Three-dimensional sonar point cloud target classification method and device based on continuous coherence

Abstract

The invention discloses a three-dimensional sonar point cloud target classification method and device based on continuous coherence. The method comprises the following steps of performing preprocessing including filtering, downsampling and noise reduction on acquired three-dimensional sonar point cloud data; clustering the preprocessed three-dimensional sonar point cloud data by adopting a clustering algorithm to obtain a point cloud cluster; extracting topological structure information of the point cloud cluster by adopting a continuous coherence method, representing the topological structure information with a persistence graph, calculating a Wasserstein distance of the persistence graph, and counting multi-feature fusion parameters including density, centroid and average scattering intensity of the point cloud cluster; and inputting the Wasserstein distance of the persistence graph and the multi-feature fusion parameter into a KNN model to predict and obtain a target classification result. According to the method, the generalization ability of KNN model prediction can be further improved, so that the model has higher adaptability and robustness in practical application, and the accuracy of three-dimensional sonar point cloud target classification is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional sonar detection, and in particular relates to a three-dimensional sonar point cloud target classification method and device based on continuous coherence. Background Art

[0002] With the development of marine resources, the demand for underwater detection has increased, and underwater target recognition has become a hot topic. Sonar, as a means of underwater acoustic detection, has received widespread attention. At present, related technologies in the field of sonar have been developed to a certain extent, but at the same time, there are still many difficulties and challenges in the research of this field.

[0003] Compared with two-dimensional images, the point cloud data obtained by three-dimensional imaging sonar has a large amount of information due to the increase in dimensions, and there are also differences in processing algorithms. At present, there are roughly three types of processing for three-dimensional imaging sonar data: two-dimensional image-based, voxel-based, and point cloud-based methods. The two-dimensional image-based method projects the three-dimensional point cloud onto a two-dimensional plane and processes it using mature image processing algorithms, but this method will lose some spatial information and affect recognition accuracy. The voxel-based method divides the three-dimensional space into multiple small cubes (voxels) and performs statistical analysis on the point cloud in each voxel. Although this method retains some spatial information, it has a large amount of calculation and is easily affected by noise. The point cloud-based method directly processes the original point cloud data, which can make full use of the spatial information and structural characteristics of the point cloud, but it requires high computing resources.

[0004] In the process of 3D point cloud processing, the following difficult issues need special attention:

[0005] (1) Small number of samples: Due to the high cost of underwater sonar equipment and the fact that data collection requires ships and technicians with professional operating skills, sonar data collection is long and costly. Therefore, the number of sonar samples actually available is often limited, and there is a lack of sufficient public data sets for algorithm verification and performance comparison.

[0006] (2) Target type imbalance: In real-world environments, the types of underwater targets are limited and are affected by different experimental environments. This leads to large differences in the number of samples of different target types in sonar data, and a small number of targets overall. This imbalance not only affects the generalization ability of the target recognition model, but also reduces the model’s recognition accuracy for rare targets.

[0007] (3) Low resolution: Compared with high-precision sensors such as lidar, the resolution of sonar images is low, and target features are difficult to clearly identify. The complexity of the underwater environment also makes the analysis of sound wave signals more difficult. Therefore, the imaging results of sonar images often have problems such as blur, high noise, and poor interpretability. These problems not only increase the difficulty of target recognition, but also reduce the accuracy and reliability of the recognition results.

[0008] Taking all the above factors into consideration, 3D imaging sonar technology has broad application prospects and important research value in the field of underwater target recognition. However, faced with challenges and difficulties such as small sample size, unbalanced target types and low resolution, researchers need to continuously explore new technical routes and solutions to promote the further development and application of 3D imaging sonar technology. Summary of the invention

[0009] In view of the above, the purpose of the present invention is to provide a three-dimensional sonar point cloud target classification method and device based on continuous coherence, which can speed up the processing speed through preprocessing and clustering, and significantly improve the accuracy of target classification by finely introducing the topological structure information and multi-feature fusion parameters of the three-dimensional sonar point cloud into the decision-making of target classification. Furthermore, the KNN model is trained by the K-fold cross-validation method to select the hyperparameters most suitable for the sample data, thereby further improving the generalization ability of the model, so that the model has stronger adaptability and robustness in practical applications.

[0010] In order to achieve the above-mentioned invention object, the technical solution provided by the present invention is as follows:

[0011] An embodiment of the present invention provides a three-dimensional sonar point cloud target classification method based on continuous coherence, comprising the following steps:

[0012] Preprocess the acquired 3D sonar point cloud data including filtering, downsampling and noise reduction;

[0013] The clustering algorithm is used to cluster the preprocessed 3D sonar point cloud data to obtain point cloud clusters;

[0014] The topological structure information of the point cloud cluster is extracted by the persistent homology method and represented by the persistence graph. The Wasserstein distance of the persistence graph is calculated, and the multi-feature fusion parameters of the point cloud cluster including density, centroid and average scattering intensity are counted.

[0015] The Wasserstein distance of the persistence graph and the multi-feature fusion parameters are input into the KNN model to predict the target classification result.

[0016] Specifically, the method of extracting the topological structure information of the point cloud cluster by using the persistent homology method and representing it by a persistence graph, and calculating the Wasserstein distance of the persistence graph, includes:

[0017] Encode the point clouds in the same point cloud cluster and initialize the filtering parameter ε;

[0018] The complex is formed by adding data points through a simplex, that is, when the diameter between two data points is less than 2ε, an edge is added between them to construct a VR complex;

[0019] For the filter parameter ε at different scales, the above VR complex construction process is repeated to obtain a series of nested subcomplexes of the simplicial complex. These subcomplexes constitute the filter flow, and the topological features in them are identified through calculation.

[0020] During the change of the filtering parameter ε, the generation and disappearance of topological features are recorded and represented in the form of a persistence graph;

[0021] The Wasserstein distance between the persistence graphs of different point clouds is calculated to evaluate the similarity of these point clouds in topological structure. The smaller the distance value, the higher the similarity of the two point clouds in topological structure.

[0022] Specifically, the Wasserstein distance of the persistence graph and the multi-feature fusion parameters are input into the KNN model to predict the target classification result, including:

[0023] The multi-feature fusion parameters of the point cloud cluster including density, center of mass and average scattering intensity are weighted and normalized with the Wasserstein distance of the persistence map before being input into the KNN model. In the KNN model, at least one neighboring point of the data point to be classified in the point cloud cluster is found and the category of the data point to be classified is determined according to the majority voting principle, and finally the target classification result of the point cloud cluster is obtained.

[0024] Specifically, the KNN model is trained using the K-fold cross-validation method based on the point cloud clusters obtained by clustering.

[0025] Specifically, the filtering in the preprocessing includes threshold filtering and maximum filtering. For the beam domain obtained by the beamforming process of the three-dimensional sonar point cloud data, the data with intensity less than the filtering threshold in the beam domain is removed by threshold filtering, and the data with the highest intensity in each beam direction in the beam domain is retained by maximum filtering.

[0026] Specifically, the downsampling in the preprocessing includes using a voxel method to divide the sonar detection space into a number of cubic units, each cubic unit as a voxel, selecting representative points of the three-dimensional sonar point cloud data in the voxel, and counting the intensity average of all data points in the voxel. The intensity average is used as the intensity of the representative point, and the representative points in all voxels are combined to form the downsampled point cloud data.

[0027] Specifically, the noise reduction in the preprocessing includes adopting a regional self-growth algorithm based on Euclidean distance. For the three-dimensional sonar point cloud data obtained after filtering and downsampling, the point with the smallest curvature in the point cloud data is selected as the seed point, and the growth algorithm is executed with the seed point as the center for expansion. In the expansion process, the neighboring points of the seed point are searched. If the Euclidean distance from the neighboring point to the seed point is less than the set distance threshold, the neighboring point is divided into a valid point. For the remaining three-dimensional sonar point cloud data except the valid points, the seed point selection, expansion, search and division of valid points are continued to be iteratively executed, and finally all the valid points are saved as the preprocessed three-dimensional sonar point cloud data.

[0028] Specifically, in the expansion process, for the three-dimensional sonar point cloud data obtained after filtering and downsampling, the kd-tree structure is used to organize the point cloud data, and the point cloud data is divided into two sets as the left subtree and the right subtree respectively. The nearest neighbor search algorithm based on Euclidean distance is used to search for neighbor points based on the left subtree and the right subtree.

[0029] Specifically, the clustering algorithm includes using a k-means++ algorithm to cluster the pre-processed three-dimensional sonar point cloud data to obtain a number of point cloud clusters.

[0030] To achieve the above-mentioned purpose of the invention, an embodiment of the present invention further provides a three-dimensional sonar point cloud target classification device based on continuous coherence, comprising: a preprocessing module, a clustering module, a topological feature extraction module and a target classification module;

[0031] The preprocessing module is used to perform preprocessing including filtering, downsampling and noise reduction on the acquired three-dimensional sonar point cloud data;

[0032] The clustering module is used to cluster the pre-processed three-dimensional sonar point cloud data using a clustering algorithm to obtain point cloud clusters;

[0033] The topological feature extraction module is used to extract the topological structure information of the point cloud cluster by using the persistent homology method and represent it with a persistence graph, calculate the Wasserstein distance of the persistence graph, and simultaneously count the multi-feature fusion parameters of the point cloud cluster including density, centroid and average scattering intensity;

[0034] The target classification module is used to input the Wasserstein distance of the persistence graph and the multi-feature fusion parameters into the KNN model to predict and obtain the target classification result.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] (1) The present invention introduces the topological structure information of the point cloud into the decision-making of target classification, fully exploits the structural information of existing data, and improves the resolution of three-dimensional sonar point cloud data and the interpretability of target classification.

[0037] (2) The present invention comprehensively considers the density, center of mass and average scattering intensity information of the point cloud, and combines the Wasserstein distance weighted calculation between persistence graphs as the input of the KNN model, and can automatically retrain parameters at idle time, thereby effectively improving the classification accuracy of the model.

[0038] (3) The present invention reduces the impact of underwater noise echoes by filtering, downsampling and denoising the point cloud data in the preprocessing stage, and accelerates the processing speed through the kd-tree structure, which can further save platform computing power.

[0039] (4) The present invention adopts the K-fold cross-validation method to train the KNN model parameters, optimize the parameter selection process, and optimize for small sample scenarios. It can effectively deal with the problems of small number and insufficient diversity of underwater acoustic samples, and further improve the generalization ability and robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0041] Figure 1 It is a flowchart of a three-dimensional sonar point cloud target classification method based on continuous coherence provided by an embodiment of the present invention;

[0042] Figure 2 It is a schematic diagram of the data preprocessing process provided by an embodiment of the present invention;

[0043] Figure 3 It is a schematic diagram of a process of constructing a kd-tree provided by an embodiment of the present invention;

[0044] Figure 4 is a schematic diagram of a continuous synchronization process provided by an embodiment of the present invention;

[0045] Figure 5 is a schematic diagram of a simplex, complex, and persistence graph provided by an embodiment of the present invention;

[0046] Figure 6 is a schematic diagram of a complex growth process provided by an embodiment of the present invention;

[0047] Figure 7 It is a schematic diagram of the process of target classification of the KNN model provided by an embodiment of the present invention;

[0048] Figure 8Schematic diagram of a multi-feature fusion parameter construction process provided by an embodiment of the present invention;

[0049] Fig. 9 It is a structural schematic diagram of a three-dimensional sonar point cloud target classification device based on continuous coherence provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0051] The inventive concept of the present invention is: in view of the challenges and difficulties faced by the three-dimensional imaging sonar technology in the field of underwater target recognition in the prior art, such as small sample size, unbalanced target types and low resolution, the embodiment of the present invention provides a three-dimensional sonar point cloud target classification method and device based on continuous coherence, which can reduce the influence of underwater noise echo and speed up the processing speed by preprocessing and clustering the three-dimensional sonar point cloud data, and improve the accuracy of target classification by introducing the topological structure information and multi-feature fusion parameters of the three-dimensional sonar point cloud into the decision of KNN model target classification. Furthermore, the KNN model is trained by the K-fold cross-validation method, which not only effectively balances the number of samples of different target types, but also enhances the diversity and representativeness of the sample data, thereby further improving the generalization ability of the model, making the model more adaptable and robust in practical applications, and supporting the model to retrain parameters in idle time if the conditions are met during the actual deployment process to better adapt to different application scenarios.

[0052] Figure 1 is a flow chart of a method for classifying three-dimensional sonar point cloud targets based on continuous coherence provided by an embodiment of the present invention. Figure 1 As shown, the embodiment provides a three-dimensional sonar point cloud target classification method based on continuous coherence, comprising the following steps:

[0053] S1, preprocessing the acquired 3D sonar point cloud data including filtering, downsampling and noise reduction. The preprocessing process is as follows: Figure 2 shown.

[0054] S1.1, filtering. After the signal received by the three-dimensional sonar undergoes the beamforming process, the data intensity of the target in the beam domain is larger than the underwater environmental noise and is concentrated in a specific frequency band. In order to speed up the subsequent processing flow, a specific threshold is set for preliminary threshold filtering to remove data with an intensity less than μ times the maximum intensity in the beam domain data, thereby achieving the purpose of removing random environmental noise and reducing the amount of data. At the same time, considering that the signal intensity received by the sonar device has a certain range limitation in general, in order to avoid being affected by strong noise or strong interference in the environment and its side lobes, which may lead to a decrease in target observation effect, the data is filtered for maximum value, and the n points with the highest intensity in each beam direction are retained.

[0055] S1.2, downsampling processing. The point cloud data is downsampled using a voxel method. The process includes: dividing the sonar detection space into a number of cubic units with specific side lengths, each cubic unit as a voxel, selecting representative points of the three-dimensional sonar point cloud data in the voxel, and counting the intensity average of all data points in the voxel, using the intensity average as the intensity of the representative point, so that the size of the cube can intuitively reflect the point cloud downsampling ratio, and finally the representative points in all voxels are combined to form the downsampled point cloud data. In order to improve the running rate of the algorithm while minimizing information loss, the voxel size in the embodiment is set as an adjustable hyperparameter and optimized according to the running results of the subsequent algorithm.

[0056] S1.3, noise reduction. Considering that suspended matter in the water body will generate echo interference to form abnormal points, it is necessary to perform noise reduction on the point cloud. In the embodiment, a regional self-growth algorithm based on Euclidean distance is used for noise reduction. The steps are as follows:

[0057] S1.3.1, for the three-dimensional sonar point cloud data obtained after filtering and downsampling, select the point with the smallest curvature in the point cloud data as the seed point, and perform the growth algorithm with the seed point as the center for expansion;

[0058] S1.3.2, during the expansion process, search for neighboring points of the seed point. If the Euclidean distance from the neighboring point to the seed point is less than the set distance threshold, the neighboring point is classified as a valid point;

[0059] S1.3.3, for the remaining 3D sonar point cloud data except the valid points, continue to select the point with the smallest curvature as the seed point;

[0060] S1.3.4, repeat steps S1.3.2-S1.3.3 until the number of remaining points is less than a certain threshold, or the number of iterations reaches an upper limit, and save all valid points as preprocessed three-dimensional sonar point cloud data.

[0061] In the expansion process, in order to speed up the process of searching for neighboring points in step S1.3.2, the point cloud data is organized in a kd-tree structure. Kd-tree is a data structure that divides data points in k-dimensional space and is often used to implement nearest neighbor searches in large-scale high-dimensional data spaces.

[0062] The process of constructing a kd-tree is the process of generating a binary tree, such as Figure 3 As shown, specifically including:

[0063] (1) Select the coordinate axis i (i is any coordinate axis in the three-dimensional coordinates) to be divided. When the data variance on a certain coordinate axis is large, it means that the data points are more dispersed on this coordinate axis, and the segmentation effect is usually better. Therefore, the coordinate axis with the largest coordinate variance is selected by calculation for division to obtain the coordinate value on the coordinate axis i.

[0064] (2) Select a median point from the coordinate values ​​on the coordinate axis i, use the data point as the parent node of the divided subtree, and use the perpendicular axis between the median point and the coordinate axis i as the dividing plane to divide the three-dimensional space, dividing the point cloud data into two sets, which are respectively used as the left subtree and the right subtree;

[0065] (3) Repeat the above steps (1)-(2) for the left subtree and the right subtree obtained by the above segmentation until the space cannot be segmented. The data point at this time is regarded as a leaf node.

[0066] After completing the construction of the kd-tree, in step S1.3.2, the nearest neighbor search algorithm based on Euclidean distance is used to search the point cloud data. The steps are as follows:

[0067] S1.3.2.1, for the seed point to be searched, start from the root node of the kd-tree and compare them one by one, calculate the Euclidean distance between the seed point and the current node, if it is less than the set distance threshold ω, then visit the left subtree, otherwise visit the right subtree, and search for the nearest neighbor point as a valid point;

[0068] S1.3.2.2, save and update the current nearest neighbor point and nearest neighbor distance during the search process until the leaf node of the kd-tree is reached;

[0069] S1.3.2.3, backtrack to determine whether there are points with updated distances among other nodes on the search path. If so, add them to the search path and continue searching until the search path is empty.

[0070] After the above preprocessing, the point cloud data is more concise and has less noise interference, which helps to improve the running speed and accuracy of subsequent algorithms.

[0071] S2, clustering algorithm is used to cluster the preprocessed 3D sonar point cloud data to obtain point cloud clusters.

[0072] The k-means clustering algorithm has the characteristics of being able to minimize the distance between data points in the same category and being simple and fast. However, considering that the accuracy of the k-means algorithm is difficult to guarantee due to its characteristic of randomly selecting the initial center point of the cluster, it may produce bad iterative results. In order to reduce the potential impact of the random initial center on the clustering results, the embodiment uses the improved k-means++ algorithm to cluster the pre-processed three-dimensional sonar point cloud data to obtain a number of point cloud clusters. The characteristic of the improved k-means++ algorithm is that the selection of the first center point is also completely random, but the selection of the subsequent n+1th center point adopts D 2 The weighted method means that the farther the point is from the first n center points, the greater the probability of being selected as the n+1th center point. Under normal circumstances, the k-means++ algorithm can achieve stable results in all aspects.

[0073] S3 uses the persistent homology method to extract the topological structure information of the point cloud cluster and represents it with a persistence graph. The Wasserstein distance of the persistence graph is calculated, and the multi-feature fusion parameters of the point cloud cluster including density, center of mass and average scattering intensity are counted.

[0074] After clustering, point clouds belonging to the same category are considered as a whole underwater target. Generally speaking, the structure of underwater objects in fixed scenes is often relatively fixed and has persistent characteristics. Therefore, the structural information of sonar point clouds plays an important role in target classification. These structural information can be expressed by topological structures. Topology is a mathematical tool that studies the invariance of geometric figures or spaces in the process of continuous shape change. It focuses on expressing the geometric features and positional relationships of objects without considering deformation and size. Topological data analysis (TDA) characterizes the global characteristics of high-dimensional data by studying the shape characteristics of data. In TDA, persistent homology is an important analytical method. It focuses on certain topological features that appear or disappear with scale changes. It can give the number of structures such as groups, rings, and cavities in the data set, so it can characterize the overall picture of the data without losing dimensional information.

[0075] In the embodiment, the topological structure information of the point cloud cluster is extracted by using the continuous homology method and represented by a persistence graph. The continuous homology process is as follows: Figure 4 As shown, the specific steps are as follows:

[0076] S3.1, initialize the filter parameter ε = 0 and encode the point cloud;

[0077] S3.2, for a given filtering parameter ε, a complex is formed by adding data points to a simplex: According to the definition of VR complex, when the diameter between two data points is less than 2ε during the growth of the simplex, an edge is added between the point clouds of the data points to construct a VR complex. The definitions of simplex and complex and an example of persistence graph are shown in Figure 5 As shown in Figure 6 As shown;

[0078] S3.3, for the filter parameter ε at different scales, repeat the above VR complex construction process to obtain a series of nested subcomplexes of the simplicial complex, which constitute the filter flow. The topological features are identified by calculating the homology group at different scales;

[0079] S3.4, record the generation and disappearance of topological features during the change of the filtering parameter ε, and use the generation time of each topological feature as the horizontal axis and the disappearance time as the vertical axis to obtain a series of point pairs, which constitute the persistence graph of the homology group;

[0080] S3.5, the Wasserstein distance between the persistence graphs of different point clouds is calculated to evaluate the similarity of these point clouds in topological structure. The smaller the distance value, the higher the similarity of the two point clouds in topological structure.

[0081] In addition to topological features, the centroid, density and average scattering intensity information of point cloud clusters of different categories are simultaneously counted during the calculation process to form multi-feature fusion parameters, which are weighted and involved in decision-making in the KNN model.

[0082] S4, the Wasserstein distance of the persistence graph and the multi-feature fusion parameters are input into the KNN model to predict the target classification result.

[0083] In the embodiment, a KNN model is constructed based on an improved KNN (K-Nearest Neighbors) algorithm. The KNN algorithm is a basic classification and regression algorithm that is suitable for situations where the initial data samples are relatively insufficient, and can automatically optimize and update the model during actual application. In order to obtain better hyperparameters, the KNN model is trained using a K-fold cross-validation method based on the point cloud clusters obtained by clustering. The data set used in the embodiment is sourced from three-dimensional sonar point cloud data obtained from offshore experiments. Parameters are optimized through K-fold cross-validation. In the embodiment, K=5 is selected, that is, the data set is divided into 5 folds for cross-validation. At the same time, due to the limited training set, the generalization ability of the obtained model may be limited. With the increase of the test data set and the addition of prior information, the performance of the original model decreases. Therefore, in the embodiment, when specific conditions are met, the parameter model is automatically retrained in idle time to automatically adapt to more data and improve the generalization ability of the model.

[0084] In the embodiment, the process of using the KNN model to classify the target is as follows: Figure 7 As shown in Figure 2, the dataset is divided into five folds for cross-validation, and the multi-feature fusion parameters of the point cloud cluster including density, centroid and average scattering intensity and the Wasserstein distance of the persistence map are weighted and normalized before being input into the KNN model (as shown in Figure 2). Figure 8 In the KNN model, the k nearest neighbor points of the data points to be classified in the point cloud cluster are found and the category of the data points to be classified is determined according to the majority voting principle, and finally the target classification result of the point cloud cluster is obtained.

[0085] At this point, through the algorithm flow of point cloud data preprocessing, clustering, topological feature extraction, and target classification, the implementation plan designed by the present invention completes the target classification of three-dimensional sonar point cloud data, and the classification results obtained can be observed by users or applied to other data post-processing processes.

[0086] In summary, the method for classifying three-dimensional sonar point cloud targets based on persistent coherence provided by the embodiment of the present invention introduces the topological structure information of the point cloud into the decision-making of target classification, fully exploits the structural information of the existing data, and improves the resolution of the three-dimensional sonar point cloud data and the interpretability of target classification. The density, centroid and average scattering intensity information of the point cloud are comprehensively considered, and the Wasserstein distance weighted calculation between the persistence graphs is used as the input of the KNN model, and the parameters can be automatically retrained at idle time, thereby effectively improving the classification accuracy of the model. By filtering, downsampling and denoising the point cloud data in the preprocessing stage, the influence of underwater noise echoes is reduced, and the processing speed is accelerated by the kd-tree structure, which can further save the computing power of the platform. The K-fold cross-validation method is used to train the KNN model parameters, optimize the parameter selection process, and optimize for small sample scenarios, which can effectively deal with the problems of small number and insufficient diversity of underwater acoustic samples, and further improve the generalization ability and robustness of the model.

[0087] Based on the same inventive concept, Fig. 9 As shown, an embodiment of the present invention further provides a three-dimensional sonar point cloud target classification device 900 based on continuous coherence, including: a preprocessing module 910, a clustering module 920, a topological feature extraction module 930 and a target classification module 940.

[0088] The preprocessing module 910 is used to perform preprocessing including filtering, downsampling and noise reduction on the acquired three-dimensional sonar point cloud data.

[0089] The clustering module 920 is used to cluster the pre-processed three-dimensional sonar point cloud data using a clustering algorithm to obtain point cloud clusters.

[0090] The topological feature extraction module 930 is used to extract the topological structure information of the point cloud cluster using the persistent homology method and represent it with a persistence graph, calculate the Wasserstein distance of the persistence graph, and simultaneously count the multi-feature fusion parameters of the point cloud cluster including density, center of mass and average scattering intensity.

[0091] The target classification module 940 is used to input the Wasserstein distance of the persistence graph and the multi-feature fusion parameters into the KNN model to predict and obtain the target classification result.

[0092] It should be noted that the three-dimensional sonar point cloud target classification device based on continuous coherence provided in the above embodiment belongs to the same inventive concept as the three-dimensional sonar point cloud target classification method based on continuous coherence. The specific implementation process is detailed in the embodiment of the three-dimensional sonar point cloud target classification method based on continuous coherence, which will not be repeated here.

[0093] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A three-dimensional sonar point cloud target classification method based on continuous coherence, characterized in that: The following steps are involved: Preprocess the acquired 3D sonar point cloud data including filtering, downsampling and noise reduction; The clustering algorithm is used to cluster the preprocessed 3D sonar point cloud data to obtain point cloud clusters; The topological structure information of the point cloud cluster is extracted by the persistent homology method and represented by the persistence graph. The Wasserstein distance of the persistence graph is calculated, and the multi-feature fusion parameters of the point cloud cluster including density, centroid and average scattering intensity are counted. The Wasserstein distance of the persistence graph and the multi-feature fusion parameters are input into the KNN model to predict the target classification result.

2. The three-dimensional sonar point cloud target classification method based on continuous coherence according to claim 1 is characterized in that: The method of extracting the topological structure information of the point cloud cluster by using the persistent homology method and representing it by a persistence graph, and calculating the Wasserstein distance of the persistence graph, includes: Encode the point clouds in the same point cloud cluster and initialize the filtering parameter ε; The complex is formed by adding data points through a simplex, that is, when the diameter between two data points is less than 2ε, an edge is added between them to construct a VR complex; For the filter parameter ε at different scales, the above VR complex construction process is repeated to obtain a series of nested subcomplexes of the simplicial complex. These subcomplexes constitute the filter flow, and the topological features in them are identified through calculation. During the change of the filtering parameter ε, the generation and disappearance of topological features are recorded and represented in the form of a persistence graph; The Wasserstein distance between the persistence graphs of different point clouds is calculated to evaluate the similarity of these point clouds in topological structure. The smaller the distance value, the higher the similarity of the two point clouds in topological structure.

3. The three-dimensional sonar point cloud target classification method based on continuous coherence according to claim 1 is characterized in that: The Wasserstein distance of the persistence graph and the multi-feature fusion parameters are input into the KNN model to predict the target classification result, including: The multi-feature fusion parameters of the point cloud cluster including density, center of mass and average scattering intensity are weighted and normalized with the Wasserstein distance of the persistence map before being input into the KNN model. In the KNN model, at least one neighboring point of the data point to be classified in the point cloud cluster is found and the category of the data point to be classified is determined according to the majority voting principle, and finally the target classification result of the point cloud cluster is obtained.

4. The method for three-dimensional sonar point cloud target classification based on continuous coherence according to claim 3, characterized in that: The KNN model is trained using the K-fold cross-validation method based on the point cloud clusters obtained by clustering.

5. The three-dimensional sonar point cloud target classification method based on continuous coherence according to claim 1 is characterized in that: The filtering in the preprocessing includes threshold filtering and maximum filtering. For the beam domain obtained by the beam forming process of the three-dimensional sonar point cloud data, the data with intensity less than the filtering threshold in the beam domain is removed by threshold filtering, and the data with the highest intensity in each beam direction in the beam domain is retained by maximum filtering.

6. The method for three-dimensional sonar point cloud target classification based on continuous coherence according to claim 1, characterized in that: The downsampling in the preprocessing includes using voxel method to divide the sonar detection space into several cubic units, each cubic unit is used as a voxel, representative points of the three-dimensional sonar point cloud data in the voxel are selected, and the intensity average of all data points in the voxel is counted. The intensity average is used as the intensity of the representative point, and the representative points in all voxels are combined to form the downsampled point cloud data.

7. The method for three-dimensional sonar point cloud target classification based on continuous coherence according to claim 1, characterized in that: The noise reduction in the preprocessing includes the use of a regional self-growth algorithm based on Euclidean distance. For the three-dimensional sonar point cloud data obtained after filtering and downsampling, the point with the smallest curvature in the point cloud data is selected as the seed point, and the growth algorithm is executed with the seed point as the center for expansion. During the expansion process, the neighboring points of the seed point are searched. If the Euclidean distance from the neighboring point to the seed point is less than the set distance threshold, the neighboring point is divided into a valid point. For the remaining three-dimensional sonar point cloud data except the valid points, the seed point selection, expansion, search and division of valid points are continued to be iteratively executed, and finally all the valid points are saved as the preprocessed three-dimensional sonar point cloud data.

8. The method for three-dimensional sonar point cloud target classification based on continuous coherence according to claim 7, characterized in that: During the expansion process, for the three-dimensional sonar point cloud data obtained after filtering and downsampling, the kd-tree structure is used to organize the point cloud data, and the point cloud data is divided into two sets as the left subtree and the right subtree respectively. The nearest neighbor search algorithm based on Euclidean distance is used to search for neighbor points based on the left subtree and the right subtree.

9. The method for three-dimensional sonar point cloud target classification based on continuous coherence according to claim 1, characterized in that: The clustering algorithm includes using the k-means++ algorithm to cluster the preprocessed three-dimensional sonar point cloud data to obtain several point cloud clusters.

10. A three-dimensional sonar point cloud target classification device based on continuous coherence, characterized in that: include: Preprocessing module, clustering module, topological feature extraction module and target classification module; The preprocessing module is used to perform preprocessing including filtering, downsampling and noise reduction on the acquired three-dimensional sonar point cloud data; The clustering module is used to cluster the pre-processed three-dimensional sonar point cloud data using a clustering algorithm to obtain point cloud clusters; The topological feature extraction module is used to extract the topological structure information of the point cloud cluster by using the persistent homology method and represent it with a persistence graph, calculate the Wasserstein distance of the persistence graph, and simultaneously count the multi-feature fusion parameters of the point cloud cluster including density, centroid and average scattering intensity; The target classification module is used to input the Wasserstein distance of the persistence graph and the multi-feature fusion parameters into the KNN model to predict and obtain the target classification result.

Images