A method for simultaneous localization and mapping of underwater scenes with uneven illumination

Abstract

The present invention discloses a method for simultaneous positioning and mapping of underwater scenes with uneven illumination, comprising the following steps: establishing a data set, collecting data for uneven illumination and performing image processing, establishing a data set for feature extraction and matching of uneven illumination images; designing a feature extractor to extract image point features; encoding and describing point features to implement a method for matching identical features of uneven illumination images; and implementing specific area matching of feature matching results to complete the screening of matching features based on the SLAM operation characteristics. The present invention utilizes the fitting ability of neural networks for specific tasks to implement feature extraction and matching of uneven illumination images, and implements robust matching of image features based on the SLAM operation characteristics. The present invention extracts and matches features based on image processing methods, and can achieve performance superior to traditional feature extraction and matching in specific scenes with uneven underwater illumination, reducing the probability of tracking loss during SLAM operation.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a simultaneous localization and mapping (SLAM) method for underwater scenes with uneven illumination. Background Art

[0002] With the continuous advancement of science and technology, countries around the world are striving to develop underwater exploration technologies to meet the needs of human exploration of underwater environments and the exploitation of abundant marine resources. However, underwater environments often suffer from limited illumination, currents, and turbidity, making most positioning methods difficult to apply. Furthermore, due to the complexity of underwater environments and the high cost of surveying and mapping, underwater environmental information cannot be obtained during underwater surveys. SLAM methods, however, are naturally well-suited for underwater positioning tasks, as they require only relative pose calculations to achieve robot positioning and consume relatively few computational resources.

[0003] However, the current feature point SLAM method based on vision usually needs to assume grayscale consistency, that is, in the process of extracting features on the image and performing feature matching, the grayscale values ​​of the same features should be the same. However, when using the SLAM method based on vision underwater, a point light source is required for illumination, which will cause the point light source of the image to illuminate the area bright and the area away from the point light source to be dark, that is, the illumination is uneven. At this time, the grayscale consistency assumption no longer holds. At the same time, the current existing underwater SLAM image processing method (CN201410833841.0) specifically discloses the specific principle steps of underwater image processing in the SLAM method, uses the underwater light model to remove the interference of the underwater environment, then corrects the underwater image through underwater image enhancement, and finally realizes the matching of feature points in the underwater image through geometric correction. The existing technology considers the influencing factors of underwater images, but ignores the fact that underwater images are usually collected using point light sources for illumination, so that underwater images often have the problem of uneven illumination, causing the feature point matching process to be disturbed. Therefore, how to accurately extract and match features under the condition of uneven illumination becomes the key to the promotion of underwater SLAM technology. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies in the prior art, the purpose of the present invention is to provide a method for feature extraction and matching of uneven image illumination when the vision-based SLAM method requires a point light source for illumination underwater; this method achieves accurate feature extraction and matching by training a feature extractor and a feature matcher for uneven illumination and performing neighborhood matching based on the real-time operating status of SLAM.

[0005] The present invention is achieved through at least one of the following technical solutions.

[0006] A method for simultaneously positioning and mapping an underwater scene with uneven illumination comprises the following steps:

[0007] S1, collecting an image containing a target object and preprocessing the image using histogram equalization;

[0008] S2. extracting point features of the object from the preprocessed image and recording the point feature information;

[0009] S3, randomly selecting a specific area of ​​the image and adjusting the brightness of the image to make it have uneven illumination;

[0010] S4, performing data enhancement on the brightness-adjusted image and the source image, expanding the original image, and generating a new image sample;

[0011] S5. Match the data-enhanced image with the point features extracted from the source image, store the matching information, and construct an uneven illumination image dataset;

[0012] S6. Build a feature extractor to extract features from images, capture stable and distinguishable feature points from images with uneven illumination, and use the dataset to test the performance of the feature extractor. The effectiveness of the feature extractor is judged by evaluating the consistency and robustness of the extracted features under different illumination conditions.

[0013] S7. Construct a feature matcher. The feature matcher performs feature matching on images by mathematically encoding and describing the features, thereby creating associations between feature points of different images under different illumination conditions.

[0014] S8. According to the real-time running status of the simultaneous localization and mapping process (SLAM), the neighborhood screening is performed on the feature matching results, and the feature points that are consistent with the matching of the adjacent areas are screened out in the matching results, thereby reducing mismatching and improving the stability and performance of the SLAM system.

[0015] Furthermore, step S1 includes the following steps:

[0016] S11. Establish a virtual scene dataset, where the images in the virtual scene dataset include basic graphics with rich corner point features;

[0017] S12, calculating the histogram of the image and performing grayscale statistics on the input image;

[0018] S13. Accumulate the histogram to obtain the cumulative distribution function, and remap the original image grayscale to obtain the equalized grayscale:

[0019]

[0020] Among them, a represents the grayscale value of 0 to 255, n aRepresents the number of pixels with gray level a in the image, N is the total number of pixels in the image, k is the maximum gray level in the image, P r (a) is the relative frequency of gray levels in the histogram, r k is the normalized gray level of the original image, T is the histogram cumulative distribution function, s k is the grayscale of the image after histogram equalization.

[0021] Furthermore, step S2 includes the following steps:

[0022] S21. Use feature detection algorithms to analyze the local structure, texture, or edges of the image and find feature points at different scales and directions.

[0023] S22. Store and record the horizontal and vertical positions of the feature points in the image.

[0024] Furthermore, step S3 includes the following steps:

[0025] S31. To simulate the movement characteristics of a point light source in an underwater image, a circular area in the image is randomly selected for processing:

[0026]

[0027] Where i and j are the horizontal and vertical positions of the selected pixel in the image, respectively, x and p y Represent the horizontal and vertical positions of the randomly selected circle center in the image, r is the radius of the randomly selected circle, R c is a randomly selected circular area, R(i,j) is the pixel value of the image at position (i,j);

[0028] S32. Based on the characteristic of underwater point light source grayscale image that the center is bright and the surroundings are dark, the center area of ​​the original image is processed, and the linear brightness change property is:

[0029] g l (i,j)=f(i,j)+b,i,j∈R c

[0030] Where i and j represent the horizontal and vertical positions of the selected pixel in the image; f(i, j) is the pixel at position (i, j) in the original image; g l (i, j) is the transformed image pixel; b is the pixel change value; R c To simulate the illumination area of ​​a point light source.

[0031] Furthermore, step S4 includes the following steps:

[0032] Apply data augmentation to the original image with uneven brightness, so that the original image generates enhanced samples:

[0033] g r =F(g l )

[0034] where g l represents the brightness-adjusted image, g r Represents the image after data enhancement, and F represents the selected data enhancement method.

[0035] Furthermore, step S5 includes the following steps:

[0036] S51, matching the feature point coordinate information stored in step S2 to achieve point feature matching between the image after data enhancement and the feature point information of the original image;

[0037] S52: storing the point feature matching information and performing supervised training as the true value of the dataset of underwater uneven illumination images.

[0038] Furthermore, in step S6, extracting features from the image using a feature extractor includes:

[0039] Downsample the image to H c ×W c , then perform feature point detection for each pixel one by one, the specific optimization function is as follows:

[0040]

[0041] Among them, H c and W c is the height and width of the image after downsampling, x hw It is H c ×W c dimensional feature vector, where each value represents the corresponding value of whether the corresponding pixel is a feature point; hw H c ×W c The eigenvector of the dimension is a real value, which is 1 at the position with the feature point and 0 at the other positions.

[0042] Furthermore, in step S7, the feature matcher performs feature matching on the image, including:

[0043] The input is two images, the original image and the target image. Since there is usually a static world assumption in the SLAM motion process, the feature matching is rigid matching, satisfying the homography assumption. The point features of the two images are matched to obtain the homography matrix H. The positions of the feature points in the original image should be able to completely match the positions of the feature points in the target image after the homography transformation, and these assumptions are not affected by uneven illumination. The feature matching function is as follows:

[0044]

[0045] x s =D(H*I s )

[0046] x t =D(I t )

[0047] Among them, H c and W c is the height and width of the image after downsampling, L match is the feature matching function, x s is the F-dimensional feature vector of the original image, x t The target image H c ×W c dimensional feature vector; I s is the original image; I t is the target image; D(·) stretches the image into H c ×W c vector.

[0048] Furthermore, step S8 includes the following steps:

[0049] S81, integrate the feature extractor and feature matcher into SLAM to replace the traditional visual feature extraction;

[0050] S82. Based on the real-time operating status of the SLAM system, when the SLAM is in a stable operating state, each continuously input image is divided into k sub-regions; the feature information in each sub-region is counted; in the process of matching the current and subsequent frames, matching point pairs are output; the output feature point pairs are filtered based on whether they are in the same sub-region as a principle; thereby improving the robustness of the final output matching point pairs. Furthermore, data enhancement means include:

[0051] (1) By rotating the image, we simulate the rotation state during underwater positioning and mapping, so that the invariance of target shape and structure in feature matching is enhanced and the target can be adapted to appear at different angles in the image. The image rotation is as follows:

[0052] g r =Rotate(g l ,θ)

[0053] where g l represents the brightness-adjusted image, g r Represents the image after data enhancement, Rotate represents the rotation function, and θ represents the clockwise rotation angle;

[0054] (2) By translating the image, we simulate the movement during underwater positioning and mapping, so that feature matching can adapt to the position change of the target in the image. The image translation is as follows:

[0055] g r =Translate(g l ,δ)

[0056] where g l represents the brightness-adjusted image, g r Represents the image after data enhancement, Translate represents the translation function, and δ represents the pixel value translated to the right;

[0057] (3) By randomly cropping the image, we simulate the target being occluded during underwater simultaneous positioning and mapping, and introduce a transformation to the diversity of local areas of the image. The image is randomly occluded as follows:

[0058] g r =Mask(g l ,γ)

[0059] where g l represents the brightness-adjusted image, g r represents the data-enhanced image, Mask represents the occlusion transformation function, and γ represents the randomly selected occlusion area in the image.

[0060] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0061] 1. The present invention constructs a diverse dataset that is suitable for model training based on existing images. Taking into account the uneven illumination of underwater images caused by point light sources, it helps SLAM adapt to changes in lighting conditions in the world.

[0062] 2. The present invention extracts and matches features based on image processing methods, which can achieve better performance than traditional feature extraction and matching in specific scenarios with uneven underwater illumination, reduce the probability of tracking loss during SLAM operation, and thus improve the robustness of SLAM operation.

[0063] 3. The present invention filters the feature matching results based on the real-time running status of SLAM, which adds a constraint condition compared to using only feature matching and provides a robust feature matching effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flow chart of a method for simultaneous positioning and mapping of underwater scenes with uneven illumination according to the present invention;

[0065] Figure 2 This is a schematic diagram of the process of constructing an uneven illumination data set according to the present invention;

[0066] Figure 3 It is a schematic diagram of removing mismatched points based on the real-time running status of SLAM. DETAILED DESCRIPTION

[0067] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] This embodiment provides a method for simultaneously positioning and mapping underwater scenes with uneven illumination. Figure 1 As shown in the figure, it includes the following 8 steps: collecting images and performing preprocessing; extracting and recording point feature information; adjusting image brightness to obtain uneven illumination characteristics; enhancing uneven illumination image data; constructing an uneven illumination image dataset; constructing a feature extractor to extract image point features; constructing a feature matcher to achieve feature association under different illumination conditions; and performing feature neighborhood matching according to the SLAM operation status.

[0069] Specifically, S1 collects and preprocesses images, selects appropriate data sources related to the underwater SLAM task, and preprocesses the collected images using the histogram equalization method to improve image quality, including the following steps:

[0070] S11. Establish a virtual scene dataset, where the images in the virtual scene dataset include basic graphics with rich corner features.

[0071] As an embodiment, the basic graphics are squares, triangles, letters, etc.

[0072] S12. Calculate the histogram of the image and perform grayscale statistics on the input image.

[0073] S13. Accumulate the histogram to obtain the cumulative distribution function, and remap the original image grayscale to obtain the equalized grayscale:

[0074]

[0075] Among them, j represents the gray level, which is generally an integer value from 0 to 255, and n j represents the number of pixels with gray level j in the image, N is the total number of pixels in the image, k is the maximum gray level in the image, P r (j) is the relative frequency of gray levels in the histogram, r k is the normalized gray level of the original image, T is the histogram cumulative distribution function, s k is the grayscale of the image after histogram equalization.

[0076] At this point, the source images and preprocessing of the dataset are completed.

[0077] Specifically, S2, extracting and recording point feature information to capture key information in the image, includes the following steps:

[0078] S21. Use feature detection algorithms to analyze the local structure, texture, or edges of the image and find significant feature points at different scales and directions.

[0079] S22. Store and record the horizontal and vertical positions of the feature points in the image.

[0080] At this point, the feature information of the original image points in the dataset has been extracted and recorded.

[0081] Specifically, S3 adjusts the image brightness to obtain the uneven illumination characteristics, and simulates the illumination conditions of the underwater image point light source by changing the image brightness value, which specifically includes:

[0082] S31. To simulate the movement characteristics of a point light source in an underwater image, a circular area in the image is randomly selected for processing:

[0083]

[0084] Where i and j are the horizontal and vertical positions of the selected pixel in the image, respectively, x and p y Represent the horizontal and vertical positions of the randomly selected circle center in the image, r is the radius of the randomly selected circle, R c is a randomly selected circular area, and R(i,j) is the pixel value of the image at position (i,j).

[0085] S32. Based on the characteristic of underwater point light source grayscale image that the center is bright and the surroundings are dark, the center area of ​​the original image is processed, and the linear brightness change property is:

[0086] g l (i,j)=f(i,j)+b,i,j∈R c

[0087] Where i and j represent the horizontal and vertical positions of the selected pixel in the image; f(i, j) is the pixel at position (i, j) in the original image; g(i, j) is the pixel in the transformed image; b is the pixel change value; R c To simulate the illumination area of ​​a point light source.

[0088] At this point, the original image of the dataset has the characteristic of uneven illumination due to the change of grayscale value.

[0089] Specifically, S4, data enhancement for uneven illumination images, simulates the factors that affect feature matching in underwater scenes through data enhancement, including the following steps:

[0090] Data augmentation is applied to the original image with uneven brightness to improve the generalization ability of subsequent model learning, so that the original image generates enhanced samples:

[0091] g r =F(g l )

[0092] where g l represents the brightness-adjusted image, g r Represents the image after data enhancement, and F represents the selected data enhancement method.

[0093] As an embodiment, the data augmentation methods used include image rotation, image translation, and random image cropping. By rotating the image, we simulate the rotation state during underwater simultaneous positioning and mapping, so that the invariance of target shape and structure related to feature matching is enhanced, and the target can be adapted to appear at different angles in the image. The image rotation is as follows:

[0094] g r =Rotate(g l ,θ)

[0095] where g l represents the brightness-adjusted image, g r Represents a data-enhanced image, Rotate represents a rotation function, and θ represents a clockwise rotation angle, generally ranging from 0 to 360 degrees.

[0096] By translating the image, we simulate the movement during underwater positioning and mapping, so that feature matching can adapt to the position change of the target in the image. The image translation is as follows:

[0097] g r =Translate(g l ,δ)

[0098] where g l represents the brightness-adjusted image, g r Represents the image after data enhancement, Translate represents the translation function, and δ represents the pixel value translated to the right.

[0099] By randomly cropping the image, we simulate the target being occluded during underwater simultaneous positioning and mapping, and introduce a transformation that diversifies the local areas of the image. The image is randomly occluded as follows:

[0100] g r =Mask(g l ,γ)

[0101] where g l represents the brightness-adjusted image, g r represents the data-enhanced image, Mask represents the occlusion transformation function, and γ represents the randomly selected occlusion area in the image.

[0102] At this point, the original image with uneven illumination is expanded through data augmentation.

[0103] Specifically, S5, constructing an uneven illumination image dataset, collecting high-quality and diverse image data and corresponding image matching relationship truth values, includes the following steps:

[0104] S51 , according to S3 , adjusting the image brightness to obtain the feature point coordinate information stored in the illumination unevenness characteristic, so as to achieve point feature matching between the feature point information of the image after data enhancement and the original image.

[0105] S52: storing the point feature matching information and performing supervised training as the true value of the dataset of underwater uneven illumination images.

[0106] At this point, the illumination unevenness image feature extraction matching dataset is constructed, such as Figure 2 shown.

[0107] Specifically, S6 designs a feature extractor that can capture stable and distinguishable feature points from images with uneven illumination. The constructed dataset is used to test the performance of the feature extractor, and its effectiveness is determined by evaluating the consistency and robustness of the extracted features under different illumination conditions. This includes the following steps:

[0108] Build a feature extractor; downsample the image to H c ×W c , then perform feature point detection for each pixel one by one, the specific optimization function is as follows:

[0109]

[0110] Among them, H c and W c is the height and width of the image after downsampling, x hw It is H c ×W c dimensional feature vector, where each value represents the corresponding value of whether the corresponding pixel is a feature point; hw H c ×W c The eigenvector of the dimension is a real value, which is 1 at the position with the feature point and 0 at the other positions.

[0111] At this point, the feature extractor is constructed.

[0112] Specifically, S7 constructs a feature matcher to achieve feature association under different illumination conditions. The matcher can retrieve the same feature points in the objective world from the feature points extracted from the uneven illumination image, thereby establishing an effective feature matching relationship. The feature matcher creates an association relationship between feature points in different images under different illumination conditions through mathematical coding and mathematical description of the features, as follows:

[0113] Construct a feature matcher; the input is two images, the original image and the target image; since there is usually a static world assumption in the SLAM motion process, the feature matching is rigid matching, satisfying the homography assumption; match the point features of the two images to obtain the homography matrix H; the position of the feature points in the original image should be able to completely match the position of the feature points in the target image after the homography transformation, and this assumption is not affected by uneven illumination. The feature matching function is as follows:

[0114]

[0115] x s =D(H*I s )

[0116] x t =D(I t )

[0117] Among them, H c and W c is the height and width of the image after downsampling, L match is the feature matching function, x s is the F-dimensional feature vector of the original image, x t The target image H c ×W c dimensional feature vector; I s is the original image; I t is the target image; D(·) stretches the image into H c ×W c vector.

[0118] At this point, the feature matcher is constructed.

[0119] Specifically, S8 performs feature neighborhood matching based on the SLAM operating status. Based on the real-time operating status of the SLAM, the feature matching results are subjected to neighborhood screening, and feature points with high matching consistency with adjacent areas are screened out from the matching results, thereby reducing mismatches and improving the stability and performance of the SLAM system. This includes the following steps:

[0120] S81. Integrate the feature extractor and feature matcher into the SLAM system to replace the traditional visual feature extraction.

[0121] S82. According to the real-time operation status of the SLAM system, when the SLAM is in a stable operation state, each continuously input image is divided into k sub-regions; the feature information in each sub-region is counted; in the process of matching the current and subsequent two frames of images, the matching point pairs are output; the output feature point pairs are filtered based on whether they are in the same sub-region; thereby improving the robustness of the final output matching point pairs.

[0122] At this point, through the real-time operation status of SLAM, feature robustness matching under uneven underwater illumination can be achieved, such as Figure 2 As shown in Figure 1, three mismatched points obtained through nearest neighbor matching are deleted using the rotational operating state fed back by SLAM.

[0123] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, numerous modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, so that those skilled in the art can better understand and utilize the present invention.

Claims

1. A method for simultaneous positioning and mapping of underwater scenes with uneven illumination, characterized in that: The following steps are involved: S1, collecting an image containing a target object and preprocessing the image using histogram equalization; S2. extracting point features of the object from the preprocessed image and recording the point feature information; S3, randomly selecting an image area and adjusting the brightness of the image to make it have uneven illumination; S4, performing data enhancement on the brightness-adjusted image and the source image, expanding the original image, and generating a new image sample; S5. Match the data-enhanced image with the point features extracted from the source image, store the matching information, and construct an uneven illumination image dataset; S6. Build a feature extractor to extract features from images, capture stable and distinguishable feature points from images with uneven illumination, and use the dataset to test the performance of the feature extractor. The effectiveness of the feature extractor is judged by evaluating the consistency and robustness of the extracted features under different illumination conditions. Using a feature extractor to extract features from an image includes: Downsample the image to H c ×W c , then perform feature point detection for each pixel one by one, the specific optimization function is as follows: Among them, H c and W c is the height and width of the image after downsampling, x hw It is H c ×W c dimensional feature vector, where each value represents the corresponding value of whether the corresponding pixel is a feature point; hw H c ×W c The eigenvector of the dimension is a real value, which is 1 at the position with the feature point and 0 at the other positions; S7. Construct a feature matcher. The feature matcher performs feature matching on images by mathematically encoding and describing the features, thereby creating associations between feature points of different images under different illumination conditions. The feature matcher performs feature matching on images including: The input is two images, the original image and the target image. Since there is usually a static world assumption in the SLAM motion process, the feature matching is rigid matching, satisfying the homography assumption. The point features of the two images are matched to obtain the homography matrix H. The positions of the feature points in the original image should be able to completely match the positions of the feature points in the target image after the homography transformation, and these assumptions are not affected by uneven illumination. The feature matching function is as follows: x s =D(H*I s ) x t =D(I t ) Among them, H c and W c is the height and width of the image after downsampling, L match is the feature matching function, x s is the F-dimensional feature vector of the original image, x t The target image H c ×W c dimensional feature vector; I s is the original image; I t is the target image; D(·) stretches the image into H c ×W c vector, H represents the homography matrix; S8. Based on the real-time running status of the simultaneous localization and mapping process SLAM, the neighborhood screening is performed on the feature matching results, and the feature points that are consistent with the matching of the adjacent areas are screened out in the matching results, thereby reducing mismatching and improving the stability and performance of the SLAM system.

2. The method for simultaneous positioning and mapping of underwater scenes with uneven illumination according to claim 1, characterized in that: Step S1 includes the following steps: S11. Establish a virtual scene dataset, where the images in the virtual scene dataset include basic graphics with rich corner point features; S12, calculating the histogram of the image and performing grayscale statistics on the input image; S13. Accumulate the histogram to obtain the cumulative distribution function, and remap the original image grayscale to obtain the equalized grayscale: Among them, a represents the gray level of integer value from 0 to 255, n a Represents the number of pixels with gray level a in the image, N is the total number of pixels in the image, k is the maximum gray level in the image, P r (a) is the relative frequency of gray levels in the histogram, r k is the normalized gray level of the original image, T is the histogram cumulative distribution function, s k is the grayscale of the image after histogram equalization.

3. The method for simultaneous positioning and mapping of underwater scenes with uneven illumination according to claim 1, characterized in that: Step S2 includes the following steps: S21. Use feature detection algorithms to analyze the local structure, texture, or edges of the image and find feature points at different scales and directions. S22. Store and record the horizontal and vertical positions of the feature points in the image.

4. The method for simultaneous positioning and mapping of underwater scenes with uneven illumination according to claim 1, characterized in that: Step S3 includes the following steps: S31. To simulate the movement characteristics of a point light source in an underwater image, a circular area in the image is randomly selected for processing: Where i and j are the horizontal and vertical positions of the selected pixel in the image, respectively, x and p y Represent the horizontal and vertical positions of the randomly selected circle center in the image, r is the radius of the randomly selected circle, R c is a randomly selected circular area, R(i,j) is the pixel value of the image at position (i,j); S32. Based on the characteristic of underwater point light source grayscale image that the center is bright and the surroundings are dark, the center area of ​​the original image is processed, and the linear brightness change property is: g l (i,j)=f(i,j)+b,i,j∈R c Where i and j represent the horizontal and vertical positions of the selected pixel in the image; f(i, j) is the pixel at position (i, j) in the original image; g l (i, j) is the transformed image pixel; b is the pixel change value; R c To simulate the illumination area of ​​a point light source.

5. The method for simultaneous positioning and mapping of underwater scenes with uneven illumination according to claim 1, characterized in that: Step S4 includes the following steps: Apply data augmentation to the original image with uneven brightness, so that the original image generates enhanced samples: g r =F(g l ) where g l represents the brightness-adjusted image, g r Represents the image after data enhancement, and F represents the selected data enhancement method.

6. The method for simultaneous positioning and mapping of underwater scenes with uneven illumination according to claim 1, characterized in that: Step S5 includes the following steps: S51, matching the feature point coordinate information stored in step S2 to achieve point feature matching between the image after data enhancement and the feature point information of the original image; S52: storing the point feature matching information and performing supervised training as the true value of the dataset of underwater uneven illumination images.

7. The method for simultaneous positioning and mapping of underwater scenes with uneven illumination according to claim 1, characterized in that: Step S8 includes the following steps: S81, integrate the feature extractor and feature matcher into SLAM to replace the traditional visual feature extraction; S82. According to the real-time operation status of the SLAM system, when the SLAM is in a stable operation state, each continuously input image is divided into k sub-regions; the feature information in each sub-region is counted; in the process of matching the current and subsequent two frames of images, the matching point pairs are output; the output feature point pairs are filtered based on whether they are in the same sub-region; thereby improving the robustness of the final output matching point pairs.

8. The method for simultaneous positioning and mapping of underwater scenes with uneven illumination according to claim 5, characterized in that: Data augmentation methods include: (1) By rotating the image, we simulate the rotation state during underwater positioning and mapping, so that the invariance of target shape and structure in feature matching is enhanced and the target can be adapted to appear at different angles in the image. The image rotation is as follows: g r =Rotate(g l ,θ) where g l represents the brightness-adjusted image, g r Represents the image after data enhancement, Rotate represents the rotation function, and θ represents the clockwise rotation angle; (2) By translating the image, we simulate the movement during underwater positioning and mapping, so that feature matching can adapt to the position change of the target in the image. The image translation is as follows: g r =Translate(g l ,δ) where g l represents the brightness-adjusted image, g r Represents the image after data enhancement, Translate represents the translation function, and δ represents the pixel value translated to the right; (3) By randomly cropping the image, we simulate the target being occluded during underwater simultaneous positioning and mapping, and introduce a transformation to the diversity of local areas of the image. The image is randomly occluded as follows: g r =Mask(g l ,c) where g l represents the brightness-adjusted image, g r represents the data-enhanced image, Mask represents the occlusion transformation function, and γ represents the randomly selected occlusion area in the image.

Images