A training data set construction method for underwater multi-beam sonar point cloud terrain matching
Patent Information
- Application Number
- CN202611068954.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-17
- Publication Date
- 2026-09-11
AI Technical Summary
本发明的目的在于解决水下多波束声纳点云地形匹配模型训练过程中存在的任务适配训练样本不足、真实位姿标签难以获取、通用点云数据集与水下地形数据分布不一致等技术问题
1、本发明基于任务区域的海底地形网格基准图自动生成源点云与目标点云样本对,并通过模拟多波束声纳条带扫描过程生成源点云;相比通用三维物体点云数据集,所构建的数据集兼具连续水下地形表面特征与多波束测深条带的空间采样特征,在数据分布和几何形态上更贴近航行器实际搭载多波束声纳获取的数据,有利于降低训练数据与实际作业数据之间的差异。
Smart Images

Figure CN122731679A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of navigation, guidance and control technology, and particularly relates to underwater terrain-assisted navigation, three-dimensional point cloud registration and deep learning training data construction technology. Specifically, it relates to a method for constructing a training dataset for underwater multibeam sonar point cloud terrain matching. Background Technology
[0002] Underwater vehicles are widely used in marine scientific research, marine resource exploration, seabed topography mapping, and marine engineering inspection. Because electromagnetic waves attenuate significantly underwater, global navigation satellite systems (GNSS) are difficult to use directly underwater. Inertial navigation systems (INS) are typically used as the primary positioning method for autonomous navigation of underwater vehicles. INS offer advantages such as strong autonomy, good stealth, and continuous output; however, their positioning errors accumulate and diverge over time, requiring correction using other navigation information.
[0003] Underwater terrain matching positioning is an effective auxiliary inertial navigation method. This method utilizes a pre-acquired seabed topographic reference map, along with real-time depth or multibeam sonar topographic information measured by the vehicle during the mission. A terrain matching algorithm is used to estimate the vehicle's position, thereby correcting the inertial navigation output. Compared to single-beam bathymetry data, multibeam sonar can acquire bathymetry strips with a larger lateral coverage during navigation, forming underwater terrain point clouds with spatial geometric structures, providing richer observational information for 3D point cloud registration and terrain matching positioning.
[0004] In existing technologies, deep learning-based point cloud registration methods require a large number of source and target point cloud sample pairs with relative pose ground truth labels for supervised training. However, existing publicly available point cloud registration training datasets are mostly derived from indoor objects, general 3D models, or mechanical parts point clouds, which typically have clear object boundaries and complete 3D topological structures, significantly different from the geometric distribution of the continuous seabed surface underwater. Models trained directly using such general point cloud datasets present at least the following technical problems when applied to underwater multibeam terrain matching tasks:
[0005] 1. The learned feature priors are not well matched with the underwater topographic spatial pattern, making it difficult to effectively extract the geometric features of the continuous seabed surface; 2. The model is not robust enough to interference factors such as sparse features, repetitive textures, and local similarity in underwater terrain scenes; 3. The model cannot be adapted to the specific number of beams, strip scanning mechanism, inertial navigation coarse positioning error and depth sounding noise characteristics of the multi-beam sonar equipment carried by the actual aircraft.
[0006] Furthermore, acquiring large-scale underwater multibeam sonar point cloud training data under real sea trial conditions is costly, and accurate relative pose ground truth labels are often difficult to obtain from real measurement data. If the model is trained solely based on manual annotation or a small amount of measured data, not only is the data scale insufficient, but it also fails to cover actual operational factors such as different terrain features, inertial navigation errors, heading angle errors, and sonar bathymetry noise. This results in trained models that are prone to insufficient generalization ability and poor adaptability to actual mission areas.
[0007] Therefore, a training dataset construction method for underwater multibeam sonar point cloud terrain matching tasks is needed. This method should be able to automatically generate source point cloud-target point cloud training samples with registration correction ground truth labels based on the seabed topographic grid reference map of the task area, combined with the multibeam sonar scanning mechanism, inertial navigation coarse positioning error and depth sounding noise. Summary of the Invention
[0008] To address the aforementioned problems, this invention provides a method for constructing a training dataset for underwater multibeam sonar point cloud terrain matching. The purpose of this invention is to resolve technical issues encountered during the training of underwater multibeam sonar point cloud terrain matching models, such as insufficient task-adaptive training samples, difficulty in obtaining true pose labels, and inconsistencies between the distribution of general point cloud datasets and underwater terrain data.
[0009] A method for constructing a training dataset for underwater multibeam sonar point cloud terrain matching includes the following steps: Step S1: Obtain the seabed topographic grid baseline map of the target area and configure the dataset construction parameters; Step S2: Extract the target point cloud from the seabed topographic grid reference map; Step S3: Simulate the multibeam sonar strip scanning process to generate the original source point cloud corresponding to the target point cloud; Step S4: Apply a perturbation transformation, which includes horizontal translation error and heading angle error, to the original source point cloud to obtain a perturbation source point cloud. The perturbation transformation adopts a rotation strategy around the centroid of the original source point cloud. Step S5: Inject random sounding noise proportional to the depth value into the depth coordinates of each point in the disturbance source point cloud; Step S6: Generate registration correction truth labels based on the inverse correction relationship of the perturbation transformation; Step S7: Associate and store the perturbation source point cloud with injected random depth sounding noise, the target point cloud, the perturbation transformation parameters, and the registration correction ground truth label to form a training sample.
[0010] Preferably, the method for extracting the target point cloud in step S2 includes: Randomly select a center point from the seabed topographic grid reference map, extract a neighborhood grid of a preset size with the center point as the center as the target topographic map block, and map the grid index of the effective depth points in the neighborhood grid to the geographic coordinate system to obtain the target point cloud.
[0011] Preferably, the simulated multibeam sonar strip scanning process in step S3 includes: Set the track length and the number of single strip beams, move along a random heading angle with a preset step size, and obtain multiple sounding points in the vertical heading direction through grid linear interpolation each time to construct a strip point cloud that conforms to the multi-beam sampling principle as the original source point cloud.
[0012] Preferably, the rotation strategy around the centroid in step S4 includes: Calculate the centroid of the original source point cloud; the horizontal translation error follows a uniform distribution along the two horizontal coordinate axes; the heading angle error... The rotation angle about the vertical axis and follows a uniform distribution; based on the heading angle error Construct rotation matrix By combining the horizontal translation error and the centroid, a rigid body perturbation is applied to the original source point cloud to obtain the perturbed source point cloud. :
[0013] in, Represents the source point cloud. Represents source point cloud The center of mass, This represents the horizontal translation error vector; .
[0014] Preferably, the random depth sounding noise in step S5 follows a Gaussian distribution with a mean of zero and a variance proportional to the square of the depth value, and the noise amplitude increases with the increase of the depth value.
[0015] Preferably, the registration correction truth label in step S6 is the inverse transform of the perturbation transform.
[0016] Preferably, the training samples in step S7 are also associated with metadata, which includes at least one of the following: sample center point coordinates, candidate map tile size, number of multibeams, strip length, heading angle, horizontal translation error, heading angle error, depth sounding noise ratio, and topographic map index information.
[0017] Preferably, the dataset construction parameters include the total number of training samples, the number of multibeams, the candidate map tile size, the maximum planar translation error, the maximum heading angle error, the measurement noise ratio, and the strip length. Each parameter is configured according to the characteristics of the mission water area, the parameters of the multibeam sonar equipment, and the accuracy of the inertial navigation system.
[0018] Preferably, the training dataset is used for supervised training of an underwater multibeam sonar point cloud terrain matching model. The supervised training includes: inputting the disturbance source point cloud and the target point cloud into the point cloud terrain matching model to be trained, predicting the registration correction transformation, calculating the loss and updating the model parameters using the registration correction ground truth label as the supervision signal.
[0019] A system for constructing a training dataset for underwater multibeam sonar point cloud terrain matching includes: The baseline map acquisition module is used to acquire a baseline map of the seabed topography grid for the target area; The parameter configuration module is used to configure the dataset construction parameters; The target point cloud extraction module is used to extract target point clouds from the seabed topographic grid reference map; The source point cloud generation module is used to simulate the multibeam sonar strip scanning process and generate the original source point cloud corresponding to the target point cloud. An error perturbation module is used to apply a perturbation transformation, which includes horizontal translation error and heading angle error, to the original source point cloud to obtain a perturbation source point cloud. The perturbation transformation adopts a rotation strategy around the centroid of the original source point cloud. A depth sounding noise injection module is used to inject random depth sounding noise proportional to the depth value into the depth coordinates of each point in the disturbance source point cloud; The tag generation module is used to generate registration correction truth labels based on the inverse correction relationship of the perturbation transformation. The dataset storage module is used to associate and store the disturbance source point cloud, the target point cloud, and the registration correction ground truth label to form training samples.
[0020] The present invention has the following beneficial effects: 1. This invention automatically generates source point cloud and target point cloud sample pairs based on the seabed topographic grid reference map of the mission area, and generates source point clouds by simulating the multibeam sonar strip scanning process. Compared with general three-dimensional object point cloud datasets, the constructed dataset has both continuous underwater topographic surface features and spatial sampling features of multibeam bathymetry strips. In terms of data distribution and geometry, it is closer to the data obtained by the actual multibeam sonar carried by the vehicle, which helps to reduce the difference between training data and actual operation data.
[0021] 2. This invention introduces horizontal translation error and heading angle error during the generation of source point cloud, and generates a disturbed source point cloud by applying rotational perturbation around the centroid of the source point cloud. It can automatically obtain the registration correction true value label, which avoids the problem of difficulty in accurately obtaining the pose true value in real sea trial data, and also avoids the problem of unrealistic overall offset and numerical error amplification when directly rotating around the origin of the projection coordinates.
[0022] 3. This invention injects random sounding noise proportional to the depth value of each point in the generated source point cloud to simulate the characteristics of multibeam sonar sounding error changing with the measurement depth. This makes the noise distribution of the training data closer to the actual measurement environment and improves the robustness of the point cloud terrain matching model to sonar measurement errors. At the same time, the dataset construction parameters can be flexibly set according to the task water area, the resolution of the topographic reference map, the accuracy of the inertial navigation system, and the multibeam sonar parameters, which is convenient for supervised training and pre-deployment verification of underwater multibeam point cloud terrain matching models. Attached Figure Description
[0023] Figure 1 This is a flowchart illustrating the overall process of constructing the training dataset and training the model in this invention. It is used to explain the overall process from inputting the seabed topographic grid reference map and generating training samples to supervised training of the point cloud topographic matching model.
[0024] Figure 2 This diagram illustrates the generation of target point cloud blocks and multibeam strip source point clouds, used to explain how to extract target point clouds from a seabed topographic grid reference map. Source point clouds are generated based on the multibeam sonar scanning process. The process.
[0025] Figure 3 A schematic diagram illustrating the generation of source point cloud error disturbance, depth sounding noise injection, and registration correction truth labels is provided to explain the process of generating source point cloud data. By applying heading angle error, horizontal translation error, and sounding noise, the disturbance source point cloud is obtained. And according to the perturbation transformation The reverse correction relation generates registration correction truth labels. The process.
[0026] Figure 4 This is an example image of the underwater multibeam sonar point cloud terrain matching training dataset constructed in this invention, used to illustrate the disturbance source point cloud, target point cloud, and their sample pairing results generated under different strip lengths, number of beams, translation errors, and heading angle errors. Detailed Implementation
[0027] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention.
[0028] This invention provides a method for constructing a training dataset for underwater multibeam sonar point cloud terrain matching. For example... Figure 1 As shown, this method takes a seabed topographic grid reference map as input and generates a training dataset of source point cloud-target point cloud pairs with registration-corrected ground truth labels through steps such as target point cloud patch extraction, multibeam strip source point cloud generation, coarse positioning error perturbation, bathymetry noise injection, registration correction ground truth label generation, and serialization storage. The constructed dataset can be used for supervised training of underwater multibeam sonar point cloud terrain matching models.
[0029] To clearly describe the various point clouds involved in this invention and their interrelationships, the reference topographic point cloud extracted from the seabed topographic grid reference map will be referred to as the target point cloud. The point cloud generated during the simulated multibeam sonar strip scanning process, without the application of coarse positioning errors, is denoted as the original source point cloud. The original source point cloud will be analyzed. The point cloud obtained after applying coarse positioning error perturbation and injecting depth sounding noise is denoted as the perturbation source point cloud. The perturbation transformation applied to the original source point cloud is denoted as... The correction transformation required to register the disturbance source point cloud to the target point cloud is denoted as . .
[0030] In underwater multibeam terrain matching and localization scenarios, target point clouds Considered as a reference point cloud provided by a seabed topographic reference map; original source point cloud Multibeam strip point clouds used to represent a real spatial correspondence with the target point cloud; perturbation source point clouds This is used to simulate the matching strip point cloud obtained by a multibeam sonar mounted on an aircraft under the influence of coarse positioning errors in inertial navigation and depth sounding noise. The registration correction relationship between the disturbance source point cloud and the target point cloud is defined by the truth label. Characterization.
[0031] A method for constructing a training dataset for underwater multibeam sonar point cloud terrain matching includes the following steps: Step 1: Obtain the seabed topographic grid reference map and configure the parameters.
[0032] In this step, an underwater topographic reference map of the target area is obtained. This topographic reference map can be constructed by fusing multi-source bathymetry data and serves as a reference for subsequent topographic matching and positioning. The seabed topographic grid reference map can be represented as a depth matrix. , and depth matrix Correspondingly establish a horizontal coordinate matrix and , is used to represent the horizontal coordinates of each grid point in the geographic coordinate system.
[0033] Before generating the dataset, the parameters for constructing the dataset are uniformly set based on the actual hardware parameters of the shipborne multibeam bathysound sonar (such as the number of strip beams, beam opening angle, bathysound error characteristics, etc.) and the accuracy level of the shipborne inertial navigation system, so that the training dataset is consistent with the characteristics of the real operational scenario. These parameters include: total_samples (total number of training samples), num_beams (number of multibeams), patch_size (candidate map patch size), max_trans_xy (maximum planar translation error), max_rot_z (maximum heading angle error), noise_ratio (measurement noise ratio), and strip_length (strip length).
[0034] In one implementation, total_samples can be set to 500 to 20000, num_beams can be determined according to the device parameters used, patch_size can be set to 50 to 100 grid points, max_trans_xy can be set to 0 to 1000m, max_rot_z can be set to 5° to 30°, noise_ratio can be set to 0.005, and strip_length can be set to 5 to 20.
[0035] Step 2: Extract target point cloud from seabed topographic grid reference map .
[0036] In this step, the original terrain grid matrix is used... Randomly select the center point ,in u and v These represent the row and column indices in the terrain grid.
[0037] With the center point Centered on, extract size is The neighborhood grid is used as the target terrain map tile, where S The size of the candidate map patch is `patch_size`. The grid indices of the effective depth points within this neighborhood are mapped to the geographic coordinate system to obtain the target point cloud. As shown in formula (1): (1) Formula (1) represents the center point Centered on the target, the extracted row and column index offsets do not exceed [a certain value]. Furthermore, grid points with valid depth values are mapped to geographic coordinates to obtain the target point cloud. .
[0038] Step 3: Simulate the multibeam sonar strip scanning process to generate source point clouds .
[0039] In this step, based on the terrain area selected in step 2, the strip scanning process of the multibeam sonar carried by the aircraft is simulated to generate a point cloud of the target. Corresponding multibeam strip source point cloud .like Figure 2 As shown, source point cloud This is a strip point cloud generated based on the multibeam sonar scanning process.
[0040] Specifically, the track length is set to Lstrip, the number of beams per strip is Nbeams, and the movement step size is... Along a random heading angle With step size The system moves Lstrip times, acquiring Nbeams depth sounding points each time via grid linear interpolation in the vertical heading direction, thereby constructing a strip point cloud conforming to the multibeam sampling principle. .
[0041] Source Cloud The coordinates of the central beam point in the model can be used to represent the position of the corresponding waypoint of the vehicle. In subsequent model applications, the trained point cloud terrain matching model can predict the registration correction transformation based on the disturbance source point cloud and the target point cloud, and apply this correction transformation to the horizontal coordinates of the central beam of the multi-beam strip, thereby obtaining the vehicle position matching and positioning correction result.
[0042] Step 4: Process the original source point cloud Apply perturbation transformation The source cloud of the disturbance was obtained. .
[0043] like Figure 3 As shown, this step is used to simulate the coarse positioning error in underwater multibeam measurement, which corresponds to the horizontal coordinate error of the track point output by the shipborne inertial navigation system.
[0044] Considering the large coordinate values in the projected coordinate system, directly rotating around the origin of the projected coordinate system would introduce additional large-scale translations that are inconsistent with local heading errors and may reduce the stability of numerical calculations. Therefore, this invention addresses the source point cloud... The applied rigid body transformation employs a rotation strategy around the centroid of the point cloud.
[0045] First, calculate the source point cloud. center of mass As shown in formula (2): (2) Where, mean represents the source point cloud The average of the three-dimensional coordinate components of all points is calculated separately.
[0046] Coarse positioning error consists of horizontal translation error and heading angle error (around) z The equation consists of two parts: the rotation error of the axis, which correspond to the translation vector and rotation matrix in rigid body transformation, respectively. The specific generation method is as follows: Translation error components in the horizontal direction (x, y axes) and They respectively follow a uniform distribution, as shown in formulas (3) and (4): (3) (4) The depth error (z-axis) is not introduced at this stage; it is added later in step 5 based on the measurement noise characteristics of the multibeam sonar. Translation vector As shown in formula (5): (5) Heading angle error (corresponding to the winding) z The rotation angle of the axis follows a uniform distribution, as shown in formula (6).
[0047] (6) Depend on Construction around z Rotation matrix of axis As shown in formula (7): (7) The rotation matrix Translation vector and the center of rotation The commonly determined overall rigid perturbation operation is called perturbation transformation. For source point clouds Apply around the center of mass The rigid body perturbation yields the perturbation source point cloud. As shown in formula (8): (8) Therefore, formula (8) can also be written as ,in .
[0048] Step 5: Process the point cloud with error sources Inject multibeam sounding noise.
[0049] like Figure 3 As shown, this step is used to simulate the depth sounding error of multibeam sonar. Specifically, it involves the error source point cloud obtained in step 4. Random noise is added to the depth coordinates of each point, and the noise amplitude is proportional to the depth value of that point.
[0050] Let the noise proportionality factor be... Point cloud with error source The Middle i The depth value of each point is Then its depth value after adding measurement noise As shown in formula (9): (9) in, Let represent the noise figure at the i-th point, which follows a mean of 0 and a variance of . The Gaussian distribution. After completing this step, the point cloud with error sources will be... Replace the depth values at each point with the corresponding values. .
[0051] Step 6: Generate registration correction truth labels based on the perturbation transformation.
[0052] For each training sample pair Step 4 is applied to the original source point cloud. perturbation transformation These are known quantities. The perturbation transformation is used to simulate coarse positioning errors in inertial navigation, including the heading angle rotation error applied around the centroid of the source point cloud and the horizontal translation error.
[0053] Since the source and target point clouds of the disturbance are in different initial poses, the model needs to learn a rigid body correction transformation to restore the source point cloud to the reference coordinate system of the target point cloud. Therefore, the disturbance transformation... The inverse transform is used as the registration correction truth label. ,Right now:
[0054] If the perturbation transformation is as shown in Equation (8), the above perturbation relationship is solved in reverse to obtain the transformation relationship that corrects the perturbation source point cloud back to its original spatial position: (10) in, Rotation matrix The transpose of the matrix. Since the rotation matrix is an orthogonal matrix, therefore The first part is used to offset the heading angle rotation error applied in step 4; the rest of the formula is used to offset the coordinate changes and horizontal translation errors caused by rotation around the center of mass.
[0055] Therefore, the registration correction truth label is determined according to formula (10). The truth label is used to supervise the underwater multibeam sonar point cloud terrain matching model in learning the rigid body pose correction relationship from the disturbance source point cloud to the reference coordinate system where the target point cloud is located.
[0056] It should be noted that the registration corrects the truth label. This is only used to correct the heading angle error and horizontal translation error applied in step 4. The random sounding noise injected in step 5 is not part of the rigid body pose transformation and does not participate in the process. Instead of performing calculations, it is retained as observation noise in the training samples to improve the robustness of the model under multibeam sonar sounding noise conditions.
[0057] Step 7: Serialize and store the training dataset.
[0058] In this step, the source point cloud will be disturbed. Target point cloud Perturbation transformation parameters Registration and correction of truth labels The corresponding metadata is then associated and stored to form a training sample. The metadata includes at least one of the following: sample center point coordinates, candidate map tile size, number of multibeams, strip length, heading angle, horizontal translation error, heading angle error, depth sounding noise ratio, and topographic map index information.
[0059] The perturbation source point cloud in this dataset It also includes simulated inertial navigation coarse positioning error and simulated sonar depth sounding noise, target point cloud. Based on seabed topographic reference map, registration and correction of truth labels. Transformation from known perturbation The inverse is determined. This dataset covers different terrain features and error distributions, providing a data foundation with real physical meaning for supervised learning of underwater multibeam sonar point cloud terrain matching models.
[0060] The dataset construction process of the present invention will be further described below with reference to a specific embodiment. It should be understood that the parameter values in the following embodiments are merely illustrative examples, and those skilled in the art can adjust them according to factors such as the characteristics of the mission water area, the specifications of the sonar equipment, and the accuracy of the inertial navigation system. All implementations using the technical solutions described in this invention fall within the protection scope of this invention.
[0061] In this embodiment, a seabed topographic reference map of a specific mission area is selected as the data source. Based on the target multibeam sonar equipment parameters, the candidate map patch size (patch_size) is set to 100, the measurement noise ratio (noise_ratio) is set to 0.005, and the maximum heading angle error (max_rot_z) is set to 15°. Depending on different experimental requirements, the number of multibeams (num_beams) can be set to 101 or 601, the strip length (strip_length) can be set to 5 or 10, and the maximum planar translation error (max_trans_xy) can be set to 200m or 400m.
[0062] When generating a single training sample, a center point (u, v) is first randomly selected from the terrain grid matrix, and a target point cloud with a size of S × S grid points centered on this center point is extracted according to formula (1). Then along a random heading angle With step size The Lstrip is moved L times, and Nbeams bathymetry points are obtained each time in the vertical heading direction through grid linear interpolation, resulting in a multibeam strip source point cloud corresponding to the target point cloud block. .
[0063] Subsequently, the source point cloud is calculated according to formula (2). center of mass ; Generate the horizontal translation error according to formulas (3) and (4) , And construct the translation vector according to formula (5). Generate heading angle error according to formula (6) And construct the winding according to formula (7) z Rotation matrix of axis ; Apply formula (8) to the source point cloud Applying a rigid body transformation around the center of mass yields the disturbance source point cloud. Then apply formula (9) to... The depth values at each point are injected with Gaussian noise that is proportional to the water depth.
[0064] Finally, the sample pairs Perturbation transformation parameters Registration and correction of truth labels And the metadata is serialized and stored.
[0065] In one implementation, the underwater multibeam sonar point cloud terrain matching model can be trained under supervised supervision using the aforementioned training dataset. The model training process can be performed in a CUDA-accelerated environment using a batch gradient descent strategy, specifically including the following steps: Step A: Read the training dataset.
[0066] The training dataset saved in step 7 is read using a custom data loading class. The data is read in parallel, and the sample order is randomly shuffled to ensure both sample randomness and training stability. One batch of training samples is read at a time to obtain the perturbation source point cloud. Target point cloud and the corresponding registration correction truth labels .
[0067] Step B: Input the sample pairs into the underwater multibeam sonar point cloud terrain matching model to be trained.
[0068] Disturb the source cloud With target point cloud Input the point cloud terrain matching model to be trained. The model performs a registration correction transformation based on the input samples to predict the perturbation source point cloud to the target point cloud. .in, This includes predicting corrected rotation parameters and predicting corrected translation parameters.
[0069] Step C: Calculate the supervised training loss.
[0070] Correct the truth label based on registration. The registration correction transformation predicted by the computational model With truth labels The error between the values is used as the supervised training loss. The supervised training loss may include rotation error loss and translation error loss, which are used to constrain the corrected rotation parameters and corrected translation parameters of the model output to be close to their corresponding true values.
[0071] Step D: Update the model parameters and save the training results.
[0072] Based on the supervised training loss obtained in step C, backpropagation and parameter updates are performed on the point cloud terrain matching model. During training, the model weights can be saved based on the validation set or test set loss to obtain an underwater multibeam sonar point cloud terrain matching model suitable for the target water area.
[0073] In one implementation, the model training can be completed using the Adam optimizer, a learning rate decay strategy, and a periodic checkpoint saving mechanism, but the specific optimizer type, learning rate parameters, and saving period do not constitute a limitation on the scope of protection of this invention.
[0074] Through the above-mentioned supervised training process, the underwater multibeam sonar point cloud terrain matching model can learn the terrain geometric correspondence between the disturbance source point cloud and the target point cloud, and predict the registration correction transformation parameters required for the disturbance source point cloud to be registered to the target point cloud, thus providing a model basis for underwater terrain matching and positioning of multibeam sonar.
[0075] It should be noted that the above model training process is used to illustrate the application of the training dataset constructed in this invention and does not constitute a limitation on the scope of protection of the dataset construction method of this invention. Any underwater point cloud terrain matching model that requires supervised training of source point cloud, target point cloud, and their registration correction ground truth labels can be trained using the dataset constructed in this invention.
[0076] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the present invention.
Claims
1. A method for constructing a training dataset for underwater multibeam sonar point cloud terrain matching, characterized in that, Includes the following steps: Step S1: Obtain the seabed topographic grid baseline map of the target area and configure the dataset construction parameters; Step S2: Extract the target point cloud from the seabed topographic grid reference map; Step S3: Simulate the multibeam sonar strip scanning process to generate the original source point cloud corresponding to the target point cloud; Step S4: Apply a perturbation transformation, which includes horizontal translation error and heading angle error, to the original source point cloud to obtain a perturbation source point cloud. The perturbation transformation adopts a rotation strategy around the centroid of the original source point cloud. Step S5: Inject random sounding noise proportional to the depth value into the depth coordinates of each point in the disturbance source point cloud; Step S6: Generate registration correction truth labels based on the inverse correction relationship of the perturbation transformation; Step S7: Associate and store the perturbation source point cloud with injected random depth sounding noise, the target point cloud, the perturbation transformation parameters, and the registration correction ground truth label to form a training sample.
2. The method according to claim 1, characterized in that, The method for extracting the target point cloud in step S2 includes: Randomly select a center point from the seabed topographic grid reference map, extract a neighborhood grid of a preset size with the center point as the center as the target topographic map block, and map the grid index of the effective depth points in the neighborhood grid to the geographic coordinate system to obtain the target point cloud.
3. The method according to claim 1, characterized in that, The simulated multibeam sonar strip scanning process described in step S3 includes: Set the track length and the number of single strip beams, move along a random heading angle with a preset step size, and obtain multiple sounding points in the vertical heading direction through grid linear interpolation each time to construct a strip point cloud that conforms to the multi-beam sampling principle as the original source point cloud.
4. The method according to claim 1, characterized in that, The rotation strategy around the center of mass described in step S4 includes: Calculate the centroid of the original source point cloud; the horizontal translation error follows a uniform distribution along the two horizontal coordinate axes; the heading angle error... The rotation angle about the vertical axis and follows a uniform distribution; based on the heading angle error Construct rotation matrix By combining the horizontal translation error and the centroid, a rigid body perturbation is applied to the original source point cloud to obtain the perturbed source point cloud. : in, Represents the source point cloud. Represents source point cloud The center of mass, This represents the horizontal translation error vector; 。 5. The method according to claim 1, characterized in that, The random sounding noise mentioned in step S5 follows a Gaussian distribution with a mean of zero and a variance proportional to the square of the depth value. The noise amplitude increases with the increase of the depth value.
6. The method according to claim 1, characterized in that, The registration correction truth label mentioned in step S6 is the inverse transform of the perturbation transform.
7. The method according to claim 1, characterized in that, The training samples mentioned in step S7 are also associated with stored metadata, which includes at least one of the following: sample center point coordinates, candidate map tile size, number of multibeams, strip length, heading angle, horizontal translation error, heading angle error, depth sounding noise ratio, and topographic map index information.
8. The method according to claim 1, characterized in that, The dataset construction parameters include the total number of training samples, the number of multibeams, the candidate map tile size, the maximum planar translation error, the maximum heading angle error, the measurement noise ratio, and the strip length. Each parameter is configured according to the characteristics of the mission water area, the parameters of the multibeam sonar equipment, and the accuracy of the inertial navigation system.
9. The method according to claim 1, characterized in that, The training dataset is used for supervised training of the underwater multibeam sonar point cloud terrain matching model. The supervised training includes: inputting the disturbance source point cloud and the target point cloud into the point cloud terrain matching model to be trained, predicting the registration correction transformation, calculating the loss and updating the model parameters using the registration correction ground truth label as the supervision signal.
10. A system for constructing a training dataset for underwater multibeam sonar point cloud terrain matching, characterized in that, include: The baseline map acquisition module is used to acquire a baseline map of the seabed topography grid for the target area; The parameter configuration module is used to configure the dataset construction parameters; The target point cloud extraction module is used to extract target point clouds from the seabed topographic grid reference map; The source point cloud generation module is used to simulate the multibeam sonar strip scanning process and generate the original source point cloud corresponding to the target point cloud. An error perturbation module is used to apply a perturbation transformation, which includes horizontal translation error and heading angle error, to the original source point cloud to obtain a perturbation source point cloud. The perturbation transformation adopts a rotation strategy around the centroid of the original source point cloud. A depth sounding noise injection module is used to inject random depth sounding noise proportional to the depth value into the depth coordinates of each point in the disturbance source point cloud; The label generation module is used to generate registration correction truth labels based on the inverse correction relationship of the perturbation transformation. The dataset storage module is used to associate and store the disturbance source point cloud, the target point cloud, and the registration correction ground truth label to form training samples.
