination changes: Randomly adjust the brightness, contrast, and saturation of the image. Add Gaussian noise to simulate sensor noise under low illumination conditions.

[0055] (2) Simulated partial occlusion: Using techniques such as random occlusion, CutOut or MixUp, a portion of the image is randomly occluded, forcing the model to learn features that do not depend on the complete appearance of the image.

[0056] (3) Geometric transformation: Apply random rotation, scaling, cropping and shearing to make the model invariant to different viewpoints and scales of the marker.

[0057] Step 2: Model building and attention mechanism integration; 1. Backbone network selection: CSPDarknet, which uses YOLO series models (such as YOLOv5s as a lightweight deployment basis, or YOLOv5m / X to pursue higher accuracy), is used as the backbone network for efficient feature extraction.

[0058] 2. Introduce an attention mechanism; Following the critical convolutional layers of the backbone, a CBAM module is inserted. This CBAM module sequentially employs channel attention and spatial attention. Channel attention weights each channel of the feature map, guiding the model to focus on which feature channels are more important. This helps the model focus on the salient features of the markers in complex backgrounds. Spatial attention weights each spatial location of the feature map, guiding the model to focus on which locations in the feature map are more important. This directly enhances the feature response of the region where the marker is located and, when the marker is partially occluded, focuses attention on the undisturbed parts. Through this attention mechanism, the model can adaptively enhance the feature representation of the marker region while suppressing irrelevant background interference, thus effectively resisting changes in illumination and partial occlusion.

[0059] 3. Neck and Head Design; Neck (Feature Fusion Network): Employs either FPN (Feature Pyramid Network) or PANet (Path Aggregation Network) structures to fuse features from different depths of the Backbone, combining high-resolution location information from the shallow layers with strong semantic information from the deep layers to improve the detection capability of labels at different scales. Head (Detection Head): Retains YOLO's default detection head, responsible for performing bounding box regression and class prediction on the fused feature map.

[0060] Step 3: Model training and optimization; 1. Transfer learning: Initialize using model weights pre-trained on a large dataset (such as COCO). This can accelerate model convergence and improve generalization performance.

[0061] 2. Loss Function: The standard composite loss function for object detection tasks is adopted, which typically includes: bounding box regression loss, class prediction loss, and confidence loss. Bounding box regression loss, such as CIoU Loss, can better measure the overlap between the predicted box and the ground truth box, the distance between the center points, and the aspect ratio. The class prediction loss uses binary cross-entropy loss to determine whether the box contains an ArUco label. The confidence loss uses binary cross-entropy loss to determine whether the box contains an object and the reliability of the box.

[0062] 3. Training hyperparameter settings: Use an adaptive optimizer (such as AdamW or SGD with momentum). Set appropriate learning rate, batch size, and training cycles. Learning rate warm-up and cosine annealing strategies are typically used to stabilize the training process.

[0063] Step 4: Dynamic tracking and post-processing; 1. Inference and Post-processing: The video stream is fed frame by frame into the trained model for forward propagation to obtain the predicted bounding boxes; Non-maximum suppression (NMS) or its variants (such as DIoU-NMS) are applied to eliminate redundant detection boxes for the same label.

[0064] 2. Build a tracking pipeline: Combine the target detector with a standalone tracker (such as ByteTrack or DeepSORT).

[0065] 3. Workflow: The model performs detection on each frame and outputs bounding boxes labeled with ArUco. The tracker uses the target's appearance features to associate the detection boxes of the current frame with the tracking trajectories of previous frames, assigning a unique and persistent ID to each label. Lost trajectories are preserved and recovered, and new trajectories are initialized for newly added detection boxes. Even if detection fails in some frames due to severe occlusion or overexposure, the tracker can maintain the target's trajectory based on motion consistency, ensuring the continuity of tracking.

[0066] Step 5: Deployment and Performance Evaluation; 1. Model Deployment: Convert the trained model into a deployment-friendly format, such as ONNX, TensorRT, or TFLite, to enable low-latency inference on edge computing devices such as Jetson Nano and Nvidia GPUs.

[0067] 2. Performance Evaluation Metrics: Precision, Recall, and Mean Accuracy (mAP) are used to evaluate the accuracy of the detection model. Multi-Object Tracking Accuracy (MOTA) and Multi-Object Tracking Precision (MOTP) are used to evaluate the overall performance of the tracking system.

[0068] In the phased landmark recognition process, this invention employs different optimized recognition algorithms for objects at different heights, improving the accurate capture of the complete shape of the entire landmark at heights above 10 meters, enhancing the stability of ArUco marker recognition at heights below 10 meters, and reducing the impact of environmental factors on the recognition results.

[0069] Step S2: Dynamic pose estimation, the specific implementation process of which is as follows: S2.1: High-altitude corner point matching; In the first stage, for the extracted key feature information of landmarks, SIFT feature point extraction and RANSAC algorithm are used to eliminate mismatches. Based on the perspective transformation relationship between the three-dimensional coordinates of the landmark corner points and the UAV imaging coordinates, the six-degree-of-freedom pose parameters are calculated.

[0070] I. SIFT Feature Point Extraction; The SIFT (Scale Invariant Feature Transform) algorithm aims to extract local feature descriptors from images that remain invariant to changes in scale, rotation, and brightness. Its specific implementation steps are as follows: 1. Scale-space extremum detection (constructing a Gaussian difference pyramid); (1) Constructing a Gaussian pyramid: The input landmark template image and the real-time image from the UAV are convolved using Gaussian kernels with different standard deviations σ to generate multiple sets (octaves) of images at different scales. Each set typically contains several layers.

[0071] (2) Constructing the Gaussian difference pyramid: Subtract Gaussian images of adjacent scales within the same group to obtain the Gaussian difference image.

[0072] (3) Finding extreme points: Each pixel in the difference of Gaussian pyramid is compared with its 8 neighboring pixels at the same scale and 9 pixels at the adjacent scales above and below (a total of 18). If the pixel value is the maximum or minimum value among these 26 neighboring pixels, it is marked as a candidate keypoint (feature point). This step is to find stable extreme points in scale space and image space.

[0073] 2. Key point identification and screening; (1) Precise positioning: The position and scale of key points are precisely determined by fitting a three-dimensional quadratic function (in the image space x, y and the scale space σ). The sub-pixel level accuracy can improve the stability of matching.

[0074] (2) Remove low contrast points: Eliminate key points with low function values ​​(i.e., contrast) because they are sensitive to noise.

[0075] (3) Eliminate edge response: Calculate the Hessian matrix at keypoints and use the principal curvature ratio to remove unstable points located at the image edges. Edge points have small principal curvature along the edge direction, but large principal curvature perpendicular to the edge direction. Keypoints with a principal curvature ratio below a certain threshold are retained.

[0076] 3. Assign direction; Assign one or more orientations to each keypoint to achieve rotation invariance. Calculate the gradient magnitude and orientation of all pixels within a neighborhood window of the keypoint in the scale image. Generate a 36-bin gradient orientation histogram, dividing the range from 0° to 360°. The peak value of the histogram represents the principal orientation of the keypoint. Simultaneously, retain orientations with a peak value greater than 80% as secondary orientations for that keypoint.

[0077] 4. Generate key point descriptors; The coordinate axes are rotated to the principal direction of the keypoint, ensuring rotation invariance. The 16x16 pixel area surrounding the keypoint is divided into 4x4 sub-regions. For each 4x4 sub-region, its gradient orientation histogram (HHOC) in eight directions is calculated. The 8-dimensional histograms of these 16 sub-regions are concatenated to form a 4x4x8 = 128-dimensional feature vector, which is the SIFT descriptor. Finally, this descriptor is normalized to further mitigate the effects of lighting variations.

[0078] II. RANSAC algorithm for eliminating false matches; After initial matching using descriptor distances (such as nearest neighbors), a large number of mismatches will exist. RANSAC uses iterative random sampling to find a geometric model that best explains the correctly matched point pairs, thereby separating interior points from mismatched exterior points. The specific implementation steps are as follows: 1. Model Assumptions: Choose an appropriate geometric model based on whether the landmark is planar or non-planar. For planar landmarks (such as wall markings or ground patterns), the most commonly used model is the homography matrix. It is a 3x3 matrix that describes the perspective transformation relationship between two planar views. The homography matrix H has 8 degrees of freedom, and solving it requires at least 4 pairs of non-collinear matching points.

[0079] 2. Iterative sampling and model calculation; (1) Random sampling: Randomly select 4 pairs from all the initial matching point pairs (the minimum sample set that satisfies the solution of the homography matrix).

[0080] (2) Calculation model: Use these 4 pairs of points to calculate a candidate homography matrix H.

[0081] (3) Interior point detection: All preliminary matching point pairs (not just these 4 pairs) are tested using this candidate model H. Specifically, the points in the landmark image are transformed into the UAV image coordinate system through H, and the projection error (usually Euclidean distance) between them and the corresponding matching points in the UAV image is calculated.

[0082] (4) Determine interior points: If the projection error is less than a preset threshold (e.g., 2-5 pixels), the matching point pair is considered an interior point of the current model H. Otherwise, it is marked as an exterior point.

[0083] 3. Select the optimal model; Repeat step 2 (random sampling, model calculation, interior point detection) N times. The number of iterations N is adaptive and can be dynamically calculated based on the proportion of interior points in the best model found so far, ensuring that at least one sample contains all interior points with a high probability (e.g., 99%). In all these iterations, the candidate model with the highest number of interior points is retained as the optimal model.

[0084] 4. Model reestimation and output; (1) Eliminate false matches: All matching pairs that are determined to be outliers by the optimal model are eliminated as false matches.

[0085] (2) Accurate estimation: Using all interior points, the final homography matrix H is recalculated using robust estimation algorithms such as least squares. This step utilizes all the correct data, making the estimated model more accurate.

[0086] III. Calculate the six-DOF pose parameters; 1. Decompose the pose from the homography matrix; After obtaining the accurate homography matrix H and the known camera intrinsic matrix K, the rotation matrix R (its first two columns r1, r2) and translation vector t can be decomposed from H using the relationship H = K [r1 r2t]. The third column r3 of the rotation matrix can be obtained by r3 = r1 × r2 (cross product). Finally, the rotation matrix R = [r1 r2 r3] needs to be orthogonalized to ensure that it satisfies the properties of a rotation matrix. In this way, we obtain the three rotational degrees of freedom and three translational degrees of freedom of the UAV relative to the landmark plane, that is, the complete six-degree-of-freedom pose.

[0087] S2.2: Low-altitude marker resolution; In the second stage, the 3D pose of the UAV is directly obtained by using the coded information of ArUco tags, and Kalman filtering is performed to fuse the IMU inertial measurement data to improve the real-time performance and stability of UAV pose estimation.

[0088] Using the ArUco pose as the observation and the IMU measurement as the prediction, the Kalman filter algorithm is used to perform optimal weighted fusion based on the uncertainty (covariance) of the two. The specific implementation steps are as follows: Step 1: Define the state vector and covariance matrix; 1. State Vector: To fully describe the motion state of the UAV, the state vector needs to include position, velocity, attitude, and IMU sensor bias. A typical 15-dimensional state vector is defined as follows: ; p The position is a three-dimensional position (x, y, z); v For three-dimensional velocity (v) x , v y ,v z ); q The three-dimensional pose is in quaternion form; b a Zero bias for three-dimensional accelerometers; b g It is a three-dimensional gyroscope with zero bias.

[0089] 2. Error Covariance Matrix: PP is set as a 15x15 matrix to represent the uncertainty of the estimated state vector. Initially, the diagonal elements are set to larger values, indicating uncertainty in the initial estimate. PP is continuously updated as the filtering process progresses.

[0090] Step 2: System Model and State Prediction (Time Update). This step uses high-frequency data from the IMU to predict the state of the UAV at the next moment. 1. State transition equation: Based on the angular velocity ω and acceleration a measured by the IMU, the differential equation of the state vector is derived and then discretized and integrated: Location prediction: ; Speed ​​prediction: ; Attitude prediction: ; in, It is a rotation matrix composed of quaternions, used to transform the acceleration of the body coordinate system to the world coordinate system; It is the vector of gravitational acceleration; Represents quaternion multiplication; a m and ω m These are the raw measurements from the IMU; It is three-dimensional velocity. It's time.

[0091] Predicting the covariance matrix: ; F k It is the Jacobian matrix of the state transition equation (obtained by linearizing the previous state estimate), which describes how the state evolves over time. Q k It is the process noise covariance matrix, which represents the impact of IMU measurement noise and model inaccuracy, and is usually set as a parameter for filter tuning.

[0092] Step 3: Observation model and state update (measurement update). This step corrects the predicted state upon receiving pose observations provided by the ArUco markers. 1. Observation Vector ; p aruco The three-dimensional position is obtained by ArUco marker calculation; q aruco It is a quaternion attitude obtained by ArUco tagging; the superscript T indicates transpose.

[0093] 2. Observation Equation ; H H is the observation matrix. In this invention, the observation directly corresponds to the position and attitude parts of the state vector, so H is a simple selection matrix. It is observation noise. Its covariance is; It is in the observation state.

[0094] 3. Calculate the Kalman gain: Kalman gain It is a weight matrix. When the observation noise R is small (ArUco measurements are very accurate), the Kalman gain increases, and the filter trusts the observations more; when the prediction uncertainty P is large (IMU integral drift is large), the Kalman gain also increases, and the filter trusts the observations more.

[0095] 4. Status Update: The predicted and observed states are weighted and fused using Kalman gain. Known as a news item, it is the difference between the observed value and the predicted value.

[0096] 5. Covariance Update: After incorporating observational information, we update the uncertainty of our state estimate. This update reduces the uncertainty.

[0097] Step 4: Asynchronous sensor processing and implementation details; 1. Multi-rate sensor fusion: The IMU frequency (e.g., 200Hz) is much higher than the visual ArUco detection frequency (e.g., 30Hz). Implementation strategy: The system continuously performs prediction steps at the IMU frequency. An update step is triggered only when new ArUco pose data arrives. This constitutes a typical multi-rate asynchronous Kalman filter.

[0098] 2. Bias estimation: The state vector includes the zero bias of the accelerometer and gyroscope. b a and b g The Kalman filter dynamically estimates these biases. When ArUco provides an accurate pose observation, the filter works backward to calculate and correct for the sensor bias that caused the IMU prediction error. This is key to eliminating long-term IMU drift.

[0099] 3. Initialization and parameter tuning; (1) Initialization: The first ArUco observation can be used to initialize the state vector x0.

[0100] (2) Parameter tuning: The covariance matrix of process noise Q and observation noise R is the key to filter performance.

[0101] (3) Q-tuning indicates that the system model is unreliable, the filter will be more dependent on observation, the response will be faster but may be more jittery.

[0102] (4) Increasing R indicates that the observation is unreliable. The filter will rely more on the IMU prediction, resulting in a smoother result but potentially poor tracking and large drift.

[0103] This invention employs a phased landmark recognition and pose estimation method, which identifies the entire landmark at a height of 10 meters or more and identifies ArUco markers at a height of less than 10 meters. This improves the accuracy of landmark recognition and pose estimation at different heights and automatically adapts to the altitude changes required during the autonomous landing of UAVs.

[0104] Step S3: Environmental adaptive flight control, the specific implementation process of which is as follows:

[0105] By integrating data from visual sensors, millimeter-wave radar, and barometers, a dynamic weighted fusion model is constructed to correct altitude measurement errors in real time.

[0106] Based on the real-time operating status, environmental conditions, and historical performance of each sensor, their weights in the fusion are dynamically adjusted so that the system tends to trust the most reliable sensor at any given time.

[0107] The specific implementation steps for constructing the dynamic weight fusion model are as follows: Step 1: Sensor characteristic analysis and uncertainty modeling; 1. Sensor Characteristics Analysis: Visual sensors provide relative altitude, offering high accuracy under rich texture and good lighting conditions, but uncertainty increases dramatically under blurry, fast-moving, or texture-deficient conditions. Millimeter-wave radar directly measures distance to the ground, offering high accuracy and strong anti-interference capabilities, but it introduces errors on inclined or non-uniform ground surfaces. Barometers provide absolute altitude, providing smooth data but exhibiting drift and being susceptible to temperature, airflow (wind pressure), and weather changes.

[0108] 2. Establish a real-time uncertainty model: Define a real-time uncertainty measure for each sensor. (Variance), which is the core basis for weight allocation.

[0109] Visual uncertainty Image sharpness (gradient magnitude) and feature point quantity / quality are calculated based on image quality; motion blur is estimated based on motion state combined with IMU angular velocity. The specific calculation formula is as follows:

[0110] in, It is the number of feature points. Angular velocity, It is the fundamental noise. As a scaling factor, α is the sensitivity coefficient, and α is the raw measurement value of the IMU.

[0111] Radar uncertainty Based on signal quality analysis: weak echo intensity indicates high uncertainty. Based on ground consistency: analyzing the variance of multiple ranging measurements, a large variance indicates uneven ground. The specific calculation formula is as follows:

[0112] in, As a scaling factor, β is the sensitivity coefficient, and β is the echo. This represents the echo intensity value. For signal quality.

[0113] Barometer uncertainty The high-frequency noise variance of the barometer readings within the sliding window is calculated based on short-term fluctuations; dynamic compensation is used to address the error introduced by wind pressure during UAV acceleration. The specific calculation formula is as follows:

[0114] in, This represents the air pressure value within the sliding window. is the scaling factor.

[0115] Step 2: Design a dynamic weight allocation strategy; 1. Weighting principle: The weight is inversely proportional to the uncertainty, while also considering the complementarity between sensors.

[0116] 2. Calculate the base weights: The base weight of each sensor is proportional to the inverse of its uncertainty.

[0117] ,in, To prevent small constants from being divided by zero.

[0118] 3. Weight Normalization: Normalize the basic weights to ensure that the sum of all weights is 1. ,in, The weighted normalized value, Based on the weights.

[0119] Step 3: Implement the adaptive fusion core algorithm; An adaptive weighted fusion algorithm is adopted, which can dynamically adjust the fusion strategy according to the reliability of the sensors.

[0120] 1. Fusion Output Estimation: The optimal height estimate is a weighted average of the measurements from each sensor. ,in, The dynamic weights of v For the dynamic weights of r, The dynamic weight of b, One of the height measurements of the v-sensor. This is one of the height measurement values ​​of the r-sensor. This is one of the height measurement values ​​of sensor b.

[0121] 2. Estimating the fusion uncertainty: The uncertainty of the overall estimate after fusion reflects the confidence level of the final result. .

[0122] Step 4: Construct a sensor complementarity and fault tolerance mechanism; 1. Sensor failure detection; (1) Visual failure detection: When the number of feature points is lower than the threshold or the image blur exceeds the threshold, visual failure is determined.

[0123] (2) Radar failure detection: When the echo intensity is consistently too low or the ranging value changes abruptly, the radar is determined to be faulty.

[0124] (3) Barometer failure detection: When the reading change rate is abnormal or the difference from other sensors is too large, the barometer is determined to be faulty.

[0125] 2. Failure handling strategy; Soft failure handling: Instead of completely excluding the failed sensor, its weight is significantly reduced (e.g., set to 10% of the normal weight). Hard failure handling: When a sensor continues to fail, its weight is temporarily reset to zero, and an alarm is triggered. Recovery mechanism: Set recovery conditions; once the failed sensor returns to normal, its weight is gradually restored.

[0126] 3. Long-term drift online correction: Benchmark selection: When both vision and radar are operating at high confidence, their fusion result is used as the height benchmark.

[0127] Drift estimation: Calculating the long-term deviation of barometer readings from the reference. Wherein, LPF is a low-pass filter used to extract low-frequency drift components; This is the baseline value for altitude; This is the drift estimate.

[0128] Real-time compensation: Online compensation for barometer readings. ,in, This is the real-time compensation value.

[0129] Step 5: System integration and real-time operation; 1. Data preprocessing and synchronization; Coordinate unification: Convert the data from various sensors to a unified coordinate system.

[0130] Time synchronization: Timestamp all sensor data and use interpolation methods to achieve data synchronization.

[0131] 2. Real-time fusion loop; Input: Real-time acquisition of raw data and auxiliary information from each sensor; a. Calculate the real-time uncertainty of each sensor. ; b. Perform failure detection and handling; c. Calculate dynamic weights ; d. Perform fusion calculations and output the optimal height estimate. ; e. Update drift correction parameters; Output: Optimal height estimate and integration uncertainty .

[0132] 3. Parameter tuning and verification; Offline parameter tuning: Optimizing parameters in various uncertainty models using historical data. .

[0133] Online adaptive: Parameters can be fine-tuned based on long-term performance statistics.

[0134] Performance evaluation: The fusion effect is verified by comparing it with higher precision reference systems such as GPS-RTK or laser trackers.

[0135] When the drone is flying at an altitude of less than 20 centimeters, the vertical descent speed is controlled based on the PID parameter self-tuning algorithm, and a soft landing is achieved by combining the preset parameters of the buffer mechanism.

[0136] Based on real-time performance metrics during landing (such as overshoot, response speed, and steady-state error), the PID controller is dynamically adjusted. The parameters enable the control system to maintain optimal performance at different altitude stages. The specific implementation steps of the PID parameter self-tuning algorithm are as follows: Step 1: Establish a near-ground landing control model; Within a 20cm height range, altitude response data of the UAV was collected experimentally. A near-ground dynamics model was established, considering lift variations caused by ground effects. ,in Characterized by the attenuation of ground effect with altitude; m is mass. For height, Near-Earth dynamics For tension, For coefficients, This refers to the altitude near the ground.

[0137] Baseline PID parameter settings: Initial PID parameters designed based on the model These parameters serve as the starting point and fallback parameters for the self-tuning algorithm.

[0138] Step 2: Define performance evaluation metrics; 1. Real-time performance metric calculation; Overshoot : ; For maximum height, This refers to relative height.

[0139] Adjusting time Time required for the altitude to enter the target value ±2%.

[0140] steady-state error .

[0141] Error integral index: , for Square error integral index To simulate step size, It is the integral index of absolute error.

[0142] 2. Performance weight allocation, emphasizing different indicators at different stages of descent; Initial stage (20cm-10cm): Focus on response speed ; Intermediate phase (10cm-5cm): Balanced overshoot and response speed; Final stage (5cm-0cm): Focus on steady-state accuracy .

[0143] Step 3: Design parameter self-tuning strategy; 1. Self-tuning based on fuzzy logic; The input variables are the error e(t) and the rate of change of error e˙(t); the output variable is... ; 2. Construct a fuzzy rule base: When the error e is PB (positive deviation) and the error change rate de / dt is NB (negative change rate), adjust the PID parameters as follows: set ΔKp to PB (positive gain), ΔKi to NB (negative integral gain), and ΔKd to PS (positive derivative gain).

[0144] When the error e is PS (positive deviation) and the error change rate de / dt is NS (negative change rate), adjust the PID parameters as follows: set ΔKp to PS (positive gain), ΔKi to NS (negative integral gain), and ΔKd to PM (positive derivative gain).

[0145] The logic dynamically adjusts the PID controller gain by judging the error and rate of change status based on conditions, thereby realizing an adaptive control strategy.

[0146] 3. Deblurring: Use the centroid method to calculate the accurate output value.

[0147] 4. Online optimization based on gradient descent; Objective function: ; The parameter update formula is as follows:

[0148] Where η is the learning rate, and the gradient is estimated online using the perturbation method; K p For proportional gain, K d For PID differential gain, K i This is the integral gain.

[0149] Step 4: Implement a hierarchical adaptive control structure; 1. The hierarchical strategy is as follows: 20cm-15cm: Aggressive parameters - rapid descent. K p Larger K d Moderate K i Smaller; 15cm-8cm: Balance parameters - smooth transition. K p Moderate K d Larger (suppresses oscillations) K i Moderate; 8cm-0cm: Conservative parameters - precise landing. K p Smaller K d Smaller K i Larger (eliminating steady-state error).

[0150] 2. Smooth parameter transition; At the boundary of high-level hierarchical structure, a first-order low-pass filter is used to achieve a smooth transition of parameters and avoid control shocks caused by abrupt parameter changes. The specific calculation formula is as follows: ,in, For parameter smooth transition values, To ensure a smooth transition of the parameter to the new value, To smoothly transition the parameters to their old values.

[0151] Step 5: Integrate buffer mechanism constraints; 1. Buffer Dynamics Modeling: The force-displacement characteristic model of the buffer mechanism is established as follows: Where δ is the buffer stroke, To buffer travel variables, To buffer force, This is the buffer ratio coefficient. This is the buffer coefficient.

[0152] 2. Landing impact constraints; (1) Define the maximum permissible impact: , This represents the minimum permissible impact level.

[0153] (2) Add the following impact constraint to the PID output: , For a PID controller, the control quantity limit value is calculated based on the error signal. To limit the output range of the control signal, The control quantity calculated by the PID controller based on the error signal. This is the minimum control quantity calculated by the PID controller based on the error signal. This refers to the maximum value of the control quantity calculated by the PID controller based on the error signal. For rate limiting.

[0154] Step 6: System Implementation and Security Monitoring; 1. Real-time self-tuning cycle; Sampling: Read the current height h(k) and calculate the error. ; Assessment: Calculate the current control performance indicators; Decision-making: Calculate parameter adjustment based on fuzzy rules or gradient information; Update: Smoothly update PID parameters; Control: Calculates and outputs control variables; Monitoring: Checking system constraints and security conditions.

[0155] 2. Security protection mechanism; Parameter boundary protection is as follows: in, This is the minimum proportional gain. This represents the maximum proportional gain. This is the minimum integral gain. This represents the maximum integral gain. This is the minimum value of the differential gain. This represents the maximum value of the differential gain.

[0156] Anomaly detection: When a continuous oscillation or divergence trend occurs, it automatically reverts to the baseline parameters.

[0157] Emergency stop: When abnormal acceleration or attitude is detected, an emergency hover is triggered.

[0158] 3. Performance logs and learning; Record performance data and final parameters for each landing process; use reinforcement learning algorithms to optimize fuzzy rules or objective function weights offline; and establish a database of optimal parameters under different conditions.

[0159] This invention designs a vertical landing control strategy when the UAV's flight altitude is less than about 20 centimeters and the ArUco marker cannot be detected. Through a precise PID parameter self-tuning algorithm, the safety and accuracy of the UAV's vertical landing in this stage are guaranteed.

[0160] The sliding window algorithm is used to monitor key indicators such as landmark recognition confidence and pose estimation variance in real time, and multiple threshold levels are set to trigger emergency responses. The sliding window algorithm is used to monitor time-series data of key system indicators in real time, and detects abnormal patterns by analyzing the statistical characteristics of the data within the window. Its specific implementation steps are as follows: Step 1: Define and initialize window parameters; 1. Window parameter configuration: Window Size: Defines the number of data points contained in the window, for example, 50 sampling points (approximately 2.5 seconds, assuming a 20Hz update rate); Step Size: Defines the number of data points the window slides each time, usually 1 (updating point by point); Overlap rate is calculated based on window size and step size, (1 - Step Size / Window Size) * 100%.

[0161] Step 2: Real-time data stream processing; Keep the window size fixed whenever new confidence or variance data arrives; recalculate key statistics after each window update.

[0162] Step 3: Trend and Anomaly Detection; Linear regression is used to calculate data trends, and anomaly scores are calculated based on a combined anomaly score of Z-score and rate of change.

[0163] Step 4: Window performance optimization; Incremental calculation optimization: For statistics such as mean and variance, an incremental update algorithm is used to avoid repeated calculations.

[0164] A multi-level threshold mechanism triggers corresponding emergency responses based on the severity of the anomaly, ensuring system security. The specific implementation steps are as follows: Four-level threshold system: Normal level (LEVEL_NORMAL): Confidence level > 0.7, variance < 0.1; Response: Normal operation, continuous monitoring; Warning level (LEVEL_WARNING): 0.5 < confidence level ≤ 0.7, 0.1 ≤ variance < 0.3; Response: Log the information, increase monitoring frequency, and alert the operator; Level of anomaly (LEVEL_ALERT): 0.3 < confidence level ≤ 0.5, 0.3 ≤ variance < 0.5; Response: Switch to backup sensors, reduce flight speed, and prepare contingency plans; Risk level (LEVEL_CRITICAL): Confidence level ≤ 0.3, variance ≥ 0.5; Response: Immediately implement safety strategies (hover, return to base, land).

[0165] S4.2: Dynamic decision-making; A finite state machine model is constructed to automatically execute strategies such as drone hovering, adjusting flight trajectory to avoid obstacles, and emergency return to home in scenarios including landmark recognition failure, drone equipment malfunction, and sudden obstacles, ensuring drone safety. The specific implementation steps for constructing the finite state machine model are as follows: Step 1: Define the core states of the state machine; 1. Basic flight status; Initialization (INIT): The system starts up and performs self-tests and initialization; Ready: The system is ready and awaiting task instructions; Normal Flight: Flying normally along the planned route; Hover: Hold the current position while waiting. RETURN_HOME: Return to the departure point; Emergency Landing (EMERGENCY_LAND): Immediately land at the current location; Fail-safe (FAILSAFE): The final protection state in the event of a serious system failure.

[0166] 2. Abnormal handling status; Landmark Loss (LANDMARK_LOST): Visual landmark recognition has failed; Obstacle avoidance activated (OBSTACLE_AVOID): An obstacle has been detected and is being avoided; Equipment failure (SENSOR_FAILURE): Critical sensor failure; Communication interruption (COMM_LOST): Communication with the ground station is interrupted.

[0167] Step 2: Design state transition events and conditions; 1. Event type definition: system events, abnormal events, recovery events, and emergency events.

[0168] 2. Design of transfer conditions; (1) Determine whether the conditions for returning to base are met; (2) Determine if an emergency landing is necessary; (3) Determine whether the obstacle has been avoided.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for autonomous landing of unmanned aerial vehicles (UAVs) in unattended scenarios, characterized in that, Includes the following steps: Step S1: Phased landmark identification, including high-altitude global identification and low-altitude marker tracking; Step S2: Dynamic pose estimation, including high-altitude corner point matching and low-altitude marker resolution; Step S3: Environmental adaptive flight control, including multi-sensor data fusion and near-ground landing control; Step S4: Intelligent emergency response, including anomaly monitoring and dynamic decision-making.

2. The autonomous landing method for unmanned aerial vehicles (UAVs) in unattended scenarios according to claim 1, characterized in that, The specific steps of the high-altitude global recognition are as follows: for the entire landmark with a height of more than 10 meters, high-resolution imaging and feature pyramid algorithm are used to extract the contour and enhance the edge, and extract the key feature information of the landmark.

3. The autonomous landing method for unmanned aerial vehicles (UAVs) in unattended scenarios according to claim 1, characterized in that, The specific steps for tracking low-altitude markers are as follows: for ArUco markers with an altitude of less than 10 meters, dynamic tracking is performed using a deep learning-based target detection model.

4. The autonomous landing method for unmanned aerial vehicles (UAVs) in unattended scenarios according to claim 1, characterized in that, The specific steps of the high-altitude corner point matching are as follows: for the extracted key feature information of the landmark, SIFT feature point extraction and RANSAC algorithm are used to eliminate mismatches, and the six-degree-of-freedom pose parameters are calculated based on the perspective transformation relationship between the three-dimensional coordinates of the landmark corner point and the UAV imaging coordinates.

5. The autonomous landing method for unmanned aerial vehicles (UAVs) in unattended scenarios according to claim 1, characterized in that, The specific steps of the low-altitude marker parsing are as follows: the three-dimensional pose of the UAV is directly obtained using the encoded information of the ArUco marker, and Kalman filtering is performed to fuse the data with the IMU inertial measurement data.

6. The autonomous landing method for unmanned aerial vehicles (UAVs) in unattended scenarios according to claim 1, characterized in that, The specific steps of the multi-sensor data fusion are as follows: integrate data from visual sensors, millimeter-wave radar, and barometers, construct a dynamic weighted fusion model, and correct altitude measurement errors in real time.

7. The autonomous landing method for unmanned aerial vehicles (UAVs) in unattended scenarios according to claim 1, characterized in that, The specific steps of the near-ground landing control are as follows: when the UAV's flight altitude is below 20 centimeters, the vertical descent speed is controlled based on the PID parameter self-tuning algorithm, and a soft landing is achieved by combining the preset parameters of the buffer mechanism.

8. The autonomous landing method for unmanned aerial vehicles (UAVs) in unattended scenarios according to claim 1, characterized in that, The specific steps of the anomaly monitoring are as follows: real-time monitoring of landmark recognition confidence and pose estimation variance index through a sliding window algorithm, and setting multi-level thresholds to trigger emergency response.

9. A method for autonomous landing of unmanned aerial vehicles (UAVs) in unattended scenarios according to claim 1, characterized in that, The specific steps of the dynamic decision-making are as follows: construct a finite state machine model, and automatically execute strategies such as hovering and waiting, adjusting the drone's flight trajectory to avoid obstacles, or emergency return to home in the event of landmark recognition failure, drone equipment failure, or sudden obstacles, so as to ensure the safety of the drone.

10. A method for autonomous landing of a drone in an unattended scenario, applicable to an autonomous landing system for unattended scenarios as described in any one of claims 1-9, characterized in that, Includes the following modules: The landmark recognition module is used to extract the contours and enhance the edges of landmarks when the drone is flying at an altitude of more than 10 meters, and to extract key feature information of landmarks; and to recognize ArUco marks on landmarks when the drone is flying at an altitude of less than 10 meters. The pose estimation module is used to estimate the pose of the UAV based on the key feature information of the landmarks extracted by the landmark recognition module when the UAV is flying at an altitude of more than 10 meters; and to estimate the pose of the UAV directly based on the identified ArUco markers when the UAV is flying at an altitude of less than 10 meters. The flight control module is used to plan the flight trajectory of the UAV when the flight altitude is above 10 meters and below 10 meters, respectively, based on the pose information obtained by the pose estimation module, and guide the UAV to fly smoothly towards the landmark; and to control the UAV to make a vertical landing when the flight altitude is less than about 20 centimeters and the ArUco marker cannot be detected. The environmental perception module is used to monitor environmental factors in real time during the autonomous landing process of the UAV, and to transmit the monitored environmental information to the landmark recognition module and the flight control module. The emergency response module is used to monitor various emergencies during the autonomous landing process of the drone in real time, and to activate corresponding emergency measures when an emergency is detected.

Images