High-altitude positioning method based on graph neural network

Abstract

The invention discloses a high-altitude positioning method based on a graph neural network, and relates to the technical field of high-altitude positioning, and the method comprises the steps: employing an unmanned aerial vehicle equipped with multiple sensors for data collection, enabling the unmanned aerial vehicle to fly according to a predetermined route, collecting a high-resolution image, point cloud data and attitude information, carrying out the denoising processing of image data, and obtaining a high-resolution image; filtering the point cloud data, performing attitude estimation by using IMU data to obtain preprocessed high-quality environment data, then extracting SIFT feature points by using preprocessed image data to establish a neural network model, then collecting multiple groups of environment feature maps with known position information, constructing a training data set, training the neural network model, and obtaining a neural network model; and acquiring new high-altitude environment data by using the unmanned aerial vehicle, constructing a corresponding environment feature map, performing position prediction on nodes in the feature map by using the trained map neural network model, and obtaining a final high-altitude positioning result by fusing predicted position information of all the nodes.

Description

Technical Field

[0001] The present invention relates to the field of high-altitude positioning technology, and in particular to a high-altitude positioning method based on a graph neural network. Background Art

[0002] High-altitude positioning has a wide range of applications in areas such as drone navigation, building exterior wall inspection, and natural disaster monitoring. Traditional high-altitude positioning methods usually rely on technologies such as GPS, lidar, and visual SLAM. However, these methods are easily affected by factors such as satellite signal obstruction and changes in lighting conditions in complex environments (such as between high-rise buildings and forest-covered areas), resulting in reduced positioning accuracy.

[0003] In recent years, graph neural networks, as an emerging deep learning model, have shown strong capabilities in processing graph structured data. Graph neural networks can efficiently capture potential patterns and relationships in complex graph structures through the exchange and fusion of information between nodes and their neighbors. Therefore, introducing graph neural networks into the field of high-altitude positioning is expected to improve positioning accuracy and robustness. Summary of the invention

[0004] The purpose of the present invention is to overcome the above-mentioned deficiencies in the prior art and to provide a high-altitude positioning method based on graph neural network, which can improve the accuracy and robustness of high-altitude positioning.

[0005] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:

[0006] A high-altitude positioning method based on graph neural network includes the following steps:

[0007] S1. Data collection and preprocessing: A multi-sensor drone is used for data collection. The drone flies along a predetermined route to collect high-resolution images, point cloud data, and attitude information. The image data is denoised, the point cloud data is filtered, and the IMU data is used for attitude estimation to obtain high-quality environmental data after preprocessing.

[0008] S2, environmental feature map construction, using the preprocessed image data, extract SIFT feature points, and calculate the Euclidean distance between adjacent feature points to construct a visual feature map, using point cloud data, extract geometric feature points, and calculate the spatial distance between feature points to construct a geometric feature map, and fuse the visual feature map and the geometric feature map to obtain the final environmental feature map;

[0009] S3. Build a graph neural network model, which includes a graph convolution layer, an attention mechanism layer, and a fully connected layer; wherein the IMU data is the data collected by the inertial measurement unit.

[0010] Among them, the graph convolution layer is used to extract the feature information of nodes and their neighbors, and Chebyshev polynomial approximation is used to implement efficient graph convolution operations;

[0011] The attention mechanism layer is used to dynamically adjust the influence weight of different neighbor nodes on the current node, and calculates the attention weight by learning the similarity of node pairs;

[0012] The fully connected layer is used to output the final position information of the node, and uses the ReLU activation function and Dropout regularization to improve the generalization ability of the model;

[0013] S4, model training, collect multiple sets of environmental feature maps with known location information, build a training data set, use the Adam optimizer and MSE loss function to train the graph neural network model, and gradually reduce the node position prediction error by iteratively updating the model parameters until the model converges;

[0014] S5. High-altitude positioning: Use drones to collect new high-altitude environmental data and construct the corresponding environmental feature map. Use the trained graph neural network model to predict the positions of the nodes in the feature map. By fusing the predicted position information of all nodes, the final high-altitude positioning result of the drone is obtained.

[0015] Furthermore, filtering the point cloud data in step S1 and using the IMU data for attitude estimation includes the following steps:

[0016] The collected point cloud data is transformed into coordinates and invalid points are removed. The statistical filtering method is used to calculate the distance mean and standard deviation of the neighborhood points of each point, and the points that deviate too much from the mean are removed to reduce noise interference. The point cloud data is divided into several voxels, and the points in each voxel are downsampled to reduce the amount of data and improve calculation efficiency.

[0017] Use moving average filtering or Gaussian filtering to smooth the point cloud data to further remove noise and small fluctuations, and remove zero bias, temperature compensate and calibrate the IMU data;

[0018] Using the pre-processed IMU data, the attitude change of the aircraft is calculated through integration operation;

[0019]

[0020] Among them, q k represents the attitude quaternion at the kth moment, w k represents the angular velocity at the kth moment, and Δt represents the time interval. The calculated result is corrected by combining the magnetometer and GPS data to improve the accuracy of attitude estimation.

[0021] Furthermore, in step S2, the preprocessed image data is used to extract SIFT feature points, and the Euclidean distance between adjacent feature points is calculated. Constructing a visual feature map includes the following steps:

[0022] Detect the extreme points of the image in different scale spaces. These extreme points are potential feature points.

[0023] Accurately locate the detected extreme points and remove those with low contrast and strong edge response;

[0024] Assign one or more directions to each key point to ensure the rotation invariance of the feature;

[0025] In the local neighborhood of the key point, the gradient direction histogram is calculated to form a feature descriptor; after the SI FT feature points are extracted, the Euclidean distance between adjacent feature points is calculated;

[0026] According to the calculated Euclidean distance, an undirected weighted graph G = (V, E) is constructed, where:

[0027] V is a set of nodes, each node corresponds to a SIFT feature point;

[0028] E is a set of edges, each edge corresponds to the connection between two adjacent feature points, and the weight of the edge is the Euclidean distance between the feature points.

[0029] Furthermore, the specific steps for calculating the Euclidean distance between adjacent feature points are as follows:

[0030] For each feature point, find other feature points in its neighborhood. The neighborhood here can be defined as a fixed radius range or dynamically adjusted according to the number of feature points; calculate the feature point p i and p j The Euclidean distance between them is given by the following formula:

[0031]

[0032] Among them, (x i -x j ), (y i -y j ) are the feature points p i and p j The coordinates of .

[0033] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0034] Based on the trained graph neural network model, new high-altitude image data is input, and after the above-mentioned preprocessing, feature point extraction, Euclidean distance calculation, visual feature map construction and other steps, the high-altitude positioning result is finally output through the graph neural network. The present invention can improve the accuracy and robustness of high-altitude positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0036] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0037] This embodiment provides a high-altitude positioning method based on a graph neural network. Figure 1 As shown, the following steps are included:

[0038] S1. Data collection and preprocessing: A multi-sensor drone is used for data collection. The drone flies along a predetermined route to collect high-resolution images, point cloud data, and attitude information. The image data is denoised, the point cloud data is filtered, and the IMU data is used for attitude estimation to obtain high-quality environmental data after preprocessing.

[0039] S2, environmental feature map construction, using the preprocessed image data, extracting SI FT feature points, and calculating the Euclidean distance between adjacent feature points to construct a visual feature map, using point cloud data, extracting geometric feature points, and calculating the spatial distance between feature points to construct a geometric feature map, and fusing the visual feature map and the geometric feature map to obtain the final environmental feature map;

[0040] S3. Build a graph neural network model, which includes a graph convolution layer, an attention mechanism layer, and a fully connected layer.

[0041] Among them, the graph convolution layer is used to extract the feature information of nodes and their neighbors, and Chebyshev polynomial approximation is used to implement efficient graph convolution operations;

[0042] The attention mechanism layer is used to dynamically adjust the influence weight of different neighbor nodes on the current node, and calculates the attention weight by learning the similarity of node pairs;

[0043] The fully connected layer is used to output the final position information of the node, and uses the ReLU activation function and Dropout regularization to improve the generalization ability of the model;

[0044] S4, model training, collect multiple sets of environmental feature maps with known location information, build a training data set, use the Adam optimizer and MSE loss function to train the graph neural network model, and gradually reduce the node position prediction error by iteratively updating the model parameters until the model converges;

[0045] S5. High-altitude positioning: Use drones to collect new high-altitude environmental data and construct the corresponding environmental feature map. Use the trained graph neural network model to predict the positions of the nodes in the feature map. By fusing the predicted position information of all nodes, the final high-altitude positioning result of the drone is obtained.

[0046] Based on the trained graph neural network model, new high-altitude image data is input, and after the above-mentioned preprocessing, feature point extraction, Euclidean distance calculation, visual feature map construction and other steps, the high-altitude positioning result is finally output through the graph neural network. The present invention can improve the accuracy and robustness of high-altitude positioning.

[0047] The above embodiment is further optimized, and filtering the point cloud data in step S1, and using the IMU data for attitude estimation includes the following steps:

[0048] The collected point cloud data is transformed into coordinates and invalid points are removed. The statistical filtering method is used to calculate the distance mean and standard deviation of the neighborhood points of each point, and the points that deviate too much from the mean are removed to reduce noise interference. The point cloud data is divided into several voxels, and the points in each voxel are downsampled to reduce the amount of data and improve calculation efficiency.

[0049] Use moving average filtering or Gaussian filtering to smooth the point cloud data to further remove noise and small fluctuations, and remove zero bias, temperature compensate and calibrate the IMU data;

[0050] Using the pre-processed IMU data, the attitude change of the aircraft is calculated through integration operation;

[0051]

[0052] Among them, q k represents the attitude quaternion at the kth moment, w k represents the angular velocity at the kth moment, and Δt represents the time interval. The calculated result is corrected by combining the magnetometer and GPS data to improve the accuracy of attitude estimation.

[0053] The above embodiment is further optimized. In step S2, the preprocessed image data is used to extract SI FT feature points, and the Euclidean distance between adjacent feature points is calculated. The visual feature map is constructed by the following steps:

[0054] Detect the extreme points of the image in different scale spaces. These extreme points are potential feature points.

[0055] Accurately locate the detected extreme points and remove those with low contrast and strong edge response;

[0056] Assign one or more directions to each key point to ensure the rotation invariance of the feature;

[0057] In the local neighborhood of the key point, the gradient direction histogram is calculated to form a feature descriptor; after the SIFT feature points are extracted, the Euclidean distance between adjacent feature points is calculated;

[0058] According to the calculated Euclidean distance, an undirected weighted graph G = (V, E) is constructed, where:

[0059] V is a set of nodes, each node corresponds to a SIFT feature point;

[0060] E is a set of edges, each edge corresponds to the connection between two adjacent feature points, and the weight of the edge is the Euclidean distance between the feature points.

[0061] The above embodiment is further optimized, and the specific steps of calculating the Euclidean distance between adjacent feature points are as follows:

[0062] For each feature point, find other feature points in its neighborhood. The neighborhood here can be defined as a fixed radius range or dynamically adjusted according to the number of feature points; calculate the feature point p i and p j The Euclidean distance between them is given by the following formula:

[0063]

[0064] Among them, (x i -x j ), (y i -y j ) are the feature points p i and p j The coordinates of .

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A high-altitude positioning method based on graph neural network, characterized in that: The steps include: S1. Data collection and preprocessing: A multi-sensor drone is used for data collection. The drone flies along a predetermined route to collect high-resolution images, point cloud data, and attitude information. The image data is denoised, the point cloud data is filtered, and the IMU data is used for attitude estimation to obtain high-quality environmental data after preprocessing. S2, environmental feature map construction, using the preprocessed image data, extract SIFT feature points, and calculate the Euclidean distance between adjacent feature points to construct a visual feature map, using point cloud data, extract geometric feature points, and calculate the spatial distance between feature points to construct a geometric feature map, and fuse the visual feature map and the geometric feature map to obtain the final environmental feature map; S3. Build a graph neural network model, which includes a graph convolution layer, an attention mechanism layer, and a fully connected layer. Among them, the graph convolution layer is used to extract the feature information of nodes and their neighbors, and Chebyshev polynomial approximation is used to implement efficient graph convolution operations; The attention mechanism layer is used to dynamically adjust the influence weight of different neighbor nodes on the current node, and calculates the attention weight by learning the similarity of node pairs; The fully connected layer is used to output the final position information of the node, and uses the ReLU activation function and Dropout regularization to improve the generalization ability of the model; S4, model training, collect multiple sets of environmental feature maps with known location information, build a training data set, use the Adam optimizer and MSE loss function to train the graph neural network model, and gradually reduce the node position prediction error by iteratively updating the model parameters until the model converges; S5. High-altitude positioning: Use drones to collect new high-altitude environmental data and construct the corresponding environmental feature map. Use the trained graph neural network model to predict the positions of the nodes in the feature map. By fusing the predicted position information of all nodes, the final high-altitude positioning result of the drone is obtained.

2. The high-altitude positioning method based on graph neural network according to claim 1 is characterized in that: In step S1, the point cloud data is filtered and the posture estimation is performed using the IMU data, which includes the following steps: The collected point cloud data is transformed into coordinates and invalid points are removed. The statistical filtering method is used to calculate the distance mean and standard deviation of the neighborhood points of each point, and the points that deviate too much from the mean are removed to reduce noise interference. The point cloud data is divided into several voxels, and the points in each voxel are downsampled to reduce the amount of data and improve calculation efficiency. Use moving average filtering or Gaussian filtering to smooth the point cloud data to further remove noise and small fluctuations, and perform zero bias removal, temperature compensation and calibration operations on the IMU data; Using the pre-processed IMU data, the attitude change of the aircraft is calculated through integration operation; Among them, q k represents the attitude quaternion at the kth moment, w k represents the angular velocity at the kth moment, and Δt represents the time interval. The calculated result is corrected by combining the magnetometer and GPS data to improve the accuracy of attitude estimation.

3. The high-altitude positioning method based on graph neural network according to claim 1 is characterized in that: In step S2, the preprocessed image data is used to extract SIFT feature points, and the Euclidean distance between adjacent feature points is calculated. The construction of the visual feature map includes the following steps: Detect the extreme points of the image in different scale spaces. These extreme points are potential feature points. Accurately locate the detected extreme points and remove those with low contrast and strong edge response; Assign one or more directions to each key point to ensure the rotation invariance of the feature; In the local neighborhood of the key point, the gradient direction histogram is calculated to form a feature descriptor; after the SIFT feature points are extracted, the Euclidean distance between adjacent feature points is calculated; According to the calculated Euclidean distance, an undirected weighted graph G = (V, E) is constructed, where: V is a set of nodes, each node corresponds to a SIFT feature point; E is a set of edges, each edge corresponds to the connection between two adjacent feature points, and the weight of the edge is the Euclidean distance between the feature points.

4. The high-altitude positioning method based on graph neural network according to claim 3 is characterized in that: The specific steps for calculating the Euclidean distance between adjacent feature points are as follows: For each feature point, find other feature points in its neighborhood. The neighborhood here can be defined as a fixed radius range or dynamically adjusted according to the number of feature points; calculate the feature point p i and p j The Euclidean distance between them is given by the following formula: Among them, (x i -x j ), (y i -y j ) are the feature points p i and p j The coordinates of .

Images