A method and device for visual navigation of a drone in a denied state

Abstract

The application provides a kind of rejection state unmanned aerial vehicle visual navigation method based on region matching and feature point extraction, comprising: reading image sequence to obtain each frame image, and obtaining the parameter information of visual load and the attitude information corresponding to each frame image;The obtained each frame image is subjected to histogram equalization pretreatment and denoising treatment;The attitude information corresponding to the image obtained by using camera inner participation is used as input parameter, and the processed image is subjected to image orthographic correction;Based on the image matching of region, the feature point transformation relationship and the latitude and longitude information of feature point on the image to be measured are obtained;The latitude and longitude information of feature point in the image to be measured is combined with fiber-optic inertial navigation for navigation, and the back transmission to unmanned aerial vehicle flight control system is corrected.The method based on region matching in the application combines image pyramid and regular grid, has high precision, fast running speed and strong algorithm robustness, outputs stable aircraft speed position information for combined inertial navigation, corrects inertial navigation data cumulative error, and is easy to implement in engineering.

Description

Technical Field

[0001] This invention belongs to the field of UAV navigation, and specifically relates to a visual navigation method for UAVs in a denied state based on region matching and feature point extraction. Background Technology

[0002] Under denial conditions, the satellite navigation systems of UAVs, such as BeiDou or GPS, may be subject to external interference or fail directly. Traditional autonomous navigation methods cannot continuously update accurate positioning and attitude data. Inertial navigation systems use angular velocity information measured by gyroscopes to calculate the attitude matrix and accelerometer information for navigation calculations through attitude transformation. The errors obtained in this way, as well as the initial accuracy errors, will propagate in the system, affecting navigation accuracy. More importantly, this leads to the characteristic of error accumulation over time. Without external auxiliary positioning correction, the error will accumulate over time until the end. Secondly, inertial devices also generate errors during the carrier's movement, which accumulate over time.

[0003] To address the passive navigation error problem under denied conditions, other sensors are used for active detection to acquire the UAV's navigation information, which is then fused with inertial sensors to correct the accumulated errors introduced by the inertial devices. Common correction methods under denied conditions include using lidar sensors for navigation and positioning to acquire UAV navigation information, and using visual sensors, including visible light and infrared sensors, to fuse visual sensor and inertial navigation information to correct the UAV's navigation information. However, lidar sensors are bulky, difficult to install, and expensive. In existing technologies, using visual methods to calculate navigation information offers greater adaptability. In particular, acquiring visible light and infrared video through visible light and infrared sensors allows for visual navigation and positioning even in low-light environments, which can be combined with inertial devices to achieve the correction purpose. Summary of the Invention

[0004] To address the problems of existing technologies, a visual navigation method for UAVs in a denied state based on region matching and feature point extraction is proposed. This method preprocesses the acquired images, including histogram stretching, filtering and denoising, and orthorectification to obtain orthorectified image maps. A region matching-based approach is used to downsample the images, employing image pyramids and regular grids for rapid, coarse-to-fine matching layer by layer. This reduces the impact of sparse texture features and complex lighting conditions, maintaining matching accuracy while optimizing efficiency. The latitude and longitude information of the output image map is provided to the integrated navigation system for correction, offering a novel and highly adaptable method for visual integrated navigation.

[0005] The technical solution provided by this invention is as follows:

[0006] Firstly, a visual navigation method for UAVs in a denied state based on region matching and feature point extraction includes the following steps:

[0007] Read the image sequence to obtain each frame image, and obtain the parameter information of the visual payload and the pose information corresponding to each frame image;

[0008] Each acquired image frame is preprocessed with histogram equalization and then denoised.

[0009] Using the pose information of the images acquired by the camera as input parameters, image orthorectification is performed on the processed image.

[0010] The pyramid-scale spatial sequence image is obtained by downsampling the image pyramid. The pyramid-scale spatial sequence image is divided into grids layer by layer using a regular grid. Feature points are extracted by the root mean square of gray level in each direction of the image pixels. Matching is performed under layer-by-layer control from coarse to fine to obtain the matching points corresponding to the feature points and the transformation relationship between the pyramid-scale spatial sequence image and the reference image, thereby obtaining the latitude and longitude information of the feature points on the image to be tested.

[0011] The latitude and longitude information of feature points in the image under test is combined with inertial navigation data to obtain the UAV's position and velocity information, which is then transmitted back to the UAV flight control system.

[0012] Secondly, a visual navigation device for a denied-state unmanned aerial vehicle (UAV) based on region matching and feature point extraction includes:

[0013] The image acquisition module is used to read the image sequence to acquire each frame of the image, and to obtain the parameter information of the visual load and the pose information corresponding to each frame of the image.

[0014] The image preprocessing module is used to perform histogram equalization preprocessing and noise reduction on each frame of the acquired image.

[0015] The orthorectification module is used to perform image orthorectification on the processed image by using the pose information corresponding to the images acquired by the camera as input parameters.

[0016] The image matching module is used to obtain a pyramid-scale spatial sequence image by downsampling the image pyramid, divide the pyramid-scale spatial sequence image into a grid layer by layer using a regular grid, extract feature points using the root mean square of gray level in each direction of the image pixel, and perform matching under layer-by-layer control from coarse to fine to obtain the matching points corresponding to the feature points and the transformation relationship between the pyramid-scale spatial sequence image and the reference image, thereby obtaining the latitude and longitude information of the feature points on the image to be tested.

[0017] The combined return module is used to combine the latitude and longitude information of feature points in the image under test with inertial navigation data to obtain the UAV's position and velocity information, and then return it to the UAV flight control system.

[0018] The visual navigation method and apparatus for a UAV in a denied state based on region matching and feature point extraction provided by the present invention have the following beneficial effects:

[0019] (1) The present invention provides a visual navigation method and device for a UAV in a denial state based on region matching and feature point extraction. The image preprocessing includes histogram equalization and noise reduction filtering. These operations optimize image quality and, while removing noise, minimize the impact of image processing on texture features and feature point extraction, which is beneficial for the extraction of feature corner points based on regions.

[0020] (2) The present invention provides a visual navigation method and device for a UAV in a denial state based on region matching and feature point extraction. The reference map can be downloaded from Google Maps or a high-precision orthophoto after correction and framing can be used. The reference map is collected under different seasonal lighting conditions and different resolutions, and a scale space sequence is constructed. Finally, labels are established based on these seasonal characteristics as backups, which can be more conducive to the success rate of visual matching and the effective output of visual navigation.

[0021] (3) The present invention provides a visual navigation method and device for a UAV in a denied state based on region matching and feature point extraction. Based on the image orthorectification operation of the intrinsic parameters of the visible light camera, the transformation model is obtained by using the intrinsic parameters of the camera, the imaging geometric relationship is established, and the spatial relationship is constructed by using the attitude angle of the visible light camera. The orthorectification of the projection transformation is then performed, which is more conducive to matching with the reference image.

[0022] (4) The present invention provides a visual navigation method and device for a denied state UAV based on region matching and feature point extraction. It uses a regular grid to optimize the distribution of feature points and matches the pyramid image layer by layer from coarse to fine, which greatly improves the efficiency of the matching process and is conducive to improving the efficiency and accuracy of matching without reducing the quality of feature point extraction.

[0023] (5) The present invention provides a visual navigation method and device for UAVs in denial state based on region matching and feature point extraction. Visual navigation and fiber optic inertial navigation are combined for correction. The latitude and longitude obtained by visual matching calculation are used as an active navigation means and fused with the passive fiber optic inertial navigation for offset correction. It is a highly adaptable UAV vision and inertial navigation combination navigation method that can be realized. Attached Figure Description

[0024] Figure 1 This is a flowchart of the visual navigation method of the present invention.

[0025] Figure 2 This is a flowchart of wavelet transform combined with median filtering in the image preprocessing of this invention.

[0026] Figure 3 This is a flowchart of the present invention based on region matching and feature point extraction.

[0027] Figure 4 This is a diagram showing the matching results of the method of the present invention in actual flight. Detailed Implementation

[0028] The features and advantages of the present invention will become clearer and more apparent from the following detailed description.

[0029] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments. Although various aspects of embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated otherwise.

[0030] According to a second aspect of the present invention, a visual navigation method for UAVs in denied conditions based on region matching and feature point extraction is provided for localization of the UAV under denied conditions and drift correction of inertial navigation system (INS) positioning, such as... Figure 1 As shown, it includes the following steps:

[0031] Step 1: Obtain a high-precision reference map to prepare for Step 5. The reference map can cover the ground area within the flight range and can be obtained from Google Maps. Through downsampling processing, obtain enough scale-space images as backup images for matching in Step 5 to form a reference map library.

[0032] The reference map can also use high-precision orthophotos after correction and framing, collect reference maps under different seasonal lighting conditions and different resolutions, construct a scale spatial sequence, and finally establish labels based on these time and seasonal characteristics as backups, which can be more conducive to the success rate of visual matching.

[0033] To ensure that the inertial navigation attitude offset range remains within acceptable limits, an image spanning 2 to 3 times the real-time field of view of the planned UAV flight path should be extracted as a reference image. In practice, a conditional check can be added; if the reference image library cannot be loaded correctly, the navigation method will fail to start.

[0034] Step 2: Read the image sequence to obtain each frame image, and obtain the visual payload such as the parameter information of the visible light camera and the pose information corresponding to each frame image.

[0035] By using drones to capture images of different areas from the air, visual payloads such as visible light cameras are employed. Control commands from the optoelectronic payload pod are used to adjust the visual payload, such as the visible light camera, to a horizontal, downward-looking orientation, reducing edge distortion caused by central projection. The payload pod transmits video streams via an SDI interface and reads the image sequence to acquire each frame.

[0036] Step 3: Perform histogram equalization preprocessing on each frame of the image obtained in Step 2, and use wavelet transform combined with median filtering to improve the quality of the image to be processed, which will be used as the input for Step 4.

[0037] Step 3.1: Convert the image to a single-channel grayscale image, and count the number of gray levels in the single-channel grayscale image to obtain the percentage of each gray level:

[0038]

[0039] Where, r k n represents different gray levels k This represents the total number of pixels at gray level k, while N represents the total number of pixels in the entire image.

[0040] The discrete probability distribution function of gray levels can be obtained from the histogram:

[0041]

[0042] Where, r i Indicates each less than r k grayscale levels;

[0043] Based on the number of grayscale quantization levels M (typically 256 grayscale levels for bitmaps), the quantization values ​​for each grayscale level are obtained:

[0044]

[0045] Where j represents a certain gray level;

[0046] Then, the discrete probability distribution function F(k) of each gray level is re-divided according to the following rule: if F(k) is closest to G(j), then all original gray levels are set to r. k The pixels are converted to grayscale levels r j According to this rule, pixels at each gray level are re-divided into gray levels to complete gray-level histogram equalization.

[0047] Step 3.2: Denoising method combining wavelet transform and median filtering as follows Figure 2 As shown, the specific steps include the following:

[0048] First, the noisy image is decomposed by wavelet decomposition, and the image is divided into three sub-bands: HL, LH, and HH.

[0049] Then, soft thresholding denoising was performed on the three frequency bands respectively;

[0050] Finally, median filtering was applied to each of the three sub-bands, and the image was reconstructed.

[0051] By using grayscale histogram analysis and histogram equalization, the impact of different lighting conditions on the image quality captured by the visible light camera can be reduced to some extent. In terms of image denoising, compared with traditional denoising algorithms, wavelet transform-based image filtering can more effectively preserve the texture features and edge information in the original image while removing noise from visible light images, achieving a better signal-to-noise ratio, which is helpful for subsequent algorithm processing.

[0052] Step 4: Using the pose information of the image obtained in step 2 within the camera as input parameters, perform image orthorectification on the image obtained in step 3 preprocessing.

[0053] Step 5: Obtain the feature point transformation relationship and feature point latitude and longitude information based on region-based image matching, and use the latitude and longitude information as input for step 6.

[0054] Specifically, pyramid-scale spatial sequence images are obtained using image pyramid downsampling. A regular grid is then used to progressively divide the pyramid-scale spatial sequence images into grids. Feature points are extracted using the root mean square of gray levels in each direction of the image pixels. Matching is then performed under coarse-to-fine layer-by-layer control to obtain matching points corresponding to the feature points and the transformation relationship between the pyramid-scale spatial sequence images and the reference image. This allows the acquisition of the latitude and longitude information of the feature points on the image under test. The specific process is as follows: Figure 3 As shown.

[0055] Step 5.1: Based on the different flight altitudes of the drone, select the reference map with different resolutions from Step 1, and choose the one with the closest resolution as the optimal option.

[0056] Step 5.2: Use Gaussian downsampling to downsample the acquired images to obtain a pyramid-scale spatial sequence image. Select 4-6 downsampling times according to the original image size. For example, if the SDI interface transmits a visible light image with an original size of 1920*1080, the scale spatial sequence image obtained after four downsampling times will have 5 images including the original image, with sizes of 1920*1080, 960*540, 480*270, 240*135 and 120*67, corresponding to the original resolution, 1 / 2 resolution, 1 / 4 resolution, 1 / 8 resolution and 1 / 16 resolution.

[0057] Step 5.3: Divide the pyramid-scale spatial sequence image into a grid using a regular grid. The grid division varies for different scales, with the main purpose of evenly distributing all regions of the image as much as possible. This allows for adaptive acquisition of the highest quality feature points in each region during feature point extraction, facilitating matching calculations and ensuring robustness across the entire image. For example, the top layer image uses a 10*6 grid. Therefore, the grid division needs to divide pixels in all directions as evenly as possible.

[0058] Step 5.4: In order to extract feature points more quickly and conveniently, in conjunction with the requirement in Step 5.3 that each grid needs to obtain a certain number of feature points so that the feature points can be evenly distributed, the root mean square of gray level of the grid points is calculated in four directions, and the minimum root mean square of gray level in the four directions is used as the root mean square of gray level of the point and compared with the threshold (the commonly used threshold is 20). If it is greater than the threshold, it can be determined as a feature point. Then, the local maximum suppression method is used to obtain a certain number of feature points that are as evenly distributed as possible in the grid.

[0059]

[0060]

[0061]

[0062]

[0063] Among them, V i This refers to the grayscale variance in all directions, where 2k represents the size of the template and g represents the grayscale value of a specific pixel.

[0064] If the minimum mean square error of gray levels for all points within a grid is less than 20, then based on the average number of feature points extracted from other grids, select the same number of points with the largest minimum mean square error of gray levels that also satisfy local maximum suppression as feature points. For each layer of the scale-space sequence image, extract the feature point sequence for each layer.

[0065] Step 5.5: For the topmost image (the image with the lowest resolution) in the pyramid-scale spatial sequence image, the matching points of the corresponding feature points are obtained on the reference image of the corresponding resolution by using the full-image traversal matching method. Through the same point pairs formed by the feature points and the matching points, the initial transformation model of the geographical location of the image to be tested and the reference image is obtained.

[0066] For the lower layer image outside the top layer image, the transformation model of the upper layer image is used as the position prediction model of the feature points of the layer. The predicted position of the corresponding feature point is obtained on the reference image of the corresponding resolution. Then, a small-range local traversal is performed near the predicted position of the feature point to obtain the matching point corresponding to the feature point of the layer and the matching model of the layer. This process is repeated layer by layer from coarse to fine to the bottom layer (original image layer). The matching model obtained by the matching relationship between the feature points of the bottom layer image and the matching points can be used as the transformation model between the original image and the corresponding reference image.

[0067] Step 5.6: Using the transformation model of the lowest layer image, obtain the latitude and longitude values ​​of the feature points in the image to be tested through the point correspondence relationship, and use them as the input for step 6.

[0068] Step 6: Combine the latitude and longitude information of the feature points in the image to be tested from Step 5 with the data to obtain the UAV's position and speed information, and send it back to the UAV flight control system.

[0069] According to a second aspect of the present invention, a visual navigation device for a UAV in a denied state based on region matching and feature point extraction is provided, comprising:

[0070] The image acquisition module is used to read the image sequence to acquire each frame of the image, and to obtain the parameter information of the visual load and the pose information corresponding to each frame of the image.

[0071] The image preprocessing module is used to perform histogram equalization preprocessing and noise reduction on each frame of the acquired image.

[0072] The orthorectification module is used to perform image orthorectification on the processed image by using the pose information corresponding to the images acquired by the camera as input parameters.

[0073] The image matching module is used to obtain a pyramid-scale spatial sequence image by downsampling the image pyramid, divide the pyramid-scale spatial sequence image into a grid layer by layer using a regular grid, extract feature points using the root mean square of gray level in each direction of the image pixel, and perform matching under layer-by-layer control from coarse to fine to obtain the matching points corresponding to the feature points and the transformation relationship between the pyramid-scale spatial sequence image and the reference image, thereby obtaining the latitude and longitude information of the feature points on the image to be tested.

[0074] The combined return module is used to combine the latitude and longitude information of feature points in the image under test with inertial navigation data to obtain the UAV's position and velocity information, and then return it to the UAV flight control system.

[0075] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0076] This invention has been successfully applied to relevant engineering projects for actual flight and data processing. To verify the feasibility of the visual navigation method for UAVs in denied states based on region matching and feature point extraction involved in this invention, a flight experiment was conducted on a fixed-wing UAV of a certain team. The experiment mainly verified the accuracy of the matching output of this invention during the downward look-down process. A photoelectric payload pod was used for visible light video acquisition, with a visible light focal length of 6mm, a pixel size of 3.4μm, and a resolution of 1920*1080, capable of movement in both pitch and azimuth directions; visible light video was transmitted to the visual navigation solution board via an SDI RF cable to push the H.264 bitstream; the visual navigation module and the fiber optic inertial navigation system were connected to the navigation board to complete the fusion and correction of the navigation data of visual navigation and inertial navigation.

[0077] The matching output of the actual flight data was analyzed. The data recorded the UAV's latitude and longitude calculated by the matching method of this invention, and an error analysis was performed between this data and the latitude and longitude obtained from GPS. In actual flight, a successful matching result was output on average every three to five seconds. Figure 4 As shown in the figure, the accuracy output results indicate that over 95% of the output results have an error within 200 meters, and over 90% of the output accuracy is within 100 meters. Due to the inherent error in GPS positioning, the average matching output error is 30 meters. Therefore, the matching method of this invention can be applied to UAV visual navigation in denied conditions.

[0078] The present invention has been described in detail above with reference to specific embodiments and exemplary examples; however, these descriptions should not be construed as limiting the present invention. Those skilled in the art will understand that various equivalent substitutions, modifications, or improvements can be made to the technical solutions and embodiments of the present invention without departing from the spirit and scope of the invention, and all such modifications and improvements fall within the scope of the present invention. The scope of protection of the present invention is defined by the appended claims.

[0079] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A visual navigation method for UAVs in a denied state based on region matching and feature point extraction, characterized in that, Includes the following steps: Read the image sequence to obtain each frame image, and obtain the parameter information of the visual payload and the pose information corresponding to each frame image; Each acquired image frame is preprocessed with histogram equalization and then denoised. Using the pose information of the images acquired by the camera as input parameters, image orthorectification is performed on the processed image. The pyramid-scale spatial sequence image is obtained by downsampling the image pyramid. The pyramid-scale spatial sequence image is divided into grids layer by layer using a regular grid. Feature points are extracted by the root mean square of gray level in each direction of the image pixels. Matching is performed under layer-by-layer control from coarse to fine to obtain the matching points corresponding to the feature points and the transformation relationship between the pyramid-scale spatial sequence image and the reference image, thereby obtaining the latitude and longitude information of the feature points on the image to be tested. The latitude and longitude information of feature points in the image under test is combined with inertial navigation data to obtain the UAV's position and velocity information, which is then transmitted back to the UAV flight control system. The step of obtaining the matching points corresponding to the feature points and the transformation relationship between the pyramid-scale spatial sequence image and the reference image, thereby obtaining the latitude and longitude information of the feature points on the image to be tested, includes the following sub-steps: Gaussian downsampling was used to downsample the acquired images to obtain a spatial sequence of images at the pyramid scale. Grid division of pyramid-scale spatial sequence images is performed using regular grids; Feature points are extracted from the grid of images at the pyramid scale spatial sequence to obtain the feature point sequence of each layer of the images at the pyramid scale spatial sequence. For the topmost image in the pyramid-scale spatial sequence image, a full-image traversal matching method is used to obtain matching points of corresponding feature points on the reference map of the corresponding resolution. Through the same-name point pairs formed by feature points and matching points, the initial transformation model of matching the geographical location of the image to be tested and the reference map is obtained. For the lower layer image outside the top layer image, the transformation model of the upper layer image is used as the position prediction model of the feature points of the layer. The predicted position of the corresponding feature point is obtained on the reference image of the corresponding resolution. Then, a small local traversal is performed near the predicted position of the feature point to obtain the matching point corresponding to the feature point of the layer and the matching model of the layer. The matching model is obtained layer by layer from coarse to fine to the bottom layer. The matching model obtained by the matching relationship between the feature points of the bottom layer image and the matching points can be used as the transformation model between the original image and the corresponding reference image. The transformation model of the lowest layer image is used to obtain the latitude and longitude values ​​of feature points in the original image through the point correspondence.

2. The visual navigation method for UAVs in a rejected state based on region matching and feature point extraction according to claim 1, characterized in that, In the step of reading the image sequence to obtain each frame image and obtaining the parameter information of the visual payload and the attitude information corresponding to each frame image, the flight image is obtained by using the visual payload in a horizontal downward attitude through drone shooting.

3. The visual navigation method for UAVs in a rejected state based on region matching and feature point extraction according to claim 1, characterized in that, In the step of performing histogram equalization preprocessing and noise reduction on each acquired image frame, the histogram equalization preprocessing is implemented in the following manner: The image is converted into a single-channel grayscale image. The number of grayscale levels in the single-channel grayscale image is counted, and the proportion of each grayscale level is obtained. in, Representing different gray levels, express k Total number of pixels in grayscale N This represents the total number of pixels in the entire image; The discrete probability distribution function of gray levels can be obtained from the histogram: in, Indicates each is less than grayscale levels; Based on the gray quantization level M, the quantization values ​​of each gray level are obtained: in, j Represents a specific gray level; Discrete probability distribution function for each gray level The grading will be redefined according to the following rules: If and The closest one is to set all original gray levels to Pixels are converted to grayscale levels According to this rule, pixels at each gray level are re-divided into gray levels to complete gray-level histogram equalization.

4. The visual navigation method for UAVs in a denied state based on region matching and feature point extraction according to claim 1, characterized in that, In the steps of performing histogram equalization preprocessing and denoising on each acquired image frame, a denoising method combining wavelet transform and median filtering is used.

5. The visual navigation method for UAVs in a rejected state based on region matching and feature point extraction according to claim 1, characterized in that, In the step of using Gaussian downsampling to downsample the acquired images and obtain a pyramid-scale spatial sequence image, downsampling is performed 4-6 times based on the original image size.

6. The visual navigation method for UAVs in a rejected state based on region matching and feature point extraction according to claim 1, characterized in that, The step of extracting feature points from the grid of the pyramid-scale spatial sequence image to obtain the feature point sequence of each layer of the pyramid-scale spatial sequence image is implemented as follows: the root mean square of gray level of the point in the grid is determined by four directions, and the minimum root mean square of gray level in the four directions is used as the root mean square of gray level of the point and compared with a threshold. If it is greater than the threshold, it can be determined as a feature point. Then, the local maximum suppression method is used for filtering to obtain the final feature points in the grid.

7. The visual navigation method for UAVs in a rejected state based on region matching and feature point extraction according to claim 1, characterized in that, In the step of extracting feature points from the grid of the pyramid-scale spatial sequence image to obtain the feature point sequence of each layer of the pyramid-scale spatial sequence image, if the minimum root mean square value of gray level of all points in the grid is less than 20, then based on the average number of feature points extracted in other grids, the points with the largest minimum root mean square value of gray level and satisfying local maximum suppression are selected as feature points.

8. The visual navigation method for UAVs in a denied state based on region matching and feature point extraction according to claim 1, characterized in that, The reference map is an image covering the ground area within the flight range. By acquiring a reference map with the same or similar resolution as the original image, and then downsampling it, a reference map with the same or similar resolution as each layer of the pyramid-scale spatial sequence image is obtained.

9. A visual navigation device for a UAV in a denied state based on region matching and feature point extraction, characterized in that, include: The image acquisition module is used to read the image sequence to acquire each frame of the image, and to obtain the parameter information of the visual load and the pose information corresponding to each frame of the image. The image preprocessing module is used to perform histogram equalization preprocessing and noise reduction on each frame of the acquired image. The orthorectification module is used to perform image orthorectification on the processed image by using the pose information corresponding to the images acquired by the camera as input parameters. The image matching module is used to obtain a pyramid-scale spatial sequence image by downsampling the image pyramid, divide the pyramid-scale spatial sequence image into a grid layer by layer using a regular grid, extract feature points using the root mean square of gray level in each direction of the image pixel, and perform matching under layer-by-layer control from coarse to fine to obtain the matching points corresponding to the feature points and the transformation relationship between the pyramid-scale spatial sequence image and the reference image, thereby obtaining the latitude and longitude information of the feature points on the image to be tested. The combined return module is used to combine the latitude and longitude information of feature points in the image under test with inertial navigation data to obtain the UAV's position and velocity information, and then return it to the UAV flight control system. The image matching module obtains the matching points corresponding to the feature points and the transformation relationship between the pyramid-scale spatial sequence image and the reference image, thereby acquiring the latitude and longitude information of the feature points on the image to be tested, including the following sub-steps: Gaussian downsampling was used to downsample the acquired images to obtain a spatial sequence of images at the pyramid scale. Grid division of pyramid-scale spatial sequence images is performed using regular grids; Feature points are extracted from the grid of images at the pyramid scale spatial sequence to obtain the feature point sequence of each layer of the images at the pyramid scale spatial sequence. For the topmost image in the pyramid-scale spatial sequence image, a full-image traversal matching method is used to obtain matching points of corresponding feature points on the reference map of the corresponding resolution. Through the same-name point pairs formed by feature points and matching points, the initial transformation model of matching the geographical location of the image to be tested and the reference map is obtained. For the lower layer image outside the top layer image, the transformation model of the upper layer image is used as the position prediction model of the feature points of the layer. The predicted position of the corresponding feature point is obtained on the reference image of the corresponding resolution. Then, a small local traversal is performed near the predicted position of the feature point to obtain the matching point corresponding to the feature point of the layer and the matching model of the layer. The matching model is obtained layer by layer from coarse to fine to the bottom layer. The matching model obtained by the matching relationship between the feature points of the bottom layer image and the matching points can be used as the transformation model between the original image and the corresponding reference image. The transformation model of the lowest layer image is used to obtain the latitude and longitude values ​​of feature points in the original image through the point correspondence.

Images