Unmanned aerial vehicle visual navigation and positioning method, system and related equipment

Abstract

The invention relates to the technical field of unmanned aerial vehicle navigation control, and discloses an unmanned aerial vehicle visual navigation and positioning method, system and related equipment, and the method comprises the following steps: obtaining an image in real time; performing joint correction on the current image, and determining a corrected image; high-dimensional semantic features in the corrected image are obtained, and a candidate area is rapidly and roughly matched from the reference database; performing cross-modal comparison on the depth local feature of the corrected image and the depth local feature of each candidate region to determine a target region; and determining current positioning information of the unmanned aerial vehicle based on the target area and the attitude data of the unmanned aerial vehicle, and performing visual navigation based on the current positioning information. According to the method, the matching range is narrowed through joint correction and rapid rough matching, secondary matching is performed on the narrowed range through cross-modal matching, so that the positioning precision is improved, and finally, the matched target area is combined with the flight data (attitude data) of the unmanned aerial vehicle, so that the positioning precision is further improved.

Description

Background Art

[0002] The satellite signal landing power of the Global Navigation Satellite System (GNSS) is only -130 dbm, and the frequency band is public, making it extremely vulnerable to interference. Once interfered, the drone may not be able to complete the established tasks or even have difficulty returning smoothly. Although when the GNSS signal is interfered, the dead reckoning can be carried out using the data of the Inertial Measurement Unit (IMU), but as time goes on, the problem of error accumulation is relatively serious. Therefore, it is crucial to obtain the accurate position of the drone (that is, the accurate longitude and latitude coordinates) in the case of GNSS interruption or inaccessibility for subsequent related tasks.

[0003] Vision-Based Navigation (VBN) has received extensive attention due to its many advantages such as strong anti-interference ability, low power consumption, low cost, small size, simple device structure, passive positioning, and high positioning accuracy. At the same time, with the continuous development of remote sensing and mapping technologies, high-resolution ortho-satellite or mapping images almost cover all regions of the earth, and each pixel is marked with accurate coordinates. Based on this, matching the top view of the ground taken by the drone with the corresponding remote sensing image or mapping image can achieve accurate and fast drone positioning without cumulative error, which can be used as an important supplementary part of the current drone integrated navigation system.

[0004] However, at present, the traditional vision navigation system has the problem of insufficient positioning accuracy. For example, the traditional scene matching navigation has poor navigation effect for the perspective changes caused by low-altitude flight and large undulating terrain. Summary of the Invention

[0005] In order to overcome the problem of insufficient positioning accuracy in the traditional vision navigation system, the present invention provides a drone vision navigation and positioning method, system and related equipment.

[0006] In a first aspect, to solve the above technical problems, the present invention provides a drone vision navigation and positioning method, including:

[0007] Obtaining an image in real time; wherein, the image is a monocular image taken by a single-spectral camera or a monocular image taken by a multi-spectral camera;

[0008] Performing joint correction on the current image to determine a corrected image;

[0009] Obtaining high-dimensional semantic features in the corrected image, and quickly and roughly matching at least one candidate region similar to the high-dimensional semantic features from a reference database;

[0010] Obtaining depth local features of the corrected image, and performing cross-modal comparison between the depth local features of the corrected image and the depth local features of each candidate region to determine a target region;

[0011] Based on the target area and the attitude data of the drone, determine the current positioning information of the drone, and perform visual navigation based on the current positioning information.

[0012] In a second aspect, the present invention provides a drone vision navigation and positioning system, including:

[0013] An image acquisition module, configured to acquire images in real time; wherein, the image is a monocular image captured by a single-spectral camera or a monocular image captured by a multi-spectral camera;

[0014] An image correction module, configured to perform joint correction on the current image to determine a corrected image;

[0015] A candidate area acquisition module, configured to acquire high-dimensional semantic features in the corrected image, and quickly and roughly match at least one candidate area approximate to the high-dimensional semantic features from a reference database;

[0016] A target area acquisition module, configured to acquire depth local features of the corrected image, perform cross-modal comparison between the depth local features of the corrected image and the depth local features of each candidate area, and determine the target area;

[0017] A positioning and navigation module, configured to determine the current positioning information of the drone based on the target area and the attitude data of the drone, and perform visual navigation based on the current positioning information.

[0018] In a third aspect, the present invention provides a computing device, including a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps of a drone vision navigation and positioning method as described above.

[0019] In a fourth aspect, the present invention provides a computer-readable storage medium, in which instructions are stored. When the instructions run on a terminal device, the terminal device is enabled to execute the steps of a drone vision navigation and positioning method as described above.

[0020] The beneficial effects of the present invention are as follows: perform joint correction on the acquired image to obtain a corrected image, then acquire high-dimensional semantic features in the corrected image, and thus perform rough matching from a reference database to obtain multiple candidate areas. Then, through the cross-domain matching method of a graph neural network, obtain the target area, and thus obtain the current positioning information through the target area and the attitude data of the drone, so as to perform visual navigation of the drone. Through joint correction and quick rough matching, the matching range is reduced in this application, and then the reduced range is secondarily matched through cross-modal matching, thereby improving the positioning accuracy. Finally, the matched target area is combined with the drone flight data (attitude data), further improving the positioning accuracy. Description of the Drawings

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0022] Figure 1 It is a schematic flow diagram of a method for visual navigation and positioning of an unmanned aerial vehicle according to an embodiment of the present invention;

[0023] Figure 2 It is a schematic structural diagram of a system for visual navigation and positioning of an unmanned aerial vehicle according to an embodiment of the present invention. Specific embodiments

[0024] The following embodiments are further explanations and supplements to the present invention and do not constitute any limitation to the present invention.

[0025] The following describes a method, system and related devices for visual navigation and positioning of an unmanned aerial vehicle according to an embodiment of the present invention in conjunction with the accompanying drawings.

[0026] As Figure 1 shown, an embodiment of the present invention provides a method for visual navigation and positioning of an unmanned aerial vehicle, including:

[0027] S1. Obtain images in real time; wherein, the images are monocular images taken by a single-spectral camera or monocular images taken by a multi-spectral camera.

[0028] In this embodiment, the single / multi-spectral camera selects a global shutter camera, which supports the HDR mode to cope with light changes.

[0029] S2. Perform joint correction on the current image to determine the corrected image.

[0030] S3. Obtain high-dimensional semantic features in the corrected image, and quickly and roughly match at least one candidate region similar to the high-dimensional semantic features from the reference database.

[0031] In this embodiment, the reference database generates a multi-scale feature pyramid offline, accelerates retrieval by combining hash coding, stores hierarchical geographical reference data, that is, satellite images or mapping images, and then slices the hierarchical geographical reference data into multiple candidate regions.

[0032] S4. Obtain the depth local features of the corrected image, perform cross-modal comparison between the depth local features of the corrected image and the depth local features of each candidate region, and determine the target region.

[0033] S5. Determine the current positioning information of the unmanned aerial vehicle based on the target region and the attitude data of the unmanned aerial vehicle, and perform visual navigation based on the current positioning information.

[0034] In this embodiment, the attitude data of the UAV is provided by an IMU + barometer. The IMU provides short-term pose prediction, and the barometer assists in altitude estimation.

[0035] In this embodiment, the acquired images are jointly corrected to obtain corrected images, and then high-dimensional semantic features in the corrected images are obtained. Thus, rough matching is performed in the reference database to obtain multiple candidate regions. Then, through cross-domain matching using a graph neural network, the target region is obtained. The current positioning information can be obtained through the target region and the attitude data of the UAV, thereby performing visual navigation of the UAV. Through joint correction and fast rough matching, this application narrows the matching range, and then performs secondary matching on the narrowed range through cross-modal matching, thereby improving the positioning accuracy. Finally, the matched target region is combined with the UAV flight data (attitude data), further improving the positioning accuracy.

[0036] In this embodiment, the hardware on the UAV can use the Jetson Orin NX embedded platform with 100 TOPS computing power support. The TensorRT quantization model (FP16 / INT8) can be deployed on the inference engine. Therefore, various deep learning neural networks can be embedded and deployed on the platform to execute a UAV visual navigation and positioning method as described above, improving the computing efficiency. At the same time, this algorithm also supports acceleration schemes such as GPU, CPU, and NPU.

[0037] Optionally, jointly correcting the current image to determine the corrected image includes:

[0038] Preprocessing the current image through adaptive histogram equalization and dark channel defogging algorithm to determine the preprocessed image;

[0039] Obtain the three-axis acceleration of the IMU on the UAV, and determine the motion blur degree of the preprocessed image based on the three-axis acceleration and the acceleration mean within the exposure time window;

[0040] If the motion blur degree meets the preset requirements, the preprocessed image is used as the corrected image;

[0041] If the motion blur degree does not meet the preset requirements, obtain the current rotation angle of the UAV through the IMU, and perform motion blur compensation on the preprocessed image based on the current rotation angle to determine the corrected image.

[0042] In this embodiment, the adaptive histogram equalization aims to improve the contrast of an image, especially in local regions. It processes the image by applying histogram equalization to local regions (usually small windows) of the image. It calculates the local histogram of each small window and performs equalization to enhance the local contrast, making the details of the image more obvious. To avoid over-enhancement, smoothing is usually also performed to obtain CLAHE (Contrast Limited AHE, Adaptive Histogram Equalization with Limited Contrast).

[0043] In this embodiment, the dark channel defogging algorithm is an image processing method for defogging, based on a simple but effective assumption: in a hazy image, some color channels (such as red, green, and blue) in most image regions will have a very small value, which is called the dark channel. Even in the presence of fog, the value of this dark channel is usually very small. The core idea of dark channel defogging is to estimate the atmospheric light and transmission map (i.e., the density distribution of fog) in the image to restore the fog-free image. It goes through the following steps: 1. Calculate the dark channel image. 2. Estimate the atmospheric light based on the minimum value of the dark channel. 3. Use this estimated result to infer the transmission map of each pixel in the image. 4. Restore the original fog-free image based on the atmospheric light and transmission map.

[0044] In this embodiment, the motion blur degree meeting the preset requirement means that the motion blur degree is less than the threshold.

[0045] Optionally, based on the three-axis acceleration and the average acceleration within the exposure time window, determine the motion blur degree of the preprocessed image. The formula is as follows:

[0046]

[0047] where K represents the motion blur degree, a x 、a y 、a z represent the three-axis acceleration, μ α represents the average acceleration within the exposure time window, N represents the length of the exposure time window, σ motion represents the motion blur degree evaluation function, and σ noise is the standard deviation of Gaussian noise.

[0048] In this embodiment, by constructing a motion blur degree evaluation function, the motion blur degree of the current image is calculated, and the blur degree of the current image is numericalized, which is beneficial to determining whether motion blur compensation is required for the current image based on the motion blur degree, improving the positioning accuracy of the drone.

[0049] Optionally, perform motion blur compensation on the preprocessed image based on the rotation angle to determine the corrected image, including:

[0050] Based on the rotation angle, determine the frequency-domain compensation direction weight matrix, and the formula is as follows:

[0051]

[0052] Where W(u, v) represents the frequency-domain compensation direction weight matrix, u and v are frequency-domain coordinates, representing two orthogonal direction vectors in the frequency domain, are the horizontal and vertical frequency-domain axes of the image, θ represents the rotation angle, and σ represents a preset value;

[0053] Perform motion blur compensation on the preprocessed image based on the frequency-domain compensation direction weight matrix to determine the corrected image.

[0054] In this embodiment, the degree of motion blur of the UAV depends on the tilt angle during UAV flight. Therefore, the current rotation angle of the UAV can be obtained by means of the IMU, and then the frequency-domain compensation direction weight matrix can be calculated through the current rotation angle, so as to perform motion blur compensation on the current image and improve the positioning accuracy of the UAV.

[0055] Optionally, obtain the high-dimensional semantic features in the corrected image, and quickly and roughly match at least one candidate region similar to the high-dimensional semantic features from the reference database, including:

[0056] Extract the high-dimensional semantic features in the corrected image through a lightweight convolutional neural network;

[0057] Use the random hyperplane projection method to map the high-dimensional semantic features into binary hash codes;

[0058] Through the locality-sensitive hashing algorithm, quickly and roughly match at least one candidate region in the reference database that is similar to the binary hash code corresponding to the high-dimensional semantic features.

[0059] In this embodiment, the lightweight convolutional neural network can be a series of backbone networks such as EfficientNet, MobileNet, or ResNet. Perform lightweight reconstruction to remove the last classification layer of the original network, and use global average pooling to obtain the high-dimensional semantic features of the corrected image.

[0060] In this embodiment, pre-compute the hash codes for each candidate region in the reference database and store them in buckets according to the hash values. At this time, the binary hash codes obtained by mapping the high-semantic features can be quickly compared with the hash code values corresponding to all candidate regions, and the Hamming distance is used to measure the similarity of the hash codes, so as to screen out at least one candidate region that is the closest.

[0061] Optionally, perform cross-modal comparison between the depth local features of the corrected image and the depth local features of each candidate region to determine the target region, including:

[0062] Match the depth local features of the corrected image with those of each candidate region, eliminate the feature points with incorrect matching relationships, and determine the target candidate region;

[0063] Take the target candidate region as the target region.

[0064] In this embodiment, depth local features are extracted from the corrected image and candidate regions through technologies such as SIFT (Scale-Invariant Feature Transform), DISK (DIScrete Keypoints), or SuperPoint (a deep learning algorithm for feature detection and matching).

[0065] In this embodiment, neural network algorithms such as SuperGlue, LightGlue, or GIM-LightGlue are used to complete the approximation of the depth local features of the query corrected image and the depth local features of the candidate regions, and obtain the most similar candidate region.

[0066] In this embodiment, neural network algorithms such as SuperGlue, LightGlue, or GIM-LightGlue will compare the similarity of each feature point in the depth local features with each feature point in the candidate region, so as to obtain the feature points with correct matching relationships (feature points with high similarity) and the feature points with incorrect matching relationships (feature points with low similarity).

[0067] In this embodiment, algorithms such as RANSAC (Random Sample Consensus), PROSAC (Progressive Sampling Consensus), or GC-RANSAC (Graph-Cut RANSAC robust estimator) are used to eliminate false matches, that is, feature points with incorrect matching relationships, and obtain a homography matrix to calculate the perspective transformation, complete the pose estimation at the image level, and obtain the target candidate region.

[0068] Optionally, based on the target region and the attitude data of the UAV, determine the current positioning information of the UAV. The formula is as follows:

[0069] Lat = Lattl + Cy·(Latbr - Lattl)

[0070] Lon = Lontl + Cx·(Lonbr - Lontl);

[0071] Among them, Lat and Lon represent the latitude coordinate and the progress coordinate respectively, and represent the current positioning information of the UAV, C x and Cy respectively represent the normalized position information of the UAV in the current image coordinate system, Lat tl , Lon tl represent the latitude and longitude of the upper left corner of the target area, Lat br , Lon br represent the latitude and longitude of the lower right corner of the target area.

[0072] In this embodiment, by corresponding the metadata (target area) in the reference database, the normalized position information in the current image coordinate system is converted into data in the corresponding spatial three-dimensional coordinate system, and the attitude and altitude data (attitude data) of the UAV obtained by the IMU are used to limit the conversion result, so as to obtain high-precision positioning information.

[0073] As Figure 2 shown, the present invention provides a UAV vision navigation and positioning system, including:

[0074] An image acquisition module, configured to acquire images in real time; wherein, the image is a monocular image captured by a single-spectral camera or a monocular image captured by a multi-spectral camera;

[0075] An image correction module, configured to perform joint correction on the current image to determine a corrected image;

[0076] A candidate area acquisition module, configured to acquire high-dimensional semantic features in the corrected image, and quickly and roughly match at least one candidate area approximate to the high-dimensional semantic features from the reference database;

[0077] A target area acquisition module, configured to acquire the depth local features of the corrected image, perform cross-modal comparison between the depth local features of the corrected image and the depth local features of each candidate area, and determine the target area;

[0078] A positioning and navigation module, configured to determine the current positioning information of the UAV based on the target area and the attitude data of the UAV, and perform vision navigation based on the current positioning information.

[0079] Optionally, the image correction module is specifically configured to:

[0080] Perform preprocessing on the current image through adaptive histogram equalization and dark channel dehazing algorithm to determine a preprocessed image;

[0081] Obtain the three-axis acceleration of the IMU on the UAV, and determine the motion blur degree of the preprocessed image based on the three-axis acceleration and the acceleration mean value within the exposure time window;

[0082] If the motion blur degree meets the preset requirements, use the preprocessed image as the corrected image;

[0083] If the degree of motion blur does not meet the preset requirements, obtain the current rotation angle of the UAV through the IMU, and perform motion blur compensation on the preprocessed image based on the current rotation angle to determine the corrected image.

[0084] Optionally, the image correction module is specifically configured to:

[0085] Determine the degree of motion blur of the preprocessed image based on the triaxial acceleration and the average acceleration within the exposure time window. The formula is as follows:

[0086]

[0087] where K represents the degree of motion blur, a x 、a y 、a z represent triaxial acceleration, μ α represents the average acceleration within the exposure time window, N represents the length of the exposure time window, σ motion represents the motion blur degree evaluation function, and σ noise is the standard deviation of Gaussian noise.

[0088] Optionally, the image correction module is specifically configured to:

[0089] Determine the frequency domain compensation direction weight matrix based on the rotation angle. The formula is as follows:

[0090]

[0091] where W(u, v) represents the frequency domain compensation direction weight matrix, u and v are frequency domain coordinates, representing two orthogonal direction vectors in the frequency domain, are the horizontal and vertical frequency domain axes of the image, θ represents the rotation angle, and σ represents a preset value;

[0092] Perform motion blur compensation on the preprocessed image based on the frequency domain compensation direction weight matrix to determine the corrected image.

[0093] Optionally, the candidate region acquisition module is specifically configured to:

[0094] Extract high-dimensional semantic features in the corrected image through a lightweight convolutional neural network;

[0095] Adopt the random hyperplane projection method to map the high-dimensional semantic features into binary hash codes;

[0096] Quickly and roughly match at least one candidate region approximate to the binary hash code corresponding to the high-dimensional semantic features from the reference database through the local sensitive hashing algorithm.

[0097] Optionally, the target region acquisition module is specifically configured to:

[0098] Match the depth local features of the corrected image with the depth local features of each candidate region, eliminate the feature points with incorrect matching relationships, and determine the target candidate region;

[0099] Use the target candidate region as the target region.

[0100] Optionally, the positioning and navigation module is specifically used for:

[0101] Based on the target region and the attitude data of the UAV, determine the current positioning information of the UAV. The formula is as follows:

[0102] Lat = Lattl + Cy·(Latbr - Lattl)

[0103] Lon = Lontl + Cx·(Lonbr - Lontl);

[0104] Wherein, Lat and Lon respectively represent the latitude coordinate and the progress coordinate, and represent the current positioning information of the UAV. C x and C y respectively represent the normalized position information of the UAV in the current image coordinate system. Lat tl 、Lon tl represent the longitude and latitude of the upper left corner of the target region. Lat br 、Lon br represent the longitude and latitude of the lower right corner of the target region.

[0105] The embodiment of the present invention also provides a computing device, including a memory, a manager, and a program stored on the memory and running on the manager. When the manager executes the program, it implements some or all of the steps of the above-mentioned UAV vision navigation and positioning method.

[0106] Among them, the computing device can be a computer. Correspondingly, its program is computer software. And for the parameters and steps in the computing device of the present invention above, reference can be made to the parameters and steps in the embodiments of the UAV vision navigation and positioning method in the above text, which will not be elaborated here.

[0107] Those skilled in the art of the present technology know that the present invention can be implemented as a system, a method, or a computer program product. Therefore, the present disclosure can be specifically implemented in the following forms: it can be completely hardware, can also be completely software (including firmware, resident software, microcode, etc.), or can also be in the form of a combination of hardware and software, which is generally referred to as "circuit", "module", or "system" in this article. In addition, in some embodiments, the present invention can also be implemented in the form of a computer program product in one or more computer-readable media, which contain computer-readable program codes. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination of the above.

[0108] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0109] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limitations of the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A UAV visual navigation and positioning method, characterized in that: include: Acquire an image in real time; wherein the image is a monocular image taken by a single-spectrum camera, or a monocular image taken by a multi-spectrum camera; Perform joint correction on the current image to determine a corrected image; Acquire high-dimensional semantic features in the corrected image, and quickly and roughly match at least one candidate region that is similar to the high-dimensional semantic features from a reference database; Acquire the depth local features of the corrected image, perform cross-modal comparison between the depth local features of the corrected image and the depth local features of each candidate region, and determine the target region; Based on the target area and the posture data of the drone, current positioning information of the drone is determined, and visual navigation is performed based on the current positioning information.

2. The method according to claim 1, characterized in that: The step of performing joint correction on the current image to determine the corrected image includes: Preprocessing the current image by using an adaptive histogram equalization and a dark channel defogging algorithm to determine a preprocessed image; Obtain the three-axis acceleration of the IMU on the drone, and determine the degree of motion blur of the preprocessed image based on the three-axis acceleration and the average acceleration within the exposure time window; If the degree of motion blur meets the preset requirement, the pre-processed image is used as the corrected image; If the degree of motion blur does not meet the preset requirements, the current rotation angle of the drone is obtained through the IMU, and motion blur compensation is performed on the pre-processed image based on the current rotation angle to determine a corrected image.

3. The method according to claim 2, characterized in that Based on the three-axis acceleration and the average acceleration in the exposure time window, the motion blur degree of the pre-processed image is determined, and the formula is as follows: Among them, K represents the degree of motion blur, a x 、a y 、a z represents the three-axis acceleration, μ α represents the mean acceleration within the exposure time window, N represents the length of the exposure time window, σ motion represents the motion blur evaluation function, σ noise is the standard deviation of Gaussian noise.

4. The method according to claim 2, characterized in that: The performing motion blur compensation on the pre-processed image based on the rotation angle to determine the corrected image comprises: Based on the rotation angle, the frequency domain compensation direction weight matrix is ​​determined, and the formula is as follows: Wherein, W(u,v) represents the frequency domain compensation direction weight matrix, u and v are frequency domain coordinates, representing two orthogonal direction vectors in the frequency domain, which are the horizontal and vertical frequency domain axes of the image, θ represents the rotation angle, and σ represents the preset value; Motion blur compensation is performed on the preprocessed image based on the frequency domain compensation direction weight matrix to determine a corrected image.

5. The method according to claim 1, characterized in that The step of acquiring the high-dimensional semantic features in the corrected image and quickly and roughly matching at least one candidate region similar to the high-dimensional semantic features from a reference database comprises: Extracting high-dimensional semantic features in the corrected image through a lightweight convolutional neural network; The random hyperplane projection method is used to map high-dimensional semantic features into binary hash codes; By using a local sensitive hashing algorithm, at least one candidate region that is approximately equal to the binary hash code corresponding to the high-dimensional semantic feature is quickly and roughly matched from a reference database.

6. The method according to claim 1, characterized in that The cross-modal comparison of the depth local features of the corrected image with the depth local features of each candidate region to determine the target region includes: Matching the depth local features of the corrected image with the depth local features of each candidate area, and removing feature points with incorrect matching relationships to determine the target candidate area; The target candidate area is used as the target area.

7. The method according to claim 1, characterized in that The current positioning information of the UAV is determined based on the target area and the UAV's attitude data. The formula is as follows: Lat=Lattl+Cy·(Latbr-Lattl) Lon=Lontl+Cx·(Lonbr-Lontl); Among them, Lat and Lon represent the latitude coordinate and the distance coordinate respectively, and represent the current positioning information of the drone, C x and C y They represent the normalized position information of the drone in the current image coordinate system, Lat tl , Lon tl Indicates the latitude and longitude of the upper left corner of the target area, Lat br , Lon br Indicates the latitude and longitude of the lower right corner of the target area.

8. A UAV visual navigation and positioning system, characterized in that: include: An image acquisition module, used to acquire an image in real time; wherein the image is a monocular image taken by a single-spectrum camera, or a monocular image taken by a multi-spectrum camera; An image correction module, used to perform joint correction on the current image and determine a corrected image; A candidate region acquisition module, used to acquire high-dimensional semantic features in the corrected image, and quickly and roughly match at least one candidate region similar to the high-dimensional semantic features from a reference database; A target region acquisition module is used to acquire the deep local features of the corrected image, perform cross-modal comparison between the deep local features of the corrected image and the deep local features of each candidate region, and determine the target region; The positioning and navigation module is used to determine the current positioning information of the drone based on the target area and the posture data of the drone, and perform visual navigation based on the current positioning information.

9. A computing device comprising a memory, a processor, and a program stored in the memory and running on the processor, characterized in that: When the processor executes the program, the steps of the unmanned aerial vehicle visual navigation and positioning method as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: Instructions are stored in the computer-readable storage medium. When the instructions are executed on the terminal device, the terminal device executes the steps of the unmanned aerial vehicle visual navigation and positioning method as described in any one of claims 1 to 7.

Images