Unmanned aerial vehicle navigation decoy defense method based on visual map matching

Abstract

The invention relates to an unmanned aerial vehicle navigation decoy defense method based on visual map matching, and solves the technical problems that a traditional relative positioning mode is easy to cause positioning drift and an unmanned aerial vehicle is easy to decoy and interfere under long-term and large-scene conditions. The absolute positioning method based on the scene coordinate regression network is adopted, positioning drift is avoided in mechanism, and the scene coordinate regression network based on deep learning is used for performing positioning information mapping on the visual image and the scene coordinates. And inputting a visual image shot by the unmanned aerial vehicle into the optimized scene coordinate regression network to obtain a scene coordinate predicted value, and calculating the scene coordinate predicted value by using an RANSAC algorithm to obtain a pose estimation result. And comparing the pose estimation result with global navigation satellite system (GPS) positioning information received by the unmanned aerial vehicle to complete positioning signal decoy recognition, so that the unmanned aerial vehicle can effectively defend against GPS decoy signals.

Description

Technical Field

[0001] This invention relates to a method for defending against drone navigation deception, specifically a method for defending against drone navigation deception based on visual map matching. Background Technology

[0002] Unmanned Aerial Vehicles (UAVs) are unmanned aircraft controlled by radio remote control equipment and their own program control devices, or unmanned aircraft that are fully or intermittently autonomously operated by an onboard computer via radio remote control. UAVs are characterized by their maneuverability and flexibility, playing an irreplaceable role, especially in dangerous or inaccessible locations. They are widely used in aerial photography, agriculture, plant protection, express delivery, disaster relief, surveying, news reporting, power line inspection, disaster relief, and film and television production.

[0003] Current drones primarily rely on global navigation satellite systems, such as GPS and BeiDou, to estimate their attitude using their location information and internal inertial navigation systems. Remotely controlled drones, upon receiving position commands, reach designated locations to complete tasks, or their onboard computers autonomously execute missions. The fundamental capability of drones to complete designated tasks requires reliable positioning information; however, methods of jamming and deceiving drone positioning signals are rapidly evolving. To improve the reliability of drone applications, it is necessary to improve the positioning methods that drones rely on global navigation satellite systems.

[0004] Due to the lightweight, small size, and low cost of visual sensors, using visual information collected by camera visual sensors mounted on drones for visual positioning is an effective solution to defend against drone positioning spoofing and interference. However, because traditional visual matching methods use relative positioning, drones are prone to positioning drift over long periods and in large-scale scenarios, making it impossible to obtain reliable positioning information. Therefore, this method cannot be directly applied to drones that use the Global Navigation Satellite System for positioning. Summary of the Invention

[0005] The purpose of this invention is to address the technical problems of existing technologies, such as the tendency for positioning drift to occur over long periods and in large-scale scenarios, resulting in the inability to obtain reliable positioning information, and the vulnerability of drones to deception and interference. Therefore, this invention proposes a drone navigation deception defense method based on visual map matching.

[0006] To solve the above-mentioned technical problems, the technical solution provided by the present invention is as follows:

[0007] A method for defending against drone navigation deception based on visual map matching is characterized by the following steps:

[0008] Step 0: Construct a scene coordinate regression network, which includes an input module, a processing module, and a prediction module connected in sequence; the input module is used to extract image features, the processing module is used to extract and refine features, and the prediction module is used to obtain pose estimation results;

[0009] Step 1: Obtain a 3D simulation model of the area to be located by combining satellite orthophoto maps with digital terrain models, or by using drone cameras to collect 3D mapping models of the area to be located.

[0010] Step 2: Extract and process the 3D simulation model to establish a simulation dataset for visual positioning; the simulation dataset includes one-to-one corresponding visual images, scene coordinates, and pose data;

[0011] Step 3: Preprocess the simulation dataset: obtain the one-to-one normalized image, scene coordinate ground truth, and UAV camera coordinates;

[0012] Step 4: Read the one-to-one corresponding normalized image, scene coordinate ground truth, and UAV camera coordinates into the scene coordinate regression network to obtain the scene coordinate prediction value that meets the expectations, complete the training of the scene coordinate regression network, and obtain the optimized scene coordinate regression network.

[0013] Step 5: The drone begins to fly. First, set the drone to low-frequency deception positioning and identification mode. The drone's satellite positioning and navigation module uses satellite positioning and navigation to obtain the drone's location information.

[0014] Step 6: The drone camera captures visual images, and after normalization, the visual images are input into the optimized scene coordinate regression network to obtain pose estimation results; the pose estimation results are then transformed from the ECEF coordinate system to the WGS84 coordinate system to obtain the drone's visual positioning information.

[0015] Step 7: Compare the visual positioning information with the location information obtained by the UAV's satellite positioning and navigation module to determine whether it has been spoofed. If so, trigger a spoofing alarm, switch the UAV to high-frequency visual positioning mode, use visual positioning information for navigation, and return to step 6; if not, no spoofing has occurred, return to step 5, and maintain the low-frequency spoofing positioning identification mode.

[0016] Furthermore, step 3, preprocessing the simulation dataset, specifically includes:

[0017] Step 3.1: Normalize the visual image to obtain a normalized image;

[0018] Step 3.2: Perform coordinate transformation on the scene coordinates, then downsample to obtain the true value of the scene coordinates corresponding to the pose data;

[0019] Step 3.3: Convert the attitude data into camera coordinates corresponding to the UAV. The camera coordinates are used as the pose data of the UAV.

[0020] Furthermore, step 3.3 specifically includes:

[0021] Step 3.3.1: Perform coordinate transformation on the three-dimensional coordinate data in the attitude data, converting it from the WGS84 coordinate system to the ECEF coordinate system; set a fixed spatial coordinate, and calculate the difference between the transformed three-dimensional coordinate data and the fixed spatial coordinate to obtain the translation component;

[0022] Step 3.3.2: Convert the yaw, roll, and pitch data in the attitude data into rotation matrices;

[0023] Step 3.3.3: Combine the translation components and the rotation matrix to obtain the camera coordinates of the UAV.

[0024] Furthermore, the input module includes a Stem module and a first module connected in sequence. The Stem module and the first module each contain a convolution block for convolving the visual image and extracting its features.

[0025] The processing module includes a second module, a third module, and a fourth module connected in sequence. Each of the second, third, and fourth modules includes a PatchEmbed module and a Transformer submodule. The PatchEmbed module divides the image after feature extraction into blocks to match the format requirements of the Transformer submodule. The Transformer submodule outputs the processed feature map.

[0026] The prediction module includes a convolution module, a scene coordinate filtering module, and a RANSAC method module connected in sequence. The feature map is reduced in dimension by the convolution module, and then the scene coordinate filtering module is used to obtain the scene coordinate prediction value. The scene coordinate prediction value and the UAV pose data are processed by the RANSAC method module to obtain the pose estimation result.

[0027] Furthermore, step 4 specifically includes:

[0028] Step 4.1: Input the normalized image into the Stem module. Set the kernel size of the convolution block of the Stem module to 3, perform convolution on the normalized image, obtain the feature extraction map, and input it into the first module.

[0029] Step 4.2: The convolutional block of the first module uses three layers of convolution to extract features and performs downsampling to obtain a refined feature map, which is then input into the second module.

[0030] Step 4.3: The Transformer submodule includes a local attention module and a global attention module; the PatchEmbed module of the second module divides the feature refinement map into blocks to match the format requirements of the Transformer submodule; the blocks are input to the local attention module, output from the local attention module, and then enter the global attention module, output from the global attention module, and finally enter the third module; the local attention module is used to locate and infer local key elements in the feature refinement map, and the global attention module is used to integrate global information on the overall distribution information of the feature refinement map;

[0031] Step 4.4: The third and fourth modules repeat the processing of the second module, and finally the global attention module of the fourth module outputs the processed feature map; from the second module to the fourth module, the number of heads used in the local attention module increases sequentially, gradually improving the high-level scene abstraction capability of the corresponding module;

[0032] Step 4.5: Input the feature map into the convolution module. After dimensionality reduction by the convolution module, it passes through the scene coordinate filtering module and then undergoes inverse transformation of the camera coordinates to obtain the scene coordinate prediction value. Compare the scene coordinate prediction value with the scene coordinate ground truth value obtained in Step 3.2 to calculate the loss function. If the loss function calculation result meets the specified expected standard, proceed to Step 4.6. If the loss function calculation result does not meet the specified expected standard, reselect the normalized image and return to Step 4.1 to increase the number of training rounds until the scene coordinate prediction value that meets the expectations is obtained.

[0033] Step 4.6: Process the scene coordinate prediction values ​​and the camera coordinates obtained in Step 3.3 through the RANSAC method module to obtain the pose estimation results, complete the training of the scene coordinate regression network, and obtain the optimized scene coordinate regression network.

[0034] Furthermore, in step 7, determining whether a location-based deception has occurred specifically involves:

[0035] The visual positioning information is compared with the location information obtained by the UAV satellite module. If the error between the two values ​​is greater than the maximum error tolerance, it is judged as yes; otherwise, it is judged as no. The maximum error tolerance refers to the maximum positioning error obtained in the scene coordinate regression network evaluation stage.

[0036] Furthermore, in step 2, the extraction and processing of the three-dimensional simulation model specifically includes:

[0037] The 3D simulation model was loaded into the Cesium open-source engine, and the TOPO-DataGen data generation method was used to randomly sample visual images, scene coordinates, and attitude data to establish a simulation dataset for visual positioning. The visual images are RGB images, the scene coordinates are the point cloud coordinate information corresponding to the attitude data, and the attitude data are the six degrees of freedom attitude information of the UAV camera.

[0038] Further, step 3.2 specifically involves performing coordinate transformation on the scene coordinates, followed by 8x downsampling to obtain the true scene coordinate values ​​for the corresponding pose of the UAV, in order to reduce the amount of data for the true scene coordinate values.

[0039] Step 4.2 specifically involves the first module's convolutional block using three layers of convolution to extract features and performing 8x downsampling to ensure that the predicted scene coordinates have the same shape as the true scene coordinates. After obtaining the refined feature map, it is input into the second module.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] 1. This invention employs an absolute positioning method based on a scene coordinate regression network, which avoids positioning drift in its mechanism. It utilizes a deep learning-based scene coordinate regression network to map positioning information between visual images and scene coordinates. The visual images captured by the UAV are input into the optimized scene coordinate regression network to obtain predicted scene coordinate values. The RANSAC algorithm is then used to calculate the pose estimation result. This pose estimation result is compared with the GPS positioning information received by the UAV to complete the identification of GPS spoofing signals, enabling the UAV to effectively defend against GPS spoofing signals.

[0042] 2. This invention relies on a scene coordinate regression network to enable drones to autonomously identify deception interference, solving the problem of spoofing that relies on traditional drone positioning signals. The drone can flexibly switch between low-frequency deception positioning and identification modes and high-frequency visual positioning methods, achieving a balance between reliability and low power consumption in deception defense methods.

[0043] 3. This invention can utilize publicly available or other methods to acquire 3D simulation models, and use the TOPO-DataGen data generation method to establish a simulation dataset. It can obtain more reliable absolute positioning information of UAVs through visual images, and has great application potential in ensuring the safety and reliability of UAVs.

[0044] 4. Compared with traditional visual positioning methods, the present invention has advantages such as low space occupation, low computational complexity, and data security.

[0045] 5. Based on the backbone network of the Transformer submodule and the corresponding training techniques, this invention achieves more accurate positioning in large scenes and sparse texture regions. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the training process of the scene coordinate regression network in an embodiment of a drone navigation deception defense method based on visual map matching according to the present invention.

[0047] Figure 2 This is a flowchart illustrating the drone deception defense method based on visual map matching in an embodiment of the present invention.

[0048] Figure 3 This is a schematic diagram of the scene coordinate regression network in an embodiment of a drone navigation deception defense method based on visual map matching according to the present invention;

[0049] Figure 4 This is an example of a map scene information image in an embodiment of a drone navigation deception defense method based on visual map matching according to the present invention;

[0050] Figure 5 for Figure 4 The corresponding simulation dataset image;

[0051] Figure 6 This is an error probability density distribution diagram of the scene coordinate prediction values ​​corresponding to four sets of simulation datasets in an embodiment of the UAV navigation deception defense method based on visual map matching of the present invention.

[0052] Figure 7 This is a comparison chart of scene coordinate regression errors between an embodiment of the UAV navigation deception defense method based on visual map matching of the present invention and similar methods. Detailed Implementation

[0053] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0054] A method for preventing drone navigation deception based on visual map matching mainly includes constructing a scene coordinate regression network and training the network. After training, the drone captures visual images, which are then input into the optimized scene coordinate regression network to obtain visual positioning information. The training process of the scene coordinate regression network in this invention is as follows: Figure 1 As shown, the drone navigation deception defense process is as follows: Figure 2 As shown. Specifically, it includes the following steps:

[0055] Step 0: Construct a scene coordinate regression network. The scene coordinate regression network includes an input module, a processing module, and a prediction module connected in sequence, such as... Figure 3 As shown.

[0056] The input module is used to extract image features, including a Stem module and a first module connected in sequence. The Stem module and the first module each contain a convolutional block for convolving the visual image.

[0057] The processing module is used to extract and refine features, including a second module, a third module, and a fourth module connected in sequence. The number of processing modules or the corresponding hyperparameters in the modules can be adjusted according to the scene localization scale. The second, third, and fourth modules all include a PatchEmbed module and a Transformer sub-module. The PatchEmbed module is a convolutional module that divides the image after feature extraction into blocks to match the format requirements of the Transformer submodule. In this embodiment, the PatchEmbed module divides the input image into 7*7 small image blocks, with overlap between each small image block. The Transformer submodule outputs the processed feature map. The Transformer submodule includes a local attention module and a global attention module. Both the local attention module and the global attention module enter through LayerNorm and exit through DropPath, using residual connections. The local attention module uses a Multi-Head Self-Attention (MSA) network for local attention, while the global attention module uses a fully connected network (MLP) for global attention integration. The local attention module is used to locate and infer local key elements in the feature refinement map, while the global attention module is used to integrate global information about the overall distribution of the feature refinement map.

[0058] The prediction module includes a convolution module, a scene coordinate filtering module, and a RANSAC method module connected in sequence. The feature map is reduced in dimension by the convolution module. After reduction, it is filtered by the scene coordinate filtering module to remove scene coordinate prediction values ​​that deviate significantly and data bits that have no corresponding true values, thereby obtaining scene coordinate prediction values. The scene coordinate prediction values ​​and the UAV pose data are processed by the RANSAC method module to obtain pose estimation results.

[0059] Step 1: Collect a 3D simulation model of the area to be located. This can be done by combining digital orthophotos of satellite maps with digital terrain models. The positioning accuracy of the UAV is related to the resolution of the 3D simulation model. In this embodiment, the resolution is 10cm. To further improve the positioning accuracy, the 3D mapping model of the area to be located can also be collected directly by using the UAV camera for aerial photography, thereby obtaining the 3D simulation model of the area to be located.

[0060] Step 2: Perform feature extraction and processing on the 3D simulation model, load the 3D simulation model into the Cesium open-source engine, and use the TOPO-DataGen data generation method to randomly sample visual image-scene coordinates-attitude data in a one-to-one correspondence. The visual image is an RGB image, the scene coordinates are the point cloud coordinates corresponding to the attitude data, and the attitude data is the six-degree-of-freedom attitude information of the UAV camera, thus establishing a simulation dataset for visual positioning. For example... Figure 4 The image shown is an example of a visual image captured by a drone camera. Figure 5 for Figure 4 The simulated dataset image corresponding to the pose shown in the visual image;

[0061] Step 3: Preprocess the simulation dataset: Obtain the normalized image, ground truth scene coordinates, and UAV camera coordinates. This includes the following steps:

[0062] Step 3.1: Normalize the visual image to obtain a normalized image; to reduce the parameters and training time of the network to be trained, the visual image can be grayscaled at the expense of some localization accuracy.

[0063] Step 3.2: Perform coordinate transformation on the scene coordinates to obtain the true scene coordinate values ​​corresponding to the pose data; to reduce the amount of true scene coordinate data, 8x downsampling is used to process the true scene coordinate values;

[0064] Step 3.3: Convert the attitude data into camera coordinates corresponding to the UAV. The camera coordinates serve as the UAV's pose data. Specifically:

[0065] Step 3.3.1: Perform coordinate transformation on the three-dimensional coordinate data in the attitude data, from the WGS84 coordinate system to the ECEF coordinate system; since the values ​​in the ECEF coordinate system are large, a fixed spatial coordinate is set, and the difference between the transformed three-dimensional coordinate data and the fixed spatial coordinate is calculated to obtain the translation component; the transformed three-dimensional coordinate data are all located on a coordinate system with the fixed spatial coordinate as the origin, and the original value range of the ECEF coordinate system is adjusted to a smaller value;

[0066] Step 3.3.2: Convert the yaw, roll, and pitch data in the attitude data into rotation matrices;

[0067] Step 3.3.3: Combine the translation components and the rotation matrix to obtain the camera coordinates of the UAV.

[0068] The camera coordinates, as the pose data of the UAV, are input into the RANSAC method module along with the predicted scene coordinates in the prediction module of the scene coordinate regression network to obtain the pose estimation result.

[0069] As can be deduced from the following reasoning, effective scene coordinate regression network design is one of the key steps in estimating camera pose. The specific reasoning process is as follows:

[0070] The yaw, pitch, and roll angles in the attitude data are converted into rotation matrices. The spatial translation coordinates and rotation matrices are combined to obtain a homogeneous transformation matrix. This matrix is ​​then used to rotate the preprocessed normalized image, while simultaneously processing the corresponding homogeneous transformation matrix. The preprocessed data is then randomly enhanced in terms of brightness and contrast. The preprocessed normalized image, ground truth scene coordinates, and camera coordinates are then fed into the network f to be trained. Specifically, the input is a normalized image I, and the corresponding set of scene coordinates is... Pixel p in normalized image I i Corresponding scene coordinates y i Where i is the pixel p i Number, set of scene coordinates The regression method is expressed by the following formula:

[0071]

[0072] Where w represents the network parameters to be trained;

[0073] From scene coordinates y i and pixel p i Given the camera coordinates, the following homogeneous transformation relationship exists:

[0074] y i =he i ;

[0075] e i Let I represent the corresponding camera coordinates, and h be the homogeneous transformation matrix between the camera coordinate space and the scene coordinate space. From the above relationship, it can be seen that the scene coordinates corresponding to the normalized image I can be estimated by training the network, and the corresponding homogeneous transformation matrix h can be solved to estimate the camera pose. Therefore, a more effective scene coordinate regression network is one of the key steps in estimating camera pose, and the network to be trained needs to be trained and its parameters optimized to adapt to the actual positioning area application. The training process is step 4.

[0076] Step 4: Input the normalized image, ground truth scene coordinates, and UAV camera coordinates into the scene coordinate regression network to obtain the expected scene coordinate prediction values, thus completing the training of the scene coordinate regression network and obtaining the optimized scene coordinate regression network; specifically including the following steps:

[0077] Step 4.1 Input the normalized image into the Stem module. The convolutional block of the Stem module is set with a kernel size of 3. Convolve the normalized image to obtain the feature extraction map and input it into the first module. The normalized image may be a grayscale image or a color image. When the normalized image is a grayscale image, the input channel of the convolution in the Stem module is switched from 3 to 1 before feature extraction is performed.

[0078] Step 4.2: The convolutional block of the first module uses three convolutional layers to extract features, with each layer undergoing a 2x downsampling. Therefore, the feature extraction map in the first module is downsampled a total of 8x, corresponding to the size after sampling in Step 3.2, thus ensuring that the shape of the scene coordinate prediction value is consistent with the true scene coordinate value. After obtaining the refined feature map, it is input into the second module.

[0079] Step 4.3, as follows Figure 3 As shown, the PatchEmbed module of the second module divides the feature refinement map into blocks and then inputs them into the local attention module. After being output from the local attention module, the map enters the global attention module, and is output from the global attention module and then enters the third module.

[0080] Step 4.4: The third and fourth modules repeat the processing of the second module, and finally the feature map is output by the global attention module of the fourth module. From the second module to the fourth module and the local attention module, the number of heads used as hyperparameters increases in turn, gradually improving the network's ability to abstract high-level scenes. The number of modules in each stage can be increased or decreased according to the difficulty of the feature map being processed.

[0081] Step 4.5: Input the feature map into the convolution module. After dimensionality reduction by the convolution module, it passes through the scene coordinate filtering module to obtain the scene coordinate prediction value. The scene coordinate prediction value and the ground truth scene coordinate value obtained in Step 3.2 are then compared using a loss function. The loss function is the 2-norm of the scene coordinate prediction value and the ground truth scene coordinate value, plus the uncertainty loss obtained by fusing the uncertainty map during scene coordinate filtering. AdamW optimization technology and OneCycleLR learning rate scheduling technology are used to optimize network training. If the loss function calculation result meets the specified expected standard, proceed to Step 4.6; if the loss function calculation result does not meet the specified expected standard, a normalized image is reselected, and the process returns to Step 4.1, increasing the number of training epochs until the specified expected standard is met, obtaining scene coordinate prediction values ​​that meet expectations, i.e., the training loss decreases slowly and the validation accuracy reaches saturation.

[0082] Step 4.6: Pass the scene coordinate predictions and camera coordinates through the RANSAC method module to obtain the pose estimation results; complete the training of the scene coordinate regression network; obtain the optimized scene coordinate regression network.

[0083] like Figure 6 The image shows the error probability density distribution of the scene coordinate prediction values ​​obtained after inputting four different simulation datasets into the scene coordinate regression network. This is achieved using the method proposed in this invention. Figure 6 The UVL plot represents the overall prediction error distribution closest to 0. The scene coordinate filtering module filters out scene coordinate predictions that deviate significantly and data bits without corresponding true values. For example... Figure 7 As shown, this represents the error between the scene coordinate prediction value obtained based on the scene coordinate regression network and its corresponding pose estimation result.

[0084] Step 5: The drone begins flight. The drone is set to a default low-frequency decoy positioning and identification mode, with the low frequency set to 1Hz-2Hz. This can be adjusted based on the actual flight conditions of the drone. The drone's satellite positioning and navigation module uses satellite positioning and navigation to obtain the drone's location information.

[0085] Step 6: The UAV captures visual images, normalizes the visual images, and inputs them into the optimized scene coordinate regression network to obtain pose estimation results; the pose estimation results are then transformed from the ECEF coordinate system to the WGS84 coordinate system to obtain the UAV's visual positioning information.

[0086] Step 7: Compare the visual positioning information with the location information obtained by the UAV's GPS positioning and navigation module to determine whether it has been spoofed. The determination method is as follows: compare the visual positioning information with the location information obtained by the UAV's GPS module. If the error between the two values ​​is greater than the maximum allowable error value, it is determined as yes; otherwise, it is determined as no. The maximum allowable error value refers to the maximum positioning error obtained during the scene coordinate regression network evaluation stage.

[0087] like Figure 2 The diagram shows the flowchart for drone navigation deception defense. If the judgment result is yes, a deception alarm is triggered, and the drone switches to a high-frequency visual positioning mode (20Hz in this embodiment). Navigation is performed using visual positioning information, and the process returns to step 6 until the judgment result is no. In this case, the program automatically switches back to the low-frequency deception positioning recognition mode, using the drone's GPS for navigation. Otherwise, it remains in the high-frequency visual positioning mode. If the judgment result is no, no positioning deception has occurred, and the low-frequency deception positioning recognition mode is maintained, using the drone's GPS for navigation. The deception defense method only switches to the high-frequency visual positioning mode when positioning deception is detected, to save drone energy. Flexible switching between the low-frequency deception positioning recognition mode and the high-frequency visual positioning mode achieves a balance between reliability and low power consumption in the deception defense method.

[0088] This method is a visual map matching-based anti-drone navigation deception method. Compared with traditional visual positioning methods, it has advantages such as low space occupation, low computational complexity, and data security. Based on effective global scene coordinate information of the map, the method obtains absolute positioning capability and can obtain more reliable visual positioning information. It has great application potential in drone safety and reliability assurance.

Claims

1. A method for defending against drone navigation deception based on visual map matching, characterized in that, Includes the following steps: Step 0: Construct a scene coordinate regression network, which includes an input module, a processing module, and a prediction module connected in sequence; the input module is used to extract image features, the processing module is used to extract and refine features, and the prediction module is used to obtain pose estimation results; Step 1: Obtain a 3D simulation model of the area to be located by combining satellite orthophoto maps with digital terrain models, or by using drone cameras to collect 3D mapping models of the area to be located. Step 2: Extract and process the 3D simulation model to establish a simulation dataset for visual positioning; the simulation dataset includes one-to-one corresponding visual images, scene coordinates, and pose data; Step 3: Preprocess the simulation dataset: obtain the one-to-one normalized image, scene coordinate ground truth, and UAV camera coordinates; Step 4: Read the one-to-one corresponding normalized image, scene coordinate ground truth, and UAV camera coordinates into the scene coordinate regression network to obtain the scene coordinate prediction value that meets the expectations, complete the training of the scene coordinate regression network, and obtain the optimized scene coordinate regression network. Step 5: The drone begins to fly. First, set the drone to low-frequency deception positioning and identification mode. The drone's satellite positioning and navigation module uses satellite positioning and navigation to obtain the drone's location information. Step 6: The drone camera captures visual images, and after normalizing the visual images, they are input into the optimized scene coordinate regression network to obtain pose estimation results. The pose estimation results are transformed from the ECEF coordinate system to the WGS84 coordinate system to obtain the visual positioning information of the UAV. Step 7: Compare the visual positioning information with the location information obtained by the UAV's satellite positioning and navigation module to determine whether it has been spoofed. If so, trigger a spoofing alarm, switch the UAV to high-frequency visual positioning mode, use visual positioning information for navigation, and return to step 6; if not, no spoofing has occurred, return to step 5, and maintain the low-frequency spoofing positioning identification mode.

2. The method for defending against drone navigation deception based on visual map matching according to claim 1, characterized in that, Step 3, preprocessing the simulation dataset, specifically includes: Step 3.1: Normalize the visual image to obtain a normalized image; Step 3.2: Perform coordinate transformation on the scene coordinates, then downsample to obtain the true value of the scene coordinates corresponding to the pose data; Step 3.3: Convert the attitude data into camera coordinates corresponding to the UAV. The camera coordinates are used as the pose data of the UAV.

3. The method for defending against drone navigation deception based on visual map matching according to claim 2, characterized in that, Step 3.3 specifically involves: Step 3.3.1: Perform coordinate transformation on the three-dimensional coordinate data in the attitude data, converting it from the WGS84 coordinate system to the ECEF coordinate system; set a fixed spatial coordinate, and calculate the difference between the transformed three-dimensional coordinate data and the fixed spatial coordinate to obtain the translation component; Step 3.3.2: Convert the yaw, roll, and pitch data in the attitude data into rotation matrices; Step 3.3.3: Combine the translation components and the rotation matrix to obtain the camera coordinates of the UAV.

4. A method for defending against drone navigation deception based on visual map matching as described in claim 3, characterized in that: The input module includes a Stem module and a first module connected in sequence. The Stem module and the first module each contain a convolutional block for convolving the visual image and extracting the features of the visual image. The processing module includes a second module, a third module, and a fourth module connected in sequence. Each of the second, third, and fourth modules includes a PatchEmbed module and a Transformer submodule. The PatchEmbed module divides the image after feature extraction into blocks to match the format requirements of the Transformer submodule. The Transformer submodule outputs the processed feature map. The prediction module includes a convolution module, a scene coordinate filtering module, and a RANSAC method module connected in sequence. The feature map is reduced in dimension by the convolution module, and then the scene coordinate filtering module is used to obtain the scene coordinate prediction value. The scene coordinate prediction value and the UAV pose data are processed by the RANSAC method module to obtain the pose estimation result.

5. A method for defending against drone navigation deception based on visual map matching according to claim 4, characterized in that, Step 4 specifically includes: Step 4.1: Input the normalized image into the Stem module. Set the kernel size of the convolution block of the Stem module to 3, perform convolution on the normalized image, obtain the feature extraction map, and input it into the first module. Step 4.2: The convolutional block of the first module uses three layers of convolution to extract features and performs downsampling to obtain a refined feature map, which is then input into the second module. Step 4.3: The Transformer submodule includes a local attention module and a global attention module; the PatchEmbed module of the second module divides the feature refinement map into blocks to match the format requirements of the Transformer submodule; the blocks are input to the local attention module, output from the local attention module, and then enter the global attention module, output from the global attention module, and finally enter the third module; the local attention module is used to locate and infer local key elements in the feature refinement map, and the global attention module is used to integrate global information on the overall distribution information of the feature refinement map; Step 4.4: The third and fourth modules repeat the processing of the second module, and finally the global attention module of the fourth module outputs the processed feature map; from the second module to the fourth module, the number of heads used in the local attention module increases sequentially, gradually improving the high-level scene abstraction capability of the corresponding module; Step 4.5: Input the feature map into the convolution module. After dimensionality reduction by the convolution module, it passes through the scene coordinate filtering module and then undergoes inverse transformation of the camera coordinates to obtain the scene coordinate prediction value. Compare the scene coordinate prediction value with the scene coordinate ground truth value obtained in Step 3.2 to calculate the loss function. If the loss function calculation result meets the specified expected standard, proceed to Step 4.

6. If the loss function calculation result does not meet the specified expected standard, reselect the normalized image and return to Step 4.1 to increase the number of training rounds until the scene coordinate prediction value that meets the expectations is obtained. Step 4.6: Process the scene coordinate prediction values ​​and the camera coordinates obtained in Step 3.3 through the RANSAC method module to obtain the pose estimation results, complete the training of the scene coordinate regression network, and obtain the optimized scene coordinate regression network.

6. A method for defending against drone navigation deception based on visual map matching according to claim 5, characterized in that, Step 7, determining whether a location scam has occurred, specifically involves: The visual positioning information is compared with the location information obtained by the UAV satellite module. If the error between the two values ​​is greater than the maximum allowable error value, it is judged as yes; otherwise, it is judged as no. The maximum allowable error value refers to the maximum positioning error obtained in the scene coordinate regression network evaluation stage.

7. A method for defending against drone navigation deception based on visual map matching according to claim 6, characterized in that, Step 2, the extraction and processing of the three-dimensional simulation model specifically includes: The 3D simulation model was loaded into the Cesium open-source engine, and the TOPO-DataGen data generation method was used to randomly sample visual images, scene coordinates, and attitude data to establish a simulation dataset for visual positioning. The visual images are RGB images, the scene coordinates are the point cloud coordinate information corresponding to the attitude data, and the attitude data are the six degrees of freedom attitude information of the UAV camera.

8. A method for defending against drone navigation deception based on visual map matching according to claim 7, characterized in that: Step 3.2 specifically involves performing coordinate transformation on the scene coordinates, followed by 8x downsampling to obtain the true scene coordinate values ​​for the corresponding pose of the UAV, in order to reduce the amount of true scene coordinate data. Step 4.2 specifically involves the first module's convolutional block using three layers of convolution to extract features and performing 8x downsampling to ensure that the predicted scene coordinates have the same shape as the true scene coordinates. After obtaining the refined feature map, it is input into the second module.

Images