A hull recognition method based on single-beam sonar cluster data processing
By using single-beam sonar cluster data processing and deep neural network in multi-beam sonar, the problem of insufficient accuracy of multi-beam sonar is solved, high-precision hull recognition is achieved, and the needs of high-precision detection are met.
Patent Information
- Application Number
- CN202010303930.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-04-15
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2040-04-15
AI Technical Summary
The existing multi-beam sonar processed a large amount of sound wave data in hull detection, resulting in limited accuracy, making it difficult to achieve quantitative detection, and cannot meet high-precision security and counter-terrorism needs.
Using a single-beam sonar cluster data processing method, by installing a single-beam sonar cluster on an arched underwater structure, collect and preprocess sonar data, remove water surface data, and use SqueezeNet network to train the hull recognition model to achieve high-precision hull Identify.
It improves the accuracy and economicality of hull recognition, can achieve recognition accuracy of more than 0.95, and meets the needs of high-precision detection.
Smart Images

Figure CN111650592B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of deep neural networks, and in particular to a hull recognition method based on single-beam sonar cluster data processing. Background Art
[0002] Sonar is one of the important equipment for underwater detection and inspection. Sonar equipment has become a key research object in the development of underwater target detection and inspection industry due to its advantages such as being less affected by water environment factors and reliable acoustic performance. In recent years, sonar equipment has been gradually applied to the ship inspection industry, by installing sonar on underwater structures or working ships to conduct safety inspections on passing hulls. How to identify the hull through sonar data has become the first problem to be solved.
[0003] Sonar equipment is mainly divided into two categories: multi-beam sonar and single-beam sonar. When using multi-beam sonar for hull detection and inspection, rich sonar images can be obtained. However, since multi-beam sonar needs to process a large amount of sound wave data at the same time, the accuracy of multi-beam sonar images is greatly limited, so that the detection and inspection work can only be qualitative, not quantitative. For high-precision inspection operations such as security and anti-terrorism, the performance of multi-beam sonar is far from enough. Compared with multi-beam sonar, single-beam sonar has higher accuracy and economical cost. Therefore, the hull recognition method based on the combination of single-beam sonar cluster data processing and deep neural network learning has important research value. Summary of the invention
[0004] The present invention mainly aims at the hot issues in the prior art field and provides a hull recognition method based on single-beam sonar cluster data processing.
[0005] The technical solution of the present invention to solve the corresponding technical problem is: a hull recognition method based on single-beam sonar cluster data processing, comprising the following steps:
[0006] 1) Install single-beam sonar clusters on the surface of the arched underwater structure and arrange them into an array;
[0007] 2) Collect sonar data through a single-beam sonar cluster, number them according to the arrangement of the sonar clusters, and extract a single-beam sonar cluster data atlas;
[0008] 3) Preprocessing the single-beam sonar cluster data atlas, converting the spherical data into plane data to obtain a two-dimensional sonar data atlas;
[0009] 4) Eliminate the surface data in the two-dimensional sonar data set to reduce the impact of surface data points on target ship identification;
[0010] 5) A large amount of single-beam sonar cluster data of normal hulls is collected as data training sets and data verification sets, and the SqueezeNet network is used as the hull recognition model network to train the data set to obtain the hull recognition model.
[0011] The step 1) of installing the single-beam sonar clusters on the surface of the arched underwater structure to form an array specifically includes:
[0012] The single-beam sonar cluster is installed on the arch-shaped underwater structure. The installation curvature angle θ of any single-beam sonar n , axial installation distance l n ;
[0013] The number of single-beam sonar clusters arranged in each column along the radial direction is H, and the number of single-beam sonar clusters arranged in each column along the axial direction is W.
[0014] In the step 2), sonar data is collected by a single-beam sonar cluster, and the data is numbered according to the arrangement of the sonar clusters. The single-beam sonar cluster data atlas is extracted, which specifically includes: the single-beam sonar cluster collection update rate is 15 Hz, the beam angle is 5°, the measurement range is 20 m, the accuracy is 1 mm ± 1% of the range, and the measurement data of the single-beam sonar is the water depth value depth and the echo intensity intensity;
[0015] The single beam sonars are numbered in order from 1 to n.
[0016] The step 3) pre-processes the single-beam sonar cluster data atlas to convert the spherical data into plane data to obtain a two-dimensional sonar data atlas, which specifically includes:
[0017] The position of any point in three-dimensional space (θ n , l n ) is projected to the position of a point in the two-dimensional coordinate system (i n , j n ), the projection relationship is as follows:
[0018] i n = l n , j n =depth n ·sin θ n
[0019] We get a single-beam sonar cluster data tensor V of size (H, W, C), where C represents the number of features of the beam sonar cluster data, which is 4 here. The 4 features are (θ n , l n , depth n , intensity n ).
[0020] The step 4) of eliminating the water surface data in the two-dimensional sonar data atlas to reduce the influence of the water surface data points on the target hull identification specifically includes:
[0021] 5.1) Extract the sub-tensor representing the sonar water depth value in the tensor V obtained in step 4, and convert it into a set of water depth values in a rectangular coordinate system, as follows:
[0022]
[0023] The number of data points in this set is n;
[0024] 5.2) Introduce the average value LPR of the lowest height point of the data set depth ,as follows:
[0025]
[0026] Where num LPR Indicates that LPR is selected depth The number of data points, num LPR =0.5n;
[0027] Select depth n >LPR depth +Th seed , sonar data points input water surface point cloud P sow , where Th seed represents the threshold for selecting water surface estimation data points, P sow Represents the water surface estimation data point set, select depth n ≤LPR depth +Th seed The sonar data points are classified into the non-water surface point cloud P nsow ;
[0028] 5.3) Establish a water surface estimation model as follows
[0029] n T x=-d
[0030] Where n represents the normal vector of the water surface estimation model, and d represents the offset of the water surface estimation model x = [i, j, depth] T , x∈P sow ;
[0031] Introduce the water surface estimation data point set P sow The covariance matrix of is as follows:
[0032]
[0033] In the formula is the mean of all data points;
[0034] By performing singular value decomposition on the covariance matrix C, we can solve the normal vector n perpendicular to the water surface and substitute it into Directly obtain the offset d of the water surface estimation model;
[0035] 5.4) After solving all water surface data points and obtaining the water surface estimation model in step 3), calculate the water depth set D = (i, j, depth), where the distance of each point to the orthogonal projection of the plane is depth_normal n , when depth_normal n >Th dis When , the data point belongs to the water surface, otherwise it is a non-water surface point, where Th dis Indicates the plane distance threshold;
[0036] 5.5) Based on the water surface point data position obtained in step 5.4), the corresponding sonar point in the tensor V is removed to obtain V′.
[0037] In the step 5), a large amount of single-beam sonar cluster data of normal hulls is collected as a data training set and a data verification set, and the SqueezeNet network is used as a hull recognition model network to train the data set, and the hull recognition model is obtained, which specifically includes:
[0038] 6.1) Collect a large amount of single-beam sonar cluster data of the hull and perform steps 4 and 5 to obtain a data training set S and a data verification set S′, and label each single-beam sonar cluster data to indicate whether it is hull data;
[0039] 6.2) Set the intrinsic parameters of the SqueezeNet network CNN module, set the number of channels of the firemodule module's squeeze layer to 1 / 4 of the input tensor depth size, and the number of channels of the expansion layer to 1 / 2 of the input tensor depth size;
[0040] Set the number of channels of the firedeconv module's squeeze layer to 1 / 4 of the input tensor's depth size, the channel coefficient of the deconvolution downsampling layer to 2, and the number of channels of the expansion layer to 1 / 2 of the input tensor's depth size;
[0041] The height of the input tensor is not equal to its width. The height and width of the input tensor are adjusted to be consistent by setting the number of maximum pooling layers.
[0042] 6.3) Use the CRF layer to eliminate the boundary blur phenomenon, and use the output of the CNN module as the input of the CRF module;
[0043] Two Gaussian kernels are calculated based on the single-beam sonar cluster data tensor V′. The Gaussian kernel formula is as follows:
[0044]
[0045] Where P is the curvature angle of the data point on the spherical surface and the axial installation distance, w 1 , w 2 , is the threshold value;
[0046] By calculating the probability of aggregating neighboring points with the Gaussian kernel, we get the probability distribution weight matrix; (the Gaussian kernel measures the difference in features between two points. The greater the difference between the two points, the smaller the value of the Gaussian kernel and the smaller the correlation between the two points)
[0047] Through the 1x1 convolution kernel gcov, fine-tune the probability distribution weight of each point to maximize the probability distribution weight;
[0048] 6.4) Import the data training set S into the SqueezeNet network for training, and obtain the CNN module network parameters, the CRF module convolution kernel gcov and the hull recognition model based on single-beam sonar cluster data;
[0049] 6.5) The hull recognition model and related parameters obtained in 6.4) are verified by the data verification set S′. When the recognition accuracy is greater than or equal to 0.95, the training is stopped, otherwise, it goes to step 6.4). BRIEF DESCRIPTION OF THE DRAWINGS
[0050] The present invention will be further described below with reference to the accompanying drawings, in which:
[0051] Figure 1 It is a flow chart of a hull recognition method based on single beam sonar cluster data processing of the present invention;
[0052] Figure 2 This is a stereogram of the installation of a single-beam sonar cluster;
[0053] Figure 3 It is a schematic diagram of converting the single-beam sonar cluster data atlas designed by the present invention from spherical data to plane data;
[0054] Figure 4 It is a diagram of the SqueezeNet network parameter settings used in the present invention. DETAILED DESCRIPTION
[0055] In order to make the technical solution and problem-solving ideas of the present invention clearer, the present invention is further described in detail below with reference to examples. It should be understood that the specific examples described here are only used to explain the present invention and are not used to limit the present invention.
[0056] like Figure 1 As shown, a hull recognition method based on single-beam sonar cluster data processing includes the following steps:
[0057] Step STEP1, install the single beam sonar cluster on the surface of the arched underwater structure and arrange it into an array, such as Figure 2 As shown, arranged in an array, specifically including:
[0058] The single-beam sonar cluster is installed on the arch-shaped underwater structure. The installation curvature angle θ of any single-beam sonar n , axial installation distance l n ;
[0059] The number of single-beam sonar clusters arranged in each column along the radial direction is H, and the number of single-beam sonar clusters arranged in each column along the axial direction is W.
[0060] Step STEP2, collecting sonar data through a single-beam sonar cluster, numbering them according to the arrangement of the sonar clusters, and extracting a single-beam sonar cluster data atlas, specifically including:
[0061] The single-beam sonar cluster acquisition update rate is 15Hz, the beam angle is 5°, the measurement range is 20m, and the accuracy is 1mm±1% of the range. The measurement data of the single-beam sonar are the water depth value depth and echo intensity intensity;
[0062] The single beam sonars are numbered in order from 1 to n.
[0063] Step STEP3, pre-process the single-beam sonar cluster data atlas, convert the spherical data into plane data, and obtain a two-dimensional sonar data atlas, such as Figure 3 As shown, specifically including:
[0064] Perform a planar projection on the single-beam sonar cluster data to obtain a dense, two-dimensional data matrix from the sparse and irregular 3D data point array of the point cloud data;
[0065] The position of any point in three-dimensional space (θ n , l n ) is projected to the position of a point in the two-dimensional coordinate system (i n , j n ),like Figure 3 As shown, the projection relationship is as follows:
[0066] i n = l n , j n =depth n ·sin θ n
[0067] We get a single-beam sonar cluster data tensor V of size (H, W, C), where C represents the number of features of the beam sonar cluster data, which is 4 here. The 4 features are (θ n , l n, depth n , intensity n ).
[0068] Step 4: remove the surface data in the two-dimensional sonar data set to reduce the influence of the surface data points on the target hull identification, specifically including:
[0069] a) Extract the sub-tensor representing the sonar water depth value in the tensor V and convert it into a set of water depth values in the rectangular coordinate system as follows:
[0070]
[0071] The number of data points in this set is n;
[0072] b) The average LPR of the lowest altitude point of the introduced data set depth ,as follows:
[0073]
[0074] Where num LPR Indicates that LPR is selected depth The number of data points, num LPR =0.5n;
[0075] Select depth n >LPR depth +Th seed , sonar data points input water surface point cloud P sow , where Th seed represents the threshold for selecting water surface estimation data points, P sow Represents the water surface estimation data point set, select depth n ≤LPR depth +Th seed The sonar data points are classified into the non-water surface point cloud P nsow ;
[0076] c) Establish a water surface estimation model as follows
[0077] n T x=-d
[0078] Where n represents the normal vector of the water surface estimation model, and d represents the offset of the water surface estimation model x = [i, j, depth] T , x∈P sow ;
[0079] Introduce the water surface estimation data point set P sow The covariance matrix of is as follows:
[0080]
[0081] In the formula is the mean of all data points;
[0082] By performing singular value decomposition on the covariance matrix C, we can solve the normal vector n perpendicular to the water surface and substitute it into Directly obtain the offset d of the water surface estimation model;
[0083] d) After solving all water surface data points and obtaining the water surface estimation model, calculate the water depth set D = (i, j, depth), where the distance of each point to the orthogonal projection of the plane is depth_normal n , when depth_normal n >Th dis When , the data point belongs to the water surface, otherwise it is a non-water surface point, where Th dis Indicates the plane distance threshold;
[0084] e) Based on the water surface point data position obtained in step d), the corresponding sonar point in the tensor V is removed to obtain V′.
[0085] Step STEP5, collect a large amount of single-beam sonar cluster data of normal hulls as data training sets and data verification sets, use SqueezeNet network as the hull recognition model network to train the data set to obtain the hull recognition model, specifically including:
[0086] a) Collect a large amount of single-beam sonar cluster data of the hull and perform steps 4 and 5 to obtain a data training set S and a data verification set S′, and label each single-beam sonar cluster data to indicate whether it is hull data;
[0087] b) Set the intrinsic parameters of the SqueezeNet network CNN module, such as Figure 4 As shown, the number of channels of the firemodule module's squeeze layer is set to 1 / 4 of the input tensor's depth size, and the number of channels of the expansion layer is set to 1 / 2 of the input tensor's depth size;
[0088] Set the number of channels of the firedeconv module's squeeze layer to 1 / 4 of the input tensor's depth size, the channel coefficient of the deconvolution downsampling layer to 2, and the number of channels of the expansion layer to 1 / 2 of the input tensor's depth size;
[0089] The height of the input tensor is not equal to its width. The height and width of the input tensor are adjusted to be consistent by setting the number of maximum pooling layers.
[0090] c) The CRF layer is used to eliminate the boundary blur phenomenon, and the output of the CNN module is used as the input of the CRF module;
[0091] Two Gaussian kernels are calculated based on the single-beam sonar cluster data tensor V′. The Gaussian kernel formula is as follows:
[0092]
[0093] Where P is the curvature angle of the data point on the spherical surface and the axial installation distance, w 1 , w 2 , is the threshold value;
[0094] By calculating the probability of aggregating neighborhood points using the Gaussian kernel, we get the probability distribution weight matrix;
[0095] The Gaussian kernel measures the difference in features between two points. The greater the difference between the two points, the smaller the value of the Gaussian kernel and the smaller the correlation between the two points.
[0096] Through the 1x1 convolution kernel gcov, fine-tune the probability distribution weight of each point to maximize the probability distribution weight;
[0097] d) Importing the data training set S into the SqueezeNet network for training, obtaining the CNN module network parameters, the CRF module convolution kernel gcov and the hull recognition model based on single-beam sonar cluster data;
[0098] e) Verify the hull recognition model and related parameters obtained in 6.4) through the data verification set S′. Stop training when the recognition accuracy is greater than or equal to 0.95, otherwise proceed to step d).
Claims
1. A hull recognition method based on single-beam sonar cluster data processing, characterized in that: 1) Install single-beam sonar clusters on the surface of the arched underwater structure and arrange them into an array; 2) Collect sonar data through a single-beam sonar cluster, number them according to the arrangement of the sonar clusters, and extract a single-beam sonar cluster data atlas; 3) Preprocessing the single-beam sonar cluster data atlas, converting the spherical data into plane data to obtain a two-dimensional sonar data atlas; 4) Eliminate the surface data in the two-dimensional sonar data set to reduce the impact of surface data points on target ship identification; 5) A large amount of single-beam sonar cluster data of normal hulls is collected as data training sets and data verification sets, and the SqueezeNet network is used as the hull recognition model network to train the data set to obtain the hull recognition model.
2. The hull recognition method based on single-beam sonar cluster data processing according to claim 1 is characterized in that: The step 1) of installing the single-beam sonar clusters on the surface of the arched underwater structure to form an array specifically includes: The single-beam sonar cluster is installed on the arch-shaped underwater structure. The installation curvature angle θ of any single-beam sonar n , axial installation distance l n ; The number of single-beam sonar clusters arranged in each column along the radial direction is H, and the number of single-beam sonar clusters arranged in each column along the axial direction is W.
3. The hull recognition method based on single beam sonar cluster data processing according to claim 1 is characterized in that: In the step 2), sonar data is collected by a single-beam sonar cluster, and the data is numbered according to the arrangement of the sonar clusters. The single-beam sonar cluster data atlas is extracted, which specifically includes: the single-beam sonar cluster collection update rate is 15 Hz, the beam angle is 5°, the measurement range is 20 m, the accuracy is 1 mm ± 1% of the range, and the measurement data of the single-beam sonar is the water depth value depth and the echo intensity intensity; The single beam sonars are numbered in order from 1 to n.
4. The hull recognition method based on single beam sonar cluster data processing according to claim 1 is characterized in that: The step 3) pre-processes the single-beam sonar cluster data atlas to convert the spherical data into plane data to obtain a two-dimensional sonar data atlas, which specifically includes: The position of any point in three-dimensional space (θ n , l n ) is projected to the position of a point in the two-dimensional coordinate system (i n , j n ), the projection relationship is as follows: i n =l n ,j n =depth n ·sinθ n We get a single-beam sonar cluster data tensor V of size (H, W, C), where C represents the number of features of the beam sonar cluster data, which is 4 here. The 4 features are (θ n , l n , depth n , intensity n ).
5. The hull recognition method based on single beam sonar cluster data processing according to claim 1 is characterized in that: The step 4) of eliminating the water surface data in the two-dimensional sonar data atlas to reduce the influence of the water surface data points on the target hull identification specifically includes: 5.1) Extract the sub-tensor representing the sonar water depth value in the tensor V obtained in step 4, and convert it into a set of water depth values in a rectangular coordinate system, as follows: The number of data points in this set is n; 5.2) Introduce the average value LPR of the lowest height point of the data set depth ,as follows: Where num LPR Indicates that LPR is selected depth The number of data points, num LPR =0.5n; Select depth n >LPR depth +Th seed , sonar data points input water surface point cloud P sow , where Th seed represents the threshold for selecting water surface estimation data points, P sow represents a set of water surface estimation data points; Select depth n ≤LPR depth +Th seed The sonar data points are classified into the non-water surface point cloud P nsow ; 5.3) Establish a water surface estimation model as follows n T x=-d Where n represents the normal vector of the water surface estimation model, and d represents the offset of the water surface estimation model x = [i, j, depth] T , x∈P sow ; Introduce the water surface estimation data point set P sow The covariance matrix of is as follows: In the formula is the mean of all data points; By performing singular value decomposition on the covariance matrix C, we can solve the normal vector n perpendicular to the water surface and substitute it into Directly obtain the offset d of the water surface estimation model; 5.4) After solving all water surface data points and obtaining the water surface estimation model in step 3), calculate the water depth set D = (i, j, depth), where the distance of each point to the orthogonal projection of the plane is depth_normal n , when depth_normal n >Th dis When , the data point belongs to the water surface, otherwise it is a non-water surface point, where Th dis Indicates the plane distance threshold; 5.5) Based on the water surface point data position obtained in step 5.4), the corresponding sonar point in the tensor V is removed to obtain V′.
6. The method for ship hull recognition based on single beam sonar cluster data processing according to claim 1, characterized in that: In the step 5), a large amount of single-beam sonar cluster data of normal hulls is collected as a data training set and a data verification set, and the SqueezeNet network is used as a hull recognition model network to train the data set, and the hull recognition model is obtained, which specifically includes: 6.1) Collect a large amount of single-beam sonar cluster data of the hull and perform steps 4) and 5) to obtain a data training set S and a data verification set S′, and label each single-beam sonar cluster data to indicate whether it is hull data; 6.2) Set the intrinsic parameters of the SqueezeNet network CNN module, set the number of channels of the firemodule module's squeeze layer to 1 / 4 of the input tensor depth size, and the number of channels of the expansion layer to 1 / 2 of the input tensor depth size; Set the number of channels of the firedeconv module's squeeze layer to 1 / 4 of the input tensor's depth size, the channel coefficient of the deconvolution downsampling layer to 2, and the number of channels of the expansion layer to 1 / 2 of the input tensor's depth size; The height of the input tensor is not equal to its width. The height and width of the input tensor are adjusted to be consistent by setting the number of maximum pooling layers. 6.3) Use the CRF layer to eliminate the boundary blur phenomenon, and use the output of the CNN module as the input of the CRF module; Two Gaussian kernels are calculated based on the single-beam sonar cluster data tensor V′. The Gaussian kernel formula is as follows: Where P is the curvature angle of the data point on the spherical surface and the axial installation distance, w1, w2, is the threshold value; By calculating the probability of aggregating neighborhood points using the Gaussian kernel, we get the probability distribution weight matrix; Through the 1x1 convolution kernel gcov, fine-tune the probability distribution weight of each point to maximize the probability distribution weight; 6.4) Import the data training set S into the SqueezeNet network for training, and obtain the CNN module network parameters, the CRF module convolution kernel gcov and the hull recognition model based on single-beam sonar cluster data; 6.5) The hull recognition model and related parameters obtained in 6.4) are verified by the data verification set S′. When the recognition accuracy is greater than or equal to 0.95, the training is stopped, otherwise, it goes to step 6.4).
