Aircraft visual navigation method based on double matching and dynamic satellite map loading

Abstract

The invention discloses an aircraft visual navigation method based on double matching and dynamic satellite map loading. The method comprises the following steps: firstly, preparing a satellite slice database offline for dynamic loading during subsequent online matching; when visual matching positioning is carried out, feature extraction and matching are carried out on a current frame and a satellite map which is dynamically spliced and loaded according to course angle and height information, and latitude and longitude positioning information P1 is obtained through calculation; meanwhile, the current frame and the previous frame are subjected to same feature extraction and matching processing, and latitude and longitude positioning information P2 is obtained through calculation; and finally, inputting the double-matching positioning result into a Bayesian fusion module for fusion optimization to obtain a final positioning result of the current frame. The method provided by the invention effectively solves the problem that the traditional single matching method is easy to cause unstable positioning and failure, and greatly improves the precision and robustness of matching positioning.

Description

Technical Field

[0001] The invention belongs to the technical field of aircraft, and in particular relates to an aircraft visual navigation method based on double matching and dynamic satellite image loading. Background Art

[0002] High-precision positioning and navigation are key technologies for autonomous flight of drones and other aircraft. At present, common navigation solutions mainly rely on global positioning systems (such as the US GPS and China's BeiDou system), which are widely used for their high precision, all-weather, flexible and convenient features. However, this type of navigation method is heavily dependent on satellite signals, and has limitations such as weak autonomy, limited anti-electromagnetic interference capabilities, and inability to be used in densely built areas or indoor environments. In contrast, positioning and navigation technology based on visual scene matching, with its advantages of strong anti-electromagnetic interference capabilities, low cost and good environmental adaptability, has become an important way to enhance the autonomous and reliable flight capabilities of aircraft in complex and dynamic environments. Especially in the case of satellite signal interference and other denial conditions, its application potential is huge.

[0003] Although visual image matching navigation technology has received extensive research and attention, existing methods still face the following challenges. First, the differences in heterogeneous images, as well as changes in perspective and scale, significantly affect the accuracy and robustness of matching. Current methods (such as patent numbers CN 117974791 A and CN 114199250 B) mainly use aerial photos and satellite images for matching and positioning. This single matching method is sensitive to environmental changes and is prone to problems such as unstable matching or feature loss during the matching process. Secondly, traditional methods consume huge computing resources when processing large-size satellite images, making it difficult for the matching algorithm to meet the real-time requirements of the edge computing platform in the aircraft.

[0004] In summary, in order to overcome the shortcomings of existing technologies, it is urgent to develop a more accurate and robust visual scene matching positioning and navigation technology. At the same time, this technology should be efficient and lightweight to optimize its application effect in actual positioning and navigation tasks of aircraft. Summary of the invention

[0005] In order to overcome the deficiencies of the prior art, the present invention provides an aircraft visual navigation method based on dual matching and dynamic satellite image loading. First, a satellite slice database is prepared offline for dynamic loading during subsequent online matching. When performing visual matching positioning, the current frame is subjected to feature extraction and matching with the satellite image dynamically spliced ​​and loaded according to the heading angle and altitude information, and the longitude and latitude positioning information P1 is obtained after calculation. At the same time, the current frame is subjected to the same feature extraction and matching processing as the previous frame, and the longitude and latitude positioning information P2 is obtained after calculation. Finally, the dual matching positioning result is input into the Bayesian fusion module for fusion optimization to obtain the final positioning result of the current frame. The method of the present invention effectively solves the problem that the traditional single matching method easily leads to unstable and invalid positioning, and greatly improves the accuracy and robustness of matching positioning.

[0006] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0007] Step 1: Preparation and preprocessing of satellite image slice database;

[0008] Download and prepare the original satellite image with real latitude and longitude location information, split it into several slices, prepare a satellite image slice database, and store it on the edge computing platform of the aircraft;

[0009] On the edge computing platform, pre-deploy the deep learning-based feature extraction and matching network model to prepare for subsequent image processing and matching;

[0010] Step 2: System initialization;

[0011] Initialize the visual matching positioning system using the initial position and heading angle information of the aircraft, obtain the first visual matching area, and thus start the visual matching navigation process;

[0012] Step 3: Image input and dynamic satellite image loading;

[0013] The input end receives real-time image data captured by the downward-looking camera under the aircraft body to obtain the current frame image; at the same time, according to the positioning information of the previous frame, the satellite image slice splicing and reloading strategy is adopted to load the corresponding regional satellite image to prepare for the next step of image matching;

[0014] Step 4: Double matching positioning;

[0015] Double matching is performed through the trained deep learning feature extraction and matching network model:

[0016] 1) Match and locate the aerial image of the current frame with the loaded satellite area image;

[0017] 2) Matching and positioning between the current frame of aerial image and the previous frame of aerial image;

[0018] Step 5: Positioning result fusion;

[0019] The Bayesian fusion module is used to integrate the positioning information obtained from the double matching to generate the optimized current frame positioning result;

[0020] Step 6: Save and iterate the results;

[0021] The current frame image and its positioning result are stored, and this position information is used as the constraint condition for the next frame matching positioning;

[0022] Return to step 3 to perform matching and positioning of the next frame;

[0023] This cycle continues, ensuring continuous positioning and navigation throughout the flight.

[0024] Preferably, the step 1 is specifically:

[0025] First, the original satellite image is downloaded to the local computer and split into N*N satellite sub-image slices. The size of each slice is much smaller than the original satellite image. Each slice after splitting has accurate longitude and latitude coordinate information for subsequent dynamic splicing and reorganization. Finally, the prepared slice database is stored in the aircraft edge computing platform. This process is completed offline before the matching algorithm is deployed.

[0026] Preferably, the step 3 is specifically:

[0027] According to the position and attitude information of the previous frame, a total of 9 slice images adjacent to the position of the previous frame are selected from the satellite slice database, and a continuous satellite area map is formed through image stitching technology; then, the heading angle and flight altitude information are used to perform corresponding geometric transformations on the stitched satellite images, including rotation, scaling and cropping;

[0028] Assume that the stitched satellite area image is Is, its width and height are w*h, and the horizontal and vertical field of view angles of the camera are FOV h and FOV v , the aircraft's flight altitude is H, and the heading angle is θ, then the scaling and rotation process of the stitched image is as follows:

[0029] First, the ground cover width and height of the aerial image generated by the drone camera are calculated;

[0030]

[0031] Assume the satellite image resolution is R s , use formula (2) to calculate the factor that the satellite image needs to be scaled:

[0032]

[0033] The satellite image is scaled using the scaling factor Scale_factor, and the adjusted image size is:

[0034]

[0035] Then, the scaled satellite image is rotated using the rotation matrix shown in formula (4), and a rectangular area is cropped to obtain the satellite image to be matched:

[0036]

[0037] Preferably, the step 4 is specifically:

[0038] The XFeat network model based on deep learning is used to perform feature extraction and matching operations. In this process, the features of the current frame’s aerial image, the generated satellite area map, and the previous frame’s aerial image are extracted to achieve a dual matching mechanism of “sub-image-parent image” and “sub-image-sub-image”.

[0039] The process of feature extraction and matching is as follows:

[0040] 1) Feature extraction and matching;

[0041] For the aerial image I1 and the satellite image I2, the XFeat feature extractor is used to extract features respectively, and the extracted feature descriptors are obtained:

[0042]

[0043] After obtaining the descriptor, the features between the two images are matched to find the corresponding relationship. This process is completed by the XFeat feature matcher, as shown in formula (6):

[0044] M1_pts, M2_pts, conf=XFeat.match(Descriptors1,Descriptors2) (6)

[0045] 2) Solve the homography transformation matrix and calculate the longitude and latitude coordinates;

[0046] The robust estimation algorithm MAGSAC is used to find the homography matrix H in the matching pair and find the transformation relationship between the two images, as shown in formula (7):

[0047] H= findHomography(M1_pts, M2_pts, MAGSAC) (7)

[0048] Then, the obtained homography matrix is ​​used to transform the center coordinates of the drone image (P cx ,Pcy ) is converted to the pixel coordinates in the satellite image (P' cx ,P' cy ), as shown in formula (4):

[0049] P c ' x ,P c ' y =PerspectiveTransform(P cx ,P cy ,H) (8)

[0050] Finally, according to the known latitude and longitude range (longitude [Lon1:Lon2], latitude [Lat1:Lat2]) and pixel size (w2*h2) of the satellite image, the actual longitude and latitude are calculated by formula (9):

[0051]

[0052] Matching and positioning operations are performed on the aerial image I1 and the satellite image I2, as well as the aerial image I1 and the previous frame aerial image I0, to obtain the double matching longitude and latitude coordinate results and confidence, namely the satellite image positioning result (lon1, lat1, conf1) and the inter-frame positioning result (lon2, lat2, conf2).

[0053] Preferably, the step 5 is specifically:

[0054] A fusion module based on Bayesian theory is introduced to fuse the double matching results;

[0055] The matching positioning result between the current frame and the satellite image is recorded as P1(lon1, lat1, conf1); the inter-frame positioning result is recorded as P2(lon2, lat2, conf2). Assuming that they obey independent probability distributions, the two sets of positioning results are integrated through the Bayesian theory update mechanism to optimize the final positioning output.

[0056] Satellite image matching positioning distribution:

[0057]

[0058] Inter-frame matching positioning distribution:

[0059]

[0060] According to Bayesian theory, the fused positioning result is obtained through formula (12):

[0061]

[0062] The confidence after fusion is estimated by formula (13):

[0063]

[0064] A computer program enables a computer to execute the above-mentioned aircraft visual navigation method.

[0065] An electronic device comprises: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device executes the above-mentioned aircraft visual navigation method.

[0066] A computer-readable storage medium stores a computer program, which implements the above-mentioned aircraft visual navigation method when executed by a processor.

[0067] A chip includes: a processor for calling and running a computer program from a memory, so that a device equipped with the chip executes the above-mentioned aircraft visual navigation method.

[0068] A computer program product comprises a computer storage medium storing a computer program, wherein the computer program comprises instructions executable by at least one processor, and the above-mentioned aircraft visual navigation method is implemented when the instructions are executed by the at least one processor.

[0069] The beneficial effects of the present invention are as follows:

[0070] 1) The dual matching method in the present invention improves the positioning accuracy and robustness in complex environments: The present invention adopts a dual matching method, which includes matching the current frame with the satellite image, and matching the current frame with the previous frame, and integrating the results through a Bayesian fusion module. This method effectively solves the problem that the traditional single matching method easily leads to unstable and invalid positioning, and greatly improves the accuracy and robustness of matching positioning.

[0071] 2) The dynamic reorganization loading method in the present invention reduces the computing resource consumption of the visual matching positioning algorithm: by dividing the original satellite image into smaller slices and dynamically loading these small-sized satellite sub-images during the matching process, the present invention significantly reduces the computing resource consumption. This method effectively solves the problems of excessive computing burden and memory usage in traditional methods. While improving the real-time performance of the algorithm, it reduces the matching search space, thereby further improving the positioning accuracy.

[0072] 3) The present invention provides an efficient and adaptable visual matching positioning method, which achieves excellent performance in resource-constrained platforms. The method is suitable for long-distance positioning and navigation tasks in complex environments such as aircraft satellite denial, showing good application prospects and significant engineering practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a diagram of the implementation steps of the present invention;

[0074] Figure 2 It is a framework for visual positioning and navigation methods based on dual matching and dynamic satellite image loading;

[0075] Figure 3 This is a schematic diagram of satellite image slicing and dynamic loading process;

[0076] Figure 4 This is a schematic diagram of the dual matching positioning mechanism;

[0077] Figure 5 A schematic diagram of the matching and positioning process using the neural network model XFeat;

[0078] Figure 6 This is the processing flow chart of the Bayesian fusion module;

[0079] Figure 7 It is an experimental implementation device of the visual matching positioning navigation system according to an embodiment of the present invention;

[0080] Figure 8 This is a diagram of the flight route of a drone during the experiment of an embodiment of the present invention;

[0081] Fig. 9 This is the double matching result of the embodiment of the present invention;

[0082] Fig.10 The test results of the visual matching positioning navigation implementation of the present invention are given. DETAILED DESCRIPTION

[0083] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0084] The present invention aims to solve the problems of reduced accuracy and insufficient robustness caused by image heterogeneity, perspective and scale changes in existing single visual scene matching methods, as well as the challenge of poor real-time performance on resource-constrained edge platforms due to excessive consumption of computing resources. To this end, the present invention proposes an aircraft positioning and navigation method and device based on a dual matching strategy and dynamic satellite mother image slice reconstruction and loading. This innovative solution can significantly improve the accuracy and robustness of matching positioning, and achieve excellent real-time performance on the edge computing platform.

[0085] The present invention proposes a UAV positioning and navigation method based on a dual matching method (matching of aerial sub-images with satellite mother images and matching of sub-images between frames) and a satellite image slice splicing and loading strategy. The implementation steps are as follows: Figure 1 As shown:

[0086] Step 1: Preparation and preprocessing of satellite reference images;

[0087] To achieve efficient and accurate matching, the present invention first downloads and prepares a satellite reference image with real latitude and longitude location information, and stores it on the edge computing platform of the aircraft. To optimize computing resources and accelerate the matching process, the original satellite image is split into several small-sized slices to adaptively load the area to be processed. On the edge computing platform, a feature extraction and matching network model based on deep learning is pre-deployed to prepare for subsequent efficient image processing and matching.

[0088] Step 2: System initialization;

[0089] The visual matching positioning system is initialized using the initial position and heading angle information of the aircraft to obtain the first visual matching area, thereby starting the visual matching navigation process. This step ensures that the system can quickly enter a stable working state.

[0090] Step 3: Image input and dynamic satellite image loading;

[0091] The input end of the algorithm receives real-time image data captured by the downward-looking camera under the aircraft body to obtain the current frame image. At the same time, according to the positioning information of the previous frame (or initial frame), the satellite image slice splicing and reloading strategy designed in the present invention is applied to load the corresponding small area satellite image to prepare for the next step of image matching.

[0092] Step 4: Double matching positioning;

[0093] Double matching is performed through the trained deep learning feature extraction and matching network model. This step includes:

[0094] 1) Match and locate the aerial image of the current frame with the loaded satellite area image.

[0095] 2) Match and locate the aerial image of the current frame with the aerial image of the previous frame.

[0096] Step 5: Positioning result fusion;

[0097] The Bayesian fusion module is used to integrate the positioning information obtained from the dual matching to generate the optimized current frame positioning result. This process significantly improves the accuracy and robustness of positioning through confidence-based fusion optimization.

[0098] Step 6: Save and iterate the results;

[0099] The current frame image and its positioning result are stored, and this position information is used as the constraint condition for the next frame matching and positioning. Return to step 3 to perform the matching and positioning of the next frame. This cycle continues to ensure continuous and high-precision positioning and navigation throughout the flight.

[0100] The overall processing flow of the present invention is as follows Figure 2 As shown in the figure, first prepare the satellite slice database offline for dynamic loading during subsequent online matching. When performing visual matching positioning, the current frame is subjected to feature extraction and matching with the satellite image dynamically spliced ​​and loaded according to the heading angle and altitude information, and the longitude and latitude positioning information P1 is obtained after solution. At the same time, the current frame is subjected to the same feature extraction and matching processing as the previous frame, and the longitude and latitude positioning information P2 is obtained after solution. Then, the dual matching positioning results are input into the Bayesian fusion module for fusion optimization to obtain the final positioning result of the current frame.

[0101] The specific implementation plans involved in the above technical framework are as follows:

[0102] (1) Preparation of satellite image slice database

[0103] First, the original satellite image is downloaded to the local computer and split into N*N satellite sub-image slices. The size of each slice is much smaller than the original satellite image. Each slice after splitting has accurate latitude and longitude coordinate information, which can be used for subsequent dynamic splicing and reorganization. Finally, the prepared slice database is stored in the aircraft edge computing platform. This process is completed offline before the matching algorithm is deployed.

[0104] (2) Preprocessing and dynamic loading of satellite area images

[0105] According to the position and attitude information of the previous frame (or initial frame), a total of 9 slice images adjacent to the previous frame position are selected from the satellite slice database, and a continuous satellite area map is formed through image stitching technology. Then, in order to correct the difference in viewing angle between the satellite image and the aerial image, the heading angle and flight altitude information are used to perform corresponding geometric transformations on the stitched satellite image, including rotation, scaling and cropping. This step aims to generate a reference image that matches the current aerial image frame more closely to enhance the accuracy of the image matching process.

[0106] Assume that the stitched satellite area image is Is, its width and height are w*h, and the horizontal and vertical field of view angles of the camera are FOV h and FOV v , the aircraft's flight altitude is H, and the heading angle is θ. Then the scaling and rotation process of the stitched image is as follows:

[0107] 1) First, the ground cover width and height of the aerial image generated by the drone camera are calculated.

[0108]

[0109] 2) Assume that the satellite image resolution is Rs, and use formula (2) to calculate the scaling factor of the satellite image. In order to avoid image deformation caused by scaling, a smaller scaling factor is selected:

[0110]

[0111] 3) Use the scaling factor Scale_factor to scale the satellite image. The adjusted image size is:

[0112]

[0113] 4) Then, the scaled satellite image is rotated using the rotation matrix shown in formula (4), and a rectangular area is cropped to obtain the satellite image to be matched.

[0114]

[0115] Satellite image slice reorganization and scaling and rotation loading process Figure 3 As shown:

[0116] (3) Dual feature extraction and matching positioning;

[0117] The present invention uses the XFeat network model based on deep learning to perform feature extraction and matching operations. In this process, the features of the current frame aerial image (sub-image) and the generated satellite area image (parent image) and the previous frame aerial image (sub-image) are extracted to realize the "sub-image-parent image" and "sub-image-sub-image" dual matching mechanisms, such as Figure 4 as shown in .

[0118] The XFeat model is a highly efficient neural network architecture that uses a multi-scale feature detection framework to identify and extract significant structures in images from local to global perspectives. This multi-scale feature detection method provides robustness for image processing under different viewing angles and zoom conditions. Because XFeat uses a lightweight CNN structure, it can run efficiently on platforms with limited computing resources while meeting application requirements that require high throughput or high computing efficiency. Its matching refinement module can extract pixel-level offsets from coarse semi-dense matching to achieve more accurate matching results. In the present invention, the "double matching" step is implemented using the same shared XFeat model to simplify the system architecture and improve processing consistency. Figure 5 shown.

[0119] The process of feature extraction and matching is as follows:

[0120] 1) Feature extraction and matching;

[0121] For the aerial image I1 and the satellite image I2, the XFeat feature extractor is used to extract features respectively, and the extracted feature descriptors are obtained:

[0122]

[0123] After obtaining the descriptor, the features between the two images are matched to find the corresponding relationship. This process is completed by the XFeat feature matcher, as shown in formula (6):

[0124] M1_pts, M2_pts, conf=XFeat.match(Descriptors1,Descriptors2) (19)

[0125] 2) Solve the homography transformation matrix and calculate the longitude and latitude coordinates;

[0126] Next, a more advanced robust estimation algorithm MAGSAC (Maximum A Posteriori SAmple Consensus) is used to find the homography matrix H in the matching pair to find the transformation relationship between the two images, as shown in formula (7):

[0127] H= findHomography(M1_pts, M2_pts, MAGSAC) (20)

[0128] Then, the obtained homography matrix is ​​used to transform the center coordinates of the drone image (P cx ,P cy ) is converted to the pixel coordinates in the satellite image (P' cx ,P' cy ), as shown in formula (4):

[0129] P c ' x ,P c ' y =PerspectiveTransform(P cx ,P cy ,H) (21)

[0130] Finally, according to the known latitude and longitude range (longitude [Lon1:Lon2], latitude [Lat1:Lat2]) and pixel size (w2*h2) of the satellite image, the actual longitude and latitude are calculated by formula (9):

[0131]

[0132] By performing the above matching and positioning operations on the aerial image I1 and the satellite image I2, as well as the aerial image I1 and the previous frame aerial image I0, the double matching longitude and latitude coordinate results and confidence levels can be obtained, namely, the satellite image positioning result (lon1, lat1, conf1) and the inter-frame positioning result (lon2, lat2, conf2).

[0133] (4) Bayesian fusion module;

[0134] In order to improve the accuracy and robustness of the positioning results and avoid the failure of the visual positioning algorithm due to inaccurate matching or matching failure between aerial photos and satellite images, the present invention introduces a fusion module based on Bayesian theory to fuse the double matching results. The processing flow of the fusion module is as follows: Figure 6 shown.

[0135] The processing process of the Bayesian fusion algorithm is as follows: for the matching positioning results of the current frame and the satellite image, recorded as P1 (lon1, lat1, conf1) and the inter-frame positioning results, recorded as P2 (lon2, lat2, conf2), it is assumed that they obey independent probability distributions. The two sets of positioning results are integrated through the Bayesian theory update mechanism to optimize the final positioning output.

[0136] Satellite image matching positioning distribution:

[0137]

[0138] Inter-frame matching positioning distribution:

[0139]

[0140] According to Bayesian theory, the fused positioning result is obtained through formula (12):

[0141]

[0142] The confidence after fusion is estimated by formula (13):

[0143]

[0144] Example:

[0145] In order to verify the effectiveness of the UAV visual matching navigation method based on dual matching and dynamic loading of slices proposed in this invention, it was deployed on the embedded edge computing platform NVIDIA Orin Nano, and a flight experiment was carried out on the UAV platform. The specific implementation examples are as follows:

[0146] (1) Implementation device and composition;

[0147] The experimental device used in this embodiment consists of an eight-axis rotor drone, an edge computing platform NVIDIA Orin Nano, and a visible light pod. Figure 7 As shown in . The rotorcraft drone is equipped with a high-precision RTK positioning module to provide accurate reference positioning values. As an edge computing platform, NVIDIA Orin Nano has a deep learning computing power of 40TOPS, a power consumption range of 7.5-15W, and deploys the visual matching positioning and navigation algorithm designed by the present invention. The visible light pod is connected to NVIDIA Orin Nano, providing real-time downward viewing images to support the operation of the visual positioning algorithm. The drone is also equipped with a data transmission module that can transmit the visual matching positioning results to the ground monitoring station in real time for result analysis.

[0148] (2) Experimental environment;

[0149] The experimental verification was carried out in a place where the terrain covered a variety of scenes such as grassland, farmland, playground and buildings to fully test the accuracy and robustness of the algorithm. The drone flew along the preset route. The flight route is as follows Figure 8 As shown in the figure, the latitude and longitude of the flight area are: longitude [108.7451474°E, 108.7728831°E], latitude [34.04564629°N, 34.02483682°N]. The flight altitude is 500 meters, the average flight speed is 10.5 meters per second, and the total flight distance is about 24.5 kilometers.

[0150] (3) Experimental effects and result analysis;

[0151] Fig. 9 The figure shows the dual feature matching result of a frame during flight. It can be seen that due to the difference in viewing angles, there are fewer matching point pairs between the aerial image and the satellite image in this frame (left picture). This situation easily leads to the failure of the traditional single matching algorithm, while the inter-frame matching added in the present invention still matches enough feature point pairs (right picture). Therefore, dual matching of the current frame with the satellite image and with the previous aerial image can effectively avoid the single matching failure problem in complex scenes and effectively enhance the robustness of the matching positioning method.

[0152] Fig.10 The experimental results shown show the positioning trajectory of the visual matching flight test, where the yellow trajectory represents the actual flight path and the green trajectory represents the matching positioning path achieved by the present invention. Qualitative analysis shows that the two trajectories are highly overlapped, indicating that the present invention has excellent positioning accuracy. During the experiment, the present invention continued to maintain a stable matching positioning function under various complex turning paths, and no matching loss occurred, confirming its excellent robustness. Fig.10The real-time error output in the 3D image is quantitatively analyzed. The positioning errors of the present invention in longitude and latitude are respectively less than 2.5% and 3.0%, which is significantly improved compared with the prior art. In terms of real-time performance, the total time consumption of single-frame positioning of the present invention method is less than 30ms, which can well meet the airborne real-time requirements.

Claims

1. A method for aircraft visual navigation based on double matching and dynamic satellite image loading, characterized in that: The steps include: Step 1: Preparation and preprocessing of satellite image slice database; Download and prepare the original satellite image with real latitude and longitude location information, split it into several slices, prepare a satellite image slice database, and store it on the edge computing platform of the aircraft; On the edge computing platform, pre-deploy a deep learning-based feature extraction and matching network model to prepare for subsequent image processing and matching; Step 2: System initialization; Initialize the visual matching positioning system using the initial position and heading angle information of the aircraft, obtain the first visual matching area, and thus start the visual matching navigation process; Step 3: Image input and dynamic satellite image loading; The input end receives real-time image data captured by the downward-looking camera under the aircraft body to obtain the current frame image; at the same time, according to the positioning information of the previous frame, the satellite image slice splicing and reloading strategy is adopted to load the corresponding regional satellite image to prepare for the next step of image matching; Step 4: Double matching positioning; Double matching is performed through the trained deep learning feature extraction and matching network model: 1) Match and locate the aerial image of the current frame with the loaded satellite area image; 2) Matching and positioning between the current frame of aerial image and the previous frame of aerial image; Step 5: Positioning result fusion; The Bayesian fusion module is used to integrate the positioning information obtained from the double matching to generate the optimized current frame positioning result; Step 6: Save and iterate the results; The current frame image and its positioning result are stored, and this position information is used as the constraint condition for the next frame matching positioning; Return to step 3 to perform matching and positioning of the next frame; This cycle continues, ensuring continuous positioning and navigation throughout the flight.

2. The method for aircraft visual navigation based on double matching and dynamic satellite image loading according to claim 1, characterized in that: The step 1 is specifically as follows: First, the original satellite image is downloaded locally and split into N*N satellite sub-image slices, each of which is much smaller than the original satellite image. Each slice after splitting carries accurate latitude and longitude coordinate information for subsequent dynamic splicing and reorganization. Finally, the prepared slice database is stored in the aircraft edge computing platform. This process is completed offline before the matching algorithm is deployed.

3. The aircraft visual navigation method based on double matching and dynamic satellite image loading according to claim 2 is characterized in that: The step 3 is specifically as follows: According to the position and attitude information of the previous frame, a total of 9 slice images adjacent to the position of the previous frame are selected from the satellite slice database, and a continuous satellite area map is formed through image stitching technology; then, the heading angle and flight altitude information are used to perform corresponding geometric transformations on the stitched satellite images, including rotation, scaling and cropping; Assume that the stitched satellite area image is Is, its width and height are w*h, and the horizontal and vertical field of view angles of the camera are FOV h and FOV v , the aircraft's flight altitude is H, and the heading angle is θ, then the scaling and rotation process of the stitched image is as follows: First, the ground cover width and height of the aerial image generated by the drone camera are calculated; Assume the satellite image resolution is R s , use formula (2) to calculate the factor that the satellite image needs to be scaled: The satellite image is scaled using the scaling factor Scale_factor, and the adjusted image size is: Then, the scaled satellite image is rotated using the rotation matrix shown in formula (4), and a rectangular area is cropped to obtain the satellite image to be matched:

4. The aircraft visual navigation method based on double matching and dynamic satellite image loading according to claim 3 is characterized in that: The step 4 is specifically as follows: The XFeat network model based on deep learning is used to perform feature extraction and matching operations. In this process, the features of the current frame aerial image, the generated satellite area map, and the previous frame aerial image are extracted to realize the "sub-image-parent image" and "sub-image-sub-image" dual matching mechanisms; The process of feature extraction and matching is as follows: 1) Feature extraction and matching; For the aerial image I1 and the satellite image I2, the XFeat feature extractor is used to extract features respectively, and the extracted feature descriptors are obtained: After obtaining the descriptor, the features between the two images are matched to find the corresponding relationship. This process is completed by the XFeat feature matcher, as shown in formula (6): M1_pts, M2_pts, conf=XFeat.match(Descriptors1,Descriptors2) (6) 2) Solve the homography transformation matrix and calculate the longitude and latitude coordinates; The robust estimation algorithm MAGSAC is used to find the homography matrix H in the matching pair and find the transformation relationship between the two images, as shown in formula (7): H= findHomography(M1_pts, M2_pts, MAGSAC) (7) Then, the obtained homography matrix is ​​used to transform the center coordinates of the drone image (P cx ,P cy ) is converted to the pixel coordinates in the satellite image (P' cx ,P' cy ), as shown in formula (4): P c ′ x ,P c ′ y =PerspectiveTransform(P cx ,P cy ,H) (8) Finally, according to the known latitude and longitude range (longitude [Lon1:Lon2], latitude [Lat1:Lat2]) and pixel size (w2*h2) of the satellite image, the actual longitude and latitude are calculated by formula (9): Matching and positioning operations are performed on the aerial image I1 and the satellite image I2, as well as the aerial image I1 and the previous frame aerial image I0, to obtain the double matching longitude and latitude coordinate results and confidence, namely the satellite image positioning result (lon1, lat1, conf1) and the inter-frame positioning result (lon2, lat2, conf2).

5. The method for aircraft visual navigation based on double matching and dynamic satellite image loading according to claim 4, characterized in that: The step 5 is specifically as follows: A fusion module based on Bayesian theory is introduced to fuse the double matching results; The matching positioning result between the current frame and the satellite image is recorded as P1(lon1, lat1, conf1); the inter-frame positioning result is recorded as P2(lon2, lat2, conf2). Assuming that they obey independent probability distributions, the two sets of positioning results are integrated through the Bayesian theory update mechanism to optimize the final positioning output. Satellite image matching positioning distribution: Inter-frame matching positioning distribution: According to Bayesian theory, the fused positioning result is obtained through formula (12): The confidence after fusion is estimated by formula (13):

6. A computer program, characterized in that The computer program enables a computer to execute the method according to any one of claims 1 to 5.

7. An electronic device, characterized in that: include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device executes the method as claimed in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

9. A chip, characterized in that: include: A processor, configured to call and run a computer program from a memory, so that a device equipped with the chip executes a method as claimed in any one of claims 1 to 5.

10. A computer program product, characterized in that The computer program product comprises a computer storage medium storing a computer program, wherein the computer program comprises instructions executable by at least one processor, and when the instructions are executed by the at least one processor, the method according to any one of claims 1 to 5 is implemented.

Images