A method for realizing panoramic viewing of unmanned ship

Abstract

The present invention relates to a method for achieving panoramic surround vision of an unmanned vessel, comprising: collecting panoramic stitched images of the unmanned vessel, converting the panoramic stitched images from RGB color space to HSV color space, obtaining a grayscale image, preprocessing the grayscale image to obtain superpixel regions, and further generating a saliency map; weighting waves in the grayscale image according to the saliency map to generate a region of interest, performing SLIC0 superpixel processing on the region of interest to generate a superpixel map; connecting the superpixel map with a HOG feature vector to obtain a composite feature vector, calculating the gradient direction distribution within each superpixel region, obtaining a symmetric gradient feature, fusing the composite feature vector with the symmetric gradient feature, inputting the fused feature into a support vector machine for classification, and performing panoramic image stitching on the classification results. The present invention fully considers the effects of insufficient illumination and water surface ripples on the water surface, thereby improving the stitching quality of the unmanned vessel's panoramic images.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a method for achieving panoramic viewing of an unmanned ship. Background Art

[0002] A vessel surround view system (also known as a 360-degree surround view system or panoramic vessel monitoring system) is an advanced visual monitoring solution designed specifically for ships. The system captures images using multiple cameras installed around the vessel, then stitches these images together to create a seamless panoramic view. The primary purpose of this system is to provide the captain and crew with comprehensive visual information, enabling them to better assess the vessel's position and surroundings, particularly when navigating confined waters, at anchor in ports, or while underway.

[0003] Despite significant progress in many areas, ship surround view systems still face challenges in practical applications. This is particularly true in complex water environments, low lighting conditions, and the effects of surface ripples, where performance is often unsatisfactory. Waves and currents on the water surface cause the positions of objects in the image to constantly shift, complicating image stitching and target recognition. Surface fluctuations can also cause image jitter, impacting system stability and accuracy.

[0004] Among existing image processing technologies, the implementation of panoramic surround view systems for unmanned vessels is crucial for improving their environmental perception and ability to perform complex tasks. However, due to illumination variations, noise interference, water surface ripple interference, and discontinuities in the image stitching process, traditional image feature extraction methods often fail to provide stable and robust feature descriptions. Summary of the Invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes a method for realizing panoramic viewing of an unmanned boat. This method fully considers the influence of insufficient lighting and water ripples on the water surface, is more suitable for the water surface environment than the traditional method based on feature detection and feature matching, and improves the accuracy of image stitching.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] To achieve the above objectives, the present invention provides a method for achieving panoramic viewing of an unmanned vessel, comprising:

[0008] Acquire a panoramic stitched image of the unmanned vessel, convert the panoramic stitched image from an RGB color space to an HSV color space to obtain a grayscale image, preprocess the grayscale image to obtain a superpixel region, and generate a saliency map based on the superpixel region;

[0009] Weight the waves in the grayscale image according to the saliency map to generate a region of interest, and perform SLIC0 superpixel processing on the region of interest to generate a superpixel map;

[0010] Connect the superpixel map with the HOG feature vector to obtain a composite feature vector, calculate the gradient direction distribution inside each superpixel region to obtain the symmetric gradient feature, fuse the composite feature vector and the symmetric gradient feature, input the fused feature into a support vector machine for classification, and perform panoramic stitching on the classification result.

[0011] Optionally, generating the saliency map includes:

[0012] Perform noise removal and superpixel segmentation on the grayscale image to obtain the superpixel regions, extract the feature information of the superpixel regions, and perform multi-cue fusion by combining the saliency values of each superpixel region calculated from the global and local contrasts to generate a preliminary saliency map; the feature information includes: color histogram, texture feature and shape feature;

[0013] Perform smoothing processing on the preliminary saliency map, perform threshold segmentation on the smoothed preliminary saliency map, and remove the background light by adjusting the brightness and contrast of the background region to further optimize the boundary of the salient object to generate the saliency map.

[0014] Optionally, generating the region of interest includes:

[0015] Weight the waves in the grayscale image according to the saliency map to obtain the saliency-weighted wave coordinate features;

[0016] Set potential feature screening conditions:

[0017]

[0018] where Th hl is the maximum length of the ship, D l is the horizontal distance between two waves on the left and right of the same ship, that is, the length, D h is the vertical distance between two waves, that is, the spacing, L1 and L2 are the coordinates of the left wave respectively, and L'1 and L'2 are the coordinates of the right wave;

[0019] Based on the potential feature screening conditions, use the horizontal distance and horizontal position characteristics of the waves to screen potential wave pairs, and generate the region of interest through the wave coordinate features and the wave pairs.

[0020] Optionally, generating the superpixel map includes:

[0021] Obtain the pixel points of the region of interest, use the pixel points to set a number of superpixels in the region of interest, and distribute multiple initial clustering centers;

[0022] Calculate the spectral distance between the initial clustering centers and the pixels in the region, normalize the spectral distance, assign pixels to each superpixel, and determine the maximum spectral distance from all the pixels assigned to the superpixel;

[0023] Use the maximum spectral distance to update the pixel assignment of the superpixels. After the assignment is completed, merge the isolated pixels into the adjacent superpixels to generate a superpixel map.

[0024] Optionally, distributing the multiple initial clustering centers includes:

[0025] Set an initial center target neighborhood, and find the pixel with the minimum gradient within the target neighborhood of each initial center as the initial clustering center.

[0026] Optionally, the formula for calculating the spectral distance between the initial clustering centers and the pixels in the region is:

[0027]

[0028] where D s is the comprehensive distance between the clustering center and the pixel, including a weighted combination of the color distance and the spatial distance, d lab is the color distance between the pixel and the clustering center in the CIELAB color space, d xy is the spatial distance between the pixel and the clustering center on the two-dimensional image plane, m is the balance factor between the color distance and the spatial distance, and S is the scale parameter of the grid size of the region where the clustering center is located or the size of the segmentation region.

[0029] Optionally, obtaining the symmetric gradient feature includes:

[0030] Calculate the gradient direction distribution inside each superpixel region, divide the gradient feature into horizontal and vertical direction histograms, and extract the symmetric gradient features of the horizontal and vertical direction histograms.

[0031] Optionally, panorama stitching of the classification results includes:

[0032] Construct a hyperplane using a radial function, input the fused feature as an input vector into the support vector machine according to the hyperplane for classification, and obtain the classification result;

[0033] The formula of the radial function is: K(x,x') = exp(-γ||x - x'|| 2 );

[0034] Where x and x' are the feature vectors of the input samples, ||x-x'|| 2 is the Euclidean distance between two samples, γ is the hyperparameter of the kernel function, which controls the width of the radial basis function;

[0035] The classification results are stitched into a panorama, and during the panorama stitching process, the overlapping areas of the images are color fused.

[0036] Optionally, the formula for obtaining the overlapping area is:

[0037]

[0038] in, is the overlapping area, W is the window function, (x i ,y j ) is the pixel coordinate, I x , I y They are horizontal gradient and vertical gradient respectively.

[0039] The beneficial effects of the present invention are:

[0040] This method extracts image feature points by integrating a region-based superpixel HOG feature method. The preprocessing method uses the Simple Linear Iterative Clustering Zero (SLIC0) superpixel algorithm to weight each pixel in the HOG feature generation, improving adaptive feature extraction and noise reduction capabilities. This method can remove the interference of ripples on the water surface on image stitching, making it more suitable for aquatic environments. The system also optimizes image feature extraction, more effectively distinguishing between textured and non-textured areas, and improving the accuracy of image stitching. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 The present invention is a flowchart of a method for achieving panoramic viewing of an unmanned vessel according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 1 As shown, this embodiment discloses a method for realizing panoramic surround viewing of an unmanned ship, including: collecting panoramic stitching images of the unmanned ship, converting the panoramic stitching images from the RGB color space to the HSV color space, obtaining a grayscale image, preprocessing the grayscale image, obtaining a superpixel area, and generating a saliency map based on the superpixel area; weighting the waves in the grayscale image according to the saliency map to generate a region of interest, performing SLIC0 superpixel processing on the region of interest to generate a superpixel map; connecting the superpixel map with the HOG feature vector to obtain a composite feature vector, calculating the gradient direction distribution within each superpixel area, obtaining a symmetric gradient feature, fusing the composite feature vector with the symmetric gradient feature, inputting the fused feature into a support vector machine for classification, and performing panoramic image stitching on the classification results.

[0046] Specifically:

[0047] This embodiment discloses a method for achieving panoramic viewing of an unmanned vessel, including:

[0048] The unmanned vessel surround view system includes an NVIDIA development module and is connected to the camera group via a deserializer;

[0049] The image processing module in the NVIDIA development module stitches the acquired images together to form a panoramic image;

[0050] The acquired image is subjected to image distortion correction using the open source software library OpenCV, and then the image is processed using the method of the present invention:

[0051] Step 1: Convert the rectified image from RGB color space to HSV color space to obtain a grayscale image, and extract water surface ripples by combining background illumination removal and saliency model;

[0052] In the second step, these water surface ripples are combined with a template-based method to delineate the areas containing potential features;

[0053] Step 3: Superpixel and HOG (S-HOG) feature fusion is performed in these areas;

[0054] Step 4: The vertical histogram of symmetric features of directional gradients (V-HOG) is integrated, and support vector machine (SVM) is used for classification, and the classification results are spliced in the next step;

[0055] Step 5: Use gray-scale fusion technology to stitch the processed images into a panorama.

[0056] Furthermore, generating a saliency map includes: removing noise and performing superpixel segmentation on the grayscale image to obtain superpixel regions, extracting the feature information of the superpixel regions, and performing multi-cue fusion on the saliency values of each superpixel region calculated by combining global and local contrasts to generate a preliminary saliency map; performing smoothing on the preliminary saliency map, performing threshold segmentation on the smoothed preliminary saliency map, and removing the background light by adjusting the brightness and contrast of the background region to further optimize the boundaries of the salient objects and generate a saliency map; where the feature information includes: color histogram, texture feature, and shape feature.

[0057] The above superpixel segmentation is equivalent to a rough processing, aiming to make each superpixel contain one or more pixels and be relatively consistent in terms of color, texture, brightness, etc.

[0058] Specifically:

[0059] The saliency-based background light removal method significantly improves the computational efficiency and detection performance by using a region-level saliency model. This method not only reduces the computational amount because the number of regions is much less than the number of pixels, but also extracts more information features from the regions, such as color histograms, so as to better capture the salient objects in the scene. In addition, by using complementary prior knowledge, such as multiple cues like color, texture, and shape, this method shows stronger robustness and higher accuracy under complex lighting conditions. Compared with traditional patch-based methods, the region-based method can more effectively maintain the boundaries of salient objects, avoid misjudgment of high-contrast edges, and thus provide more reliable results in various application scenarios.

[0060] The saliency-based background light removal method first preprocesses the input image, including noise removal and superpixel segmentation, to generate superpixel regions with similar color and texture characteristics. Then, color histograms, texture features, and shape features are extracted from each superpixel region, and the saliency value of each region is calculated by combining global and local contrasts. Through multi-cue fusion, a preliminary saliency map is generated and smoothed to remove noise and discontinuous regions. Then, a threshold segmentation method (such as the Otsu method) is used to divide the saliency map into foreground and background, and the influence of background light is removed by adjusting the brightness and contrast of the background region. Finally, through edge detection and post-processing steps, such as morphological operations, the boundaries of the salient objects are further optimized to generate the final saliency map, ensuring that the boundaries of the salient objects are clear and the background light has been effectively removed. This series of steps ensures the efficiency and accuracy of the method under complex lighting conditions.

[0061] Further, generating the region of interest includes: weighting the waves in the grayscale image according to the saliency map to obtain the saliency-weighted wave coordinate features;

[0062] Setting potential feature screening conditions:

[0063]

[0064] where Th hl is the maximum length of the ship, D l is the lateral distance, i.e., the length, between two waves on the left and right of the same ship, D h is the longitudinal distance, i.e., the spacing, between two waves, L1 and L2 are the coordinates of the left wave, and L'1 and L'2 are the coordinates of the right wave;

[0065] Based on the potential feature screening conditions, screening potential wave pairs using the lateral distance and horizontal position characteristics of the waves, and generating the region of interest through the wave coordinate features and the wave pairs.

[0066] Specifically:

[0067] The template-based method first preprocesses the captured image to reduce noise and enhance wave features. Then, it weights the waves in the image using the saliency map to generate saliency-weighted wave coordinate features. Through these features, a region of interest (ROI) is generated, and the lateral distance and horizontal position characteristics of the waves are used to screen potential wave pairs. Specifically, the lateral distance D l between two waves on the same ship must be less than the maximum value Th l of the ship, and these two waves are basically at the same horizontal position. That is:

[0068] Although this method may result in some false detections, it can effectively avoid missed detections and provide sufficient candidate regions for the subsequent verification stage. In the verification stage, other features (such as the shape, texture, and motion characteristics of the waves) are further used to finely screen the candidate regions, and finally the accurate wave positions are determined. This series of steps ensures the robustness and accuracy of the method in a complex water surface environment.

[0069] Further, generating the superpixel map includes: obtaining the pixel points of the region of interest, setting several superpixels in the region of interest using the pixel points, and distributing multiple initial clustering centers; calculating the spectral distance between the initial clustering centers and the pixels in the region, normalizing the spectral distance, assigning pixels to each superpixel, and determining the maximum spectral distance from all the pixels assigned to the superpixel; using the maximum spectral distance to update the pixel assignment of the superpixel, and after the assignment is completed, merging the isolated pixels into the adjacent superpixels to generate the superpixel map.

[0070] Furthermore, distributing multiple initial clustering centers includes: setting an initial center target neighborhood, and finding the pixel with the minimum gradient within the target neighborhood of each initial center as the initial clustering center.

[0071] Specifically:

[0072] The S-HOG (Superpixel-based Histogram of Oriented Gradients) method significantly improves the spatial consistency, self-adaptability, discriminative ability, and robustness of features by integrating a region-based superpixel method into the HOG features. Specifically, this method uses the Simple Linear Iterative Clustering Zero (SLIC0) superpixel algorithm to weight each pixel in the generation of HOG features, without inputting a compactness factor, thus automatically calculating superpixels. This method aims to improve spatial consistency, handle changes in the ship's direction, provide adaptive feature extraction, reduce noise, enhance discriminative ability, utilize context information, and create a more comprehensive and robust feature representation for accurate feature recognition under diverse and challenging conditions.

[0073] The S-HOG method processes the image using the SLIC0 algorithm and then connects the Histogram of Oriented Gradients (HOG) method for feature extraction. Finally, the two are fused to obtain a superpixel-level feature representation. Different from the traditional method of directly applying HOG to the entire image, the present invention extracts HOG features at the superpixel level generated by SLIC0, which means that each superpixel has an independent HOG feature descriptor, enabling better capture of local structural information while reducing the computational amount.

[0074] Feature fusion: Combining the superpixel boundary information generated by SLIC0 with the HOG features to form a more abundant feature representation. Adding statistical information such as the color mean and variance of the superpixels to the HOG features, or using the shape and size of the superpixels to enhance the discriminability of the features.

[0075] In the process of implementing the S-HOG method, first, the SLIC0 algorithm is used to perform superpixel segmentation on the input image. This process adaptively adjusts the content of the superpixels, optimizes the performance of texture and non-texture regions, while maintaining low computational and memory consumption. Subsequently, size normalization is performed on the images of these two channels respectively, unified to the standard size of 32×32 pixels, and the size of each HOG cell is set to 8×8 pixels, and each block contains 4 such cells:

[0076] SLIC0 superpixel processing process:

[0077] 1. Convert the RGB image color space to HSV;

[0078] 2. Calculate the initial grid spacing. Assume the image has N pixel points and the expected number of superpixels to be generated is K. The expected side length of each superpixel is

[0079] 3. Initialize the cluster centers:

[0080] Distribute K initial cluster centers evenly in the image. To prevent these centers from falling on the image edges or noise points, SLIC0 will find the pixel with the minimum gradient within a 3x3 neighborhood around each initial center as the final cluster center.

[0081] Calculate the gradient of the image using the Sobel operator or other edge detection methods to help select the most suitable initial cluster centers. The gradient calculation only needs to consider the information of the L channel, that is, the brightness information of the image, to simplify the calculation.

[0082] 4. Distance calculation:

[0083] Each pixel is represented as a five-dimensional vector [l, a, b, x, y], where l, a, b are the components in the HSV color space, and x, y are the position coordinates of the pixel. For the distance between a given cluster center and a pixel, the following formula is used for calculation:

[0084]

[0085] where D s is the comprehensive distance between the cluster center and the pixel, which includes a weighted combination of the color distance and the spatial distance. d lab is the color distance between the pixel and the cluster center in the CIELAB color space. d xy is the spatial distance between the pixel and the cluster center on the two-dimensional image plane. m is the balance factor between the color distance and the spatial distance, and S is the scale parameter of the grid size of the area where the cluster center is located or the size of the segmentation area.

[0086]

[0087] where, (y p -y c ) are the image coordinates of the pixel, and (x p -x c ) are the image coordinates of the cluster center.

[0088] Different from the standard SLIC, SLIC0 introduces the normalization of spectral distance. After each iteration, the maximum spectral distance is determined from all the pixels assigned to that superpixel. In the next iteration, the pixel assignment of the superpixel is updated, and when evaluating the potential assignment of a pixel to a superpixel, the spectral distance from the pixel to the superpixel is divided by the maximum spectral distance of the current superpixel. This helps to more fairly compare the spectral differences between different superpixels.

[0089] 5. Handling isolated pixels. After the iteration ends, there may be some isolated pixels that do not belong to any large superpixel segments. To ensure the connectivity of each superpixel, SLIC0 processes these isolated pixels and merges them into the adjacent largest superpixel. Specifically, the algorithm scans the image from top to bottom and from left to right, tries to search for unlabeled pixels, and adds the pixels in the four-neighborhood that belong to the same old superpixel to the search queue to form a new connected superpixel.

[0090] 6. Generating a superpixel map. The output of SLIC0 is an image segmentation result containing multiple compact and regular superpixels. The color and texture within each superpixel are relatively consistent, and the boundaries fit as closely as possible to the object contours in the original image.

[0091] A 324-dimensional HOG feature vector is constructed by evenly dividing the gradient direction of pixels into 9 histogram channels. Finally, the 324-dimensional HOG feature vectors generated by the grayscale image and the red image are merged to form a 648-dimensional composite feature vector. To further improve the classification performance of the model, a support vector machine (SVM) based on the radial basis function (RBF) kernel is used to learn and construct an optimal decision surface, thereby achieving effective verification and classification of ripples.

[0092] This superpixel segmentation aims to further simplify and optimize the superpixel segmentation process, generate high-quality segmented pixels, and introduces the SLIC0 algorithm, which has the advantages of automatic parameter adjustment, emphasizing the generation of high-quality superpixels, higher computational efficiency, and better robustness.

[0093] Furthermore, obtaining symmetric gradient features includes: calculating the gradient direction distribution within each superpixel region, dividing the gradient features into horizontal and vertical direction histograms, and extracting the symmetric gradient features of the horizontal and vertical direction histograms.

[0094] Furthermore, panorama stitching of the classification results includes: constructing a hyperplane using a radial function, inputting the fused features as input vectors into a support vector machine for classification according to the hyperplane, and obtaining the classification results;

[0095] The formula for the radial function is: K(x,x')=exp(-γ||x-x'|| 2 );

[0096] where x and x' are the feature vectors of the input samples, and ||x - x'|| 2 is the Euclidean distance between the two samples, and γ is the hyperparameter of the kernel function that controls the width of the radial basis function;

[0097] Stitch the panoramic images of the classification results, and during the stitching process of the panoramic images, perform color fusion on the overlapping areas of the images.

[0098] Specifically:

[0099] Connect the average feature vector of the superpixel region with the HOG feature vector to generate a composite feature vector F S-HOG = [F S , F HOG , where F S is the superpixel feature vector and F HOG is the Hog feature vector.

[0100] V-HOG feature fusion:

[0101] Calculate the gradient direction distribution inside each superpixel region, and divide the gradient features into horizontal and vertical direction histograms. Extract the symmetric gradient feature F V-HOG .

[0102] Further fuse F S-HOG with F V-HOG : F final = [F S-HOG , F V-HOG , and reduce the dimension through the t-SNE dimensionality reduction technique.

[0103] SVM classification:

[0104] It is necessary to pre-train an SVM classifier model first, use the radial function (RBF) to construct a hyperplane, take the composite feature F S-HOG corresponding to each superpixel region as the input vector, and the class label feature points and non-feature points as the output. The specific form of the radial function is as follows: K(x, x') = exp(-γ||x - x'|| 2 );

[0105] where x and x' are the feature vectors of the input samples, and ||x - x'|| 2 is the Euclidean distance between the two samples, and γ is the hyperparameter of the kernel function that controls the width of the radial basis function. Usually γ > 0, and the larger its value, the narrower the radial function and the more sensitive it is to local data.

[0106] There may be a two-degree difference between the images to be spliced, resulting in a two-degree discontinuous phenomenon at the splicing position in the image, which will affect the visual effect. Usually, color fusion of the overlapping area of the image is used to solve this problem. The present invention uses a fade-in and fade-out method of linear interpolation to fuse the spliced images:

[0107] Let the overlapping interval of images I1 and I2 on the x-axis be [x min , x max . The present invention first matches the feature points of the classification results and uses the formula:

[0108]

[0109] where is the overlapping area, W is the window function, (x i , y j ) is the pixel coordinate, I x , I y are the horizontal gradient and the vertical gradient respectively.

[0110] Let λ1 and λ2 be the two eigenvalues of the matrix. The judgment basis for interest point detection is: where This criterion has a strong response to interest points. Generally, k = 0.04 - 0.06 is taken.

[0111] Calculate the color value I0 of the overlapping area:

[0112] I0(x, y) = βI1(x, y) + (1 - β)I2(x, y);

[0113] where β is the gradient factor, and its value β = (x min - x)(x max - x), is the trace of the matrix is the determinant of the matrix, x min , x max are the starting position and the ending position of the area on the x-axis respectively.

[0114] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for achieving panoramic viewing of an unmanned vessel, characterized in that: include: Acquire a panoramic stitched image of the unmanned vessel, convert the panoramic stitched image from an RGB color space to an HSV color space to obtain a grayscale image, preprocess the grayscale image to obtain a superpixel region, and generate a saliency map based on the superpixel region; Weighting the waves in the grayscale image according to the saliency map to generate a region of interest, performing SLIC0 superpixel processing on the region of interest to generate a superpixel map; The superpixel image is connected to the HOG feature vector to obtain a composite feature vector, the gradient direction distribution inside each superpixel area is calculated, the symmetric gradient feature is obtained, the composite feature vector is fused with the symmetric gradient feature, the fused feature is input into the support vector machine for classification, and the classification results are stitched into a panoramic image.

2. The method for achieving panoramic viewing of an unmanned vessel according to claim 1, characterized in that: Generating the saliency map includes: Denoising and superpixel segmenting are performed on the grayscale image to obtain the superpixel region, extracting feature information of the superpixel region, and performing multi-cue fusion based on the saliency value of each superpixel region calculated by combining global and local contrast to generate a preliminary saliency map; the feature information includes: color histogram, texture features, and shape features; The preliminary saliency map is smoothed, the smoothed preliminary saliency map is threshold segmented, and the background light is removed by adjusting the brightness and contrast of the background area to further optimize the boundaries of the salient objects to generate the saliency map.

3. The method for achieving panoramic viewing of an unmanned vessel according to claim 1, characterized in that: Generating the region of interest includes: weighting the waves in the grayscale image according to the saliency map to obtain saliency-weighted wave coordinate features; Set up potential feature screening conditions: Among them, Th hl is the maximum length of the ship, D l is the horizontal distance between the left and right waves on the same ship, i.e. the length, D h is the longitudinal distance between the two waves, i.e. the spacing, L1 and L2 are the coordinates of the left wave, L'1 and L'2 are the coordinates of the right wave; Based on the potential feature screening conditions, potential wave pairs are screened using the lateral distance and horizontal position characteristics of the waves, and the region of interest is generated through the wave coordinate features and the wave pairs.

4. The method for achieving panoramic viewing of an unmanned vessel according to claim 1, characterized in that: Generating the superpixel map includes: Obtaining pixel points of the region of interest, using the pixel points to set a plurality of superpixels in the region of interest, and distributing a plurality of initial cluster centers; Calculating spectral distances between the initial cluster center and pixels in the region, normalizing the spectral distances, assigning pixels to each superpixel, and determining a maximum spectral distance from all pixels assigned to the superpixel; The maximum spectral distance is used to update the pixel allocation of the superpixel. After the allocation is completed, the isolated pixels are merged into adjacent superpixels to generate a superpixel map.

5. The method for achieving panoramic viewing of an unmanned vessel according to claim 4, characterized in that: Distributing the multiple initial cluster centers includes: Set the initial center target neighborhood and find the pixel with the smallest gradient in the target neighborhood of each initial center as the initial cluster center.

6. The method for achieving panoramic viewing of an unmanned vessel according to claim 4, characterized in that: The formula for calculating the spectral distance between the initial cluster center and the pixels in the region is: Among them, D s is the comprehensive distance between the cluster center and the pixel, which includes a weighted combination of color distance and spatial distance, d lab is the color distance between the pixel and the cluster center in the CIELAB color space, d xy is the spatial distance between the pixel and the cluster center on the two-dimensional image plane, m is the balance factor between color distance and spatial distance, and S is the scale parameter of the grid size or segmentation area size of the area where the cluster center is located.

7. The method for achieving panoramic viewing of an unmanned vessel according to claim 1, characterized in that: Acquiring the symmetric gradient feature includes: The gradient direction distribution inside each superpixel region is calculated, the gradient features are divided into horizontal and vertical direction histograms, and the symmetric gradient features of the horizontal and vertical direction histograms are extracted.

8. The method for achieving panoramic viewing of an unmanned vessel according to claim 1, characterized in that: Panoramic image stitching of classification results includes: constructing a hyperplane using a radial function, and inputting the fused features as input vectors into the support vector machine for classification according to the hyperplane to obtain the classification result; The formula of the radial function is: K(x,x')=exp(-γ||x-x'|| 2 ); Where x and x' are the feature vectors of the input samples, ||x-x'|| 2 is the Euclidean distance between two samples, γ is the hyperparameter of the kernel function, which controls the width of the radial basis function; The classification results are stitched into a panorama, and during the panorama stitching process, the overlapping areas of the images are color fused.

9. The method for achieving panoramic viewing of an unmanned vessel according to claim 8, characterized in that: The formula for obtaining the overlapping area is: in, is the overlapping area, W is the window function, (x i ,y j ) is the pixel coordinate, I x , I y They are horizontal gradient and vertical gradient respectively.

Images