Method and system for integrated data preprocessing, training, and evaluation for detecting seafloor anomalies based on 3D seafloor point clouds acquired by a Multibeam Echo Sounder(MBES)
The method stabilizes seabed anomaly detection by normalizing MBES data and applying object visibility criteria, addressing challenges of varying depths and partial visibility to achieve accurate detection and classification.
Patent Information
- Application Number
- KR1020260181173
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2026-09-22
- Publication Date
- 2026-11-06
- Estimated Expiration
- 2046-09-22
AI Technical Summary
Existing methods for detecting seabed anomalies using multibeam echo sounder (MBES) data face challenges such as varying absolute depths, partial visibility, and inaccurate label representations, leading to unreliable object detection and classification.
A method and system that normalizes height relative to the seabed, corrects object labels, separates clusters, and applies object visibility criteria within tiles to generate stable detection results, using a single interconnected processing flow.
Enables reliable calculation of seabed anomaly location and type by aligning labels with actual shapes, reducing noise, and ensuring accurate detection and classification.
Smart Images

Figure R1020260181173_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to a method and system for detecting seabed anomalies, and more specifically, to a method and system for integrating data preprocessing, learning, and evaluation for detecting seabed anomalies based on a 3D seabed point cloud acquired by a multibeam echo sounder (MBES), which receives a 3D seabed point cloud acquired by a multibeam echo sounder (MBES) and a polygon or 3D box label of a seabed anomaly as input, and configures the normalization of height relative to the seabed, object label correction according to the actual point cloud, tile label selection based on object visibility, Target, Ignore, and Discard label processing, cluster label separation according to the point cloud distribution, and detection, post-processing, and evaluation based on object center distance into a single interconnected processing flow, thereby enabling the stable calculation of the location and type of the actual seabed anomaly. Background Technology
[0003] MBES (Multibeam Echo Sounder) is equipment that uses acoustics to measure the depth of the seabed over a wide range.
[0004] And by processing the sounding data acquired with MBES, a three-dimensional point cloud of the seabed can be formed, in other words, the seabed can be represented as a set of numerous three-dimensional points, so that not only the shape of the seabed but also the location and height of artificial reefs, blocks, etc. protruding or installed above the seabed can be represented.
[0005] General underwater object detection can be performed by converting sonar images or depth rasters into 2D images and then applying image classification, segmentation, and object detection models, or by directly inputting seabed point clouds into 3D object detection models used in autonomous driving, etc.
[0006] However, seafloor point clouds produced by MBES have data characteristics that differ from typical terrestrial point clouds in that absolute depth, local seafloor slope, scan density, partial visibility at tile boundaries, cluster labels where multiple objects are represented as a single area, and labels that are not currently detected but whose meaning must be preserved exist simultaneously.
[0007] Meanwhile, in seafloor point clouds, absolute depths vary depending on the survey area, the vertical range of raw labels may not match the actual point cloud, and a problem arises where only a portion of an object is visible during the process of tiling large-scale data. Additionally, training meaningful structures such as sunken ships or sparse classes as mere backgrounds may be disadvantageous for future detection expansion, and cluster labels, where multiple entities are represented as a single large polygon, are difficult to use directly for individual object detection.
[0008] The limitations of existing underwater object detection methods for detecting seabed anomalies are summarized as follows.
[0009] (1) Limitations of sonar image or depth raster-based methods
[0010] While 2D methods utilizing sonar images or depth rasters facilitate the application of existing image processing models, they reduce the actual 3D height of point clouds and the vertical structure of objects during the projection process. Since the distribution of point clouds can vary depending on the seafloor slope, absolute depth, and point density even for objects of the same shape, it is difficult to adequately reflect these differences using only 2D representation.
[0011] In addition, 2D masks or axis alignment boxes may not adequately represent the orientation of objects installed or submerged in arbitrary directions on the seabed, and a separate coordinate and height restoration process is required to utilize the detection results as actual 3D locations.
[0012] (2) Limitations of direct application of general-purpose 3D object detection models
[0013] 3D object detection models for autonomous driving generally assume fixed objects on the ground and a sensor-centered coordinate system. In contrast, in seabed point clouds, the absolute Z value varies significantly depending on the depth of the survey area, and the height of an object must be interpreted based on the local seabed at that location.
[0014] If the height or center point of the raw label does not match the actual point cloud, the training box may excessively include the area below the seafloor or fail to sufficiently include the object's summit, even for the same object. Therefore, it is difficult to reliably learn the actual shape of seafloor anomalies by simply inputting the raw point cloud and label directly into a general-purpose model.
[0015] (3) Limitations of the simple tiling method
[0016] When partitioning large-scale seafloor point clouds into fixed-size tiles, if an object crosses a tile boundary, complete object labels may be included even on tiles where only a portion of the actual object is visible. These partially visible labels introduce learning noise regarding object centers and features, and can increase redundant predictions between overlapping tiles.
[0017] Therefore, rather than a simple rule that determines only whether the object's center point is contained within the tile, it is necessary to quantitatively determine how much of the object is actually visible on that tile.
[0018] (4) Limitations of non-detection label and cluster label processing
[0019] If all labels that are not currently being detected are removed and treated as general background, the voice samples are trained to cover meaningful areas that can be expanded into future detection targets, such as sunken ships, rare artificial reefs, and linear structures. Conversely, if all non-target objects are integrated into the detection class, class definitions become unclear, and minority classes may cause training instability.
[0020] Furthermore, if a single cluster polygon contains multiple artificial reefs or blocks, converting it into a single large object box provides inaccurate centers and sizes to models aiming for individual object detection.
[0021] (5) Necessity of aligning post-processing and evaluation criteria according to operational purpose
[0022] Post-processing and evaluation of 3D object detection may utilize different criteria, such as box overlap or object center distance, depending on the applied model or dataset. Variations in box size and rotation angle may occur in seabed anomaly labels due to the survey time, point cloud density, and the operator's polygon creation method. This implies that object center position accuracy can be used as a primary criterion to suit the purposes of locating and counting seabed anomalies.
[0023] In particular, if the primary purpose is to verify the location of artificial reefs and blocks, count quantities, or manage seabed facilities, the error in the object's center coordinates can serve as a more direct operational indicator than the minute rotation angle or size of the box.
[0024] Based on the technical background described above, the applicant has proposed the present invention as a processing technology that converts MBES point clouds and polygon or box labels into three-dimensional training data suitable for detecting seabed anomalies, and links elements such as object visibility, label semantics, cluster structure, height relative to the seabed, and object center position to a training and evaluation flow.
[0025] To elaborate, the present invention proposes a technology that reconstructs training data and labels to fit the seabed point cloud, rather than relying on specific open-source AI models or simply applying other 3D detection models to the seabed, while ensuring that the processing criteria are consistently connected from training to post-processing and evaluation. Prior art literature
[0027] Korean Registered Patent No. 10-2425484 (Published July 27, 2022), “Method for processing data for 3D high-frequency geological exploration in near-shore waters”; Korean Published Patent No. 10-2025-0180121 (Published Dec. 31, 2025), “System for generating 3D data for underwater exploration using 2D sonar”; Korean Registered Patent No. 10-1571886 (Published November 25, 2015), “Bathability surveying method for verifying seabed topography in 3D”; Korean Registered Patent No. 10-2502838 (Published February 23, 2023), “System capable of acquiring depth information by verifying seabed topography in 3D” The problem to be solved
[0028] The present invention aims to provide a technology that can stably calculate the location and type of actual seabed anomalies by receiving a 3D seabed point cloud acquired by a multi-beam echo sounder (MBES) and a polygon or 3D box label of a seabed anomaly as input, and configuring the normalization of height relative to the seabed, object label correction according to the actual point cloud, tile label selection based on object visibility, Target·Ignore·Discard label processing, cluster label separation according to the point cloud distribution, and detection, post-processing, and evaluation based on object center distance into a single interconnected processing flow.
[0029] Furthermore, the present invention enables learning with the same relative height representation by normalizing point clouds in survey areas with different absolute depths based on the local seabed; allows the center and height of the labels to be closer to the actual shape of seabed anomalies by converting polygonal labels into directional boxes and correcting the vertical range with the actual point cloud; reduces noise where objects partially visible at boundaries are learned as complete positive labels by applying object visibility criteria within the tile; prevents incorrect learning against future detection targets or uncertain structures as backgrounds by distinguishing between Target, Ignore, and Discard, while allowing the actual background to be used normally for negative learning; reduces the influence of inaccurate large ground truth boxes by decomposing overclusters into individual detection units based on the uplift and spatial distribution of the point cloud and treating areas that cannot be decomposed as Ignore; generates detection results suitable for actual operational purposes, such as object count aggregation and location verification, by configuring post-processing and evaluation based on object center distance; allows the application of the same data preprocessing and evaluation pipeline to different 3D detection structures without being dependent on a specific detection model; and the minority class The aim is to provide technology that allows data to be preserved in Ignore mode and then switched to Target mode to expand to multi-class detection once sufficient data is secured. means of solving the problem
[0031] A method for integrating data preprocessing, learning, and evaluation for detecting seabed anomalies based on a 3D seabed point cloud acquired by a multi-beam echo sounder (MBES) according to an embodiment of the present invention comprises: (a) a step of inputting a 3D seabed point cloud having X, Y, Z coordinates acquired by the MBES and polygonal or 3D box-shaped object labels; (b) a step (b-1) in which, for the input data of step (a), processing is performed to standardize the names of object labels, convert polygonal objects into 3D box shapes, and correct the vertical range of objects to match the seabed surface of the actual point cloud and the top of the objects; (b-2) a step of reconstructing cluster labels among the 3D box-shaped object labels, in which a plurality of objects are included in the corresponding 3D box-shaped object, into a plurality of individual detection units based on the point cloud distribution; and (b-3) a step of dividing a 3D seabed point cloud classified as a large point cloud into a plurality of tiles having mutually overlapping regions and selecting tiles to be learned based on object visibility. The method may include a step (b-4) of normalizing the Z values of point clouds and objects to relative heights based on tiles or local seafloor surfaces based on the work of the above step (b-3); a step (b-5) of maintaining the meaning of the training data by classifying them into targets, ignore regions, and discards according to the learning purpose and label meaning; (c) a step of linking the results of preprocessing and label object processing through the above step (b) with center distance-based duplicate removal and performance evaluation; and (d) a step of restoring to coordinates in actual seafloor space and generating a final result.
[0032] And a system for integrating data preprocessing, learning, and evaluation for detecting seabed anomalies based on a 3D seabed point cloud acquired by a multi-beam echo sounder (MBES) according to an embodiment of the present invention comprises: an input unit into which a 3D seabed point cloud having X, Y, Z coordinates acquired by the MBES and polygonal or 3D box-shaped object labels are input; a vertical range (Z) correction module that performs multiple processing operations on the input data input to the input unit, wherein standardization of object label names, conversion of polygonal objects into 3D box shapes, and correction processing of the object's vertical range to match the actual point cloud seabed surface and object top; an overcluster separation module that reconstructs cluster labels containing multiple objects within the corresponding 3D box-shaped object labels into multiple individual detection units based on the point cloud distribution; a learning tile selection module that divides a 3D seabed point cloud classified as a large point cloud into multiple tiles having mutually overlapping areas and selects tiles for learning based on object visibility; and the learning tile It may include a data preprocessing unit comprising a Z-value normalization module that normalizes the Z-values of point clouds and objects to relative heights based on tiles or local seabed surfaces based on the work of a selection module, and a label classification processing module that maintains the meaning of training data by classifying them into targets, ignore regions, and discards according to the learning purpose and label meaning; a control unit that links the results of preprocessing and label object processing through the data preprocessing unit with center distance-based duplicate removal and performance evaluation; and a result output unit that restores to coordinates in actual seabed space and generates a final result. Effects of the invention
[0034] According to an embodiment of the present invention, by receiving a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) and a polygon or three-dimensional box label of a seabed anomaly as input, and configuring the normalization of height relative to the seabed, object label correction based on the actual point cloud, tile label selection based on object visibility, Target, Ignore, and Discard label processing, cluster label separation based on the point cloud distribution, and detection, post-processing, and evaluation based on object center distance into a single interconnected processing flow, the location and type of the actual seabed anomaly can be reliably calculated.
[0035] Furthermore, point clouds from survey areas with different absolute depths can be normalized based on the local seafloor to enable training with the same relative height representation; polygonal labels can be converted into directional boxes and the vertical range corrected using the actual point cloud to more closely match the center and height of the labels to the actual shape of seafloor anomalies; noise caused by objects partially visible at boundaries being trained as complete positive labels can be reduced by applying object visibility criteria within the tiles; distinguishing between Target, Ignore, and Discard prevents incorrect training on future detection targets or uncertain structures as backgrounds while allowing the actual background to be used normally for negative learning; overclustering can be decomposed into individual detection units based on the uplift and spatial distribution of the point cloud, and areas that cannot be decomposed are treated as Ignore to reduce the impact of inaccurate large ground truth boxes; post-processing and evaluation can be configured based on object center distance to generate detection results suitable for actual operational purposes such as object counting and location verification; the same data preprocessing and evaluation pipeline can be applied to different 3D detection structures without being dependent on a specific detection model; and the minority class is treated as Ignore Once sufficient data is secured after preservation, it becomes possible to provide technology that allows switching to Target and expanding to multi-class detection. Brief explanation of the drawing
[0037] FIG. 1 is a flowchart illustrating a method for integrating data preprocessing, learning, and evaluation for detecting seabed anomalies based on a 3D seabed point cloud acquired by a multi-beam echo sounder (MBES) according to an embodiment of the present invention. FIG. 2 is a flowchart illustrating the process in FIG. 1 where cluster labels are reconstructed into multiple individual detection units based on a point cloud distribution. Figure 3 is a flowchart illustrating the process of selecting training target tiles based on tile segmentation and object visibility of a large-scale 3D seafloor point cloud in Figure 1. FIG. 4 is a flowchart illustrating the process of a center-distance-based duplicate removal operation in FIG. 1 (Claim 1-8) FIG. 5 is a block diagram illustrating a system that integrates data preprocessing, learning, and evaluation for detecting seabed anomalies based on a 3D seabed point cloud acquired by a multi-beam echo sounder (MBES) according to an embodiment of the present invention. Specific details for implementing the invention
[0038] The following detailed descriptions relating to the present invention refer to the accompanying drawings, which are embodiments in which the present invention may be practiced and are illustrated as examples of such embodiments. These embodiments are described in detail to sufficiently enable those skilled in the art to practice the present invention. It should be understood that various embodiments of the present invention are different but need not be mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be implemented in other embodiments without departing from the spirit and scope of the present invention in relation to one embodiment. Furthermore, it should be understood that the location or arrangement of individual components within each described embodiment may be changed without departing from the spirit and scope of the present invention.
[0039] Accordingly, the following detailed description is not intended to be taken in a limiting sense, and the scope of the invention is limited only by the appended claims, including all equivalents to those claimed therein, provided they are properly described. Similar reference numerals in the drawings refer to the same or similar functions across various aspects.
[0040] The terms used in this invention have been selected based on currently widely used general terms, taking into account their functions within the invention; however, these terms may vary depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Additionally, in specific cases, terms have been arbitrarily selected by the applicant, and in such cases, their meanings will be described in detail in the relevant description of the invention. Therefore, the terms used in this invention should be defined not merely by their names, but based on their meanings and the overall content of the invention.
[0041] When a part of an invention is described as “comprising” a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Furthermore, terms such as “…part,” “…module,” etc., described in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware or software, or as a combination of hardware and software.
[0042] Referring to FIGS. 1 to 5, a method and system for integrated processing of data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) according to an embodiment of the present invention will be described.
[0043] First, with reference to FIGS. 1 to 4, a method for integrating data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) according to an embodiment of the present invention will be described.
[0044] FIG. 1 is a flowchart illustrating a method for integrating data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) according to one embodiment of the present invention.
[0045] As described above, a method for integrating data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) according to one embodiment of the present invention, wherein at step (S110), a three-dimensional seabed point cloud having X, Y, Z coordinates acquired by the MBES and a polygonal or three-dimensional box-shaped object label are input. Here, the three-dimensional seabed point cloud may include at least one of local coordinates and georeference information, the polygonal object label may include the position coordinates of a vertex, and the three-dimensional box-shaped object label may include center coordinates, length, width, and height.
[0046] To elaborate, spaces or detailed names in label names can be standardized before training. For example, different types of independent or clustered artificial reefs can be combined into Fish_Reef, and multiple block detailed classes can be combined into Block.
[0047] And steps (S121) to (S125) are performed sequentially and non-sequentially on the input data from the above step (S110).
[0048] First, in step (S121), standardization of object label names, conversion of polygonal objects into 3D box shapes, and correction processing of the object's vertical range to match the actual point cloud's seabed and object's top are performed.
[0049] To explain this step (S121) in more detail, for the polygon label, a rotation rectangle of the minimum area containing the object is calculated using vertices on a horizontal plane. The center, major side, minor side, and rotation angle of the selected rectangle can be used as the horizontal components of the object box.
[0050] This method can reflect the actual installation direction of the object better than an axis alignment box and is advantageous for reducing the inclusion of unnecessary surrounding seabed in the object area.
[0051] Then, a point cloud is extracted from the horizontal area of each object box, and the local seafloor height is calculated using the sub-distribution of the point cloud. The height of the object's crest can be calculated using the upper point or upper quantile value of the object's horizontal area, and if necessary, the point cloud from an area extending the object's outer boundary by a certain range can also be used.
[0052] The object height is the difference between the peak and the seafloor, and the vertical center coordinate can be set to the midpoint of the two heights. Accordingly, the height and center of the raw label are corrected to match the shape of the actual MBES point cloud.
[0053] If the number of valid points is excessively small or it is difficult to calculate the normal vertical range, the original label value can be retained or the object can be switched to Ignore.
[0054] To elaborate, the process of converting a polygonal object into a 3D box shape in step (S121) includes calculating a rotation rectangle with a minimum area containing the object using vertices on a horizontal plane, and then using the center, long side, short side, and rotation angle of the calculated rectangle as the horizontal components of the object box.
[0055] Additionally, the process of correcting the vertical range of the object in step (S121) to align with the actual point cloud's seabed and the object's top includes extracting a point cloud from the horizontal area of each object box, calculating the local seabed height using the sub-distribution of the point cloud, calculating the object's top height using the upper point or upper quantile value of the object's horizontal area, and, if necessary, using the point cloud from an area where the object's outer boundary is extended by a certain range, and setting the object height as the difference between the top and the seabed, while simultaneously setting the vertical center coordinate as the midpoint of the two heights.
[0056] In step (S122), among the 3D box-shaped object labels, cluster labels containing multiple objects within the 3D box-shaped object are reconstructed into multiple individual detection units based on a point cloud distribution.
[0057] To explain this step (S122) in more detail, if the length or width of an object box is greater than a certain standard, it may be considered that the area contains a cluster label with multiple objects. In this case, point clouds of the object area and surrounding free area are extracted, and only points that are uplifted above a certain height above the seabed are selected and clustered according to spatial density.
[0058] For each cluster, individual object boxes can be generated by calculating the directional box and vertical range again. If the generated box is appropriate as an individual detection unit, it can be used as a Target; if it is still overvalued or lacks valid points, it can be used as an Ignore; and if no valid cluster is generated, the original overvalued box can be preserved as an Ignore.
[0059] Referring to FIG. 2, the step (S122) may include: a step of selecting object boxes whose length or width is greater than a certain standard (S122a); a step of extracting point clouds of the object area and surrounding free area of the selected object boxes, and then selecting only points that are uplifted above a preset height above the seabed and clustering them according to spatial density (S122b); a step of generating individual object boxes by calculating a directional box and a vertical range again for each clustered cluster (S122c); and a step of using the generated object box as a Target if it is appropriate as an individual detection unit, using it as an Ignore if it is still excessive or lacks valid points, and preserving the original excessive box as an Ignore if no valid cluster is generated (S122d).
[0060] Referring again to FIG. 1, in step (S123), a 3D seabed point cloud classified as a large point cloud is divided into multiple tiles having mutually overlapping regions, and then the tiles to be learned are selected based on object visibility.
[0061] To explain this step (S123) in more detail, a large-scale seafloor point cloud is divided into multiple tiles that partially overlap each other. The X and Y coordinates are converted to local coordinates based on the center of each tile, and the original tile center or georeference information is preserved as metadata so that it can be restored to actual coordinates later.
[0062] Then, the intersection area between the horizontal region and the tile region of each object is calculated, and the ratio of the intersection area to the total area of the object is determined as object visibility. By using the corresponding tile as the training target label only when the object visibility is above a set threshold, the problem of tiles showing only a part of the object being trained as complete positive labels is reduced.
[0063] Referring to FIG. 3 for further explanation, step (S123) may include: a step (S123a) of dividing a 3D seabed point cloud classified as a large point cloud into a plurality of tiles having mutually overlapping areas; a step (S123b) of calculating the intersection area between the horizontal area of each object and the tile area and calculating the ratio of the intersection area to the total area of the object as object visibility; and a step (S123c) of using the tile as a learning target label only when the object visibility is greater than or equal to a preset standard.
[0064] Referring again to FIG. 1, in step (S124), based on the work of step (S123), the Z values of the point cloud and objects are normalized to relative heights relative to the tile or local seabed.
[0065] To explain this step (S124) in more detail, the seabed height can be calculated from the point group of each tile, and then normalized by subtracting the corresponding seabed height from the Z value of each point and the Z value of the object center. Accordingly, regardless of the absolute depth of the survey area, the seabed is expressed near a certain standard, and the uplift of the object is expressed as a relative height value.
[0066] This configuration enables data acquired at different depths to be trained using the same criteria, allowing the model to focus on relative shape and height from the seafloor rather than the absolute depth itself.
[0067] In step (S125), the meaning of the training data is maintained by classifying it into Target, Ignore, and Discard based on the learning purpose and label meaning.
[0068] To explain this step (S125) in more detail, labels can be classified into Target, Ignore, and Discard depending on the learning purpose and meaning.
[0069] ◎ Target: The object currently to be detected, used for class classification and local regression learning.
[0070] ◎ Ignore: Objects that are meaningful structures but are not currently used as detection targets, which can mask classification loss so that the area is not learned as a general background and can be optionally excluded from evaluation.
[0071] ◎ Discard: Items that should be trained as actual backgrounds rather than detected as objects; object labels are removed to treat them as normal backgrounds.
[0072] For example, Fish_Reef and Block can be used as Targets, sunken ships, rare classes, linear structures, etc., can be preserved as Ignore, and items that do not need to be detected as separate objects, such as rocky seabeds, can be processed as Discard.
[0073] Additionally, classes managed as Ignore can be converted to Target by updating class mapping and training information once sufficient data is secured in the future.
[0074] In step (S130), the results of preprocessing and label object processing through the aforementioned steps (S121) to (S125) constituting step (S120) are linked with center distance-based duplicate removal and performance evaluation.
[0075] To explain this step (S130) in more detail, data can be separated at the raw scan level so that tiles derived from the same raw scan are not simultaneously included in the training set and the validation set. This reduces the overestimation of performance caused by spatially adjacent portions of the same survey data being included in both training and validation.
[0076] In the case of multiple classes, the distribution by region or scan of each class can be separated so that it is appropriately reflected in the training and validation sets.
[0077] A 3D object detection model is trained using a normalized point cloud and a target object box, and the model can use various forms such as a center heatmap-based structure that directly predicts object centers, a sparse voxel-based structure, or a Transformer-based structure.
[0078] CenterPoint, VoxelNeXt, DSVT, etc. are applicable embodiments, and the present invention is not limited thereto.
[0079] During the training process, data augmentation such as left-right and front-back flipping, forward rotation, scaling, and object point cloud insertion can be applied. Additionally, Ignore areas can be masked so that the classification loss for those locations is not calculated.
[0080] Not only the intersection area of object boxes but also the accuracy of the object center location can be used for important detection purposes. During the training process, heatmaps and center offsets corresponding to the object center are predicted, and the importance of X and Y center coordinates can be set highly during position regression.
[0081] After inference, duplicate predictions can be eliminated based on the distance between prediction centers, depending on the operational purpose. Additionally, different confidence thresholds can be applied according to class-specific detection characteristics.
[0082] When evaluating, predictions are sorted by confidence, each prediction is connected to the nearest unmatched correct answer center, and precision and recall can be calculated based on whether the center distance is less than or equal to a set allowable distance. For example, the center distance-based average precision (CD-mAP) can be calculated using the average precision for allowable distances of 1m, 2m, and 3m.
[0083] Rather than presenting the center distance itself as a new evaluation criterion, the results of the preceding data preprocessing and labeling can be consistently linked through learning, deduplication, and performance evaluation to align with the operational objective of the actual positional accuracy of seabed anomalies.
[0084] Referring to FIG. 4 for further explanation, the center distance-based duplicate removal operation in step (S130) may include: a step (S131) of predicting a heatmap and center offset corresponding to the object center during the learning process and setting the importance of X,Y center coordinates high during position regression; a step (S132) of removing duplicate predictions based on the distance between prediction centers according to the operational purpose after inference; a step (S133) of applying different confidence thresholds according to the detection characteristics of each class; a step (S134) of sorting predictions in order of confidence during evaluation and connecting each prediction to the nearest unmatched correct answer center; and a step (S135) of calculating precision and recall based on whether the center distance is less than or equal to a preset allowable distance.
[0085] Referring again to Fig. 1, in step (S140), coordinates in actual seabed space are restored and the final result is generated.
[0086] To explain this step (S140) in more detail, the results detected in the tile area coordinates are restored to coordinates in actual seabed space by applying georeference information of the original tile center or scan. The final result may include object class, center coordinates, size, rotation angle, and detection reliability.
[0087] Table 1 below summarizes the main components and interrelationships of the above-described embodiment according to the present invention. Referring thereto, the components of the embodiment according to the present invention solve different problems, but these components are interconnected for the same purpose of converting MBES seabed point clouds into training data suitable for detecting seabed anomalies and stably detecting the center location of actual objects.
[0088] division composition Problems to solve Key processing details A Seafloor-based Z normalization Learning deviation due to differences in absolute depth by survey area Convert Z values of point clouds and objects to relative heights based on tiles or local seafloor B Actual point cloud-based label vertical range correction Discrepancy between the height and center of the raw label and the actual object shape Calculate the seafloor and object crest from the point cloud to correct the object height and vertical center. C Object Visibility-Based Nested Tile Label Screening The problem where objects partially visible at tile boundaries are trained as complete positive labels Select sufficiently visible objects as training targets using the intersection ratio of the object area and the tile area. D Target·Ignore·Discard label processing The problem of meaningful objects not currently detected being mistrained as general background Processing by distinguishing between detection targets, non-interest regions, and background based on learning objectives and label meanings. E Point cloud-based overcluster separation The problem where individual object detection is difficult because multiple objects are displayed under a single large label Reconstruct clusters into individual detection units using the spatial distribution of elevation point clouds, and preserve uncertain regions as Ignore. F Post-processing and evaluation tailored to the purpose of object centering Discrepancy between box shape-based criteria and operational objectives of location identification and quantity counting Linking the preceding preprocessing and label processing results with center-distance-based duplicate removal and performance evaluation
[0089] The technical connection relationships of the major components according to Table 1 above are summarized as follows.
[0090] ◎ A normalizes point groups and object heights based on the seabed to reduce differences in absolute depth between survey areas.
[0091] ◎ B corrects the vertical range of the raw label to match the seafloor and object crest of the actual point cloud.
[0092] ◎ C calculates object visibility during the tile partitioning process of a large point cloud to reduce learning noise of partially visible objects.
[0093] ◎ D maintains the meaning of the training data by distinguishing between the current detection target, the non-target whose meaning is to be preserved, and the actual background.
[0094] ◎ E reconstructs individual detection units using a point cloud distribution when multiple objects are marked with a single label.
[0095] ◎ F connects the results of the preceding processing to the post-processing and evaluation stage for the purpose of identifying the location and counting the quantity of seabed anomalies.
[0096] And the table below summarizes the differences between the above-described embodiment of the present invention and the prior art.
[0097]
[0098] As can be seen from the contents summarized in the table above, the main difference between an embodiment of the present invention and existing technologies lies in the fact that it links the entire process—from data generation to label management, training, post-processing, and evaluation—to the single objective of detecting actual object-centered locations, thereby addressing the problems occurring in MBES seabed point clouds. The following is a more summarized description of this.
[0099] ◎ Seafloor-based Z normalization and point cloud-based label correction reflect different depths and inaccurate label issues together.
[0100] ◎ Visibility-based tile selection controls the learning noise of partially visible objects generated during the large-scale point cloud segmentation process.
[0101] ◎ Target, Ignore, Discard, and cluster separation preserve and reconstruct the complex label structure of actual seabed survey data for learning purposes.
[0102] ◎ It is to consistently connect the results of preprocessing and labeling that precede the center distance standard itself with postprocessing and evaluation tailored to the purpose of verifying the location of objects and counting quantities.
[0103] The following describes an example to which an embodiment of the present invention is applied and adjustable parameters.
[0104] Table 2 below summarizes the parameters adjustable in the embodiment of the present invention and the process thereof, and may be changed depending on the type of object, MBES point density, survey area, tile size, and purpose of operation.
[0105] item Example of implementation note Tile size Approximately 50m x 50m Large-capacity point cloud partitioning unit Tile Overlapping About 10m Compensating for missing boundary objects and continuity Object Visibility Criteria About 50% or more Suppressing partial visible objects from learning complete targets Cluster separation application size Length or width of approximately 25m or more Determining overclustering candidates Based on the uplift point About 0.3m or more than the seafloor statistical value Separation of the seafloor and the object protrusion Clustering example DBSCAN radius approx. 2.0m, minimum number of points 8 Adjustable based on point cloud density Voxel size example About 0.2m Adjustable depending on model and dot density Example of applied model CenterPoint, VoxelNext, DSVT Not limited to a specific model
[0106] In the cluster separation process, points that are elevated above the seafloor within and around the object box can be selected, and density-based spatial clustering can be applied. For each generated cluster, the directional box and vertical range are recalculated, and clusters that are difficult to determine as individual objects can be kept as Ignore.
[0107] During the learning phase, horizontal and front-to-back flipping, forward rotation, scaling, and object point cloud insertion can be applied, and different object insertion amounts can be set for each class of prime numbers.
[0108] In the operational phase, different confidence thresholds can be applied by considering detection performance and false positive costs for each class, or predictions from models with superior performance for each class can be selected and combined.
[0109] The performance improvements confirmed through the above implementation examples are summarized as follows.
[0110] In an experiment where the object visibility criterion was strengthened from 20% to 50% under conditions where the training and validation sets were separated by scan, the center-distance-based mean aptitude (CD-mAP) improved from 65.68% to 76.07%. This corresponds to an improvement of 10.39%p.
[0111] In addition, the recall at 1m center distance reached 90.76%. This result demonstrates that center distance-based detection performance improved under conditions where object visibility criteria were strengthened, and confirms that a configuration controlling partial visibility labels on tile boundaries can contribute to improved detection performance.
[0112] In a multi-class embodiment that simultaneously detects Fish_Reef and Block, it was confirmed that over-clustering separation, Ignore masking, and class imbalance correction can be applied to the same training pipeline.
[0113] To elaborate, center-distance-based detection performance improved under conditions where object visibility criteria were strengthened, a result demonstrating that controlling partial visibility labels on tile boundaries can contribute to performance improvement.
[0114] As can be seen from the descriptions above, the following operations and effects can be realized by a method for integrating data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by an acoustic depth sounder (MBES) according to one embodiment of the present invention.
[0115] 1. Point groups of survey areas with different absolute depths can be normalized based on the local seabed to learn the same relative height representation.
[0116] 2. By converting polygonal labels into directional boxes and correcting the vertical range with actual point clouds, the center and height of the labels can be aligned more closely with the shape of the actual seafloor anomaly.
[0117] 3. By applying object visibility criteria within the tile, noise can be reduced where objects that are only partially visible at the boundary are learned as complete positive labels.
[0118] 4. By distinguishing between Target, Ignore, and Discard, it is possible to prevent incorrect learning against future detection targets or uncertain structures as a background, while allowing the actual background to be used normally for speech learning.
[0119] 5. Over-clustering can be broken down into individual detection units based on the elevation and spatial distribution of the point cloud, and areas that cannot be broken down can be treated as Ignore to reduce the impact of inaccurate large correct boxes.
[0120] 6. By configuring post-processing and evaluation based on object center distance, detection results suitable for actual operational purposes, such as object quantity aggregation and location verification, can be generated.
[0121] 7. The same data preprocessing and evaluation pipeline can be applied to different 3D detection structures without being dependent on a specific detection model.
[0122] 8. After preserving minority classes as Ignore, you can switch to Target and expand to multi-class detection once sufficient data is obtained.
[0124] Next, with reference to FIG. 5, a system for integrating data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by an acoustic depth sounder (MBES) according to an embodiment of the present invention will be described.
[0125] FIG. 5 is a block diagram illustrating a system that integrates data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) according to one embodiment of the present invention.
[0126] As described above, a system for integrating data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam acoustic sounder (MBES) according to one embodiment of the present invention comprises an input unit (100), a data preprocessing unit (200), a control unit (300), and a result output unit (400).
[0127] Before the explanation, the system for integrated processing of data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) according to the present embodiment corresponds to a system that performs the overall steps described with reference to FIGS. 1 to 4. Therefore, more specific configurations and operations not mentioned in the following description can be understood through the descriptions above with reference to FIGS. 1 to 4. In other words, the input unit (100), data preprocessing unit (200), control unit (300), and result output unit (400) according to the present embodiment are configurations that include the detailed configurations and operations described with reference to FIGS. 1 to 4.
[0128] The input unit (100) receives a three-dimensional seabed point cloud and a polygonal or three-dimensional box-shaped object label having X, Y, Z coordinates obtained from MBES.
[0129] The data preprocessing unit (200) performs a plurality of processing operations on input data input to the input unit (100), and the data preprocessing unit (200) includes a vertical range (Z) correction module (210), an overcluster separation module (220), a learning tile selection module (230), a Z-value normalization module (240), and a label separation processing module (250).
[0130] The vertical range (Z) correction module (210) performs the standardization of object label names, the conversion of polygonal objects into three-dimensional box shapes, and the correction of the object's vertical range to match the actual point cloud's seabed and the object's top.
[0131] The overcluster separation module (220) performs the task of reconstructing cluster labels among the 3D box-shaped object labels into multiple individual detection units based on a point cloud distribution, wherein the cluster labels containing multiple objects in the 3D box-shaped object are reconstructed.
[0132] The learning tile selection module (230) divides a 3D seabed point cloud classified as a large-scale point cloud into multiple tiles having mutually overlapping regions, and then performs the task of selecting tiles to be learned based on object visibility.
[0133] The Z-value normalization module (240) performs the task of normalizing the Z-values of point clouds and objects to relative heights based on tiles or local seabeds, based on the work of the learning tile selection module (230).
[0134] The label classification processing module (250) performs the task of maintaining the meaning of the training data by classifying it into a target, an area of non-interest (Ignore), and a background (Discard) according to the learning purpose and the meaning of the label.
[0135] The control unit (300) performs the function of linking the preprocessing and label object processing results through the data preprocessing unit (200) with center distance-based duplicate removal and performance evaluation.
[0136] The result output unit (400) performs the function of restoring to actual seabed space coordinates and generating and outputting the final result.
[0137] And since the operation of a system that integrates data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) according to one embodiment of the present invention having the above-described configuration is substantially similar to the operation of a method that integrates data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) described with reference to FIGS. 1 to 4, a detailed description thereof is omitted in this embodiment.
[0138] To elaborate, the operation of a system that integrates data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) according to an embodiment of the present invention can be understood through the operation of a method that integrates data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) according to an embodiment of the present invention.
[0139] In addition, for a more specific configuration of the input unit (100), data preprocessing unit (200), control unit (300), and result output unit (400) of a system that integrates data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) according to an embodiment of the present invention, and for a more specific configuration of the vertical range (Z) correction module (210), oversized cluster separation module (220), learning tile selection module (230), Z-value normalization module (240), and label classification processing module (250) constituting the data preprocessing unit (200), one can refer to the description of the corresponding configurations of the method that integrates data preprocessing, learning, and evaluation for detecting seabed anomalies based on a three-dimensional seabed point cloud acquired by a multi-beam echo sounder (MBES) described with reference to FIGS. 1 to 4.
[0141] As described above, although the present description has been explained with specific details such as specific components, limited embodiments, and drawings, this is provided merely to aid in a more comprehensive understanding of the invention, and the invention is not limited to the above embodiments. A person skilled in the art can make various modifications and variations from this description.
[0142] Therefore, the scope of the present invention should not be limited to the described embodiments, and all variations equivalent to or equivalent to the claims set forth below, as well as the claims themselves, shall be considered to fall within the scope of the present invention. Explanation of the symbols
[0144] 100 : Input section 200 : Data preprocessing section 210: Vertical Range (Z) Correction Module 220: Overcluster Separation Module 230: Training Tile Selection Module 240: Z-Value Normalization Module 250 : Label classification processing module 300 : Control unit 400 : Result output section
Claims
Claim 1 (a) a step in which 3D seafloor point clouds and polygonal or 3D box-shaped object labels having X,Y,Z coordinates acquired from MBES are input; (b) a step (b-1) in which, for the input data of step (a), processing is performed to standardize the names of object labels, convert polygonal objects into 3D box shapes, and correct the vertical range of objects to match the actual point cloud seafloor and object tops; (b-2) a step in which cluster labels containing multiple objects within the corresponding 3D box-shaped object labels are reconstructed into multiple individual detection units based on the point cloud distribution; (b-3) a step in which a 3D seafloor point cloud classified as a large point cloud is divided into multiple tiles having mutually overlapping regions, and then tiles for training are selected based on object visibility; and (b-4) a step in which the Z values of the point cloud and objects are normalized to relative heights based on the tiles or local seafloor based on the work of step (b-3); A method for integrating data preprocessing, learning, and evaluation for detecting seabed anomalies based on a 3D seabed point cloud acquired by a multibeam echo sounder (MBES), comprising: (b-5) a step of maintaining the meaning of the learning data by classifying it into a target, an area of non-interest (Ignore), and a background (Discard) according to the learning purpose and label meaning; (c) a step of linking the preprocessing and label object processing results through step (b) with center distance-based duplicate removal and performance evaluation; and (d) a step of restoring to coordinates in actual seabed space and generating a final result. Claim 2 An input unit (100) receiving a 3D seabed point cloud having X, Y, Z coordinates acquired from MBES and a polygonal or 3D box-shaped object label; a vertical range (Z) correction module (210) that performs a plurality of processing operations on the input data input to the input unit (100), wherein standardization of object label names, conversion of polygonal objects into 3D box shapes, and correction processing of the object's vertical range to match the actual point cloud's seabed surface and the object's top surface; an overcluster separation module (220) that reconstructs cluster labels containing multiple objects within the corresponding 3D box-shaped object into multiple individual detection units based on the point cloud distribution among the 3D box-shaped object labels; a learning tile selection module (230) that divides the 3D seabed point cloud classified as a large point cloud into multiple tiles having mutually overlapping areas and selects tiles for learning based on object visibility; and based on the operation of the learning tile selection module (230), the point cloud and based on the tile or local seabed surface A system for integrated processing of data preprocessing, learning, and evaluation for detecting seabed anomalies based on a 3D seabed point cloud acquired by a multibeam echo sounder (MBES), comprising: a data preprocessing unit (200) including a Z-value normalization module (240) that normalizes the Z-value of an object to a relative height, and a label classification processing module (250) that maintains the meaning of the learning data by classifying it into a target, an area of non-interest (Ignore), and a background (Discard) according to the learning purpose and label meaning; a control unit (300) that links the preprocessing and label object processing results through the data preprocessing unit (200) with center distance-based duplicate removal and performance evaluation; and a result output unit (400) that restores the result to coordinates in the actual seabed space and generates the final result.
Citation Information
Patent Citations
Apparatus and method for controlling USV(Unmanned Surface Vehicle) for structural monitoring of offshore power plants
KR1020230086369A
3D data generation system for underwater exploration using 2D sonar
KR1020250180121A
3D modeling method for undersea topography
KR102186733B1
