Underwater sonar three-dimensional point cloud denoising method based on pattern mining

Abstract

The invention provides an underwater sonar three-dimensional point cloud denoising method based on pattern mining, which is used for processing visual content of three-dimensional sonar original data. The method comprises the following steps of: 1, deleting unreasonable random discrete point cloud data by using a statistical abnormal value removal algorithm SOR aiming at a sonar automechanical noise common mode, and 2, automatically excavating mode characteristics of underwater object and environment noise, underwater biological noise and reverberation reflection signals by using a recursive adaptive Gaussian mixture model GMM architecture, and automatically extracting the mode characteristics of the underwater object and environment noise, the underwater biological noise and the reverberation reflection signals. In each recursive stage of the recursive self-adaptive Gaussian mixture model, the self-adaptive mode mining mode features and judging whether the denoising process is completed or not; according to the method, aiming at the common characteristics of sensor noise points, environment noise points, underwater organism noise points, suspended matter noise points and underwater sound wave reverberation reflection noise points of the underwater sonar three-dimensional point cloud, the statistical outlier removal technology and the Gaussian mixture model feature extraction technology are combined, and rapid and efficient underwater sonar three-dimensional point cloud denoising can be achieved.

Description

Technical Field

[0001] The invention relates to the technical field of equipment fault detection, in particular to an underwater sonar 3D point cloud denoising method based on pattern mining. Background Art

[0002] Although the principle of underwater 3D sonar ranging is the same as that of 3D laser ranging on land. However, compared with 3D laser sensors on land, underwater sonar point clouds based on acoustic principles are susceptible to noise interference. Underwater 3D sonar point cloud data has disadvantages such as low resolution and high noise intensity, and it is difficult to directly obtain point cloud data with low noise interference. Usually, in actual underwater measurements, 3D sonar point cloud data will be affected by multiple noises such as environmental noise, underwater biological / suspended matter noise, reverberation reflection signals and self-mechanical noise. The current standard practice for denoising is to follow the white paper guidance and software provided by the sonar manufacturer to manually identify the noise source, repeatedly select the noisy point cloud area, and manually denoise, which is labor-intensive and inefficient.

[0003] Given the complexity and diversity of underwater sonar measurement environments, simply following the denoising white paper guidance provided by sonar manufacturers cannot directly obtain low-noise point cloud data. In practice, the preprocessing of underwater point clouds requires a heavy manual denoising process. In view of this, an efficient automatic denoising algorithm is needed to preprocess multiple noises such as environmental noise, underwater biological / suspended matter noise, reverberation reflection signals, and self-mechanical noise in underwater 3D sonar point cloud data. Summary of the invention

[0004] The present invention proposes a method for denoising underwater sonar 3D point clouds based on pattern mining. Aiming at the common sensor self-noise, environmental noise, aquatic organisms, suspended noise and underwater sound wave reverberation reflection noise characteristics of underwater sonar 3D point clouds, the present invention combines statistical outlier removal (SOR) technology and Gaussian mixture model (GMM) feature extraction technology to achieve fast and efficient underwater sonar 3D point cloud denoising.

[0005] The present invention adopts the following technical solutions.

[0006] The method for denoising underwater sonar 3D point cloud based on pattern mining is directed to processing the visualization content of 3D sonar raw data of underwater sonar, and comprises the following steps:

[0007] Step 1: For the common patterns of sonar self-mechanical noise, the statistical outlier removal algorithm SOR is used to delete unreasonable random discrete point cloud data (on the water surface and under the seabed).

[0008] Step 2: Use the recursive adaptive Gaussian mixture model (GMM) architecture to automatically mine the pattern features of underwater objects and environmental noise, underwater biological noise, and reverberation reflection signals. In each recursive stage of the recursive adaptive Gaussian mixture model, the adaptive model mines the pattern features and determines whether the denoising process is completed.

[0009] The random discrete point cloud data includes the self-mechanical noise point cloud existing in the original measurement data of the underwater sonar, and also includes the discrete noise point cloud.

[0010] The main body of the noise pattern of the self-mechanical noise point cloud is a randomly distributed discrete point cloud, and the distribution area appears on the water surface and under the seabed.

[0011] The discrete noise point cloud is randomly distributed and has Gaussian distribution parameters that are significantly different from those of underwater objects.

[0012] In step 1, the method of using the statistical outlier removal algorithm includes:

[0013] Method 1: According to the Gaussian distribution statistics of the original data, the Gaussian distribution parameter threshold is set;

[0014] Method 2: Calculate the number of nearest neighbors of the original data and set the nearest neighbor parameter threshold;

[0015] In step 2, the Gaussian mixture model is composed of a mixture of different probability distribution models, and each component of the model has an independent Gaussian distribution to represent a subset of the data;

[0016] The GMM model is data-oriented at runtime. It directly inputs the original data set without reference labels. At the same time, it adaptively finds the noise pattern of the underwater sonar point cloud by directly using the unsupervised expectation maximization algorithm EM, and uses the pattern mining of the recursive architecture to mine a sufficiently detailed noise model. In each recursive stage, the EM algorithm estimates the noise component parameters and groups them. Finally, all mined GMM patterns are mapped to real noise categories.

[0017] When measuring underwater scenes in the method of step 1, the 3D sonar raw data of the underwater sonar is first visualized to form visualization content, such as Figure 2 As shown, it is determined from the visualization content whether the 3D point cloud data is affected by multiple noises such as environmental noise, suspended matter noise in water, reverberation reflection signal and self-mechanical noise. If it is determined to be affected by the above multiple noises, the nearest neighbor parameter threshold is set in the software, and the statistical outlier removal algorithm SOR is used for processing.

[0018] When the software uses matlab, if the value is set to 10 when setting the nearest neighbor parameter threshold, the relevant calling code is:

[0019] [ptCloudOut,inlierIndices,outlierIndices]=

[0020] pcdenoise(point cloud data variable name,'NumNeighbors',10);

[0021] The outlierIndices variable is set to: all discrete point clouds that do not meet the nearest neighbor parameter threshold = 10. The sub-class point cloud data mainly includes unreasonable noise point cloud data on the water surface and under the seabed;

[0022] The visualization content of the 3D sonar raw data is processed with Matlab to remove environmental noise, water suspended noise, reverberation reflection signal and self-mechanical noise; Figure 3 shown.

[0023] In step 2, the recursive adaptive Gaussian mixture model (GMM) architecture is used to automatically mine the pattern features of underwater objects and environmental noise, underwater biological noise, and reverberation reflection signals, and automatically visualize the patterns of surface environmental noise, underwater biological and suspended noise, and underwater sound wave reverberation reflection noise. The unsupervised expectation maximization algorithm (EM) of the recursive adaptive Gaussian mixture model is used to adaptively find different noise patterns of underwater sonar point clouds to automatically separate the noise set data.

[0024] In step 2, Matlab software is used to set the Gaussian mixture model parameters in the software. If the GMM model parameter is set to 4, the calling code is

[0025] gm = fitgmdist (point cloud data variable name, 4);

[0026] idx = cluster(gm,X');

[0027] The idx variable includes four Gaussian distribution point cloud patterns mined by the EM algorithm data;

[0028] In each recursive stage, all the mined patterns of the GMM model are mapped to the real noise categories. Figure 4 .

[0029] In step 2, after removing the noise point cloud of the water surface environment noise, the water suspended matter noise and the reverberation reflection signal from the visualization content, the remaining GMM model point cloud is reconstructed, that is, the denoising preprocessing of the underwater sonar point cloud is completed, as shown in Figure 5 .

[0030] The present invention proposes a method for denoising underwater sonar 3D point clouds based on pattern mining. The method aims at the common characteristics of sensor self-noise, environmental noise, aquatic organisms, suspended noise and underwater sound wave reverberation reflection noise of underwater sonar 3D point clouds, combines statistical outlier removal (SOR) technology and Gaussian mixture model (GMM) feature extraction technology, and realizes a fast and efficient underwater sonar 3D point cloud denoising method. The method can target the noise pattern of underwater sonar 3D point clouds, adopt an adaptive pattern mining architecture, and recursively perform underwater point cloud denoising.

[0031] Compared with the traditional technology, the advantages of the present invention are:

[0032] 1. Compared with the classic statistical outlier removal algorithm commonly used for terrestrial laser point cloud data, this method automatically mines the multiple noise pattern characteristics of underwater 3D sonar. It can completely remove underwater environmental noise, underwater biological / suspended matter noise, reverberation reflection signal and self-mechanical noise point cloud data.

[0033] 2. This method uses unsupervised filtering to remove noise point clouds from underwater 3D sonar raw data. No pre-trained label samples are required, and it can be applied to denoising detection data in unfamiliar waters.

[0034] 3. Combined with the original white paper of sonar equipment, it can simplify the cumbersome semi-automatic manual denoising process and quickly reconstruct clean underwater 3D point cloud data.

[0035] 4. The existing patent uses the classic statistical outlier removal SOR algorithm, which is only suitable for deleting random discrete unreasonable noise point cloud data, and cannot delete other underwater noise point clouds with non-directional density. The present invention uses a recursive adaptive Gaussian mixture model (GMM), which can automatically mine the point cloud pattern features of underwater objects and environmental noise, underwater biological / suspended noise and reverberation reflection signals.

[0036] 5. Existing patents use statistical filtering to denoise underwater biological noise and underwater environment, and specific filtering parameters must be set in the algorithm. The GMM model of the present invention uses an unsupervised expectation maximization EM algorithm, which can adaptively find different noise patterns of underwater point clouds. With a recursive pattern mining architecture, in each recursive stage, the EM algorithm adjusts the pattern parameters for each noise component and groups them, and finally connects all mined patterns to the real noise category.

[0037] 6. The existing patents require obtaining the label samples of underwater object point clouds in advance and then training the denoising neural network. The proposal described in the present invention has an unsupervised learning mechanism: that is, it does not need to know the point cloud labels of underwater structure / non-structure objects in advance, and can adaptively mine the patterns of multiple noises such as environmental noise, reverberation reflection signals and self-mechanical noise from the original underwater point cloud data.

[0038] The present invention mainly comprises the following advantages:

[0039] (1) The present invention adopts adaptive denoising technology and does not require an accurate noise model or geometric statistics of the original point cloud.

[0040] (2) The present invention can remove random noise of the sonar sensor itself and unreasonable underwater 3D point cloud data by counting outlier points SOR.

[0041] (3) The present invention uses a recursive Gaussian mixture model to adaptively mine the noise pattern characteristics of the sonar 3D point cloud, automatically visualize the patterns of water surface environmental noise, underwater organisms, suspended matter noise, and underwater sound wave reverberation reflection noise, and automatically separate the noise set data.

[0042] (4) The method of the present invention is easy to operate and can effectively improve the heavy load of manual denoising of underwater sonar 3D point clouds.

[0043] The present invention adopts a highly efficient automatic denoising algorithm for preprocessing multiple noises such as environmental noise, underwater biological / suspended matter noise, reverberation reflection signal and self-mechanical noise of underwater 3D sonar point cloud data. The denoising algorithm has the following advantages:

[0044] 1. It has an adaptive denoising mode: it does not require the precise parameters of the noise model or the statistical parameters of the underwater point cloud.

[0045] 2. It has an unsupervised learning mechanism: that is, it is not necessary to obtain the point cloud labels of underwater structures / non-structure objects in advance.

[0046] 3. Possessing diverse denoising capabilities: It can adaptively mine the patterns of multiple noises, including environmental noise, aquatic biological / suspended matter noise, reverberation reflection signals, and self-mechanical noise, from the original underwater point cloud data. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0048] Attached Figure 1 It is a schematic diagram of the process of the present invention;

[0049] Attached Figure 2 It is a schematic diagram of the raw data of underwater 3D sonar;

[0050] Attached Figure 3 This is a schematic diagram of the point cloud data after SOR removal in step 1;

[0051] Attached Figure 4 is a schematic diagram of noise pattern characteristics of water surface environmental noise, suspended matter noise in water and reverberation reflection signal in a port environment in an embodiment;

[0052] Attached Figure 5 It is for Figure 4 Schematic diagram of underwater sonar 3D point cloud after reconstruction and denoising. DETAILED DESCRIPTION

[0053] like Figure 1 As shown, a method for denoising underwater sonar 3D point cloud based on pattern mining, the denoising method processes the visualization content of 3D sonar raw data of underwater sonar, and includes the following steps;

[0054] Step 1: For the common patterns of sonar self-mechanical noise, the statistical outlier removal algorithm SOR is used to delete unreasonable random discrete point cloud data (on the water surface and under the seabed).

[0055] Step 2: Use the recursive adaptive Gaussian mixture model (GMM) architecture to automatically mine the pattern features of underwater objects and environmental noise, underwater biological noise, and reverberation reflection signals. In each recursive stage of the recursive adaptive Gaussian mixture model, the adaptive model mines the pattern features and determines whether the denoising process is completed.

[0056] The random discrete point cloud data includes the self-mechanical noise point cloud existing in the original measurement data of the underwater sonar, and also includes the discrete noise point cloud.

[0057] The main body of the noise pattern of the self-mechanical noise point cloud is a randomly distributed discrete point cloud, and the distribution area appears on the water surface and under the seabed.

[0058] The discrete noise point cloud is randomly distributed and has Gaussian distribution parameters that are significantly different from those of underwater objects.

[0059] In step 1, the method of using the statistical outlier removal algorithm includes:

[0060] Method 1: According to the Gaussian distribution statistics of the original data, the Gaussian distribution parameter threshold is set;

[0061] Method 2: Calculate the number of nearest neighbors of the original data and set the nearest neighbor parameter threshold;

[0062] In step 2, the Gaussian mixture model is composed of a mixture of different probability distribution models, and each component of the model has an independent Gaussian distribution to represent a subset of the data;

[0063] The GMM model is data-oriented at runtime. It directly inputs the original data set without reference labels. At the same time, it adaptively finds the noise pattern of the underwater sonar point cloud by directly using the unsupervised expectation maximization algorithm EM, and uses the pattern mining of the recursive architecture to mine a sufficiently detailed noise model. In each recursive stage, the EM algorithm estimates the noise component parameters and groups them. Finally, all mined GMM patterns are mapped to real noise categories.

[0064] When measuring underwater scenes in the method of step 1, the 3D sonar raw data of the underwater sonar is first visualized to form visualization content, such as Figure 2 As shown, it is determined from the visualization content whether the 3D point cloud data is affected by multiple noises such as environmental noise, suspended matter noise in water, reverberation reflection signal and self-mechanical noise. If it is determined to be affected by the above multiple noises, the nearest neighbor parameter threshold is set in the software, and the statistical outlier removal algorithm SOR is used for processing.

[0065] When the software uses matlab, if the value is set to 10 when setting the nearest neighbor parameter threshold, the relevant calling code is:

[0066] [ptCloudOut,inlierIndices,outlierIndices]=

[0067] pcdenoise(point cloud data variable name,'NumNeighbors',10);

[0068] The outlierIndices variable is set to: all discrete point clouds that do not meet the nearest neighbor parameter threshold = 10. The sub-class point cloud data mainly includes unreasonable noise point cloud data on the water surface and under the seabed;

[0069] The visualization content of the 3D sonar raw data is processed with Matlab to remove environmental noise, water suspended noise, reverberation reflection signal and self-mechanical noise; Figure 3 shown.

[0070] In step 2, the recursive adaptive Gaussian mixture model (GMM) architecture is used to automatically mine the pattern features of underwater objects and environmental noise, underwater biological noise, and reverberation reflection signals, and automatically visualize the patterns of surface environmental noise, underwater biological and suspended noise, and underwater sound wave reverberation reflection noise. The unsupervised expectation maximization algorithm (EM) of the recursive adaptive Gaussian mixture model is used to adaptively find different noise patterns of underwater sonar point clouds to automatically separate the noise set data.

[0071] In step 2, Matlab software is used to set the Gaussian mixture model parameters in the software. If the GMM model parameter is set to 4, the calling code is

[0072] gm = fitgmdist (point cloud data variable name, 4);

[0073] idx = cluster(gm,X');

[0074] The idx variable includes four Gaussian distribution point cloud patterns mined by the EM algorithm data;

[0075] In each recursive stage, all the mined patterns of the GMM model are mapped to the real noise categories. Figure 4 .

[0076] In step 2, after removing the noise point cloud of the water surface environment noise, the water suspended matter noise and the reverberation reflection signal from the visualization content, the remaining GMM model point cloud is reconstructed, that is, the denoising preprocessing of the underwater sonar point cloud is completed, as shown in Figure 5 .

[0077] Example:

[0078] This example uses the BlueView BV5000 underwater sonar as an example to illustrate how to use this method to denoise underwater point cloud data.

[0079] Underwater 3D sonar measurement and point cloud data collection were carried out at the port terminal. The main underwater scene at the bottom of the port includes the seabed and a row of dock support columns. The raw data of underwater 3D sonar is as follows: Figure 2 As shown. Figure 2 It can be clearly found that underwater sonar 3D point cloud data will be affected by multiple noises such as environmental noise, suspended matter noise in water, reverberation reflection signals and self-mechanical noise.

[0080] Set the nearest neighbor parameter threshold = 10. Use the classic statistical outlier removal algorithm (SOR),

[0081] Take matlab software as an example: calling code

[0082] [ptCloudOut,inlierIndices,outlierIndices]=

[0083] pcdenoise(point cloud data variable name,'NumNeighbors',10);

[0084] outlierIndices variable: all discrete point clouds that do not meet the nearest neighbor parameter threshold = 10. This type of point cloud data mainly includes unreasonable noise point cloud data on the water surface and under the sea.

[0085] Delete unreasonable noise point cloud data on the water surface and under the seabed (such as Figure 3 ).from Figure 3 It can be found that the underwater sonar 3D point cloud data is only affected by environmental noise, water suspended matter noise and reverberation reflection signal noise.

[0086] The recursive adaptive Gaussian mixture model (GMM) is used to automatically mine the pattern characteristics of underwater objects and environmental noise, suspended noise in water, and reverberation reflection signals. The unsupervised expectation maximization (EM) algorithm can adaptively find different noise patterns in underwater sonar point clouds.

[0087] Taking Matlab software as an example: when setting the Gaussian mixture model parameters, if the GMM model parameter is set to 4, the code is called

[0088] gm = fitgmdist (point cloud data variable name, 4);

[0089] idx = cluster(gm,X');

[0090] The idx variable includes four Gaussian distribution point cloud patterns mined by the EM algorithm data.

[0091] In each recursive stage, all mined patterns of the GMM model are mapped to the real noise category. Figure 4 .

[0092] After deleting the noise point cloud of the water surface environment noise, the water suspended noise and the reverberation reflection signal, the remaining GMM model point cloud is reconstructed to complete the denoising preprocessing process of the underwater sonar point cloud (such as Figure 5 ).

[0093] According to the proposed method, multiple noise patterns can be mined: (a) self-mechanical noise, (b) surface environmental noise, and (c) suspended matter / reverberation reflection noise.

[0094] In the embodiment, the original data set has 1874338 point clouds. The traditional classic SOR method only removes 934 noise point clouds. The method proposed in the present invention can remove 396036 noise point clouds, which is better than the classic SOR method. The point cloud denoising performance comparison of the embodiment is shown in Table 1.

[0095] Table 1 Comparison of point cloud denoising performance of the embodiments

[0096]

[0097] The denoising method in this example mainly includes:

[0098] (A) The statistical outlier removal algorithm (SOR) deletes unreasonable random discrete point cloud data (on the water surface and under the seabed).

[0099] For example, the original underwater sonar measurement data contains self-mechanical noise point clouds. This type of noise pattern is mainly a randomly distributed discrete point cloud, and it appears on the water surface and under the seabed (irrational underwater point cloud data). The randomly distributed discrete noise point cloud has Gaussian distribution parameters that are significantly different from those of underwater objects. The classic statistical outlier removal algorithm (SOR) can be used to delete such unreasonable point cloud data (on the water surface and under the seabed). The statistical outlier removal algorithm can be used

[0100] (1) According to the Gaussian distribution statistics of the original data, the Gaussian distribution parameter threshold is set.

[0101] (2) Calculate the number of nearest neighbors of the original data and set the nearest neighbor parameter threshold.

[0102] (ii) Use a recursive adaptive Gaussian mixture model (GMM) to automatically mine the noise pattern characteristics of underwater objects and environmental noise, underwater biological noise, and reverberation reflection signals.

[0103] For example, when the raw underwater sonar measurement data contains noise point clouds of environmental noise, underwater biological / suspended matter noise, and reverberation reflection signals, such noise patterns are similar to underwater objects and have non-directional density gradients. The biggest difference from underwater objects is that the noise point cloud is a non-structured smooth point cloud voxel. Compared with randomly distributed discrete noise point clouds, such noise point clouds have a larger density and cannot be deleted using the classic SOR algorithm.

[0104] The Gaussian mixture model is composed of a mixture of different probability distribution models. Each component of the model has an independent Gaussian distribution to represent a subset of the data. Because the GMM model is data-oriented, given the original data set, without reference labels, the unsupervised expectation maximization (EM) algorithm can be used directly to adaptively find the noise pattern of the underwater sonar point cloud. In order to mine a sufficiently fine noise model, the pattern mining of the recursive architecture can be further used. In each recursive stage, the EM algorithm estimates the parameters of the noise component and groups them. Finally, all mined GMM patterns are mapped to the real noise category. After deleting the noise point cloud of environmental noise, underwater biological noise, and reverberation reflection signal, the remaining GMM model point cloud is reconstructed to complete the data denoising preprocessing process of the underwater sonar point cloud.

Claims

1. An underwater sonar 3D point cloud denoising method based on pattern mining, characterized by: The denoising method is aimed at processing the visualization content of the 3D sonar raw data of the underwater sonar, and comprises the following steps: Step 1: Based on the common patterns of sonar self-mechanical noise, the statistical outlier removal algorithm SOR is used to delete unreasonable random discrete point cloud data. Step 2: Use the recursive adaptive Gaussian mixture model (GMM) architecture to automatically mine the pattern features of underwater objects and environmental noise, underwater biological noise, and reverberation reflection signals. In each recursive stage of the recursive adaptive Gaussian mixture model, the adaptive model mines the pattern features and determines whether the denoising process is completed.

2. The method for denoising underwater sonar 3D point cloud based on pattern mining according to claim 1, characterized in that: The random discrete point cloud data includes the self-mechanical noise point cloud existing in the original measurement data of the underwater sonar, and also includes the discrete noise point cloud.

3. The method for denoising underwater sonar 3D point cloud based on pattern mining according to claim 2, characterized in that: The main body of the noise pattern of the self-mechanical noise point cloud is a randomly distributed discrete point cloud, and the distribution area appears on the water surface and under the seabed.

4. The method for denoising underwater sonar 3D point cloud based on pattern mining according to claim 3 is characterized in that: The discrete noise point cloud is randomly distributed and has Gaussian distribution parameters that are significantly different from those of underwater objects.

5. The method for denoising underwater sonar 3D point cloud based on pattern mining according to claim 4, characterized in that: In step 1, the method of using the statistical outlier removal algorithm includes: Method 1: According to the Gaussian distribution statistics of the original data, the Gaussian distribution parameter threshold is set; Method 2: Calculate the number of nearest neighbors of the original data and set the nearest neighbor parameter threshold; In step 2, the Gaussian mixture model is composed of a mixture of different probability distribution models, and each component of the model has an independent Gaussian distribution to represent a subset of the data; The GMM model is data-oriented at runtime. It directly inputs the original data set without reference labels. At the same time, it adaptively finds the noise pattern of the underwater sonar point cloud by directly using the unsupervised expectation maximization algorithm EM, and uses the pattern mining of the recursive architecture to mine a sufficiently detailed noise model. In each recursive stage, the EM algorithm estimates the noise component parameters and groups them. Finally, all mined GMM patterns are mapped to real noise categories.

6. The method for denoising underwater sonar 3D point cloud based on pattern mining according to claim 5, characterized in that: When measuring underwater scenes in the method in step 1, the 3D sonar raw data of the underwater sonar is first visualized to form visualization content, and then it is determined from the visualization content whether the 3D point cloud data is affected by multiple noises such as environmental noise, suspended matter noise in water, reverberation reflection signal and self-mechanical noise. If it is determined that it is affected by the above multiple noises, the nearest neighbor parameter threshold is set in the software, and the statistical outlier removal algorithm SOR is used for processing.

7. The method for denoising underwater sonar 3D point cloud based on pattern mining according to claim 6, characterized in that: When the software uses matlab, if the value is set to 10 when setting the nearest neighbor parameter threshold, the relevant calling code is: [ptCloudOut,inlierIndices,outlierIndices]= pcdenoise(point cloud data variable name,'NumNeighbors',10); The outlierIndices variable is set to: all discrete point clouds that do not meet the nearest neighbor parameter threshold = 10. The sub-class point cloud data mainly includes unreasonable noise point cloud data on the water surface and under the seabed; The visualization content of the 3D sonar raw data is processed using Matlab to remove environmental noise, water suspended noise, reverberation reflection signals and self-mechanical noise.

8. The method for denoising underwater sonar 3D point cloud based on pattern mining according to claim 5, characterized in that: In step 2, the recursive adaptive Gaussian mixture model (GMM) architecture is used to automatically mine the pattern features of underwater objects and environmental noise, underwater biological noise, and reverberation reflection signals, and automatically visualize the patterns of surface environmental noise, underwater biological and suspended noise, and underwater sound wave reverberation reflection noise. The unsupervised expectation maximization algorithm (EM) of the recursive adaptive Gaussian mixture model is used to adaptively find different noise patterns of underwater sonar point clouds to automatically separate the noise set data.

9. The method for denoising underwater sonar 3D point cloud based on pattern mining according to claim 8, characterized in that: In step 2, Matlab software is used. When setting the Gaussian mixture model parameters in the software, if the GMM model parameter is set to 4, the calling code is gm=fitgmdist(point cloud data variable name, 4); idx = cluster(gm,X'); The idx variable includes four Gaussian distribution point cloud patterns mined by the EM algorithm data; In each recursive stage, all mined patterns of the GMM model are mapped to the actual noise categories.

10. The method for denoising underwater sonar 3D point cloud based on pattern mining according to claim 8, characterized in that: In step 2, after removing the noise point cloud of the water surface environment noise, suspended matter noise in the water and reverberation reflection signal from the visualization content, the remaining GMM model point cloud is reconstructed, thus completing the denoising preprocessing of the underwater sonar point cloud.

Images