Vision-based unmanned aerial vehicle autonomous obstacle avoidance and dynamic landing method and system

Abstract

The application discloses a kind of unmanned aerial vehicle autonomous obstacle avoidance and dynamic landing method and system based on vision.The method comprises the following steps: collecting environment image, identifying landing platform and obstacle and tracking and positioning landing platform;Based on obstacle information, threat assessment is carried out, and environment state description is generated by encoding vision and threat information;Based on landing platform position, reward information is constructed, and environment state description is input into autonomous decision network, after feature fusion of reward information and environment state description in horizontal control sub-network, horizontal motion decision is output, and vertical control sub-network outputs vertical motion decision;According to motion decision, control instruction is generated, unmanned aerial vehicle is controlled to avoid obstacle and land to landing platform.The application converts vision information into structured state description by threat assessment mechanism, and guides decision process by fusing reward information and state features, realizes autonomous obstacle avoidance and accurate landing in complex environment.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) navigation technology, specifically to a vision-based method and system for autonomous obstacle avoidance and dynamic landing of UAVs. Background Technology

[0002] Drones have advantages such as small size, low cost, high cost-effectiveness, long range, and unmanned operation. Compared with manned aircraft, drones are more suitable for performing tasks in dangerous and harsh environments. They are now widely used in military and civilian fields. In application scenarios such as logistics distribution, disaster relief, and agricultural and forestry plant protection, the landing platform serves as the take-off and landing point for drones, providing endurance support and data storage. This is because whether a drone can achieve precise autonomous landing determines whether the mission can be successfully completed.

[0003] Compared to inertial and satellite navigation systems, visual navigation systems have less error accumulation, offer concealment, complete autonomy, and the anti-interference capabilities of airborne cameras, providing fundamental support for autonomous landing of UAVs. However, UAVs rely on flight control algorithms for precise autonomous control throughout their flight. Visual landing methods for UAVs based on traditional control utilize conventional computer vision, path planning, and control algorithms to achieve autonomous flight and precise landing. However, in unpredictable and complex environments, modeling difficulties lead to poor robustness of traditional control algorithms.

[0004] With the continuous development of computer vision, methods based on machine learning theories such as deep learning (DL) and reinforcement learning (RL) have gradually become research hotspots. In situations where modeling is difficult, model-free reinforcement learning is an effective control method. Currently, deep reinforcement learning (DRL), combining RL and DL, has been successfully applied to many control fields, including UAV vision-based autonomous landing. However, many DRL methods do not consider the interference of external uncertainties during UAV landing. First, sudden changes in wind gusts can significantly interfere with the stability and control accuracy of the UAV. Second, considering obstacle avoidance during autonomous landing improves landing accuracy; integrating obstacle avoidance systems into the autonomous landing process greatly enhances the overall autonomy and reliability of UAV operation. However, current research on this is limited, and the need for global information in practical deployment restricts its feasibility in real-world applications. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles (UAVs). By establishing an environmental state description through a threat assessment mechanism and a reconstructed network, some data containing rewards are fused into the UAV's visual observation features in the form of a state description. An autonomous decision-making network based on deep reinforcement learning is constructed to achieve autonomous obstacle avoidance and precise landing in complex dynamic environments.

[0006] This invention is achieved through the following technical solution: Firstly, this application provides a vision-based method for autonomous obstacle avoidance and dynamic landing of unmanned aerial vehicles (UAVs), comprising the following steps: Step 1: Collect environmental images based on the UAV's onboard visual sensor, perform target detection on the environmental images, identify the landing platform and obstacles in the environment, and perform real-time tracking and positioning of the landing platform; Step 2: Conduct a threat assessment based on obstacle information, determine the threat level of each obstacle, and encode the visual information and threat assessment information to generate an environmental status description; Step 3: Construct reward information based on the real-time position information of the landing platform relative to the drone; The environmental state description generated in step 2 is input into the autonomous decision-making network, which includes a horizontal control sub-network and a vertical control sub-network. In the horizontal control subnetwork, the reward information and the environmental state description are fused to generate fused features. The horizontal control subnetwork outputs a horizontal motion decision based on the fused features, and the vertical control subnetwork outputs a vertical motion decision based on the environmental state description. Step 4: Generate horizontal and vertical control commands for the UAV based on the horizontal motion decision and the vertical motion decision, and control the UAV to avoid the obstacle and land on the landing platform.

[0007] Preferably, the real-time tracking and positioning of the landing platform in step 1 specifically includes: Using the first identified landing platform image as a static template and a dynamic template, the current environment image, static template, and dynamic template are input into the improved spatiotemporal Transformer tracking network. The search region features, static template features, and dynamic template features of the current environment image are extracted through the convolutional backbone network. The search region features, static template features, and dynamic template features are processed and then input into the Transformer encoder-decoder structure to output the predicted probability value. The predicted probability value is input into a fully convolutional network, and a probability distribution map of the diagonal of the prediction box is generated by estimating the probability distribution of corner coordinates. The expected value is calculated to obtain the coordinates of the prediction box, and the real-time position of the landing platform in the current environmental image is determined based on the coordinates of the prediction box. The confidence score is calculated based on the search area features, static template features, and dynamic template features. The landing platform height is verified by combining the UAV altitude information. The static template or dynamic template is updated based on the verification result and the confidence score. The updated template is then used to continuously track and locate the landing platform in subsequent environmental images.

[0008] Preferably, step 2, which involves threat assessment based on obstacle information to determine the threat level of each obstacle, specifically includes: The total threat score is calculated based on the category, distance, and size of the obstacle. The total threat score is obtained by weighted summation of the category threat assessment item, the distance threat assessment item, and the size threat assessment item. Among them, the category threat assessment item sets different constants according to the type of obstacle, the distance threat assessment item is calculated based on the distance from the obstacle to the center of the image, and the size threat assessment item is calculated based on the pixel size of the obstacle in the image; Obstacles are classified into threat levels based on the calculated threat score.

[0009] Preferably, the total threat score is calculated as follows:

[0010]

[0011]

[0012]

[0013] in, The total threat score, For category threat assessment, that is, determining whether the obstacle is a ground obstacle or an air obstacle. It is a constant. For distance threat assessment, the closer the obstacle is to the drone, the greater the threat. The distance is the diagonal of the image. Let be the two-dimensional Euclidean distance from the obstacle to the center point of the image. For threat assessment, the larger the obstacle, the greater the threat. The pixel size of the obstacle in the image. This represents the total pixel size of the image.

[0014] Preferably, the encoding of visual information and threat assessment information to generate an environmental state description in step 2 specifically includes: The real-time environmental images acquired in step 1 are processed using a real-time reconstruction network. Combined with the geometric representation corresponding to the threat level, obstacles in the current environment are modeled in a unified manner to generate real-time reconstruction results containing information on obstacle location, obstacle shape, and threat level. The real-time reconstruction results are jointly encoded with threat assessment information to construct an environmental state description that includes real-time environmental images, obstacle geometric representations, and threat level information.

[0015] Preferably, the step 3, which involves constructing reward information based on the real-time position information of the landing platform relative to the drone, specifically includes: Obtain the coordinates of the landing platform center in the image coordinate system; Calculate the two-dimensional Euclidean distance from the center of the drone to the center of the landing platform; The distance and coordinate information are used to construct reward information.

[0016] Preferably, step 3, which involves fusing the reward information with the environmental state description to generate fused features, specifically includes: The environmental state description is input into a three-level convolutional encoder for feature extraction, and the extracted feature map is flattened into a one-dimensional vector. The reward information is concatenated with the flattened feature vector to generate a fused feature; The fused features are input into the fully connected layer of the horizontal control subnetwork, and the action value of the horizontal motion decision is output.

[0017] Preferably, the horizontal control subnetwork outputs a horizontal motion decision based on the fused features, and the vertical control subnetwork outputs a vertical motion decision based on the environmental state description, including: The fused features are input into the fully connected layer of the horizontal control subnetwork. After processing by the fully connected layer, the action values ​​corresponding to the four action neurons in the horizontal direction are output. The horizontal motion decision is selected based on the action values. The environmental state description is input into a three-level convolutional neural network of the vertical control subnetwork for feature extraction. The extracted feature map is flattened and then input into a fully connected layer. After processing by the fully connected layer, the action values ​​corresponding to the two action neurons in the vertical direction are output. The vertical motion decision is selected based on the action values.

[0018] Preferably, the training process of the autonomous decision-making network in step 3 specifically includes: In each training step, the UAV acquires the current observation state from the environment, inputs the current observation state into the autonomous decision-making network, and generates the current action; After the drone performs the current action, the environment returns a new state, an immediate reward, and an abort flag. The (state, action, reward, new state, abort flag) is stored as an experience tuple in the experience replay pool. A batch of experience tuples is randomly sampled from the experience replay pool to update the parameters of the autonomous decision-making network; The autonomous decision-making network takes state descriptions as input and outputs state-action value functions for all possible actions. By minimizing the temporal difference error loss function, the network parameters are optimized, enabling the UAV to learn decision strategies for completing stable landing missions in complex dynamic environments. The horizontal control subnetwork and the vertical control subnetwork are trained using independent empirical replay pools and loss functions, respectively. Repeat the above training steps until the preset training termination condition is met, and the trained autonomous decision-making network is obtained.

[0019] Preferably, step 4, which involves generating horizontal and vertical control commands for the UAV based on the horizontal motion decision and the vertical motion decision, specifically includes: Based on the action type corresponding to the horizontal motion decision, and combined with the relative positional relationship between the landing platform and the UAV, an adaptive horizontal speed control command is generated. Based on the action type corresponding to the vertical motion decision, a vertical speed control command is generated; The generated horizontal and vertical speed control commands are output to the UAV flight control system to control the UAV to adjust its flight attitude and trajectory in real time.

[0020] Secondly, this application provides a vision-based autonomous obstacle avoidance and dynamic landing system for unmanned aerial vehicles (UAVs), comprising: The image acquisition module is used to acquire environmental images based on the UAV's onboard visual sensor, perform target detection on the environmental images, identify the landing platform and obstacles in the environment, and perform real-time tracking and positioning of the landing platform; The state construction module is used to perform threat assessment based on obstacle information, determine the threat level of each obstacle, and encode visual information and threat assessment information to generate an environmental state description. The decision-making module is used to construct reward information based on the real-time position information of the landing platform relative to the drone; The environmental state description is input into the autonomous decision-making network, which includes a horizontal control subnetwork and a vertical control subnetwork. In the horizontal control subnetwork, the reward information and the environmental state description are fused to generate fused features. The horizontal control subnetwork outputs a horizontal motion decision based on the fused features, and the vertical control subnetwork outputs a vertical motion decision based on the environmental state description. The control module is used to generate horizontal and vertical control commands for the UAV based on the horizontal motion decision and the vertical motion decision, and to control the UAV to avoid the obstacle and land on the landing platform.

[0021] Compared with the prior art, the present invention has the following beneficial technical effects: This application proposes a vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles (UAVs). It transforms raw visual information into a structured environmental state description and introduces a fusion mechanism of reward information and state features, enabling UAVs to achieve autonomous obstacle avoidance and dynamic landing even without global information. First, the method transforms raw visual information into a structured state description containing obstacle threat levels through a threat assessment mechanism. This addresses the problem of redundant perception information in complex environments, making it difficult to directly use in the decision-making network, and provides the subsequent decision-making network with information-rich and dimensionally controllable input features. Second, the method introduces a fusion mechanism of reward information and state features into the horizontal control sub-network, directly integrating the relative position information of the landing platform into the high-dimensional feature space. This allows the agent to use valuable information to guide decision-making at each time step, significantly enhancing the guiding role of reward signals in the decision-making process. Third, the decoupling design of horizontal and vertical control reduces the dimensionality of the action space, allowing the two sub-networks to focus on lateral obstacle avoidance and vertical descent tasks respectively. This avoids the learning difficulties of a single network in a high-dimensional action space, improving the accuracy and stability of decision-making. Finally, this method does not rely on external positioning facilities and global maps, but can complete autonomous obstacle avoidance and dynamic landing solely with airborne visual sensors, significantly improving the autonomous operation capability of UAVs in scenarios where GPS signals are missing or the environment is unknown.

[0022] This application also proposes a vision-based autonomous obstacle avoidance and dynamic landing system for unmanned aerial vehicles (UAVs), an electronic device, and a computer storage medium, which possess all the advantages of the aforementioned vision-based autonomous obstacle avoidance and dynamic landing methods for UAVs. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a logic block diagram of the UAV autonomous obstacle avoidance and dynamic landing method of the present invention.

[0025] Figure 2 This is a schematic diagram of the improved spatiotemporal Transformer tracking method of the present invention.

[0026] Figure 3 This is a schematic diagram of the threat assessment mechanism and reconstructed network framework of the present invention.

[0027] Figure 4 This is a schematic diagram of the autonomous decision-making network structure of the present invention.

[0028] Figure 5 This is a schematic diagram of reward data included in part of the present invention.

[0029] Figure 6 This is a schematic diagram of the autonomous decision-making network training framework of the present invention.

[0030] Figure 7 This is a flowchart of the UAV autonomous obstacle avoidance and dynamic landing method of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0032] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0033] See Figure 1-7 A vision-based method for autonomous obstacle avoidance and dynamic landing of unmanned aerial vehicles (UAVs) includes the following steps: Step 1: Environmental images are acquired using the UAV's bottom-mounted visual sensors. Deep learning methods are then used to identify the landing platform and obstacles. An improved spatiotemporal Transformer tracking method is employed to perform real-time localization of the landing platform. Specifically, based on real-time environmental images acquired by the UAV's onboard visual sensors, a convolutional neural network feature extraction method and an improved spatiotemporal Transformer tracking method are combined to output the location information of the landing platform and obstacles, as well as the dynamic trajectory prediction of the landing platform. This provides foundational data for subsequent threat assessment and autonomous decision-making.

[0034] Specifically, it includes the following processes: S1.1 Utilize the visual sensor mounted on the bottom of the drone to collect real-time image data of the environment below the drone.

[0035] The visual sensor can be a monocular camera, a binocular camera, or a depth camera. The acquisition frequency is set according to the UAV's flight speed and scene complexity to ensure the acquisition of a continuous and stable sequence of environmental images.

[0036] S1.2. The real-time environmental images acquired in S1.1 are input into a pre-trained convolutional neural network for feature extraction, identifying the landing platform and various obstacles in the environment from the images. The convolutional neural network can adopt an object detection network structure, outputting the bounding box positions and category information of the landing platform and obstacles.

[0037] S1.3 Using the initial position of the landing platform obtained in S1.2 as the reference frame, the improved spatiotemporal Transformer tracking method is used to continuously track the landing platform.

[0038] refer to Figure 2 The improved spatiotemporal Transformer tracking method consists of a convolutional backbone network, a Transformer encoder-decoder, a bounding box prediction head, and a confidence evaluation module. The convolutional backbone network uses ResNet101 to extract image features. The Transformer encoder and decoder each have 6 stacked layers, employing a multi-head attention mechanism with a width of 256 and 8 layers. The bounding box prediction head uses a 5-layer stacked ConvBNReLU fully convolutional network, and the score prediction head in the confidence evaluation uses a 3-layer perceptron.

[0039] During the tracking process, the static and dynamic templates are first initialized. Based on the initial perception results, the landing platform image in the initial state of the UAV is captured as the static template. and dynamic templates Then, the current bottom visual image of the drone will be displayed. Static templates Dynamic templates The input to the convolutional backbone network yields three sets of feature maps. These feature maps are preprocessed by reducing the number of channels using a bottleneck layer, flattening and concatenating them, and then adding sinusoidal positional encoding before being used as encoder input. The enhanced sequence features processed by the encoder, along with the target query, serve as decoder input, outputting predicted probability values.

[0040] To improve prediction accuracy, a method that estimates the probability distribution of the corner coordinates of the predicted bounding box is used instead of direct coordinate regression. The encoder output sequence is multiplied by the decoder output probability to obtain the similarity score for each location. Then, the similarity score is multiplied by the corresponding element of the encoder output sequence to enhance the features of important regions. Finally, the processed feature map is input into a fully convolutional network to obtain the probability distribution map of the top-left corner coordinates of the predicted bounding box. Probability distribution diagram of the bottom right corner coordinates The coordinates of the prediction box are obtained by calculating the expected value. and :

[0041] S1.4 Calculate the confidence score based on the search region features, static template features, and dynamic template features output by the encoder.

[0042] When the landing platform appears within the detection range of the laser sensor, the altitude of the landing platform is calculated by combining the difference between the UAV's altimeter and the laser range. If the altitude estimation is correct, the dynamic template is updated based on the confidence score; if the altitude estimation is incorrect, it indicates that the static template may be unreliable or the landing platform has left the field of view. In this case, the UAV's attitude is adjusted and it accelerates to expand the search field of view, the landing platform is re-sensed, the static template is updated using the re-sensing results, and the dynamic template continues to be updated based on the confidence score.

[0043] This step achieves real-time and accurate localization of the dynamic landing platform by constructing a visual perception architecture for target recognition, continuous tracking, and confidence evaluation. The improved spatiotemporal Transformer tracking method utilizes a dual mechanism of static and dynamic templates, combined with multi-head attention feature enhancement and corner probability distribution prediction, effectively suppressing interference from similar objects in the tracking process and improving the robustness of platform tracking in complex environments. The confidence evaluation and template update mechanism ensures that the tracker can adaptively cope with anomalies such as platform occlusion and loss of field of view, providing continuous and reliable platform location information for subsequent threat assessment and autonomous decision-making.

[0044] In some embodiments, the vision sensor may employ the following sensors in addition to a monocular camera: A binocular vision camera can simultaneously acquire environmental depth information to assist in the three-dimensional positioning of obstacles; Depth cameras directly acquire pixel depth values, simplifying the coordinate transformation process; Infrared thermal imaging cameras are suitable for platform identification in low-light or nighttime environments.

[0045] In some embodiments, the target recognition network may employ the following methods in addition to using a convolutional neural network: Transformer-based object detection networks (such as DETR) output detection results end-to-end. A lightweight detection network based on the YOLO series, suitable for airborne platforms with limited computing resources; Networks based on instance segmentation (such as Mask R-CNN) can simultaneously acquire the contour information of the target.

[0046] Step 2: Based on the obstacle and landing platform information obtained in Step 1, a unified environmental state description is generated by combining the threat assessment mechanism and the reconstruction network. Specifically, using the obstacle and landing platform information output in Step 1 as input, the threat assessment mechanism calculates the threat score and determines the threat level. The reconstruction network is used to perform geometric modeling of the obstacles, and the original visual information and threat assessment information are encoded to output an environmental state description in a unified feature space for use by the subsequent autonomous decision-making network.

[0047] See Figure 3 Specifically, it includes the following processes: S2.1 Based on the obstacle and landing platform information identified in Step 1, calculate the total threat score for each target using a threat assessment mechanism. The threat score consists of three parts: category threat assessment, distance threat assessment, and size threat assessment. The calculation formula is as follows:

[0048] in, The total threat score, This is a category-based threat assessment item, which determines whether the obstacle is a ground obstacle or an air obstacle. For distance threat assessment, For size threat assessment, and These are the weighting coefficients.

[0049] Category Threat Assessment Different constants are set according to the target type:

[0050] Distance Threat Assessment Based on the distance from the obstacle to the image center, the closer the obstacle, the greater the threat.

[0051] in, The distance is the diagonal of the image. Let be the two-dimensional Euclidean distance from the obstacle to the center point of the image.

[0052] Size Threat Assessment Based on the pixel size of the obstacle in the image, the larger the area, the greater the threat:

[0053] in, This represents the number of pixels representing the obstacle in the image. This represents the total number of pixels in the image.

[0054] S2.2 Total threat score calculated based on S2.1 Environmental threats are classified into three levels: low-level threats, medium-level threats, and high-level threats.

[0055] Different threat levels correspond to different geometric representations: low-level threats are modeled using elliptical shapes, while medium and high-level threats are modeled using rectangular shapes.

[0056] S2.3. The original visual information is processed using a reality reconstruction network, and the obstacles are modeled in a unified manner by combining the threat level determined in S2.2 and the corresponding geometric representation.

[0057] The reconstructed network is built based on the YOLOv5 target detection model, with two sets of reconstructed networks built for simulation and real-world environments, respectively.

[0058] During the algorithm training phase, Reconstruction Network 1 is enabled. This network receives image information collected from the first-person perspective of the UAV in the simulation environment, extracts key targets in the images through scene perception, and reconstructs them into highly general semantic descriptions based on threat levels, including the location, shape, and threat level information of the landing platform and various obstacles.

[0059] During the algorithm migration and deployment phase, Reconstruction Network 2 is enabled. This network processes first-view images captured by drones in real-world environments, generating reconstruction results with features similar to those in the simulation environment. This effectively reduces the difference between simulation training and real-world applications, enhancing the algorithm's transferability.

[0060] In the simulation environment, image data of the landing platform and obstacles in different scenarios were collected using the CoppeliaSim platform to establish and annotate the simulation environment dataset, which was then used to train the reconstruction network 1. In the real environment, actual scene samples were collected by UAVs and annotated, which were then used to train the reconstruction network 2.

[0061] S2.4. The reconstructed result output from S2.3 and the threat assessment information calculated in S2.1 are jointly encoded to construct a unified feature space. The unified feature space includes the original visual information, the geometric representation of the target, and the threat level information, forming an environmental state description for input to the autonomous decision-making network.

[0062] This step, through a process of geometric modeling feature encoding based on threat quantification levels, achieves a structured representation of environmental perception information. The threat assessment mechanism calculates threat scores based on three dimensions: target category, distance, and size, effectively quantifying the impact of each target on UAV safety. Different geometric representation strategies are employed for different threat levels, ensuring both environmental modeling accuracy and computational efficiency. The dual-reconstruction network architecture effectively solves the migration problem from simulation to real-world environments, enabling the algorithm to adapt to complex real-world scenarios. The construction of a unified feature space provides information-rich and structurally clear state inputs for subsequent autonomous decision-making.

[0063] In some embodiments, the threat assessment mechanism may employ the following methods in addition to linear weighting: Threat assessment based on fuzzy logic addresses uncertainty and ambiguous information. End-to-end threat assessment based on neural networks, automatically learning feature weights; Threat estimation based on Bayesian inference, incorporating multi-source uncertainty information.

[0064] In some embodiments, the geometric representation, in addition to using ellipses and rectangles, can also employ the following modeling methods: A circle is suitable for obstacles that are approximately symmetrical; Polygon representation accurately describes irregularly shaped obstacles; Voxel mesh representation: suitable for representing complex and irregular environment boundaries.

[0065] In some embodiments, the reconstructed network may employ the following network structures in addition to YOLOv5: A semantic segmentation network based on Transformer enables pixel-level scene understanding; A scene reconstruction method based on generative adversarial networks generates more realistic environmental representations; Based on an autoencoder-based feature extraction network, it learns a compact representation of the environment.

[0066] Step 3: Based on the environmental state description generated in Step 2, an autonomous decision-making network is constructed using deep reinforcement learning methods to generate motion decisions for the UAV's autonomous landing. Specifically, the environmental state description output in Step 2 is used as input. Horizontal and vertical motion decisions are processed by horizontal and vertical control networks, respectively. Dense reward and state feature fusion representations are introduced into the horizontal control network to output horizontal and vertical control commands for the UAV, guiding it to complete autonomous obstacle avoidance and landing tasks.

[0067] Specifically, it includes the following processes: S3.1 Construct an autonomous decision-making network consisting of a horizontal control network and a vertical control network.

[0068] See Figure 4 The inputs to both the horizontal and vertical control networks are A state description composed of reconstructed stacked images.

[0069] The horizontal control network output contains four action neurons, corresponding to the UAV's lateral motion decisions (left, right, forward, and backward). The vertical control network output contains two action neurons, corresponding to the UAV's longitudinal motion decisions (descent, hold).

[0070] The network structure uses a three-layer convolutional neural network for feature extraction: the first layer uses... The convolution kernel generates 32 Feature map; the second layer uses Convolution kernels, generating 32 Feature map; the third layer uses Convolution kernels, generating 32 Feature maps. All convolutional layers use the ReLU activation function. After the convolution operation, the feature maps are flattened into one-dimensional vectors, passed through three fully connected layers, and the final output layer contains the same number of neurons as the action dimension, outputting the state action value for each action.

[0071] S3.2, Reference Figure 5 A fusion representation method of dense reward and state features is introduced into the horizontal control network. This part includes reward data. For the relative position information of the landing platform, including the two-dimensional Euclidean distance from the center of the drone to the center of the landing platform. For the center coordinates of the landing platform The expression is as follows:

[0072]

[0073] At time step At that time, the state description of the horizontal control network Features are generated through convolution operations New fusion features are generated by merging with some reward data. :

[0074] in, This indicates a three-level convolutional encoder, where the superscript indicates the number of convolutional layers. Expand the features into a one-dimensional vector. This indicates a vector concatenation operation.

[0075] Based on fusion features The horizontal control network outputs state action values. A fully connected network is defined as a composite function:

[0076] in, Represents function composition. , If the activation function is nonlinear, then the formula for calculating the value function is:

[0077] in, To control the network parameters of the agent at the horizontal level.

[0078] S3.3, see also Figure 6 In each training step, the UAV acquires observational states from the environment and inputs them into the autonomous decision-making network to generate the current action. After the action is executed, a new state, reward, and termination flag are generated. The <state, action, reward, new state, termination flag> tuple generated from each interaction is stored in the experience replay pool.

[0079] During training, a batch of data is randomly sampled from the experience replay pool to update the network parameters. The decision network takes state descriptions as input and outputs state-action values ​​for all possible actions to guide action selection. The network parameters are optimized by minimizing the TD error loss function to guide the UAV in learning to complete stable landing tasks in complex dynamic environments.

[0080] The expression for the loss function is as follows:

[0081] in, It is the target Q value. It is the Q-value of the current network. These are the parameters of the current network. It is the set of samples stored in the experience replay buffer.

[0082] The horizontal control network is based on the input state. and actions Generate a new state The data is stored in the experience replay buffer 1 and processed by the reward function. The loss function L1 is calculated. The vertical control network directly calculates the loss function L1 based on the input state. and actions Generate a new state The data is stored in the experience replay buffer 2 and processed through the reward function. Calculate the loss function L2. Optimize the respective agent networks using the two loss functions to enable the UAV to learn a stable landing strategy in complex dynamic environments.

[0083] S3.4 Based on the trained autonomous decision-making network, the optimal action selection is generated according to the current environmental state description, and the horizontal and vertical control commands of the UAV are determined.

[0084] This step constructs a deep reinforcement learning framework based on dual-network collaboration. This framework introduces reward feature fusion and experience replay training mechanisms, ultimately enabling autonomous decision-making by the UAV in complex environments. The separate design of the horizontal and vertical control networks reduces the dimensionality of the action space, allowing each agent to focus on a specific dimension of control tasks. The fusion representation method of dense rewards and state features directly integrates the relative position information of the landing platform into the high-dimensional feature space, enabling the agent to utilize value information to guide decision-making at each time step, enhancing the influence of reward signals on the decision-making process, and significantly improving training convergence speed. The experience replay mechanism breaks the temporal correlation between samples, improving training stability. This method enables the UAV to achieve autonomous obstacle avoidance and dynamic landing relying solely on onboard visual perception, even without global information.

[0085] In some embodiments, in addition to employing a value function-based approach, the deep reinforcement learning algorithm may also employ the following methods: Reinforcement learning algorithms based on stochastic policies (such as PPO) are suitable for continuous action spaces; Deterministic policy-based reinforcement learning algorithms (such as DDPG) combine value functions and policy optimization; Reinforcement learning algorithms based on maximum entropy (such as SAC) balance exploration and exploitation.

[0086] In some embodiments, in addition to employing three convolutional layers, the convolutional neural network structure may also employ the following network structures: Residual networks (ResNet) alleviate the gradient vanishing problem in deep networks; DenseNet networks enhance feature reuse; Lightweight networking (MobileNet) is suitable for airborne platforms with limited computing resources.

[0087] In some embodiments, the fusion representation method, in addition to vector concatenation, can also employ the following fusion methods: Attention mechanisms are integrated, and feature weights are learned adaptively. Gating mechanisms are integrated to control the flow of information; Bilinear pooling captures higher-order interaction information between features.

[0088] Step 4: This step aims to enable the UAV to autonomously avoid obstacles and dynamically land in complex environments, based on the horizontal and vertical control commands generated in Step 3. Specifically, using the optimal action selection determined in Step 3 as input, the horizontal movement speed is adaptively adjusted according to the relative position relationship between the landing platform and the UAV, and the vertical movement is controlled according to a fixed descent speed. Horizontal and vertical speed control commands are output to drive the robot to avoid obstacles and land on the dynamic platform.

[0089] Specifically, it includes the following processes: S4.1. Based on the center coordinates of the landing platform and the center coordinates of the UAV's bottom-view image obtained in step 1, calculate the horizontal speed command. The horizontal speed adopts an adaptive adjustment method, and the speed magnitude is proportional to the horizontal distance between the UAV and the landing platform. The calculation formula is as follows:

[0090]

[0091] in, The linear velocity is in the left and right directions. , The linear velocity in the forward and backward directions. , For speed coefficient, and This is the gain coefficient. The function limits the speed to a reasonable range.

[0092] Horizontal control agent Action space is defined as These correspond to four movements: left, right, forward, and backward. Based on the speed value calculated in S4.1 and the action selection output in step 3, the final horizontal control command is generated.

[0093] S4.2, Vertical Control Agent Action space is defined as These correspond to the two motions of descent and hold, respectively. Indicates a decrease, Indicates holding. The vertical speed is set to a fixed value; in this embodiment, the descent speed is 0.15 m / s. When the action is selected... At that time, the drone descends at a fixed speed; when an action is selected... At that time, the drone maintained its current altitude.

[0094] S4.3 The UAV adjusts its flight attitude and trajectory in real time based on the horizontal and vertical control commands generated by S4.1 and S4.2.

[0095] The expression for the horizontal control command is as follows:

[0096] The expression for the vertical control command is as follows:

[0097] During execution, environmental perception, status updates, and decision optimization are continuously performed through steps 1 to 3, forming a closed-loop control process of perception, decision, and control. The mission is completed when the UAV successfully lands on the dynamic platform and a landing impact is detected; if an abnormal situation is detected (such as platform loss or approaching obstacles), the corresponding emergency response mechanism is triggered.

[0098] This step achieves complete closed-loop control from decision-making to physical motion. Adaptive adjustment of horizontal speed allows the UAV to dynamically adjust its approach speed based on the distance to the target; faster movement at greater distances improves efficiency, while deceleration at closer distances ensures safe and accurate landing. A fixed descent speed simplifies vertical control and ensures a smooth descent. The decoupling design of horizontal and vertical control allows for parallel execution of motion in both dimensions, improving the control system's response speed. Through continuous environmental perception and decision optimization, the UAV can adjust its attitude in real time in dynamically changing and complex environments, effectively avoiding aerial and ground obstacles, ultimately achieving precise autonomous landing.

[0099] In some embodiments, the horizontal speed control, in addition to function regulation, may also employ the following control methods: Proportional-integral-derivative (PID) control enables more precise position tracking; Model predictive control (MPC) optimizes control sequences over a future period of time. Adaptive control dynamically adjusts control parameters according to changes in the environment.

[0100] In some embodiments, the vertical velocity, in addition to being a fixed velocity, can also be controlled using the following methods: Altitude proportional control adjusts the descent speed based on the difference between the current altitude and the target altitude; Segmented speed control, with different descent speeds used for different altitude ranges; Continuous velocity planning generates a smooth descent velocity curve.

[0101] In some embodiments, the emergency response mechanism may include, in addition to the default obstacle avoidance strategy, the following measures: Emergency hover, waiting for the obstacle to pass or the platform to reappear; Independent advancement, broadening horizons and repositioning the platform; Return mode: Return to the takeoff point or a preset safe location.

[0102] This application presents a vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles (UAVs). This method enables UAVs to accurately land from the takeoff point to a dynamic platform in complex environments with gusts of wind and aerial obstacles, even without global information, through real-time environmental perception and autonomous decision-making. First, it proposes a fusion representation method combining dense rewards and state features. Based on a threat assessment mechanism, it evaluates the environmental threat level in real time, generates an environmental state description by reconstructing the network and combining the assessment information. It also integrates some reward-related data into the UAV's visual observation features, thus constructing an efficient value-oriented state description. A value function guides the learning of state features, optimizing the convergence direction of deep reinforcement learning. While ensuring the transferability of original observation features, it fully utilizes the intrinsic correlation between state features and rewards, significantly improving the stability and convergence speed of the training process. Second, it constructs a deep reinforcement learning-based autonomous decision-making framework for UAVs. Based on the state description, it calculates the UAV's autonomous landing motion decision and generates UAV speed control commands based on the decision results to achieve autonomous obstacle avoidance and precise dynamic landing in complex environments, significantly improving the system's efficiency and robustness.

[0103] Based on the above-mentioned autonomous obstacle avoidance and dynamic landing method for UAVs, this application also provides a vision-based autonomous obstacle avoidance and dynamic landing system for UAVs, including: The image acquisition module is used to acquire environmental images based on the UAV's onboard visual sensor, perform target detection on the environmental images, identify the landing platform and obstacles in the environment, and perform real-time tracking and positioning of the landing platform; The state construction module is used to perform threat assessment based on obstacle information, determine the threat level of each obstacle, and encode visual information and threat assessment information to generate an environmental state description. The decision-making module is used to construct reward information based on the real-time position information of the landing platform relative to the drone; The environmental state description is input into the autonomous decision-making network, which includes a horizontal control subnetwork and a vertical control subnetwork. In the horizontal control subnetwork, the reward information and the environmental state description are fused to generate fused features. The horizontal control subnetwork outputs a horizontal motion decision based on the fused features, and the vertical control subnetwork outputs a vertical motion decision based on the environmental state description. The control module is used to generate horizontal and vertical control commands for the UAV based on the horizontal motion decision and the vertical motion decision, and to control the UAV to avoid the obstacle and land on the landing platform.

[0104] It should be noted that, in the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another device, or some features may be ignored or not executed. The modules described as separate components may or may not be physically separated. The components shown as modules may be one or more physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs.

[0105] Furthermore, in the various embodiments of the present invention, the modules can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0106] An electronic device provided in this application includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles as described in any of the above embodiments.

[0107] Another electronic device provided in this application embodiment may further include: an input port connected to a processor for transmitting multimodal data collected by an external acquisition device to the processor; a display unit connected to the processor for displaying the processor's processing results to the outside world; and a communication module connected to the processor for enabling communication between the electronic device and the outside world. The display unit may be a display panel, a laser scanning display, etc.; the communication method adopted by the communication module includes, but is not limited to, Mobile High Definition Link (HML), Universal Serial Bus (USB), High Definition Multimedia Interface (HDMI), and wireless connection (including Wi-Fi, Bluetooth, Bluetooth Low Energy, and IEEE 802.11s-based communication technology).

[0108] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles as described in any of the above embodiments.

[0109] For descriptions of relevant parts of the vision-based UAV autonomous obstacle avoidance and dynamic landing system, electronic device, and computer-readable storage medium provided in this application's embodiments, please refer to the detailed descriptions of the corresponding parts in the vision-based UAV autonomous obstacle avoidance and dynamic landing method provided in this application's embodiments; they will not be repeated here. Furthermore, parts of the technical solutions provided in this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0110] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: Step 1: Collect environmental images based on the UAV's onboard visual sensor, perform target detection on the environmental images, identify the landing platform and obstacles in the environment, and perform real-time tracking and positioning of the landing platform; Step 2: Conduct a threat assessment based on obstacle information, determine the threat level of each obstacle, and encode the visual information and threat assessment information to generate an environmental status description; Step 3: Construct reward information based on the real-time position information of the landing platform relative to the drone; The environmental state description generated in step 2 is input into the autonomous decision-making network, which includes a horizontal control sub-network and a vertical control sub-network. In the horizontal control subnetwork, the reward information and the environmental state description are fused to generate fused features. The horizontal control subnetwork outputs a horizontal motion decision based on the fused features, and the vertical control subnetwork outputs a vertical motion decision based on the environmental state description. Step 4: Generate horizontal and vertical control commands for the UAV based on the horizontal motion decision and the vertical motion decision, and control the UAV to avoid the obstacle and land on the landing platform.

2. The vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles according to claim 1, characterized in that, Step 1, which involves real-time tracking and positioning of the landing platform, specifically includes: Using the first identified landing platform image as a static template and a dynamic template, the current environment image, static template, and dynamic template are input into the improved spatiotemporal Transformer tracking network. The search region features, static template features, and dynamic template features of the current environment image are extracted through the convolutional backbone network. The search region features, static template features, and dynamic template features are processed and then input into the Transformer encoder-decoder structure to output the predicted probability value. The predicted probability value is input into a fully convolutional network, and a probability distribution map of the diagonal of the prediction box is generated by estimating the probability distribution of corner coordinates. The expected value is calculated to obtain the coordinates of the prediction box, and the real-time position of the landing platform in the current environmental image is determined based on the coordinates of the prediction box. The confidence score is calculated based on the search area features, static template features, and dynamic template features. The landing platform height is verified by combining the UAV altitude information. The static template or dynamic template is updated based on the verification result and the confidence score. The updated template is then used to continuously track and locate the landing platform in subsequent environmental images.

3. The vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles according to claim 1, characterized in that, Step 2, which involves threat assessment based on obstacle information to determine the threat level of each obstacle, specifically includes: The total threat score is calculated based on the category, distance, and size of the obstacle. The total threat score is obtained by weighted summation of the category threat assessment item, the distance threat assessment item, and the size threat assessment item. Among them, the category threat assessment item sets different constants according to the type of obstacle, the distance threat assessment item is calculated based on the distance from the obstacle to the center of the image, and the size threat assessment item is calculated based on the pixel size of the obstacle in the image; Obstacles are classified into threat levels based on the calculated threat score.

4. The vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles according to claim 1, characterized in that, The total threat score is calculated as follows: in, The total threat score, For category threat assessment, that is, determining whether the obstacle is a ground obstacle or an air obstacle. It is a constant. For distance threat assessment, the closer the obstacle is to the drone, the greater the threat. The distance is the diagonal of the image. Let be the two-dimensional Euclidean distance from the obstacle to the center point of the image. For threat assessment, the larger the obstacle, the greater the threat. The pixel size of the obstacle in the image. This represents the total pixel size of the image.

5. The vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles according to claim 1, characterized in that, Step 2, which involves encoding visual information and threat assessment information to generate an environmental state description, specifically includes: The real-time environmental images acquired in step 1 are processed using a real-time reconstruction network. Combined with the geometric representation corresponding to the threat level, obstacles in the current environment are modeled in a unified manner to generate real-time reconstruction results containing information on obstacle location, obstacle shape, and threat level. The real-time reconstruction results are jointly encoded with threat assessment information to construct an environmental state description that includes real-time environmental images, obstacle geometric representations, and threat level information.

6. The vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles according to claim 1, characterized in that, Step 3, which involves constructing reward information based on the real-time position information of the landing platform relative to the UAV, specifically includes: Obtain the coordinates of the landing platform center in the image coordinate system; Calculate the two-dimensional Euclidean distance from the center of the drone to the center of the landing platform; The distance and coordinate information are used to construct reward information.

7. The vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles according to claim 1, characterized in that, Step 3, which involves fusing reward information with environmental state descriptions to generate fused features, specifically includes: The environmental state description is input into a three-level convolutional encoder for feature extraction, and the extracted feature map is flattened into a one-dimensional vector. The reward information is concatenated with the flattened feature vector to generate a fused feature; The fused features are input into the fully connected layer of the horizontal control subnetwork, and the action value of the horizontal motion decision is output.

8. The vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles according to claim 1, characterized in that, The horizontal control subnetwork outputs a horizontal motion decision based on the fused features, and the vertical control subnetwork outputs a vertical motion decision based on the environmental state description, including: The fused features are input into the fully connected layer of the horizontal control subnetwork. After processing by the fully connected layer, the action values ​​corresponding to the four action neurons in the horizontal direction are output. The horizontal motion decision is selected based on the action values. The environmental state description is input into a three-level convolutional neural network of the vertical control subnetwork for feature extraction. The extracted feature map is flattened and then input into a fully connected layer. After processing by the fully connected layer, the action values ​​corresponding to the two action neurons in the vertical direction are output. The vertical motion decision is selected based on the action values.

9. A vision-based autonomous obstacle avoidance and dynamic landing method for unmanned aerial vehicles according to claim 1, characterized in that, The training process of the autonomous decision-making network described in step 3 specifically includes: In each training step, the UAV acquires the current observation state from the environment, inputs the current observation state into the autonomous decision-making network, and generates the current action; After the drone performs the current action, the environment returns a new state, an immediate reward, and an abort flag. The (state, action, reward, new state, abort flag) is stored as an experience tuple in the experience replay pool. A batch of experience tuples is randomly sampled from the experience replay pool to update the parameters of the autonomous decision-making network; The autonomous decision-making network takes state descriptions as input and outputs state-action value functions for all possible actions. By minimizing the temporal difference error loss function, the network parameters are optimized, enabling the UAV to learn decision strategies for completing stable landing missions in complex dynamic environments. The horizontal control subnetwork and the vertical control subnetwork are trained using independent empirical replay pools and loss functions, respectively. Repeat the above training steps until the preset training termination condition is met, and the trained autonomous decision-making network is obtained.

10. A vision-based autonomous obstacle avoidance and dynamic landing system for unmanned aerial vehicles (UAVs), characterized in that, include: The image acquisition module is used to acquire environmental images based on the UAV's onboard visual sensor, perform target detection on the environmental images, identify the landing platform and obstacles in the environment, and perform real-time tracking and positioning of the landing platform; The state construction module is used to perform threat assessment based on obstacle information, determine the threat level of each obstacle, and encode visual information and threat assessment information to generate an environmental state description. The decision-making module is used to construct reward information based on the real-time position information of the landing platform relative to the drone; The environmental state description is input into the autonomous decision-making network, which includes a horizontal control subnetwork and a vertical control subnetwork. In the horizontal control subnetwork, the reward information and the environmental state description are fused to generate fused features. The horizontal control subnetwork outputs a horizontal motion decision based on the fused features, and the vertical control subnetwork outputs a vertical motion decision based on the environmental state description. The control module is used to generate horizontal and vertical control commands for the UAV based on the horizontal motion decision and the vertical motion decision, and to control the UAV to avoid the obstacle and land on the landing platform.

Images