Unmanned aerial vehicle autonomous navigation method and device based on cross-view orientation regression network

Abstract

A method and apparatus for UAV autonomous navigation based on a cross-view orientation regression network are presented. The method includes: constructing a training set, validation set, and test set of cross-view data for orientation information; constructing a UAV position and heading regression model (CVPHR) for cross-view scenarios; periodically training and validating the CVPHR using the training and validation sets to obtain the final CVPHR; inputting the test set into the final CVPHR model for testing; constructing a base map feature vector table; using the trained CVPHR, inferring about reference tiles in a 2D remote sensing map according to row and column indices, and constructing a base map feature vector table (FVT) from the output set of reference tile feature vectors and their corresponding row and column indices; transforming the CVPHR into a UAV position and heading regression model (FVT_CVPHR) using a base map feature vector table lookup mode; building an UAV autonomous navigation model (CVPHR_Naver) based on the cross-view orientation regression network; and the UAV using CVPHR_Naver, carrying the base map feature vector table and navigation sequence information, for cross-view autonomous positioning and navigation.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) navigation technology, and more specifically to UAV autonomous navigation methods and devices. Background Technology

[0002] Drones are widely used in security inspections, disaster relief, high-altitude operations, freight delivery, and agricultural monitoring, permeating various industries and becoming closely intertwined with human life. With the support of the future low-altitude economy and artificial intelligence technology, they have even broader development and application prospects. Drones primarily rely on global positioning systems such as GNSS and GPS for positioning and navigation. However, in certain special scenarios, such as disaster relief in mountainous areas, when obstructed by terrain, or when there is signal interference, or in some extreme environments, drones may be unable to receive stable positioning or navigation signals, leading to mission failure and even significant safety hazards.

[0003] To address the aforementioned issues, UAV positioning and navigation technology under GNSS signal-free conditions has been extensively studied. This technology aims to acquire one or more types of information through airborne sensors (such as visual odometry, inertial measurement units, cameras, and lidar), and achieve UAV aerial positioning and navigation through methods such as motion estimation, map building, visual feature matching, and multimodal guidance. This ensures that UAVs can autonomously locate and navigate using their own onboard sensors when there are no external positioning and navigation signals, perform obstacle avoidance and landing maneuvers to reduce safety risks, and even continue to autonomously complete their missions.

[0004] Visual navigation is one of the key technologies for autonomous navigation algorithms of UAVs in complex environments and under GNSS denial conditions. It mainly relies on visual perception of the environment, acquires image information through airborne cameras, and uses various algorithms and techniques based on knowledge from multiple fields such as computer vision, machine learning, and cybernetics to estimate the position, attitude, and motion state of the UAV in three-dimensional space. It is the foundation of the aerial perception capability of embodied intelligent agents.

[0005] In point-to-point navigation tasks, common algorithms include multi-sensor-based methods and vision-language-based methods. The former typically combines visual signals with inertial sensors for continuous point-to-point navigation. These algorithms use sensors such as magnetometers for orientation and odometry to calculate displacement, making them prone to cumulative errors, which are often eliminated periodically using visual image matching methods. The latter often uses verbal commands to guide the UAV towards the target and uses visual signals for positioning and feedback. These algorithms require communication between the commander and the UAV and are susceptible to communication interference and flight distance limitations. These algorithms do not fully utilize visual image information. If UAVs could achieve autonomous positioning and orientation solely based on visual images in the air, and then achieve autonomous navigation based on this, they would have greater survivability in extreme environments. However, current research in this area is not yet in-depth, and the potential of pure vision-based UAV navigation remains to be further explored and utilized.

[0006] Unmanned aerial vehicles (UAVs) using purely visual information for autonomous navigation require addressing issues such as continuous cross-view simulation environments and navigation-driven strategies. Due to the difficulty in obtaining real-world scene data, many current aerial visual navigation algorithms use 3D virtual environments for simulation, resulting in performance deviations when migrating from simulation to real-world scenarios, and generally lacking access to absolute latitude, longitude, and heading angle information. Using remote sensing maps as the UAV positioning and navigation simulation environment provides the conditions for obtaining absolute latitude, longitude, and heading angle information, but requires the design of continuous maps and cross-view image simulation environments. Currently, most UAV positioning and navigation algorithms use discrete remote sensing image samples, lacking the conditions for continuous point-to-point navigation, and very few methods can achieve continuous autonomous navigation under purely visual conditions; some purely visual navigation algorithms achieve precise positioning or navigation, but use 2D UAV view images, lacking the cross-view conditions of 2D remote sensing images plus 3D UAV views, resulting in a significant discrepancy with real-world scenarios. Summary of the Invention

[0007] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provide a method and apparatus for UAV position and heading regression in cross-view scenarios.

[0008] This invention addresses the aforementioned pure vision-based UAV positioning and navigation problem by constructing a more realistic cross-view pose regression dataset and a 3D view navigation simulation environment using a gridded, continuous 2D remote sensing map and 3D UAV view samples collected from Google Maps. It designs a UAV position heading-angle regression model (CVPHR) based on cross-view scenarios. CVPHR is used to process the baseline tiles of the 2D Google Maps in the task scenario into a Feature Vector Table (FVT) for flexible carrying and lookup by the UAV. Furthermore, CVPHR is transformed into a UAV position heading-angle regression model under the Feature Vector Table lookup mode (CVPHR under the Feature Vector Table, FVT_CVPHR), reducing storage and computational load during flight navigation to achieve lightweight design. Based on this, a lightweight UAV autonomous navigation method based on a cross-view orientation regression network (CVPHR based Autonomous Navigation Methods for UAV, CVPHR_Naver) is proposed.

[0009] The UAV autonomous navigation method based on cross-view orientation regression network of the present invention includes the following steps:

[0010] Step 1: Collect adjacent environmental tile samples and 3D UAV view samples using remote sensing maps to construct a cross-view dataset containing orientation information, which is then divided into training set, validation set and test set;

[0011] Step 2: Construct a UAV position and heading regression model (CVPHR) for cross-view scenarios;

[0012] Step 3: Use the training set and validation set to periodically train and validate the CVPHR model, and obtain the final CVPHR model weight file;

[0013] Step 4: Input the test set into the final CVPHR model and test the regression performance of the CVPHR model using test samples not found in the test set;

[0014] Step 5: Use the trained CVPHR model to infer the environment tiles and construct the environment tile feature vector table FVT from the output set of environment tile feature vectors and their corresponding row and column indices.

[0015] Step 6: Transform CVPHR into a UAV position and heading regression model FVT_CVPHR under FVT lookup mode; FVT_CVPHR uses the feature vectors corresponding to a set of two-dimensional environment tiles arranged in a square matrix and a three-dimensional UAV view of the same area to predict the absolute position and heading angle of the UAV; FVT_CVPHR searches for a set of feature vectors from FVT based on the index information of the environment tile where the UAV is located.

[0016] Step 7: Build an autonomous navigation model CVPHR_Naver for UAVs based on a cross-view orientation regression network. It consists of FVT_CVPHR and a step navigation module. The FVT_CVPHR module predicts the position and heading, and the step navigation module calculates the step vector and updates the position and orientation of the UAV. The navigation process is completed in a loop.

[0017] Step 8: The UAV uses CVPHR_Naver, carrying FVT and navigation sequence information, to conduct cross-view autonomous positioning and navigation tests.

[0018] Preferably, the steps for constructing the cross-perspective dataset in step 1 are as follows:

[0019] Step 1-1: Download the side length based on the latitude and longitude range of the task area. Square remote sensing map of pixels Simultaneously record the latitude and longitude values ​​of its center pixel. unit pixel distance value ;

[0020] Step 1-2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require Divided into grids with side lengths of Small square environment tiles of pixels Any set of square-shaped environmental tiles forms a square environmental block. Each environment block has a unique row index and column index number;

[0021] Steps 1-3: For each environmental block in Steps 1-2, establish a system with the block center as the origin and horizontal... Axis and Vertical The axes are relative Cartesian coordinate systems. The line connecting the center points of the outermost environment tiles of each environment block forms a square effective range, and the square effective ranges of adjacent environment blocks are seamlessly connected.

[0022] Steps 1-4: Define the due east direction on the remote sensing map as... And corresponding In the positive direction of the axis, within the effective coordinate range of each environment block, generate A set of samples with random latitude and longitude coordinates and random heading angles is used to collect data in a three-dimensional remote sensing map. A sample of drone views ;

[0023] Steps 1-5: Following the left-to-right and top-to-bottom process, repeat steps 1-3 and 1-4 for each environmental block until the remote sensing map has been traversed. All environmental blocks are used to ultimately obtain the CVPHR cross-view dataset CVPHR_Dataset.

[0024] Preferably, in the CVPHR model described in step 2, the environment tile coordinate encoder encodes the relative coordinates of a set of square-arrayed two-dimensional environment tiles into a high-dimensional vector; the VGG backbone is used to extract feature maps of the environment tiles or the UAV view; the CSMG feature vector extractor is used to extract feature vectors from the feature maps; the cross-attention module is used to extract cross-view correlation features between the UAV view and the environment tiles; the tile similarity guidance module is used to estimate the similarity between the UAV view and a set of environment tiles and convert it into a weighted coordinate vector; the position regression module and the orientation regression module predict the UAV's position and heading angle from the fused features, respectively.

[0025] The structure of the CVPHR model described in step 2 further includes: the environment tile coordinate encoder is composed of cascaded fully connected layers, each using the ReLU activation function; in the cross-attention module, the UAV view features are processed through fully connected layers to obtain a query vector, the base map features are processed through fully connected layers and transposed to obtain a key vector, the two are multiplied, scaled, and processed through softmax to obtain attention weights, which are then multiplied with the value matrix generated by the fully connected layers to output the dimensional cross-attention features; the block similarity guidance module is composed of a cosine similarity calculation unit, a softmax unit, a column broadcasting unit, a matrix element multiplier, and a column vector summing unit connected in sequence, with the matrix element multiplier connected to the base block relative coordinate matrix input; the position regression module is composed of cascaded fully connected layers, with the last fully connected layer outputting the predicted coordinates; the direction regression module is composed of cascaded fully connected layers, with the last fully connected layer outputting the predicted direction;

[0026] The process of constructing the CVPHR model also includes: designing the input and output dimensions, kernel size, stride, and padding size of all convolutional layers, pooling layers, and normalization layers in the VGG backbone network and CSMG feature vector extractor, based on the dimensions of the input image and its coordinates of the CVPHR model, as well as the input and output signal dimensions of other sub-modules and the input and output dimensions of fully connected layers.

[0027] Preferably, step 2 specifically includes:

[0028] Step 2-1: For a specific environmental block, collect a 3D drone view of that environmental block. All environment tiles Its relative coordinates are used as input;

[0029] Step 2-2: Extract UAV view feature vectors using the VGGNet backbone network and the CSMG feature extractor, respectively. and environmental tile feature vectors ;

[0030] Steps 2-3: Calculate the relative coordinate encoding vector of the environment tile. Adding the corresponding environmental tile feature vectors yields the enhanced neighborhood feature vector. ;

[0031] Steps 2-4: First, input the UAV view feature vector and the enhanced neighborhood feature vector into the cross-attention module to obtain the cross feature vector. Then, the block similarity guidance module is used to obtain a two-dimensional weighted position estimation vector. ;

[0032] Steps 2-5: Concatenate the UAV view feature vector, cross feature vector, and weighted position estimation vector to obtain the fused feature vector. ;

[0033] Steps 2-6: Input the fused feature vector into the position regression module and the heading regression module respectively to obtain the predicted relative position and the predicted heading angle, and then calculate the absolute latitude and longitude position and heading angle of the UAV.

[0034] Preferably, in step 2, the VGGNet backbone network and CSMG feature extractor network included in the core functional module of the CVPHR model are open source models. The VGGNet backbone network in the CVPHR model adopts a parameter freezing method, and the parameters of the VGGNet backbone network, the parameters of the CSMG feature extractor network, and other sub-modules are all set according to the designed structure and parameters.

[0035] Preferably, step 3 specifically includes:

[0036] Step 3-1: Divide the dataset into training, validation and test sets according to the proportions, and set the hyperparameters of the training model, such as batch size, initial learning rate and number of training rounds;

[0037] Step 3-2: Periodically train and validate the CVPHR model using the training and validation sets. Input the samples into the CVPHR model in batches and calculate the predicted location of the UAV according to steps 2-1 to 2-6. and predicted heading angle ;

[0038] Step 3-3: Use the predicted position obtained in step 3-2 and predicted heading angle with position truth value True value of heading angle Position error and heading error are calculated separately. The loss value is calculated using the Smooth L1 Loss function, and the adaptive moment estimator optimizer Adam is used as the gradient descent algorithm to realize backpropagation and iterative training of the CVPHR model.

[0039] Steps 3-4: The learning rate optimization method uses the plateau learning rate decay strategy ReduceLROnPlateau. In each training round, the training set is used for training and the validation set is used for validation. The validation results are used to adjust the variable learning rate. After completing the total number of training rounds, the weight file of the trained CVPHR model is obtained.

[0040] Preferably, step 4 specifically includes: when using the CVPHR model for testing, loading the weight file described in steps 3-4 into the CVPHR model, then inputting the samples in the test set into the CVPHR model one by one, and calculating the predicted position of the UAV view according to steps 2-1 to 2-6. and predicted course The prediction results are compared with the true values ​​of position and heading labeled for each sample in the test set, and the regression prediction accuracy of position and heading is calculated.

[0041] Preferably, step 5 specifically includes:

[0042] All environmental map tiles obtained in steps 1-2 are processed one by one through the VGG16 backbone network and CSMG feature extractor according to their row and column position indices to obtain their corresponding feature vectors. The rows and columns correspond to latitude lines and longitude lines, respectively. The row and column index numbers of these environmental map tiles and their corresponding feature vectors are used to construct the Environmental Map Patch Feature Vector Table (FVT). Each continuous map is processed once.

[0043] Preferably, step 6, which constructs the FVT_CVPHR structure, includes three parallel branches: an environment map feature vector table lookup module, an environment base map coordinate encoder, and a VGG backbone network connected to a CSMG feature vector extractor, which are then connected to a cross-attention module and a block similarity guidance module, and finally output by a position regression module and a direction regression module.

[0044] The specific steps to set up FVT_CVPHR include:

[0045] Step 6-1: Collect remote sensing maps Latitude and longitude values ​​of the center pixel unit pixel distance value Environmental Blocks Index;

[0046] Step 6-2: Based on the latitude, longitude, and heading angle of the current location of the UAV, acquire the UAV view from the 3D remote sensing map;

[0047] Step 6-3: Input the UAV view into the VGGNet backbone network plus the CSMG feature extractor branch to obtain the feature vector corresponding to the UAV view. ;

[0048] Step 6-4: Obtain the row and column index numbers of the corresponding set of environment tiles based on the row and column index numbers of the environment tiles. Then, retrieve the corresponding set of feature vectors from the environment tile feature vector table FVT based on these row and column index numbers. ;

[0049] Step 6-5: Utilize the feature vectors of the drone view The relative coordinates and feature vectors of a set of adjacent environmental tiles retrieved. The predicted position and predicted heading angle are obtained by following steps 2-3 to 2-6.

[0050] Preferably, step 7 specifically includes:

[0051] Step 7-1: On the mission remote sensing map Mark the latitude and longitude of the starting point, waypoints and the destination, and save them as a navigation waypoint sequence. In the 3D remote sensing map, initialize the UAV's initial true position = initial nominal position = coordinates of the waypoint sequence starting point;

[0052] Step 7-2: Set the initial state of the FVT_CVPHR model and obtain the UAV view from the 3D remote sensing map based on the UAV's actual position. Based on the UAV's nominal location and its environmental block index number, a set of environmental block feature vectors is obtained from the Feature Vector Table (FVT). ;

[0053] Step 7-3: Perform drone positioning and orientation according to steps 6-2 to 6-5, from the drone view. and a set of environmental block feature vectors Obtain the predicted position and predicted heading angle;

[0054] Step 7-4: Calculate the UAV step vector based on the predicted UAV position and the current target waypoint, and adjust the UAV orientation by predicting the angular difference between the heading angle and the step vector direction;

[0055] Step 7-5: The drone moves forward one step and calculates its real position and nominal position. It then calculates the distances between its real position and the target waypoint and destination, respectively.

[0056] Step 7-6: Compare the waypoint threshold or destination threshold to determine whether the target waypoint or destination has been reached. If it is below the threshold, it means that the destination has been reached. If it is above the threshold, it means that the flight has gone out of bounds or exceeded the step limit. If so, it means that the navigation has failed. Otherwise, repeat steps 7-3 to 7-6 until the destination is reached.

[0057] The present invention also relates to an autonomous navigation device for unmanned aerial vehicles (UAVs) based on a cross-view orientation regression network, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the autonomous navigation method for UAVs based on a cross-view orientation regression network of the present invention.

[0058] The present invention also relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the UAV autonomous navigation method based on a cross-view orientation regression network of the present invention.

[0059] The working principle of this invention is:

[0060] This invention collects 2D environmental baseline tile samples and 3D drone view samples from publicly available remote sensing maps (such as Google Maps) in both 2D and 3D perspectives to construct a dataset for training and testing regression models. The 3D drone view has significant visual differences from its corresponding 2D image; the former is a drone flight view that closely resembles the real-world scene and texture, while the latter was captured by a near-Earth orbit remote sensing satellite. This is key to this cross-perspective dataset.

[0061] Secondly, to improve the perception capabilities of UAVs, the global and local structural texture features contained in remote sensing maps are fully utilized to convert visual images into accurate UAV positions and headings. This invention designs and implements a cross-view position and heading regression model (CVPHR) for UAVs based on a deep convolutional neural network. The algorithm can predict the UAV's position and heading angle from a set of four 2D environmental base maps and a 3D UAV view of the same region. This design, utilizing an adjacency-grid map for orientation regression, can better adapt to UAV view localization and orientation under different intersection-union ratios.

[0062] Furthermore, this invention utilizes a UAV cross-view position and heading regression model to infer the feature vectors of the environmental baseline map tiles, constructing a remote sensing map feature vector table (FVT). During flight, the UAV quickly searches for the feature vectors of the baseline map in the FVT using a region index, reducing storage and computational loads during flight navigation and making the algorithm lightweight.

[0063] Finally, the navigation strategy designed in this invention is to deploy a lightweight CVPHR_Naver model for the UAV. The UAV carries an environmental tile feature vector table and navigation waypoint sequence information to perform cross-view autonomous localization and point-to-point navigation simulation on Google 3D map.

[0064] The innovation of this invention is:

[0065] (1) Using pure visual information to regress absolute position: Unlike existing technologies that only use remote sensing map retrieval and matching for coarse positioning, this invention focuses on extracting and utilizing global and local features of the image through cross-attention mechanism, and makes fuller use of pure visual information to achieve accurate prediction of absolute position and heading angle.

[0066] (2) Continuous remote sensing map and cross-view simulation environment: This invention uses two-dimensional remote sensing map and three-dimensional remote sensing map with geographic information to construct training and testing dataset and simulation environment for UAV positioning and navigation. The regression model is trained with cross-view dataset, and the performance is closer to the real scene. This can reduce the performance deviation of the algorithm from simulation to real scene.

[0067] (3) Lightweight design is more suitable for deployment on the UAV side with limited computing power: This invention uses a two-dimensional remote sensing map with geographic information to construct an environmental tile feature vector table. The feature information of the map can be calculated in advance, and the feature vector can be obtained by looking up the table during flight, saving computing time. This design makes the algorithm more lightweight and efficient.

[0068] (4) Flexibility and scalability: The technical solution of the present invention can be flexibly applied to various terrains and landforms. The remote sensing map can be converted into an environmental tile feature vector table in advance to facilitate data compression and carrying. The task granularity can be expanded by increasing the area of ​​the remote sensing map and adjusting specific parameters of the model according to the task requirements.

[0069] The present invention has the following advantages:

[0070] (1) Absolute positioning and orientation capability of cross-view orientation regression model. The established orientation regression network model can directly and accurately predict the latitude and longitude coordinates and heading angle of UAV, which is more accurate than the coarse positioning achieved by the general remote sensing map retrieval and matching method, and can reduce the cumulative error of pure visual navigation scheme. This provides important technical support for the continuous navigation of UAV.

[0071] (2) Regression network structure design with cross-view feature fusion capability. Considering the significant visual errors and feature differences in cross-view images, the CSMG module is first used to extract information containing global and local features from the feature map output by the backbone network, which is more conducive to localization and orientation of cross-domain images, especially improving localization and orientation accuracy when the intersection and union of cross-domain images is small. Secondly, the designed cross-attention module can more efficiently and accurately extract the intrinsic connections between features of cross-view images. Finally, the cross-feature vectors are integrated. Weighted position estimation vector Obtain the fused feature vector This provides sufficient feature information for position and heading regression, further improving cross-view positioning accuracy.

[0072] (3) The dataset and simulation environment are closer to the real scene. Using continuous remote sensing maps as the cross-view dataset and environmental tile feature vector table information source, and using 3D remote sensing maps as the simulation environment, this dataset and simulation environment are more realistic than general three-dimensional virtual scenes in terms of visual effects and ground texture. Moreover, the use of continuous remote sensing maps as the reference environment can better restore the flight navigation process of UAVs in the real environment and reduce the performance deviation when migrating from the simulation environment to the real environment.

[0073] (4) Lightweight design of FVI_CVPHR model. The model parameters are about 140MB. The designed environmental tile feature vector table can pre-calculate the feature information of the mission remote sensing map. Only the table needs to be looked up to quickly obtain the baseline tile features, which reduces the computational burden of the model during the flight of the UAV. It is more suitable for deployment in resource-constrained aerial UAV navigation operation scenarios. The current single-step time is 35.7ms. Attached Figure Description

[0074] Figure 1 This is a flowchart of the method of the present invention.

[0075] Figure 2 This is a schematic diagram of the method for constructing the cross-view remote sensing map dataset of the present invention.

[0076] Figure 3 This is a flowchart of the CVPHR model of the present invention.

[0077] Figure 4 This is a flowchart of the CVPHR core module of the present invention.

[0078] Figure 5a To provide a distribution map and statistical results of the distance error between the predicted and actual locations, Figure 5b This document presents a diagram showing the distribution of the angle error between the predicted heading and the actual heading, along with statistical results.

[0079] Figure 6 This is the flowchart of the FVI_CVPHR model of the present invention.

[0080] Figure 7 This is a diagram of the UAV autonomous positioning and navigation algorithm based on a cross-view orientation regression model, as described in this invention.

[0081] Figure 8 This is a diagram of the UAV autonomous positioning and navigation method based on cross-view orientation regression network of the present invention.

[0082] Figure 9 This is a screenshot of the simulation results of the UAV autonomous positioning and navigation of the present invention. Detailed Implementation

[0083] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0084] Example 1

[0085] like Figure 1 Based on the aforementioned objective of UAV cross-view autonomous navigation under GNSS-free conditions, the UAV autonomous navigation method based on cross-view orientation regression network of the present invention includes the following steps:

[0086] Step 1: Collect adjacent environmental remote sensing tile samples using Google 2D remote sensing maps and collect UAV 3D view samples using Google 3D maps to construct a cross-view dataset containing orientation information. Divide the dataset into training set, validation set and test set according to the set ratio.

[0087] Step 2: Construct a Cross-View PositionHeading-angle Regression (CVPHR) model for UAVs in cross-view scenarios. The CVPHR model predicts the absolute position and heading angle of the UAV from a set of four 2D environmental base maps and a 3D UAV view of the same area. The structure of the CVPHR model is as follows: starting from the input, there are two parallel branches, namely the environmental base map coordinate encoder and the VGG backbone network connected to the CSMG feature vector extractor. Both branches are connected to the cross attention module and the block similarity guidance module to obtain the fused feature vector. Finally, the position regression module and the orientation regression module output the position and orientation respectively.

[0088] The environment base map coordinate encoder encodes the relative coordinates of the four base maps into a high-dimensional vector; the VGG backbone is used to extract feature maps from the environment base map or the UAV view; the CSMG feature vector extractor is used to extract feature vectors from the feature maps; the cross-attention module is used to extract cross-view correlation features between the UAV view and the environment base map; the block similarity guidance module is used to estimate the similarity between the UAV view and the four environment base maps and convert it into a weighted coordinate vector; the position regression module and the orientation regression module predict the UAV's position and heading angle from the fused features, respectively.

[0089] Step 3: Use the training set and validation set to periodically train and validate the CVPHR model, and obtain the final CVPHR model weight file;

[0090] Step 4: Input the test set into the final CVPHR model and test the regression performance of the CVPHR model using different test samples in the test set;

[0091] Step 5: Construct an Environmental Feature Vector Table (FVT). Using the CVPHR model trained in Step 3, infer all the baseline tiles in the 2D remote sensing map according to their row and column indices, and construct an Environmental Feature Vector Table (FVT) from the output set of baseline tile feature vectors and their corresponding row and column indices.

[0092] Step 6: Transform CVPHR into a UAV position and heading regression model under the Feature Vector Table (FVT_CVPHR) in an environment map feature vector table lookup mode to achieve lightweight design. Figure 6 As shown, the FVT_CVPHR uses the feature vectors corresponding to a set of four 2D environment base maps and a 3D UAV view of the same area to predict the absolute position and heading angle of the UAV. The structure of FVT_CVPHR is as follows: the input has three parallel branches, namely the environment map feature vector table lookup module, the environment base map coordinate encoder and the VGG backbone network connected to the CSMG feature vector extractor, which are then connected to the cross attention module and the block similarity guidance module respectively. Finally, the output is provided by the position regression module and the orientation regression module (except for the addition of one branch for looking up the environment map feature vector table in the input, the rest of the structure is the same as that of CVPHR in step 2).

[0093] The lookup vector table module, which searches for the feature vectors corresponding to the four base maps in the environment block based on the index information of the environment block where the UAV is located, is a key step in building the lightweight algorithm.

[0094] Step 7: Build an autonomous navigation model for UAV based on cross-view orientation regression network (CVPHR_Naver). Figure 6 As shown, CVPHR_Naver consists of FVT_CVPHR and a step navigation module. First, the FVT_CVPHR module predicts the position and heading, and then the step navigation module calculates the step vector and updates the UAV's position and orientation, completing the navigation process in a loop.

[0095] The navigation strategy designed in this invention is for deploying lightweight systems on drones. Models, such as Figure 6 As shown in the solid gray box, the step navigation module defines the drone's true position in this step. Predicted location and nominal position . It indicates the drone's actual location, used to extract 3D views from the map; This indicates that the drone's calculated position by the regression model deviates from its actual position; the drone's position from... After taking a step, the point reaches the next real position, and from the real position... The position reached after stepping is .here and This is equivalent to the drone's perceived current location and next destination; nominal position. Its function is to help determine the index number of the area that the drone believes it is currently in.

[0096] Step 8: The UAV uses CVPHR_Naver, carrying the environmental tile feature vector table and navigation sequence information, to conduct cross-view autonomous positioning and navigation tests.

[0097] like Figure 2 As shown, the steps for constructing the cross-perspective dataset in step 1 are as follows:

[0098] Step 1-1: Determine the latitude and longitude range corresponding to the task area on Google Maps and download the remote sensing map. Set appropriate resolution parameters to make The side length has 256 pixels. times ( (For natural numbers), and record at the same time Latitude and longitude values ​​corresponding to the center pixel and the distance value corresponding to each pixel ;

[0099] Steps 1-2: Divide the task area map using a square grid with a side length of 256 pixels, and obtain... A 2D environmental base patch (Bpatch) with a size of 256×256 pixels; remote sensing map All four adjacent base tiles in the map can be combined to form a base block (Bblock). We start with the first base block in the top left corner, using... This indicates the sequence number of the base block, and assigns row and column index numbers to all base blocks. ;

[0100] Steps 1-3: For each baseline block Establish with center The point is the origin, and the horizontal lines at the intersection of the four environmental base maps are... Axis and longitudinal Using a Cartesian coordinate system with axes as the coordinate axes and a unit of 128 pixels, the four adjacent base map indices (top left, bottom left, top right, and bottom right) can be represented as follows: , The relative coordinates of the four base map center points in this coordinate system are [-1,1], [-1,-1], [1,1], and [1,-1], respectively. Connecting these points with black dashed lines forms a... The square area is the reference block. The effective range of all reference blocks is such a square coordinate system region that is seamlessly connected to each other.

[0101] Steps 1-4: In the baseline block The relative coordinates of the center point of a drone view (256×256 pixels) are randomly generated within the valid coordinate range. and heading angle The heading angle converted to sine and cosine form is ,in This stipulates Corresponding coordinate axes The positive direction of the axis corresponds to due east on the remote sensing map. This indicates the sequence number of the drone view sample collected in this block; then according to exist Latitude and longitude coordinates (can be used) and (Calculated) , Calculate the latitude and longitude coordinates and heading angle of the 3D drone view sample in Google 3D Maps and collect and save it as a sample image. ;

[0102] Steps 1-5: Data collection Then, its corresponding , ,as well as View, base map in this reference block The paths of a total of 5 images constitute... The first in The basic information of each sample is collected to complete the collection of a single sample. This indicates the sequence number of the four base maps in the sample;

[0103] Steps 1-6: Following the row-first, column-later process, select the next adjacent reference block. Repeat steps 1-3, 1-4, and 1-5 until all have been traversed. The final result was the CVPHR cross-perspective dataset CVPHR_Dataset, which collected a total of [number missing] data. One sample, including entity images A 2D environment base map and A 3D drone view.

[0104] The structure of the CVPHR model described in step 2 further includes: the environment base map coordinate encoder consists of three cascaded fully connected layers, denoted as fully connected layer 1, fully connected layer 2, and fully connected layer 3, each using the ReLU activation function; the cross-attention module consists of three fully connected layers, two matrix multipliers, one scaling unit, and one softmax unit, the three fully connected layers being denoted as fully connected layer 4, fully connected layer 5, and fully connected layer 6, wherein fully connected layer 4 is multiplied by the output of fully connected layer 5 plus matrix transpose, then scaled and softmax processed, and finally multiplied by the output of fully connected layer 6 to obtain the cross-attention feature; the block similarity guidance module consists of sequentially connected cosine similarity calculation units, The module consists of a softmax unit, a column broadcasting unit, a matrix element multiplier, and a column vector summing unit. The matrix element multiplier is connected to the relative coordinate matrix input of the basis blocks. The position regression module consists of four fully connected layers connected in series, denoted as fully connected layer 7, fully connected layer 8, fully connected layer 9, and fully connected layer 10. Fully connected layers 7, 8, and 9 all use the ReLU activation function, and fully connected layer 10 outputs the predicted coordinates. The direction regression module consists of four fully connected layers connected in series, denoted as fully connected layer 11, fully connected layer 12, fully connected layer 13, and fully connected layer 14. Fully connected layers 11, 12, and 13 all use the ReLU activation function, and fully connected layer 14 outputs the predicted direction.

[0105] The process of constructing the CVPHR model also includes: designing the input and output dimensions, kernel size, stride, and padding size of all convolutional layers, pooling layers, and normalization layers in the VGG backbone network and CSMG feature vector extractor, based on the dimensions of the input image and its coordinates of the CVPHR model, as well as the input and output signal dimensions of other sub-modules and the input and output dimensions of fully connected layers.

[0106] like Figure 3 and Figure 4 Step 2 shown specifically includes:

[0107] Step 2-1: Input including remote sensing map center point latitude and longitude values unit pixel distance value , benchmark block index A 3D drone view and its corresponding set of four reference blocks In the training mode The sample number is indicated by a different subscript symbol than that used in steps 1-5, but the essence is the same. Here, the emphasis is on independent samples that already exist in the dataset, while the subscript used when collecting samples is different. Emphasis on regional division of remote sensing images. Indicates the sequence number of the four base graphs in the sample;

[0108] Step 2-2: In the relative Cartesian coordinate system of the reference area established in Steps 1-3, the relative position coordinates of the four base maps are [-1,1], [-1,-1], [1,1], and [1,-1], respectively. (UAV view) The true values ​​of the center coordinates and the heading angle in sine and cosine form are respectively and ,in , heading angle ;

[0109] Steps 2-3: The VGGNet backbone network is frozen and shares weights when processing each environmental base map or UAV view. The clustering number of the CSMG feature extractor is set to... Each cluster feature dimension takes a value The output CSMG feature dimension Drone View and four environmental base maps D-dimensional feature vectors are extracted from the backbone network respectively. and 4 D-dimensional eigenvectors The latter is a 4×D dimensional matrix;

[0110] Steps 2-4: Encode the coordinates [-1,1], [-1,-1], [1,1], and [1,-1] of the four base maps using an environment base map coordinate encoder to obtain four D-dimensional coordinate encoding vectors. The four D-dimensional enhanced neighborhood feature vectors are obtained by summing these features with the corresponding four base graph feature vectors. ; use and They represent the first The weight and bias matrices of each fully connected layer, where and Let represent the input and output dimensions of the fully connected layer, respectively, and set... and Let represent the input and output vectors of the fully connected layer, respectively. The calculation process of adding the ReLU function to the fully connected layer in this step can be written as follows: ;

[0111] Steps 2-5: Convert the UAV view feature vector (Equivalent to query Q) and enhanced neighborhood feature vectors (Equivalent to key values ​​K and V) are input into the cross-attention module to obtain... Cross-feature vectors of dimension ,in The compression ratio is set to 1, 2, or 4 to adjust the output dimension of the cross-attention module; simultaneously, and 4 base map feature vectors The input block similarity guidance module yields a 2D weighted position estimation vector. Consistent with the assumptions made in steps 2-4, the calculation process for the fully connected layer without the ReLU function in this step is written as follows: ;

[0112] Steps 2-6: Convert the UAV view feature vector , Cross-feature vectors of dimension and a 2D weighted position estimation vector Get the connection dimensional fusion feature vector ; The predicted positions are obtained by inputting the data into the position regression module and the heading regression module, respectively. and predicted heading angle The calculation of the fully connected layer is consistent with steps 2-4 or 2-5.

[0113] Steps 2-7: Using remote sensing maps center point latitude and longitude values unit pixel distance value and benchmark blocks index The latitude and longitude of the center point of the reference block can be calculated. And then according to and Calculate the absolute latitude and longitude of the drone ,according to Calculate the heading angle of the drone ;

[0114] Steps 2-8: Combining the input image size and coordinate dimensions of the CVPHR model, as well as the functions of each sub-module of the model, the input and output dimensions of each sub-module, the size and parameters of the fully connected layer, convolutional layer, pooling layer, and normalization layer inside the sub-module are reasonably designed. This allows the acquisition of feature vectors and fused features from the UAV view and the environmental base map, ultimately regressing the latitude and longitude position and heading angle of the UAV.

[0115] Furthermore, in steps 2-8, the network parameters of the core functional modules of the CVPHR model are shown in Table 1, where bias=True indicates that the fully connected layer contains a bias; the VGGNet backbone network and CSMG feature extractor network are open source models. In the CVPHR model, the VGGNet backbone network is used with frozen parameters. Table 2 gives the default VGGNet backbone network parameters, and Table 3 gives the default CSMG feature extractor network parameters. The inplace of the ReLU function indicates that the calculation is performed in place without adding extra memory to store the calculation results.

[0116] Table 1

[0117]

[0118] Table 2

[0119]

[0120] Table 3

[0121]

[0122] Step 3 specifically includes:

[0123] Step 3-1: Set the ratio of training set, validation set, and test set to 7:2:1, and set the batch hyperparameters for training the model. Initial learning rate Training rounds ;

[0124] Step 3-2: Periodically train and validate the CVPHR model using the training and validation sets, and distribute the samples according to... Input the data into the CVPHR model in batches, and calculate the predicted location of the UAV view according to steps 2-1 to 2-6. and predicted course ;

[0125] Step 3-3: Use the predicted position obtained in step 3-2 and predicted course with position truth value True value of heading angle The position error and heading error are calculated separately. The Smooth L1 Loss function is selected as the loss function to calculate the loss value. The Adaptive Moment Estimator Optimizer (Adam) is used as the gradient descent algorithm to realize backpropagation and iterative training of the CVPHR model.

[0126] Steps 3-4: The learning rate optimization method uses a plateau-period learning rate decay strategy (ReduceLROnPlateau). Each training epoch uses the training set for training and the validation set for validation. The validation results are used to adjust the variable learning rate. Training can be stopped when the loss function value does not exceed a set range for 10 consecutive epochs, or when training is completed within the total number of training epochs. The trained CVPHR model weight file is then obtained. After training, the trained model parameters are saved to the weight file. .

[0127] Step 4 specifically includes:

[0128] When testing using the CVPHR model, the weight file described in steps 3-4 will be used. Load the data into the CVPHR model, then input each sample from the test set into the CVPHR model, and calculate the predicted location of the UAV view according to steps 2-1 to 2-6. and predicted course The predicted results are compared with the true values ​​of position and heading labeled for each sample in the test set, and the regression prediction accuracy of position and heading is calculated; these samples are unseen samples of the model, which can test the model's generalization ability; due to the acquisition of remote sensing data... Figure 2 When the D base map sample contains latitude and longitude information, the CVPHR model can obtain the latitude and longitude position and heading angle of the UAV with high accuracy. Figure 5a To predict the absolute position error distribution map and statistical results, Figure 5b This is a map showing the distribution of heading angle errors and the statistical results.

[0129] Step 5 specifically includes:

[0130] The segments obtained in steps 1-2 A 2D environmental baseline map of size 256×256 pixels is passed through the VGG16 backbone network and CSMG feature extractor one by one according to its row and column position index to obtain... D-dimensional feature vectors of an environmental base map Serial number is The row and column indices of the feature vector are (Note that if base map numbers are used here) As a subscript, when representing a group of base maps, this subscript sometimes uses the base block number plus the numbers of the four base maps in this group. (Represented by) the relative position of the reference tile in the remote sensing map. The rows and columns correspond to latitude lines and longitude lines, respectively. These row and column index numbers and the corresponding feature vectors constitute the Environmental Tile Feature Vector Table (FVT). Each continuous map only needs to be processed once.

[0131] like Figure 6 As shown in the flowchart within the dashed box, step 6, building the FVT_CVPHR, includes the following steps:

[0132] Step 6-1: Input including remote sensing map center point latitude and longitude values unit pixel distance value , benchmark block index and a 3D drone view ,in Indicates the reference area number. Indicates the drone view frame number. From the initial state Initially, the algorithm acquires a 3D drone view from the starting point using Google Maps 3D mode. and from Obtain its corresponding reference block The reference block consists of a set of four adjacent reference tiles. , ,in Indicates the first The serial numbers of the four base maps in this group;

[0133] Step 6-2: In the relative Cartesian coordinate system of the reference area established in Step 1-3, the relative position coordinates of the four base maps are [-1,1], [-1,-1], [1,1], and [1,-1], respectively. (UAV view) The true values ​​of the center coordinates and the heading angle in sine and cosine form are respectively and ,in , heading angle ;

[0134] Step 6-3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require After inputting the VGGNet backbone network and adding the CSMG feature extractor branch, we obtain the D-dimensional feature vector of the UAV view. Here, the cluster number of the CSMG feature extractor is set to... Each cluster feature dimension takes a value The output CSMG feature dimension ;

[0135] Step 6-4: Based on the baseline block The row and column index numbers can be used to obtain the baseline tile. row and column index numbers (The row and column index numbers of the base map block here are different from those in step 5, mainly emphasizing that this base map belongs to the first...) The first benchmark block (a base map), and its corresponding feature vector is obtained by querying the Environmental Patch Feature Vector Table (FVT) based on its row and column index numbers. , ;

[0136] Step 6-5: Utilize the feature vectors of the UAV's 3D view The coordinates [-1,1], [-1,-1], [1,1], and [1,-1] of the four queried environmental base maps and their eigenvectors. , The predicted position is obtained by following steps 2-4, 2-5, and 2-6. and predicted heading angle .

[0137] like Figure 7 As shown, step 7 specifically includes:

[0138] Step 7-1: On the mission remote sensing map Mark the starting and ending points, as well as the waypoints visited, and save their latitude and longitude as a navigation waypoint sequence. ,in and These are the starting point coordinates and the ending point coordinates, respectively. For process waypoints, among which Indicates the waypoint number.

[0139] Step 7-2: Initial Navigation State. The UAV locates itself on the Google 3D remote sensing map based on its starting coordinates, and prepares the initial state of the FVI_CVPHR model according to Step 6-1. In the step navigation module, the UAV's initial true position... =Initial nominal position =Drone starting coordinates Initialize the drone's current real position and nominal position as follows: and Initialize target waypoint ;

[0140] Step 7-3: Based on the actual location of the drone Get a 3D view of a drone from Google Maps , Indicates the current step number. Based on the drone's nominal position. From the environmental tile feature vector table Obtain its reference area A set of feature vectors corresponding to environmental blocks (The subscript form of the reference area here is different from that in steps 6-1 and 6-4, but is essentially the same, with the purpose of keeping it consistent with the number of UAV flight steps; the feature vector symbol of a set of environmental reference blocks is different from that in step 6-4, but is essentially the same. Here, adding 's' to the end of the variable indicates a set of feature vectors.)

[0141] Step 7-4: Use the FVI_CVPHR module to perform drone positioning and orientation, from and Obtain the predicted location and prediction angle That is, performing steps 6-2 to 6-5, this step can be represented as ;

[0142] Step 7-5: Based on the current target waypoint and the predicted location of the drone Calculate the drone's next step vector Then, the angle difference between the UAV's heading angle and step vector is calculated. Used to adjust the fuselage orientation;

[0143] Step 7-6: The UAV moves forward one step based on the step vector and calculates its true position after the step. and nominal position Then calculate the distances between the drone and the target waypoint and destination. and Update the drone's real and nominal locations. , ;

[0144] Step 7-7: Distance With the arrival threshold The comparison is performed. If the value is greater than the threshold, it is further determined whether the flight has gone out of bounds or exceeded the step limit. If so, it means that the navigation has failed. Otherwise, steps 7-3 to 7-6 are repeated. If the target waypoint is reached, the target waypoint is updated and the process proceeds to step 7-8.

[0145] Steps 7-8: Distance With the arrival threshold The comparison is performed. If the value is greater than the threshold, it is further determined whether the plane has flown out of bounds or exceeded the step limit. If so, it means that the navigation has failed. Otherwise, steps 7-3 to 7-6 are repeated. If the value is less than the threshold, it means that the destination has been reached.

[0146] Step 8 specifically includes:

[0147] like Figure 8 As shown, after completing the construction, training, and testing of CVPHR, the FVT is generated and FVT_CVPHR is built, and finally CVPHR_Naver is constructed. Based on this, according to the latitude and longitude information of the navigation task's start and end points, a navigation path waypoint sequence is generated based on the task's remote sensing map (this can be done by manually specifying the waypoint coordinates, randomly generating coordinates, or using specific ground feature coordinates, etc.). The UAV carries the navigation path waypoint sequence table. Using the Environmental Feature Vector Table (FVT) and CVPHR_Naver, cross-view autonomous navigation flight tests were conducted in a Google 3D map simulation environment. The simulation results are as follows: Figure 9 As shown, Figure 9 The image shows the UAV completing an autonomous navigation flight path of approximately 1 km in 10-meter increments, following a sequence of waypoints.

[0148] The time consumption test of the algorithm of this invention is based on the following configuration:

[0149] a) Hardware environment: Intel Core i9-9880H CPU@2.30GHz, NVIDIA Quadro RTX 4000 GPU (8GB VRAM, 2560 CUDA cores), 64GB DDR4-2667 memory, 1TB SCSI HDD;

[0150] b) Software environment: Windows 10 64bit, MSVC (compilation parameter -O3), PyTorch 2.0.1, CUDA Toolkit 11.8;

[0151] c) Test data: Download a 5120×5120 pixel 2D remote sensing map from Google Maps, collect 18050 256×256 pixel 3D view drone views in the same area to build a dataset, complete model training and conduct drone point-to-point navigation tests.

[0152] d) Time measurement: Using Python time.time() to time the process, after 5 warm-up runs, the drone took 70 steps, and the average navigation step processing time was 35.7ms.

[0153] Example 2

[0154] This embodiment relates to an autonomous navigation method and apparatus for unmanned aerial vehicles (UAVs) based on a cross-view orientation regression network. It is characterized by including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the autonomous navigation method for UAVs based on a cross-view orientation regression network as described in Embodiment 1.

[0155] Example 3

[0156] This embodiment relates to a computer-readable storage medium storing a program that, when executed by a processor, implements the UAV autonomous navigation method based on a cross-view orientation regression network as described in Embodiment 1.

Claims

1. An autonomous navigation method for unmanned aerial vehicles (UAVs) based on cross-view orientation regression networks, comprising the following steps: Step 1: Collect adjacent environmental tile samples and 3D UAV view samples using remote sensing maps to construct a cross-view dataset containing orientation information, which is then divided into training set, validation set and test set; Step 2: Construct a UAV position and heading regression model (CVPHR) for cross-view scenarios; Step 3: Use the training set and validation set to periodically train and validate the CVPHR model, and obtain the final CVPHR model weight file; Step 4: Input the test set into the final CVPHR model and test the regression performance of the CVPHR model using test samples not found in the test set; Step 5: Use the trained CVPHR model to infer the environment tiles and construct the environment tile feature vector table FVT from the output set of environment tile feature vectors and their corresponding row and column indices. Step 6: Transform CVPHR into a UAV position and heading regression model FVT_CVPHR under FVT lookup mode; FVT_CVPHR uses the feature vectors corresponding to a set of two-dimensional environment tiles arranged in a square matrix and a three-dimensional UAV view of the same area to predict the absolute position and heading angle of the UAV; FVT_CVPHR searches for a set of feature vectors from FVT based on the index information of the environment tile where the UAV is located. Step 7: Build an autonomous navigation model CVPHR_Naver for UAVs based on a cross-view orientation regression network. The FVT_CVPHR module predicts the position and heading, and the step navigation module calculates the step vector and updates the position and orientation of the UAV. The navigation process is completed in a loop. Step 8: The UAV uses CVPHR_Naver, carrying FVT and navigation sequence information, to conduct cross-view autonomous positioning and navigation tests.

2. The UAV autonomous navigation method based on cross-view orientation regression network as described in claim 1, characterized in that, The steps for constructing the cross-perspective dataset in step 1 are as follows: Step 1-1: Download the side length based on the latitude and longitude range of the task area. Square remote sensing map of pixels Simultaneously record the latitude and longitude values ​​of its center pixel. unit pixel distance value ; Step 1-2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require Divided into grids with side lengths of Small square environment tiles of pixels Any set of square-shaped environmental tiles forms a square environmental block. Each environment block has a unique row index and column index number; Steps 1-3: For each environmental block in Steps 1-2, establish a system with the block center as the origin and horizontal... Axis and Vertical The axes are relative Cartesian coordinate systems. The line connecting the center points of the outermost environment tiles of each environment block forms a square effective range, and the square effective ranges of adjacent environment blocks are seamlessly connected. Steps 1-4: Define the due east direction on the remote sensing map as... And corresponding In the positive direction of the axis, within the effective coordinate range of each environment block, generate A set of samples with random latitude and longitude coordinates and random heading angles is used to collect data in a three-dimensional remote sensing map. A sample of drone views ; Steps 1-5: Following the left-to-right and top-to-bottom process, repeat steps 1-3 and 1-4 for each environmental block until the remote sensing map has been traversed. All environmental blocks are used to ultimately obtain the CVPHR cross-view dataset CVPHR_Dataset.

3. The UAV autonomous navigation method based on cross-view orientation regression network according to claim 1, characterized in that, In the CVPHR model described in step 2, the environment tile coordinate encoder encodes the relative coordinates of a set of square-arrayed two-dimensional environment tiles into high-dimensional vectors; the VGG backbone network is used to extract feature maps of environment tiles or UAV views; the CSMG feature vector extractor is used to extract feature vectors from the feature maps; and the cross-attention module is used to extract cross-view correlation features between UAV views and environment tiles. The block similarity guidance module is used to estimate the similarity between the UAV view and a set of environmental tiles and convert it into a weighted coordinate vector; the position regression module and the orientation regression module predict the UAV's position and heading angle from the fused features, respectively.

4. The UAV autonomous navigation method based on cross-view orientation regression network as described in claim 3, characterized in that, Step 2 specifically includes: Step 2-1: For a specific environmental block, collect a 3D drone view of that environmental block. All environment tiles Its relative coordinates are used as input; Step 2-2: Extract UAV view feature vectors using the VGGNet backbone network and the CSMG feature extractor, respectively. and environmental tile feature vectors ; Steps 2-3: Calculate the relative coordinate encoding vector of the environment tile. Adding the corresponding environmental tile feature vectors yields the enhanced neighborhood feature vector. ; Steps 2-4: First, input the UAV view feature vector and the enhanced neighborhood feature vector into the cross-attention module to obtain the cross feature vector. Then, the block similarity guidance module is used to obtain a two-dimensional weighted position estimation vector. ; Steps 2-5: Concatenate the UAV view feature vector, cross feature vector, and weighted position estimation vector to obtain the fused feature vector. ; Steps 2-6: Input the fused feature vector into the position regression module and the heading regression module respectively to obtain the predicted relative position and the predicted heading angle, and then calculate the absolute latitude and longitude position and heading angle of the UAV.

5. The UAV autonomous navigation method based on cross-view orientation regression network as described in claim 4, characterized in that, In step 2, the VGGNet backbone network and CSMG feature extractor network included in the core functional module of the CVPHR model are open source models. The VGGNet backbone network in the CVPHR model adopts the parameter freezing method. The parameters of the VGGNet backbone network, the parameters of the CSMG feature extractor network, and other sub-modules are all set according to the designed structure and parameters.

6. The UAV autonomous navigation method based on cross-view orientation regression network as described in claim 4, characterized in that, Step 3 specifically includes: Step 3-1: Divide the dataset into training, validation and test sets according to the proportions, and set the hyperparameters of the training model, such as batch size, initial learning rate and number of training rounds; Step 3-2: Periodically train and validate the CVPHR model using the training and validation sets. Input the samples into the CVPHR model in batches and calculate the predicted location of the UAV according to steps 2-1 to 2-7. and predicted heading angle ; Step 3-3: Use the predicted position obtained in step 3-2 and predicted heading angle with position truth value True value of heading angle Position error and heading error are calculated separately. The loss value is calculated using the Smooth L1 Loss function, and the adaptive moment estimator optimizer Adam is used as the gradient descent algorithm to realize backpropagation and iterative training of the CVPHR model. Steps 3-4: The learning rate optimization method uses the plateau learning rate decay strategy ReduceLROnPlateau. In each training round, the training set is used for training and the validation set is used for validation. The validation results are used to adjust the variable learning rate. After completing the total number of training rounds, the weight file of the trained CVPHR model is obtained.

7. The UAV autonomous navigation method based on cross-view orientation regression network as described in claim 6, characterized in that, Step 4 specifically includes: when using the CVPHR model for testing, loading the weight file described in steps 3-4 into the CVPHR model, then inputting each sample in the test set into the CVPHR model, and calculating the predicted position of the UAV view according to steps 2-1 to 2-6. and predicted course The prediction results are compared with the true values ​​of position and heading labeled for each sample in the test set, and the regression prediction accuracy of position and heading is calculated.

8. The UAV autonomous navigation method based on cross-view orientation regression network as described in claim 1, characterized in that, Step 5 specifically includes: taking all the environmental map tiles obtained in steps 1-2, and passing them one by one through the VGG16 backbone network and CSMG feature extractor according to their row and column position indices to obtain their corresponding feature vectors. The rows and columns correspond to latitude lines and longitude lines, respectively. The row and column index numbers of these environmental map tiles and their corresponding feature vectors are used to construct an environmental map tile feature vector table FVT. Each continuous map is processed once.

9. The UAV autonomous navigation method based on cross-view orientation regression network as described in claim 4, characterized in that, Step 6, building the FVT_CVPHR structure, includes: the input has three parallel branches, namely the environment map feature vector table lookup module, the environment base map coordinate encoder, and the VGG backbone network connected to the CSMG feature vector extractor, which are then connected to the cross attention module and the block similarity guidance module, and finally output by the position regression module and the orientation regression module; The specific steps to set up FVT_CVPHR include: Step 6-1: Collect remote sensing maps Latitude and longitude values ​​of the center pixel unit pixel distance value Environmental Blocks Index; Step 6-2: Based on the latitude, longitude, and heading angle of the current location of the UAV, acquire the UAV view from the 3D remote sensing map; Step 6-3: Input the UAV view into the VGGNet backbone network plus the CSMG feature extractor branch to obtain the feature vector corresponding to the UAV view. ; Step 6-4: Obtain the row and column index numbers of the corresponding set of environment tiles based on the row and column index numbers of the environment tiles. Then, retrieve the corresponding set of feature vectors from the environment tile feature vector table FVT based on these row and column index numbers. ; Step 6-5: Utilize the feature vectors of the drone view The relative coordinates and feature vectors of a set of adjacent environmental tiles retrieved. The predicted position and predicted heading angle are obtained by following steps 2-3 to 2-6.

10. The UAV autonomous navigation method based on cross-view orientation regression network as described in claim 9, characterized in that, Step 7 specifically includes: Step 7-1: On the mission remote sensing map Mark the latitude and longitude of the starting point, waypoints and the destination, and save them as a navigation waypoint sequence. In the 3D remote sensing map, initialize the UAV's initial true position = initial nominal position = coordinates of the waypoint sequence starting point; Step 7-2: Set the initial state of the FVT_CVPHR model and obtain the UAV view from the 3D remote sensing map based on the UAV's actual position. Based on the UAV's nominal location and its environmental block index number, a set of environmental block feature vectors is obtained from the Feature Vector Table (FVT). ; Step 7-3: Perform drone positioning and orientation according to steps 6-2 to 6-5, from the drone view. and a set of environmental block feature vectors Obtain the predicted position and predicted heading angle; Step 7-4: Calculate the UAV step vector based on the predicted UAV position and the current target waypoint, and adjust the UAV orientation by predicting the angular difference between the heading angle and the step vector direction; Step 7-5: The drone moves forward one step and calculates its real position and nominal position. It then calculates the distances between its real position and the target waypoint and destination, respectively. Step 7-6: Compare the waypoint threshold or destination threshold to determine whether the target waypoint or destination has been reached. If it is below the threshold, it means that the destination has been reached. If it is above the threshold, it means that the flight has gone out of bounds or exceeded the step limit. If so, it means that the navigation has failed. Otherwise, repeat steps 7-3 to 7-6 until the destination is reached.

Images