An unmanned underwater vehicle detection method based on a water acoustic transducer and a side-scan sonar

Abstract

The application discloses an unmanned underwater vehicle detection method based on a water acoustic transducer and a side-scan sonar. The water acoustic transducer and the side-scan sonar are carried on an underwater robot. The sound source level of the water acoustic transducer is greater than 140 db, the frequency is 400-1000 Hz, and the frequency is adjustable. The water acoustic transducer can direct the fish school to disperse and will not drive the bionic unmanned underwater vehicle hidden in the fish school. The side-scan sonar can collect sonar images. Whether there is an unknown underwater unmanned vehicle is judged according to the obtained sonar images. The method is characterized by combining the water acoustic transducer and the side-scan sonar to realize detection on the unknown underwater vehicle. The method can select different sound wave frequencies according to different operation environments, is suitable for various ships and underwater robots, and improves the adaptability of the method to different aquatic environments.

Description

TECHNICAL FIELD

[0001] The present application relates to the underwater detection technical field, especially to an unmanned underwater vehicle detection method based on a hydroacoustic transducer and a side-scan sonar. BACKGROUND

[0002] With the development of intelligent technology and the need to protect national security, how to detect underwater robots is increasingly valued by domestic and foreign researchers. The existing underwater detection technology mostly adopts the form of optical acoustic image combined with a machine vision system. This way can only determine static targets more, and has the problem of low success rate in distinguishing between fish-shaped robots and mobile unmanned underwater vehicles. The method of detecting by using a high-definition camera needs to be observed and identified at close range, and the detection range is small, time-consuming and laborious. If it is necessary to monitor whether there is a possible unknown unmanned underwater vehicle in a large range of underwater dynamic environment, a large-scale underwater detection network needs to be laid out, and such a method is difficult to be applied to the water environment in daily life such as inland rivers. SUMMARY

[0003] The unmanned underwater vehicle detection method based on a hydroacoustic transducer and a side-scan sonar proposed by the present application solves the existing problems. By driving fish in a directional manner and cooperating with the underwater image returned by the side-scan sonar, the different reaction states of biological fish groups and underwater vehicles are combined with a machine vision recognition method to improve the success rate of distinguishing fish-shaped robots.

[0004] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0005] An unmanned underwater vehicle detection method based on a hydroacoustic transducer and a side-scan sonar comprises the following steps:

[0006] Step 1: set a selected ultrasonic frequency, the hydroacoustic transducer works, and fish driving is performed by a fish driver;

[0007] Step 2: start the side-scan sonar to obtain the acoustic vision image at the current time;

[0008] Step 3: based on the processing of the acoustic vision system, the original sonar image is used for data augmentation, and the sample set required for training and testing the model is obtained. The fish groups and fish-shaped robots in the sample set are artificially labeled and divided according to the different reactions of the fish groups and unmanned underwater vehicles caused by the fish driving principle of sound waves, and the label image required for testing and training the model is obtained.

[0009] Step 4: processing and analyzing the sonar image, including removing noise points by using median filtering; reducing the data size while retaining the image features by using the Canny edge detection algorithm and extracting features by using the Hough transform; extracting all suspicious moving targets in the sonar image by using the K-Means clustering analysis algorithm, at this time the suspicious targets still include fish schools and unmanned underwater vehicles; then tracking the suspicious targets by using the Kalman filtering algorithm combined with the target detector; judging the tracked objects, as fish schools will move away from the water acoustic transducer emitting ultrasonic waves, extracting suspicious targets; finally, the detection machine carrying a high-definition camera performs target recognition by using the convolutional neural network (CNN), and further judges whether the suspicious target is an unknown unmanned underwater vehicle.

[0010] Preferably, the fish expeller in step 1 is composed of a water transducer, which generates ultrasonic waves at a specific frequency, and uses the characteristics that different fish schools are sensitive to specific ultrasonic frequencies to disperse fish schools.

[0011] Preferably, the acquisition of the acoustic-visual image in step 2 is based on a side-scan sonar device, and the task of the acoustic-visual system is to acquire a two-dimensional acoustic image for post-processing in the third step.

[0012] Preferably, the median filtering is also used in step 3 to remove noise points; specifically, a 3*3 mask is used to avoid the blurring effect caused by filtering, which can protect the relatively sharp edges in the image while processing smooth impulse noise; the step is to move the 3*3 template in the image in turn, read the gray value of the corresponding pixel point and arrange it in ascending order, and assign the median value to the pixel point corresponding to the center of the template;

[0013] The Canny edge detection algorithm is used to reduce the data size while retaining the image features and extract features by using the Hough transform; first, the image is smoothed by using Gaussian filtering to eliminate noise points, then the Sabel operator is used to process the image gradient and direction, the larger the gradient, the more likely it is an edge, then non-maximum suppression processing is performed to determine whether a pixel point is a boundary, and finally a double-threshold detection is used to further determine the boundary, and finally the Hough transform is used to extract straight lines, and the expression of the extracted straight line cluster is:

[0014] r=x0cosθ+y0sinθ.

[0015] Preferably, all suspicious moving targets in the sonar image, including fish schools and unmanned underwater vehicles, are extracted by using the K-Means clustering algorithm in step 4; the step is: data is divided into two groups in advance, then K objects are randomly selected as initial cluster centers, then the distance between each object and each seed cluster center is calculated, and each object is assigned to the cluster center closest to it; the cluster centers and the objects assigned to them represent a cluster; the cluster centers are recalculated according to the existing objects in the cluster after each sample is assigned; this process will be repeated until a certain termination condition is met; the termination condition can be that no or a minimum number of objects are reassigned to different clusters, no or a minimum number of cluster centers change, or the local minimum of the sum of squared errors; in this case, the two groups divided are suspicious moving targets including fish schools and underwater unmanned vehicles, and other static detection objects including terrain, underwater sediments, and subsequent tracking targets only need to be performed on suspicious moving targets;

[0016] Randomly select k cluster centroids as μ1, μ2,..., μ k ∈R n ;

[0017] Repeat the following process until convergence;

[0018] For each sample i, calculate which class it should belong to

[0019]

[0020] For each class j, recalculate the centroid of the class

[0021]

[0022] Use the Kalman filter algorithm to track suspicious targets; traditional target tracking algorithms segment different regions of the image, classify the segmented regions, and then use the target template to match the regions in adjacent images to achieve region-based tracking of the target; since the feature point-based method is not suitable for noisy and low-resolution sonar images, the region-based matching effect is poor, so the Kalman filter method is selected to achieve target tracking;

[0023] All suspicious targets of the tracked targets are locked by calculating the pixel points in the image; first, record the number of all suspicious target pixel points detected by the image before the operation of the underwater acoustic transducer, and record it in the array aim[], after the operation of the underwater acoustic transducer, continuously scan the frames of the acoustic image, track all targets and calculate the number of pixel points, and compare the value in the aim[] array, if the number of pixel points decreases by 40% and remains for a certain period of time, it will be identified as a fish school and will no longer be tracked; the target still tracked after the detection time t will be marked as a suspicious target;

[0024] Finally, the visual image of the specific suspicious target is recognized and classified through a convolutional neural network (CNN).

[0025] Preferably, the convolutional neural network model is built by using pytorch, and the order from bottom to top is as follows: an input image is firstly processed through a convolutional layer, then the processed information is processed through a pooling method, a maxpooling method is used here, then the processed information is transmitted to two fully connected neural layers, which are also general two-layer neural network layers, and finally, a classifier is connected for classification and prediction.

[0026] Compared with the prior art, the present application has the following beneficial effects:

[0027] The method can be widely applied to detection and identification of movable underwater fish-like robots, so that the AUV can autonomously perceive the surrounding environment.

[0028] The overall device combining detection and ultrasonic fish driving is used to realize detection and identification of underwater moving or static fish-like robots in a large range, the fusion framework can not only drive fish away to ensure that no organisms interfere with the detection process, but also realize detection and identification of fish-like robots by combining side scan sonar images and visual images.

[0029] The method proposes an underwater robot detection method based on a hydroacoustic transducer and a side scan sonar device, drives fish away in a specific direction, and cooperates with the underwater image returned by the side scan sonar, locates suspicious targets through different reaction states of biological fish groups and underwater submersibles, combines a machine vision recognition method, trains a neural network, and further screens suspicious targets in the range to improve the success rate of distinguishing fish-like robots. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 It is a structural diagram of the detection method of the present application.

[0031] Figure 2 It is an implementation block diagram of the Kalman filtering algorithm of the present application.

[0032] Figure 3 It is a schematic diagram of an existing CNN network structure of the present application.

[0033] Figure 4 It is a detector device diagram of the present application.

[0034] Marked numbers in the figure: 1, hydroacoustic transducer; 2, side scan sonar. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application.

[0036] Reference Figures 1-4 A kind of unmanned underwater vehicle detection method based on underwater acoustic transducer and side scan sonar, through underwater acoustic transducer 1 and side scan sonar 2 carried on underwater robot, the sound source level of the underwater acoustic transducer 1 is greater than 140db, frequency 400-1000Hz, frequency adjustable, underwater acoustic transducer 1 will direct dispersion fish school and will not drive the bionic unmanned underwater vehicle possibly hidden in fish school, the side scan sonar 2 can gather sonar image, whether there is unknown underwater unmanned vehicle according to the sonar image obtained is judged, it is to realize the detection of unknown submarine by the combination of underwater acoustic transducer 1 and side scan sonar 2, steps are as follows:

[0037] Step 1, set selected ultrasonic frequency, underwater acoustic transducer works, carries out fish driving operation;Fish driver is composed of water transducer 1, set specific frequency to generate ultrasonic wave, realize the function of dispersing fish school by using the characteristics that different fish school is sensitive to specific ultrasonic frequency.

[0038] Step 2, start side scan sonar, obtain the sound visual image at the current time;The acquisition of sound visual image is mainly based on side scan sonar device, and the task of sound visual system is to obtain two-dimensional acoustic image for post-processing in the third step.

[0039] Step 3: based on the processing of sound visual system, data augmentation is carried out using original sonar image, and the sample set required for training and testing model is obtained, the fish school and bionic fish robot in sample set are artificially marked by different reactions of fish school and unmanned underwater vehicle caused by the principle of ultrasonic fish driving, and are divided, to obtain the label image required for testing and training model;

[0040] Noise points are removed by median filtering;Specifically, 3*3 mask is used to avoid the blurring effect caused by filtering, which can protect the relatively sharp edges in the image while processing smooth impulse noise. The main steps are to move the 3*3 template in the image in turn, read the gray value of the corresponding pixel point and arrange it in ascending order, and assign the median value to the pixel point corresponding to the center of the template.

[0041] This paper utilizes the Canny edge detection algorithm to reduce data size while preserving image features and extracts features through Hough transform. First, the image is smoothed using Gaussian filtering to eliminate noise points. Then, the Sabel operator is used to process the image gradient and direction; a larger gradient is more likely to be an edge. Next, non-maximum suppression is performed to determine whether a pixel is a boundary. Double threshold detection is then used to further confirm the boundary. Finally, Hough transform is used to extract lines; the expression for the extracted line family is as follows:

[0042] r = x0cosθ + y0sinθ

[0043] Step 4: Process and analyze the sonar images, including: removing noise points using median filtering; reducing the data size while preserving image features using the Canny edge detection algorithm and extracting features using Hough transform; extracting all suspicious moving targets in the sonar images using the K-Means clustering analysis algorithm, which still includes schools of fish and unmanned underwater vehicles (UUVs); then tracking suspicious targets using the Kalman filter algorithm combined with a target detector (such as YOLO); judging the tracked objects, since schools of fish will directionally move away from the ultrasonic transducer emitting ultrasonic waves, suspicious targets are extracted; finally, the detection machine carrying a high-definition camera uses a convolutional neural network (CNN) to identify targets and further determine whether the suspicious targets are unknown underwater UUVs.

[0044] The K-Means clustering algorithm is used to extract all suspicious moving targets in sonar images, including schools of fish and unmanned underwater vehicles (UUVs). The steps are as follows: the data is pre-divided into two groups. K objects are randomly selected as initial cluster centers. The distance between each object and each seed cluster center is calculated, and each object is assigned to the nearest cluster center. Each cluster center and the objects assigned to it represent a cluster. Each time a sample is assigned, the cluster centers are recalculated based on the existing objects in the cluster. This process is repeated until a termination condition is met. The termination condition can be that no (or a minimum number) objects are reassigned to different clusters, no (or a minimum number) cluster centers change, or the sum of squared errors reaches a local minimum. In this example, the two groups are suspicious moving targets (including schools of fish and UUVs) and other static detection objects (including terrain, underwater sediments, etc.). Subsequent target tracking only needs to target the suspicious moving targets.

[0045] 1. Randomly select k cluster centroids as μ1, μ2, ..., μ k ∈R n ;

[0046] 2. Repeat the following process until convergence.

[0047] For each sample i, calculate the class it should belong to.

[0048]

[0049] For each class j, recalculate the centroid of that class.

[0050]

[0051] The Kalman filter algorithm is used to track suspicious targets. Traditional target tracking algorithms segment different regions of an image, classify the segmented regions, and then use the target's template in adjacent images to perform region matching to achieve region-based target tracking. Since feature-point-based methods are unsuitable for noisy, low-resolution sonar images, and region-based methods have poor matching results, the Kalman filter method is chosen for target tracking. The specific process of the Kalman filter algorithm is as follows: Figure 2 As shown.

[0052] Suspicious targets are identified by calculating the number of pixels in the image. First, the number of pixels of all suspected targets detected in the image before the underwater acoustic transducer operates is recorded and stored in the array `aim[]`. After the underwater acoustic transducer operates, continuous acoustic images are scanned frame by frame, tracking all targets and comparing the number of pixels with the values ​​in the `aim[]` array. If the number of pixels decreases by 40% and remains so for a certain period, the target is identified as a fish school and tracking ceases. Targets still being tracked after a detection time `t` are marked as suspicious targets.

[0053] Finally, a convolutional neural network (CNN) is used to identify and classify the visual images of specific suspicious targets. Since the CNN model is not the focus of this invention, an existing structure is used. In this example, we choose to use PyTorch to build the CNN. A popular structure is as follows, from bottom to top: first, the input image passes through a convolutional layer, then the information from the convolution is processed using pooling (here, max pooling). This is followed by another round of processing, and the resulting information is fed into two fully connected neural layers (typically two layers in a neural network). Finally, a classifier is added for classification prediction. The flowchart of this popular structure is shown below. Figure 3 As shown.

[0054] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A detection method for unmanned underwater vehicles based on underwater acoustic transducers and side-scan sonar, characterized in that, Includes the following steps: Step 1: Set the selected ultrasonic frequency, the underwater acoustic transducer will work, and the fish repellent will be used to drive the fish away. Step 2: Activate the side-scan sonar to obtain the acoustic-visual image at the current moment; Step 3: Based on the processing of the acoustic vision system, data augmentation is performed using the original sonar images to obtain the sample set required for training and testing the model. The fish and fish-like robots in the sample set are manually labeled and divided according to the different reactions of fish and unmanned underwater vehicles caused by the principle of fish repelling by sound waves, so as to obtain the label map required for testing and training the model. Step 4: Process and analyze the sonar images, including removing noise points using median filtering; The Canny edge detection algorithm is used to reduce the data size while preserving image features, and features are extracted through Hough transform. K-Means clustering analysis is used to extract all suspicious moving targets in the sonar image, including schools of fish and unmanned underwater vehicles (UUVs). Subsequently, Kalman filtering combined with a target detector is used to track the suspicious targets. The tracked objects are then judged; since schools of fish tend to move away from the ultrasonic transducer emitting the waves, suspicious targets are extracted. Finally, a detection device equipped with a high-definition camera uses a convolutional neural network (CNN) for target recognition to further determine whether the suspicious target is an unknown underwater UUV. The Convolutional Neural Network (CNN) is built using PyTorch. From bottom to top, the input image goes through a convolutional layer, then the information from the convolution is processed by pooling (max pooling is used here), and then the image goes through the same processing again. The information from the second processing is then fed into two fully connected neural layers, and finally a classifier is added for classification and prediction.

2. The method for detecting unmanned underwater vehicles based on an underwater acoustic transducer and side-scan sonar according to claim 1, characterized in that, In step 1, the fish repellent device consists of an aquatic transducer that generates ultrasonic waves at a specific frequency. By utilizing the characteristic that different fish groups are sensitive to specific ultrasonic frequencies, the fish groups are dispersed.

3. The method for detecting unmanned underwater vehicles based on an underwater acoustic transducer and side-scan sonar according to claim 1, characterized in that, In step 2, the acquisition of acoustic-visual images is based on a side-scan sonar device. The task of the acoustic-visual system is to acquire two-dimensional acoustic images for post-processing in the third step.

4. The method for detecting unmanned underwater vehicles based on an underwater acoustic transducer and side-scan sonar according to claim 1, characterized in that, In step 3, median filtering is also used to remove noise points. Specifically, a 3*3 mask is used to avoid the blurring effect caused by filtering. While processing smooth impulse noise, it can protect the sharp edges in the image. The 3*3 template is moved sequentially in the image, the gray values ​​of the corresponding pixels are read and sorted in ascending order, and the median value is assigned to the pixel corresponding to the center of the template. This paper utilizes the Canny edge detection algorithm to reduce data size while preserving image features and extracts features through Hough transform. First, the image is smoothed using Gaussian filtering to eliminate noise points. Then, the Sabel operator is used to process the image gradient and direction; a larger gradient is more likely to be an edge. Next, non-maximum suppression is performed to determine whether a pixel is a boundary. Double threshold detection is then used to further determine the boundary. Finally, Hough transform is used to extract straight lines. The expression for the extracted line family is as follows: 。 5. The method for detecting unmanned underwater vehicles based on an underwater acoustic transducer and side-scan sonar according to claim 1, characterized in that, In step 4, the K-Means clustering algorithm is used to extract all suspicious moving targets in the sonar image, including schools of fish and unmanned underwater vehicles (UUVs). The data is pre-divided into two groups: suspicious moving targets and other static detection objects. The suspicious moving targets include schools of fish and UUVs, while the other static detection objects include terrain and underwater sediments. Subsequent target tracking only needs to target the suspicious moving targets. K objects are randomly selected as initial cluster centers, and the distance between each object and each seed cluster center is calculated. Each object is assigned to the cluster center closest to it. The cluster centers and the objects assigned to them represent a cluster. Each time a sample is assigned, the cluster centers are recalculated based on the existing objects in the cluster. This process is repeated until a certain termination condition is met. The termination condition can be that no or a minimum number of objects are reassigned to different clusters, no or a minimum number of cluster centers change, or the sum of squared errors reaches a local minimum. Randomly select k cluster centroids as … ; Repeat the following process until convergence; For each example Calculate the class it should belong to. For each class j, recalculate the centroid of that class. The Kalman filter algorithm is used to track suspicious targets. Traditional target tracking algorithms segment different regions of an image, classify the segmented regions, and then use the target's template in adjacent images to perform region matching to achieve region-based target tracking. However, feature point-based methods are not suitable for noisy and low-resolution sonar images, and region-based methods have poor matching results. Therefore, the Kalman filter method is chosen to achieve target tracking. Suspicious targets of all tracked targets are identified by calculating the number of pixels in the image. First, the number of pixels of all suspicious targets detected in the image before the underwater acoustic transducer is working is recorded and recorded in the array aim[]. After the underwater acoustic transducer is working, the continuous acoustic image is scanned frame by frame to track all targets and calculate the number of pixels and compare it with the value in the aim[] array. If the number of pixels decreases by 40% and remains so for a certain period of time, it will be identified as a school of fish and the tracking will stop. Targets that are still being tracked after the detection time t will be marked as suspicious targets. Finally, a convolutional neural network (CNN) is used to identify and classify visual images of specific suspicious targets.

Images