A method for detecting and tracking submarine cables based on semantic segmentation of sonar images

Abstract

A method of seabed cable detection and tracking based on sonar image semantic segmentation is proposed. A method of adding shadow information and distance information of sonar image to the sonar image semantic segmentation dataset is designed. The storage and use method of shadow information and distance information in the dataset are demonstrated. A semantic segmentation model ARRDs-Unet is designed, which fuses the shadow information and distance information of sonar image. The model enhances the feature extraction ability by using recursive residual convolution and visual attention mechanism. A loss function is designed to fuse the distance information of sonar image. The training weight of sonar image under different distances is set to realize the accurate detection of seabed cable. A six-degree-of-freedom AUV model controlled by PID is used. Specifically, the coordinate screening in the x-axis direction is used to reduce the computational complexity of the system, and the Theil-Sen estimator is used to realize the fitting of the seabed cable trajectory. According to the fitting result of the trajectory, a tracking navigation method is designed to realize the stable tracking of the AUV to the seabed cable.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marine information processing, and particularly relates to a seabed cable detection and tracking method based on sonar image semantic segmentation. BACKGROUND

[0002] The seabed cable is wrapped by insulating material, is related to the national communication, energy network security, and is an important part of national security; although the seabed cable is designed considering the harsh underwater conditions such as large seabed pressure and seawater corrosion, due to the situations such as ship anchor pulling, natural disasters and marine biological activities, the seabed cable failure often occurs, and the communication interruption or power interruption caused by the failure will bring a major impact on the economy and safety of the related area.

[0003] Therefore, the seabed cable must be regularly maintained and repaired to reduce the risk of serious damage, since the workload of the seabed cable repair is large, and the depth of part of the seabed cable reaches 2700 meters, the diver cannot bear the high water pressure and extremely low temperature of the deep sea, so the diver cannot be arranged to visually inspect the seabed cable. The sonar is an important tool for a series of underwater activities such as marine resource exploration, offshore defense and salvage, and the autonomous underwater vehicle (AUV) carrying a side scan sonar system can provide high-resolution, efficient and reliable seabed cable position and state information in the deep sea environment, which meets the requirements of the seabed cable maintenance. In the process of tracking and maintaining the seabed cable by the AUV, the AUV can use sensor information to autonomously select a running route, but the seabed environment is complex, and the sensor is difficult to directly obtain the position and trajectory of the seabed cable, therefore, the sonar system on the AUV can be used to help the AUV complete the positioning and tracking of the seabed cable, which requires the system to accurately detect the seabed cable from the sonar image, and to judge the position of the seabed cable according to the detection result, and then to provide a guide for the tracking movement of the AUV.

[0004] Sonar images suffer from problems such as low resolution, little feature information, difficulty in data acquisition, and small data volume. Traditional semantic segmentation models are not effective when applied to sonar images, and the accuracy of semantic segmentation is relatively limited. In addition, traditional semantic segmentation models do not fully utilize the information in sonar images based on their characteristics. For example, in optical images, factors such as the position and intensity of the light source need to be considered. In an optical image, the distance from each object to the imaging system often varies greatly, resulting in large differences in the scale of the same object in the image under different shooting conditions. Sonar images have the characteristic of relatively fixed distances from various target objects to the imaging system. For sonar images generated by the same sonar scanning mission, the scales of objects of the same type in the image are relatively close. Furthermore, due to the unique acoustic propagation characteristics of sonar imaging, sonar images contain richer shadow information than optical images. Object shadows in sonar images have been shown to effectively reflect object outlines and provide feature information. This is particularly true for submarine cables, where the complex seabed environment presents numerous objects that interfere with cable detection. This shadow information is a valuable supplement for distinguishing these objects. Furthermore, unlike objects with complex contours such as aircraft wreckage and shipwrecks, whose shadow features vary significantly at different imaging angles, this can actually hinder model learning. Submarine cable shadows, on the other hand, have relatively consistent structures, and different imaging angles do not result in varying shadow features. Therefore, this shadow information offers greater and more stable value. Traditional semantic segmentation models are primarily designed for optical images and lack the ability to leverage the distance and shadow information in sonar images, resulting in a lack of specificity. Based on the above analysis, submarine cable detection and tracking in sonar images is of great practical significance and a practical need in engineering applications.

[0005] A sonar image detection algorithm was designed using shadow information in the document Reed S, Petillot Y, Bell J. An automatic approach to the detection and extraction of mine features in sidescan sonar [J]. IEEE journal of oceanic engineering, 2003, 28(1): 90-105. It is shown that shadow information can improve the detection accuracy of objects in sonar images; In order to solve the problem of unbalanced number of objects of each category in the semantic segmentation of sonar images, the weighted loss function method is used in the documents Wang H, Gao N, Xiao Y, et al. Image feature extraction based on improved FCN for UUV side-scan sonar [J]. Marine Geophysical Research, 2020, 41: 1-17. and Wu M, Wang Q, Rigall E, et al. ECNet: Efficient convolutional networks for side scan sonar image segmentation [J]. Sensors, 2019, 19(9): 2009; In the document Zhao D, Ge W, Chen P, et al. Feature Pyramid U-Net with Attention for Semantic Segmentation of Forward-Looking Sonar Images [J]. Sensors, 2022, 22(21): 8468. Based on the U-Net model, residual structure, feature pyramid module and visual attention mechanism are added to the network, which shows good performance in the semantic segmentation of few sample classes and can be applied to sonar images with different noise levels; SUMMARY

[0006] To solve the above technical problems, the present application provides a seabed cable detection and tracking method based on sonar image semantic segmentation, which effectively utilizes the shadow information and distance information of the sonar image, realizes accurate detection of the seabed cable in the sonar image, calculates the navigation information according to the detection result, and uses the AUV to track the seabed cable.

[0007] To achieve the above purpose, the technical scheme adopted by the present application is:

[0008] A submarine cable detection and tracking method based on sonar image semantic segmentation specifically includes the following steps:

[0009] Step 1. Storage and use of shadow information and distance information in sonar images:

[0010] When creating a sonar image semantic segmentation dataset, the shadow and distance information of the sonar image are stored as part of the dataset. These two pieces of information are then used during the training and operation of the semantic segmentation model to improve the accuracy of semantic segmentation.

[0011] Step 2. Submarine cable detection;

[0012] Based on the sonar image semantic segmentation dataset containing shadow information and distance information obtained in step 1, a submarine cable semantic segmentation model is designed to achieve accurate detection of submarine cables;

[0013] Step 3. Submarine cable tracking:

[0014] The navigation information is calculated using the results of submarine cable detection in step 2, and a PID-controlled six-degree-of-freedom AUV model is used to stably track the submarine cable.

[0015] As a further improvement of the present invention, the step 1 is specifically as follows:

[0016] When annotating the sonar image semantic segmentation dataset, the shadow part of each object in the sonar image is annotated as part of the object, and the highlight area and shadow area of ​​the object are considered as a whole for training and semantic segmentation. As shown in formula (1), the imaging distance of each sonar image in the dataset is normalized:

[0017]

[0018] In formula (1), D i is the distance of the i-th image, d min is the minimum distance among all images, d max is the maximum distance in all images. After distance normalization, the normalized distance of the sonar image with a short imaging distance is close to 1, and the normalized distance of the sonar image with a long imaging distance is close to 0. The normalized distance information corresponding to each sonar image is stored in a CSV file. The file contains a header line. After the header line, the name of the sonar image and its corresponding normalized distance are separated by commas. The normalized distance information is saved in a CSV file with the same name as the corresponding sonar image. When training the semantic segmentation model, a cross entropy loss function DAWCE that integrates distance information is designed. Assume that the four categories in the semantic segmentation dataset are C1, C2, C3 and C4, where C1 represents the submarine cable category and will be given a higher weight. The function calculation method is shown in formula (2);

[0019]

[0020] In formula (2), Loss is the DAWCE loss value, N is the number of pixels in the image, and y ic is the true label of pixel i and category c, which is 1 if the pixel belongs to category c and 0 otherwise. ic is the predicted probability of pixel i and category c, ω c is the weight of category c, ω1>ω2, ω3, ω4, D i is the normalized distance information of pixel i, and p is the distance weighting coefficient, which controls the influence of distance information on the DAWCE loss function. For each pixel i, the cross entropy loss of all categories is calculated, the loss of each category is multiplied by the corresponding weight, and then the weighted loss of each pixel is multiplied by the normalized distance information. Finally, the losses of all pixels and categories are added together to obtain the calculation result of the DAWCE loss function.

[0021] As a further improvement of the present invention, in step 2, a semantic segmentation model ARRDs-Unet is designed for the data set obtained in step 1, which integrates shadow information and distance information in the sonar image to achieve accurate detection of submarine cables through semantic segmentation;

[0022] The network structure of the ARRDs-Unet model is divided into two parts: encoder and decoder;

[0023] The encoder consists of one input layer and four downsampling layers. Each layer of the encoder uses a recursive residual convolution unit connected to a ReLU activation function for feature extraction. The combination of recursive convolution and residual structure improves the network's ability to capture image spatial and semantic information. At this time, the size of the feature map remains unchanged, and the number of channels is doubled. The feature map is input into the SE module. The global average pooling layer and two fully connected layers in the SE module generate adaptive selection of feature map channel weights, improving the network's ability to focus on the most relevant features at different levels of abstraction. The feature map output by the SE module is used as the output of each layer and input into the next layer through 2×2 maximum pooling. The number of channels of the feature map remains unchanged, and the width and height are reduced to half of the original.

[0024] In the decoder part, the input is the feature map of the encoder's last feature extraction. After a 3×3 upsampling operation, the width and height of the feature map are doubled, and the number of channels is reduced to half of the original; the upsampled feature map is spliced ​​with the feature map of the corresponding level of the encoder to obtain a new feature map, thereby realizing the fusion of feature information at different levels. At this time, the size of the new feature map remains unchanged and the number of channels is doubled; the spliced ​​new feature map is input into the recursive residual convolution unit connected to the ReLU activation function for feature extraction. The size of the feature map remains unchanged and the number of channels is reduced to half of the original; the last part of the decoder is a 1×1 convolution with a ReLU activation function, which outputs the semantic segmentation result of the network, which is an output image with the same size as the input image, and each pixel represents the probability of the category to which it belongs.

[0025] As a further improvement of the present invention, in step 3, based on the result of submarine cable semantic segmentation in step 2, a coordinate system is established to correspond the position in the sonar image to the actual position on the seabed, and the computational complexity of the system is reduced by a single coordinate screening in the x-axis direction:

[0026] In the result map generated by semantic segmentation, the pixel value of the submarine cable area remains unchanged at (128, 0, 0), and the pixel values ​​of the other types of object areas are set to (0, 0, 0). After the pixel values ​​are adjusted, the target point in the semantic segmentation result map is read. The coordinate system is established with the lower left corner of the image as the origin, the horizontal axis as the x-axis, and the vertical axis as the y-axis. i ,y i ) to represent the position of each pixel point; for the pixel point with the pixel value (128, 0, 0), extract all coordinate points with the same y value, calculate the average x value of these coordinate points and round it to obtain the unique average coordinate corresponding to each y value, save the processed coordinates to the array points and input it into the Theil-Sen estimator, use the Theil-Sen estimator to fit the submarine cable trajectory, and the fitted straight line expression is y = mx + b. Since the angle α between the submarine cable and the y-axis in the sonar image and the actual angle β between the AUV heading and the submarine cable direction are approximately equal, the heading error angle β′ that the AUV should adjust is calculated by calculating the angle α′ between the submarine cable fitting trajectory line in the sonar image and the y-axis. This angle β′ is input into the AUV control system to enable the AUV to track the submarine cable in the horizontal direction; then read the sonar image distance information stored in step 1, perform an inverse normalization operation to obtain the actual distance, calculate the vertical depth error based on the preset submarine cable tracking vertical distance, and input the depth error Δh into the AUV control system to enable the AUV to track the submarine cable in the vertical direction.

[0027] This application has the following benefits:

[0028] The present invention rationally adds the shadow information and distance information of sonar images to the sonar image semantic segmentation dataset, and integrates these two types of information into the semantic segmentation of sonar images, effectively improving the accuracy of sonar image semantic segmentation; based on the results of semantic segmentation, a trajectory fitting method is designed and navigation information is calculated, realizing the stable tracking of submarine cables by AUV. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a workflow diagram of the present invention;

[0030] Figure 2 This is the network structure diagram of the ARRDs-Unet model;

[0031] Figure 3 This is the ARRDs-Unet model training flow chart;

[0032] Figure 4 It is the overall structural diagram of the submarine cable tracking method;

[0033] Figure 5 This is a schematic diagram of the sonar image-seabed coordinate transformation;

[0034] Figure 6 This is the result of submarine cable semantic segmentation;

[0035] Figure 7 This is the fitting effect diagram of the submarine cable trajectory;

[0036] Figure 8 This is the result of submarine cable tracking. DETAILED DESCRIPTION

[0037] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0038] Figure 1 This is a flow chart of a specific embodiment of the present invention, which is as follows;

[0039] Storage and use of shadow information and distance information in sonar images:

[0040] When annotating the sonar image semantic segmentation dataset, the shadows of each object in the sonar image are labeled as part of the object. The highlight and shadow areas of the object are treated as a whole for training and semantic segmentation, thus making use of the shadow information in the sonar image. This is because the shadows of objects in sonar images have been proven to effectively reflect the object's outline and provide feature information. At the same time, submarine cables are different from objects with complex outlines such as aircraft wreckage and shipwrecks. The shadow characteristics of these complex objects vary greatly at different imaging angles, which may actually cause difficulties for model learning. The shadow structure of submarine cables is relatively consistent, and different imaging angles do not result in submarine cable shadow characteristics with different shapes. Therefore, the shadow information of submarine cables has higher and more stable utilization value.

[0041] When processing the distance information of sonar images, as shown in formula (1), the imaging distance of each sonar image in the dataset is first normalized:

[0042]

[0043] In formula (1), D i is the distance of the i-th image, d min is the minimum distance among all images, d max is the maximum distance in all images. After distance normalization, the normalized distance of sonar images with close imaging distances is close to 1, and the normalized distance of sonar images with long imaging distances is close to 0. The normalized distance information corresponding to each sonar image is stored in a CSV file. The file contains a header row with column names "image_id" and "normalized_distance". The header row is followed by the name of the sonar image and its corresponding normalized distance, separated by a comma. The normalized distance information is saved in a CSV file with the same name as the corresponding sonar image. When training the semantic segmentation model, a cross-entropy loss function DAWCE (Distance-Aware Weighted Cross-Entropy) is designed to integrate distance information. Assume that the four categories in the semantic segmentation dataset are C1, C2, C3, and C4, where C1 represents the submarine cable category and will be given a higher weight. The function calculation method is shown in Equation (2).

[0044]

[0045] In formula (2), Loss is the DAWCE loss value, N is the number of pixels in the image, and y ic is the true label of pixel i and category c (1 if the pixel belongs to category c, 0 otherwise), p ic is the predicted probability of pixel i and category c, ω cis the weight of category c (ω1>ω2,ω3,ω4), D i is the normalized distance information of pixel i, and p is the distance weighting coefficient, which controls the influence of distance information on the DAWCE loss function. For each pixel i, the cross entropy loss of all categories is calculated, the loss of each category is multiplied by the corresponding weight, and the weighted loss of each pixel is multiplied by the normalized distance information. Finally, the loss of all pixels and categories is added together to obtain the calculation result of the DAWCE loss function.

[0046] Submarine cable detection:

[0047] Based on a dataset containing sonar image shadow information and distance information, a semantic segmentation model ARRDs-Unet is designed. The shadow information and distance information in the sonar image are integrated to achieve accurate detection of submarine cables through semantic segmentation.

[0048] like Figure 2 As shown in Figure 1, the network structure of the ARRDs-Unet model can be roughly divided into two parts: the encoder and the decoder. The encoder consists of one input layer and four downsampling layers. Each layer of the encoder uses a recursive residual convolutional unit (RRCU) connected to a ReLU activation function for feature extraction. The combination of recursive convolution and residual structure improves the network's ability to capture spatial and semantic information in images. Each RRCU consists of two recursive convolutional layers and a residual structure. The pseudocode for the recursive convolutional layer algorithm is shown in Table 1.

[0049] Table 1 Recursive convolutional layer

[0050]

[0051] At this time, the size of the feature map remains unchanged and the number of channels becomes twice as large as before; the feature map is input into the SE module (Squeeze-and-Excitation Module), and the global average pooling layer and two fully connected layers in the SE module generate adaptive selection of the feature map channel weights, thereby improving the network's ability to focus on the most relevant features at different levels of abstraction. The SE module is as follows: the feature map output by the SE module is used as the output of each layer and input into the next layer structure through 2×2 maximum pooling. The number of channels of the feature map remains unchanged, and the width and height become half of the original.

[0052] In the decoder part, the input is the feature map of the encoder's last feature extraction. After a 3×3 upsampling operation, the width and height of the feature map are doubled, and the number of channels is reduced to half of the original; the upsampled feature map is spliced ​​with the feature map of the corresponding level of the encoder to obtain a new feature map, thereby realizing the fusion of feature information at different levels. At this time, the size of the new feature map remains unchanged and the number of channels is doubled; the spliced ​​new feature map is input into the recursive residual convolution unit connected to the ReLU activation function for feature extraction. The size of the feature map remains unchanged and the number of channels is reduced to half of the original; the last part of the decoder is a 1×1 convolution with a ReLU activation function, which outputs the semantic segmentation result of the network, which is an output image with the same size as the input image, and each pixel represents the probability of the category to which it belongs.

[0053] The training process of the ARRDs-Unet model is as follows Figure 3 As shown in the figure, at the beginning of training, a sonar image (512×512×3) with shadow information annotated is input into the semantic segmentation network of the model for forward propagation. After the network performs semantic segmentation, the output is obtained. The loss value is calculated using a weighted cross-entropy loss function that incorporates the distance information of the sonar image. The loss value is then back-propagated to the model. The model will update the learnable parameters in the network using the ADAM optimizer, searching for the optimal parameters with the minimum loss value over multiple iterations.

[0054] Submarine cable tracking:

[0055] According to the results of submarine cable semantic segmentation in step 2, a method for AUV tracking submarine cables is designed. The overall structure of the method is as follows: Figure 4 As shown in Figure 1, navigation information is calculated from sonar images, thereby enabling the AUV to stably track the submarine cable.

[0056] When extracting submarine cable location information, such as Figure 5 The coordinate system is established as shown in the figure to correspond the position in the sonar image to the real position in the seabed, and the computational complexity of the system is reduced by filtering the coordinates in the x-axis direction once: in the result image generated by semantic segmentation, the pixel value of the submarine cable area remains unchanged at (128,0,0), and the pixel values ​​of the other types of object areas are set to (0,0,0). After the pixel value is adjusted, the target point in the semantic segmentation result image is read, and a coordinate system is established with the lower left corner of the image as the origin, the horizontal axis as the x-axis, and the vertical axis as the y-axis. The coordinates (x i,yi) to represent the position of each pixel point; for the pixel point with pixel value (128,0,0), extract all coordinate points with the same y value, calculate the average value of the x values ​​of these coordinate points and round them up to get the unique average coordinate corresponding to each y value, save the processed coordinates to the array points and input them into the Theil-Sen estimator, use the Theil-Sen estimator to fit the submarine cable trajectory, the Theil-Sen estimator first calculates the slopes between all possible pairs of data points, and then calculates the median of the overall slope to get the slope m of the trajectory fitting line. The calculation process is shown in formula (3).

[0057]

[0058] In formula (3), (x i ,y i ) and (x j ,y i ) are the coordinates of different data points. To calculate the intercept of the trajectory fitting line, the median can also be used for calculation, as shown in formula (4).

[0059] b=median(y i -mx i ) (4)

[0060] After trajectory fitting, a straight line y = mx + b is obtained. Since the angle α between the submarine cable and the y-axis in the sonar image and the actual angle β between the AUV heading and the submarine cable direction are approximately equal, by calculating the angle α′ between the submarine cable fitting trajectory line in the sonar image and the y-axis, the heading error angle β′ that the AUV should adjust can be approximately obtained. The angle β′ is input into the AUV control system to enable the AUV to track the submarine cable in the horizontal direction. Then, the sonar image distance information stored in step 1 is read and an inverse normalization operation is performed to obtain the true distance. The depth error in the vertical direction is calculated based on the preset submarine cable tracking vertical distance. The depth error Δh is input into the AUV control system to enable the AUV to track the submarine cable in the vertical direction.

[0061] Overall process:

[0062] In the process of preparing the sonar image semantic segmentation dataset, the shadow information and distance information of the sonar image are stored as part of the dataset, and these two types of information are used during the training and operation of the semantic segmentation model to improve the semantic segmentation accuracy. Based on the sonar image semantic segmentation dataset containing shadow information and distance information, the submarine cable semantic segmentation model ARRDs-Unet is designed to achieve accurate detection of submarine cables. According to the submarine cable detection results, a coordinate system is established to extract position information, and a coordinate screening in the x-axis direction is performed. The Theil-Sen estimator is then used for trajectory fitting. Based on the trajectory fitting results, a navigation information calculation method is designed, and a PID-controlled six-degree-of-freedom AUV model is used to stably track the submarine cable.

[0063] Example 1:

[0064] The following will describe in detail the process of submarine cable detection from sonar images to submarine cable tracking according to the present invention with reference to the accompanying drawings and examples.

[0065] After inputting the sonar image into the ARRDs-Unet model, the semantic segmentation results of the submarine cable are obtained, including the original sonar image of the submarine cable and the semantic segmentation results of the ARRDs-Unet model. Figure 6 As shown in the figure, although there are a few cases of incorrect segmentation and missed segmentation, the semantic segmentation task of submarine cables in sonar images is generally completed well, and the accurate detection of submarine cables is achieved. A coordinate system is established for the result image to extract the position information, and a coordinate screening is performed in the x-axis direction. Then, the Theil-Sen estimator is used for trajectory fitting. The fitting effect is shown in the figure. Figure 7 As shown in Figure 2, the fitted trajectory line can well reflect the actual trajectory of the submarine cable and its position in the sonar image; finally, the navigation information is calculated based on the trajectory fitting result, and the PID-controlled six-degree-of-freedom AUV model is used to stably track the submarine cable. The tracking effect is shown in Figure 2. Figure 8 As shown, stable tracking of submarine cables is achieved.

[0066] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A submarine cable detection and tracking method based on sonar image semantic segmentation, characterized in that: The specific steps include: Step 1. Storage and use of shadow and distance information in sonar images: When creating a sonar image semantic segmentation dataset, the shadow and distance information of the sonar image are stored as part of the dataset. These two pieces of information are then used during the training and operation of the semantic segmentation model to improve the accuracy of semantic segmentation. The step 1 is specifically as follows: When annotating the sonar image semantic segmentation dataset, the shadow part of each object in the sonar image is annotated as part of the object, and the highlight area and shadow area of ​​the object are considered as a whole for training and semantic segmentation. As shown in formula (1), the imaging distance of each sonar image in the dataset is normalized: ； In formula (1), is the distance of the ith image, is the minimum distance among all images, is the maximum distance in all images. After distance normalization, the normalized distance of the sonar image with a short imaging distance is close to 1, and the normalized distance of the sonar image with a long imaging distance is close to 0. The normalized distance information corresponding to each sonar image is stored in a CSV file. The file contains a header line. After the header line, the name of the sonar image and its corresponding normalized distance are separated by commas. The normalized distance information is saved in a CSV file with the same name as the corresponding sonar image. When training the semantic segmentation model, a cross entropy loss function DAWCE that integrates distance information is designed. Assume that the four categories in the semantic segmentation dataset are C1, C2, C3 and C4, where C1 represents the submarine cable category and will be given a higher weight. The function calculation method is shown in formula (2); ； In formula (2), Loss is the DAWCE loss value, N is the number of pixels in the image, is the true label of pixel i and category c, which is 1 if the pixel belongs to category c and 0 otherwise. is the predicted probability of pixel i and category c, is the weight of category c, > , , , is the normalized distance information of pixel i, p is the distance weighting coefficient, which controls the influence of distance information on the DAWCE loss function. For each pixel i, the cross entropy loss of all categories is calculated, the loss of each category is multiplied by the corresponding weight, and then the weighted loss of each pixel is multiplied by the normalized distance information. Finally, the loss of all pixels and categories is added together to obtain the calculation result of the DAWCE loss function. Step 2. Submarine cable detection; Based on the sonar image semantic segmentation dataset containing shadow information and distance information obtained in step 1, a submarine cable semantic segmentation model is designed to achieve accurate detection of submarine cables; Step 3. Submarine cable tracking: The navigation information is calculated using the results of submarine cable detection in step 2, and a PID-controlled six-degree-of-freedom AUV model is used to stably track the submarine cable.

2. The submarine cable detection and tracking method based on sonar image semantic segmentation according to claim 1, characterized in that: In step 2, a semantic segmentation model ARRDs-Unet is designed for the data set obtained in step 1, which integrates shadow information and distance information in the sonar image to achieve accurate detection of submarine cables through semantic segmentation; The network structure of the ARRDs-Unet model is divided into two parts: encoder and decoder; The encoder consists of one input layer and four downsampling layers. Each layer of the encoder uses a recursive residual convolution unit connected to a ReLU activation function for feature extraction. The combination of recursive convolution and residual structure improves the network's ability to capture image spatial and semantic information. At this time, the size of the feature map remains unchanged, and the number of channels is doubled. The feature map is input into the SE module. The global average pooling layer and two fully connected layers in the SE module generate adaptive selection of feature map channel weights, improving the network's ability to focus on the most relevant features at different levels of abstraction. The feature map output by the SE module is used as the output of each layer and input into the next layer through 2×2 maximum pooling. The number of channels of the feature map remains unchanged, and the width and height are reduced to half of the original. In the decoder part, the input is the feature map of the encoder's last feature extraction. After a 3×3 upsampling operation, the width and height of the feature map are doubled, and the number of channels is reduced to half of the original; the upsampled feature map is spliced ​​with the feature map of the corresponding level of the encoder to obtain a new feature map, thereby realizing the fusion of feature information at different levels. At this time, the size of the new feature map remains unchanged and the number of channels is doubled; the spliced ​​new feature map is input into the recursive residual convolution unit connected to the ReLU activation function for feature extraction. The size of the feature map remains unchanged and the number of channels is reduced to half of the original; the last part of the decoder is a 1×1 convolution with a ReLU activation function, which outputs the semantic segmentation result of the network, which is an output image with the same size as the input image, and each pixel represents the probability of the category to which it belongs.

3. The submarine cable detection and tracking method based on sonar image semantic segmentation according to claim 1, characterized in that: In step 3, based on the results of submarine cable semantic segmentation in step 2, a coordinate system is established to correspond the position in the sonar image to the actual position on the seabed. The computational complexity of the system is reduced by filtering the coordinates in the x-axis direction once: In the result image generated by semantic segmentation, the pixel value of the submarine cable area remains unchanged at (128, 0, 0), and the pixel values ​​of the other types of object areas are set to (0, 0, 0). After the pixel values ​​are adjusted, the target point in the semantic segmentation result image is read. The coordinate system is established with the lower left corner of the image as the origin, the horizontal axis as the x-axis, and the vertical axis as the y-axis. The coordinates (x i ,y i ) to represent the position of each pixel point; for the pixel point with the pixel value (128, 0, 0), extract all coordinate points with the same y value, calculate the average value of the x values ​​of these coordinate points and round them up to get the unique average coordinate corresponding to each y value, save the processed coordinates to the array points and input them into the Theil-Sen estimator, use the Theil-Sen estimator to fit the submarine cable trajectory, and the fitted straight line expression is Since the angle α between the submarine cable and the y-axis in the sonar image is approximately equal to the actual angle β between the AUV heading and the submarine cable direction, the angle between the submarine cable fitting trajectory line and the y-axis in the sonar image is calculated. , get the heading error angle that the AUV should adjust , the angle Input into the AUV control system to enable the AUV to track the submarine cable horizontally; Then read the sonar image distance information stored in step 1, perform an inverse normalization operation to obtain the true distance, calculate the depth error in the vertical direction based on the preset submarine cable tracking vertical distance, and input the depth error Δh into the AUV control system to enable the AUV to track the submarine cable in the vertical direction.

Images