A segmentation method and system

Abstract

The application discloses a segmentation method and system, acquires three-dimensional space data and performs space discretization processing to generate a three-dimensional voxel grid; based on the voxel grid, a three-dimensional convolutional neural network is used to detect a device object to obtain category information and a three-dimensional bounding box of the device object; then, instance point clouds are cropped from the three-dimensional space data according to the three-dimensional bounding box, and corresponding standard template point clouds are acquired from a preset template library according to the category information; iterative closest point registration is performed on the standard template point clouds and the instance point clouds to solve a transformation matrix for converting the standard template point clouds to a space pose of the instance point clouds; according to the transformation matrix, the standard template point clouds are used to segment the instance point clouds to obtain target point clouds of the device object. Efficient detection of devices in a large-scale point cloud scene of a substation is realized, accurate three-dimensional space positions and attitude information of the devices are synchronously acquired, and high-precision point cloud segmentation is completed, thereby providing accurate geometric pose support for digital twinning and intelligent operation and maintenance of the substation.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a segmentation method and system. Background Technology

[0002] With the rapid development of smart grids and digital twin substations, high-precision perception and intelligent inspection of power equipment based on 3D point clouds have become key technologies for ensuring the safe and stable operation of the power grid. Substation environments are characterized by a wide variety of equipment, compact structures, strong reflectivity of metal surfaces, and complex backgrounds, posing significant challenges to the accurate detection and spatial pose estimation of 3D targets. Currently, the maturity of 3D perception technologies such as LiDAR and depth cameras has made it possible to acquire high-density point cloud data. Accurately identifying key equipment such as insulators, circuit breakers, and disconnectors from massive point clouds and precisely estimating their 3D spatial position and orientation has become a core technological foundation supporting autonomous robotic arm operation, live-line work simulation, and equipment condition assessment.

[0003] In existing technologies, some solutions directly extract features from raw point clouds without effective structured preprocessing of large-scale point clouds. This results in enormous computational resource consumption when dealing with massive point cloud data at the substation level, and the inference efficiency is insufficient to meet practical engineering needs. Furthermore, their output only includes the target category and a coarse bounding box, lacking complete roll, pitch, and yaw attitude information of the equipment in 3D space. Additionally, the point cloud segmentation using cuboid bounding boxes has low precision. Other solutions fuse three modalities of data—LiDAR point clouds, visible light images, and infrared thermal images—for target detection. However, this method strictly requires precise alignment of the three heterogeneous data in 3D space. In practical engineering, multi-sensor calibration errors can significantly reduce detection accuracy. Moreover, its output also only includes target category and 3D position information, failing to provide the equipment's attitude parameters in 3D space.

[0004] In summary, existing methods for detecting 3D equipment in substations either struggle with large-scale scenarios due to a lack of efficient point cloud structuring, or are sensitive to data alignment errors due to reliance on multi-sensor fusion. More importantly, existing methods only output the target category and a rough bounding box position, failing to provide precise six-DOF pose information of the equipment in 3D space. This results in a lack of accurate geometric pose support for subsequent fine segmentation, state assessment, and digital twin applications. Therefore, how to achieve efficient equipment detection in large-scale point cloud scenarios in substations without relying on multi-sensor fusion, and simultaneously acquire precise 3D spatial position and pose information to support accurate point cloud segmentation, is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a segmentation method and system.

[0006] The embodiments of this application disclose the following technical solutions: The first aspect of this application provides a segmentation method, including: Acquire three-dimensional spatial data of the substation scene; The three-dimensional spatial data is spatially discretized, and each data point in the three-dimensional spatial data is assigned to a corresponding voxel unit to generate a three-dimensional voxel mesh. Based on the voxel grid, a three-dimensional convolutional neural network is used to detect device objects in the three-dimensional spatial data to obtain at least one device object's category information and corresponding three-dimensional bounding box. Based on the three-dimensional bounding box, the corresponding instance point cloud is cropped from the three-dimensional spatial data; Based on the category information of the device object, obtain the corresponding standard template point cloud from the preset template library; Iterative nearest point registration is performed between the standard template point cloud and the instance point cloud to solve for the transformation matrix used to transform the standard template point cloud to the spatial pose of the instance point cloud; Based on the transformation matrix, the instance point cloud is segmented using the contour of the standard template point cloud to obtain the target point cloud of the device object.

[0007] In one possible implementation, the step of spatially discretizing the three-dimensional spatial data, assigning each data point in the three-dimensional spatial data to a corresponding voxel unit, and generating a three-dimensional voxel mesh includes: Based on the spatial range of the three-dimensional spatial data and the preset voxel side length, determine the number of grids in each coordinate axis direction of the three-dimensional voxel mesh; A three-dimensional voxel grid is constructed according to the number of grids, and the voxel index coordinates corresponding to each data point in the three-dimensional spatial data are calculated with the origin of the three-dimensional voxel grid as the reference. Data points with the same voxel index coordinates are merged into the same voxel cell to generate a 3D voxel mesh.

[0008] In one possible implementation, detecting device objects in the three-dimensional spatial data using a three-dimensional convolutional neural network based on the voxel grid includes: The mean value of the attribute information of the data points contained in each non-empty voxel unit is used as the voxel feature of that voxel unit. The voxel features of each non-empty voxel unit are arranged in the corresponding positions in the three-dimensional voxel grid to form a three-dimensional feature tensor, which is then input into a three-dimensional convolutional neural network. The three-dimensional convolutional neural network performs convolution operations on the three-dimensional feature tensor through stacked three-dimensional convolutional layers, gradually extracts semantic features at different levels of abstraction, and outputs candidate three-dimensional bounding boxes through multi-scale feature fusion. Non-maximum suppression is performed on the candidate 3D bounding boxes to remove overlapping candidate boxes, thus obtaining the 3D bounding boxes.

[0009] In one possible implementation, segmenting the instance point cloud using the contour of the standard template point cloud includes: The standard template point cloud is spatially transformed according to the transformation matrix so that the contour of the standard template point cloud matches the instance point cloud in space. Using the contour of the transformed standard template point cloud as the segmentation boundary, points located inside the contour boundary are extracted from the instance point cloud and used as the target point cloud of the device object.

[0010] In one possible implementation, the three-dimensional convolutional neural network is determined in the following ways: Construct a training set, which contains voxel mesh samples labeled with device object categories and 3D bounding box labels; The voxel grid samples are input into a three-dimensional convolutional neural network, and the prediction results are obtained through forward propagation. Calculate the loss function value between the prediction result and the label. The loss function includes classification loss and regression loss. The classification loss is used to determine whether the anchor box contains the target object. The regression loss is used to adjust the position and size of the anchor box to make it approximate the labeled 3D bounding box. The network parameters of the three-dimensional convolutional neural network are updated through backpropagation based on the loss function value, iterating until convergence.

[0011] In one possible implementation, after obtaining the category information and corresponding 3D bounding box of the at least one device object, the method further includes: The candidate device objects corresponding to the three-dimensional bounding box are visualized and highlighted. In response to a manual confirmation command, retain or delete the 3D bounding box corresponding to the candidate device object; in response to a manual supplement command, add a 3D bounding box for the missed device object.

[0012] In one possible implementation, the iterative nearest-point registration of the standard template point cloud and the instance point cloud, and the solving of the transformation matrix for converting the standard template point cloud to the spatial pose of the instance point cloud, includes: Initialize the transformation parameters using the standard template point cloud as the source point cloud and the instance point cloud as the target point cloud; In each iteration, the source point cloud is transformed to the space where the target point cloud is located according to the current transformation parameters. The nearest corresponding point to each point in the transformed source point cloud is searched in the target point cloud, and the correspondence between the nearest points is established. Based on the nearest point correspondence, the update transformation parameters that minimize the distance metric between the source point cloud and the target point cloud are solved by singular value decomposition, and the transformation parameters are iteratively updated with the update transformation parameters. The iteration stops when the preset convergence condition is met, and the current transformation parameters are output as the transformation matrix.

[0013] In one possible implementation, the method further includes: In each iteration, for each point in the transformed source point cloud, the sum of squared Euclidean distances between the point and the nearest corresponding point in the target point cloud is calculated, and minimizing the sum of squared Euclidean distances is taken as the iteration objective.

[0014] In one possible implementation, the method further includes: In each iteration, for each point in the transformed source point cloud, the nearest corresponding point in the target point cloud is determined; the normal of the tangent plane containing the corresponding point is obtained, and the distance from the point along the normal direction to the tangent plane is calculated. The sum of the squares of the distances is minimized as the iteration objective.

[0015] A second aspect of this application provides a segmentation system, including: The first acquisition unit is used to acquire three-dimensional spatial data of the substation scene; The generation unit is used to perform spatial discretization processing on the three-dimensional spatial data, assign each data point in the three-dimensional spatial data to the corresponding voxel unit, and generate a three-dimensional voxel mesh. The detection unit is used to detect device objects in the three-dimensional spatial data based on the voxel grid using a three-dimensional convolutional neural network, and obtain the category information of at least one device object and its corresponding three-dimensional bounding box. The cropping unit is used to crop out the corresponding instance point cloud from the three-dimensional spatial data based on the three-dimensional bounding box; The second acquisition unit is used to acquire the corresponding standard template point cloud from the preset template library according to the category information of the device object; The solving unit is used to perform iterative nearest point registration between the standard template point cloud and the instance point cloud, and to solve for the transformation matrix used to transform the standard template point cloud to the spatial pose of the instance point cloud. The segmentation unit is used to segment the instance point cloud according to the transformation matrix and the contour of the standard template point cloud to obtain the target point cloud of the device object.

[0016] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the segmentation method described in the first aspect above.

[0017] A fourth aspect of this application provides a computer program product that, when run on a computer, executes the segmentation method described in the first aspect above.

[0018] A fifth aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the segmentation method described in the first aspect above.

[0019] Compared with the prior art, this application has the following advantages: By transforming the 3D spatial data of a substation scene into a regular voxel mesh, the data scale and processing complexity of massive point clouds are significantly reduced, enabling 3D convolutional neural networks to efficiently complete equipment detection with acceptable computing resources. Building upon this, addressing the deficiency of existing methods that only output coarse bounding boxes and lack fine poses, this application further utilizes the detected bounding boxes to crop instance point clouds and introduces standard template point clouds of equipment types for iterative nearest-point registration with the instance point clouds. By solving the transformation matrix, the six-degree-of-freedom pose of the equipment in 3D space is accurately obtained. Finally, the transformed template contour is used as the precise segmentation boundary to achieve fine segmentation of the instance point clouds. These techniques work together to balance detection efficiency and pose estimation accuracy in large-scale scenes without relying on multi-sensor fusion, fundamentally solving the problems of low computational efficiency, insufficient segmentation accuracy, and inability to support digital twin applications caused by the lack of efficient structured processing and pose estimation capabilities in existing technologies. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a segmentation method provided in an embodiment of this application; Figure 2 A flowchart of the three-dimensional device object target detection process provided in the embodiments of this application; Figure 3This is a schematic diagram of the pose estimation and point cloud segmentation process for a 3D device object provided in an embodiment of this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0023] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0024] As mentioned earlier, substation equipment detection methods based on 3D point clouds suffer from the following technical shortcomings. The first type of approach directly extracts features from the raw point cloud for target detection without effective structured preprocessing of large-scale point clouds. This results in enormous computational resource consumption when dealing with massive point cloud data across the entire substation, making model training and inference inefficient for practical engineering needs. The second type of approach integrates three heterogeneous modal data—LiDAR point clouds, visible light images, and infrared thermal images—for target detection. However, this method strictly requires precise alignment of these three types of heterogeneous data in 3D space. In practical engineering, multi-sensor calibration errors directly lead to a significant decrease in detection accuracy. More critically, the aforementioned existing methods only output the target category and a rough cuboid bounding box position, failing to provide complete roll, pitch, and yaw attitude information of the equipment in 3D space. Furthermore, the point cloud segmentation using cuboid bounding boxes has low precision, resulting in a lack of accurate geometric pose support for subsequent fine segmentation, state assessment, and digital twin applications.

[0025] Furthermore, in practical substation engineering applications, after completing 3D target detection, it is usually necessary to associate the detected equipment objects with ledger information and real-time operating data to construct computable and analyzable 3D digital assets. However, existing methods only focus on geometric target recognition and localization, failing to bridge the gap between realistic 3D models and semantic twins, and thus struggling to support higher-level application requirements such as equipment status assessment, defect detection, and maintenance decision-making. Therefore, how to achieve efficient equipment detection in large-scale point cloud scenarios in substations without relying on multi-sensor fusion, simultaneously acquire accurate 3D spatial position and attitude information of the equipment to support precise point cloud segmentation, and ultimately construct highly integrated geometric-physical-behavioral digital assets, is a pressing technical problem that needs to be solved.

[0026] To address the aforementioned issues, firstly, this application transforms the originally disordered and massive raw point cloud into a regular and compact voxel mesh representation by voxelizing the 3D spatial data, significantly compressing the data size and drastically reducing the computational load of subsequent neural networks. For substation-level scenarios, hash tables or octree storage structures can be used to retain only non-empty voxel units, further reducing storage and computational overhead. Thus, this application overcomes the shortcomings of existing technologies that consume enormous computational resources due to directly processing large-scale raw point clouds, achieving efficient target detection in large-scale substation point cloud scenarios.

[0027] Second, this application uses a three-dimensional convolutional neural network based on voxel grids for target detection. Only a single-modal three-dimensional spatial data is needed to complete device identification and positioning. There is no need for cross-modal precise alignment of point cloud, visible light image and infrared thermal image. This fundamentally avoids the impact of multi-sensor calibration error on detection accuracy. While reducing the difficulty of data acquisition and preprocessing, it ensures the accuracy and robustness of target detection.

[0028] Third, this application uses iterative nearest-point registration technology to accurately align the standard template point cloud with the instance point cloud. The transformation matrix output after iterative convergence includes a rotation matrix and a translation vector. When the device instance size differs from the standard template, it can also include a scaling factor. Thus, this application can accurately obtain the six-degree-of-freedom pose information of the device in three-dimensional space, including the rotation angles around the three coordinate axes and the translation amounts along the three coordinate axes, overcoming the shortcomings of existing technologies that only output coarse bounding boxes and lack complete pose information.

[0029] Fourth, this application uses the three-dimensional contour of the transformed standard template point cloud as the segmentation boundary to extract points located inside the contour from the instance point cloud. Compared with the prior art that uses a rough cuboid bounding box for segmentation, the contour of the standard template point cloud is an accurate description of the actual geometry of the device, which can accurately distinguish the point cloud belonging to the device from the surrounding background point cloud, effectively remove mixed background noise, and significantly improve the precision of point cloud segmentation.

[0030] Fifth, after the candidate device objects are output by the 3D convolutional neural network, this application visualizes and highlights them. Through a human-machine collaborative mechanism involving manual confirmation or deletion of detection results and manual supplementation of missed objects, it balances the efficiency of automatic detection with the reliability of manual review. Simultaneously, each confirmed or supplemented device object is bound with a unique identifier and interval information, giving each device object a unique identity in the digital twin environment, facilitating device query, location, and management at the plant / site level.

[0031] Sixth, this application uniquely encodes and binds the precisely segmented target point cloud with the equipment object's ledger information and real-time operating data. The ledger information includes static attributes such as equipment model, manufacturer, commissioning date, and rated parameters, while the real-time operating data includes dynamic status information such as voltage, current, temperature, and number of equipment actions. By uniquely encoding and structurally associating the three types of heterogeneous data—geometric shape, physical attributes, and operating status—a highly integrated digital asset of geometry, physics, and behavior is constructed. This enables the segmented geometric entities to possess complete semantic information dimensions, directly supporting upper-level applications such as equipment status assessment, defect detection, and maintenance decision-making. This represents a leap from real-world 3D to computable and analyzable semantic twins.

[0032] It should be noted that the object-oriented segmentation method, system, product, equipment, and medium for substation 3D scenes provided in this application can be applied to the field of computer technology. The above are merely examples and do not limit the application areas of the object-oriented segmentation method, system, product, equipment, and medium for substation 3D scenes provided in this application. Furthermore, the embodiments of this application may not limit the executing entity of the segmentation method. For example, the segmentation method of this application embodiment can be applied to data processing devices such as terminal devices or servers. The terminal device can be an electronic device such as a computer or personal digital assistant. The server can be a standalone server, a cloud server, or a cluster server composed of multiple servers.

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0034] The following embodiment illustrates a segmentation method provided in this application. See also... Figure 1 ,Should Figure 1 A flowchart of a segmentation method provided in this application embodiment, the method comprising: S101. Obtain the three-dimensional spatial data of the substation scene.

[0035] The three-dimensional spatial data of the substation scenario is the fundamental input for all subsequent processing steps, and its acquisition method determines the form and quality of the data. In this embodiment, the three-dimensional spatial data can take one of two forms: three-dimensional LiDAR point cloud data or three-dimensional Gaussian splashed Gaussian sphere data.

[0036] The 3D spatial data consists of either 3D LiDAR point cloud data or 3D Gaussian splashed Gaussian sphere data. 3D LiDAR point cloud data is acquired through LiDAR scanning, composed of a large number of 3D coordinate points, offering high accuracy and being unaffected by lighting conditions. 3D Gaussian splashed Gaussian sphere data is trained based on multi-view 2D images, encoding the scene as a collection of numerous Gaussian ellipsoids, simultaneously representing both geometric structure and surface texture.

[0037] In one possible implementation, when lidar scanning equipment is already deployed at the substation site, 3D lidar point cloud data of the substation scene can be directly acquired through lidar scanning. LiDAR scanning typically employs either fixed ground-based scanning or mobile scanning. Fixed ground-based scanning is suitable for acquiring localized high-precision point clouds, while mobile scanning is suitable for acquiring point clouds over large areas. After scanning, the raw point cloud needs to be preprocessed, including removing outlier noise points, filtering out ground points, and eliminating irrelevant data at the substation boundaries to reduce the amount of data required for subsequent processing.

[0038] In another possible implementation, when the substation site lacks LiDAR equipment, or when simultaneous acquisition of geometric and textural information of the scene is required, a 3D Gaussian splashing method can be used to construct scene data. Specifically, multiple cameras are used to acquire image sequences of the substation scene from different angles, ensuring that each device has sufficient visual coverage in the images. Subsequently, based on these multi-view images, the parameters of each Gaussian sphere are optimized using a 3D Gaussian splashing training algorithm. The resulting set of Gaussian spheres can then be used as 3D spatial data. This set of Gaussian spheres can not only reconstruct the spatial location, size, and relative layout of each device in the substation, but also the appearance texture and surface material information of the devices, providing a rich data foundation for subsequent voxelization processing.

[0039] S102. The three-dimensional spatial data is spatially discretized, and each data point in the three-dimensional spatial data is assigned to the corresponding voxel unit to generate a three-dimensional voxel mesh.

[0040] A voxel is short for a volume pixel in three-dimensional space. It refers to a small, uniformly sized cubic unit that regularly divides three-dimensional space. Spatial discretization is the process of mapping continuous or irregularly distributed three-dimensional spatial data onto these regular voxel units. Its core purpose is to transform the originally disordered, sparse, and large-scale raw data into a regular and compact structured representation, thereby significantly reducing the data size and improving the processing efficiency of subsequent neural networks.

[0041] In one possible implementation, the boundaries and resolution of the 3D voxel mesh are first determined. The maximum bounding box of the substation's 3D spatial data is calculated. The maximum bounding box is a minimum cuboid capable of containing all data points, with spans along the three coordinate axes denoted as Dx, Dy, and Dz, respectively. Irrelevant point cloud data at the substation boundaries can be appropriately truncated to reduce the processing scope and improve computational efficiency, depending on actual needs. Setting the voxel side length as v, the spatial span along each coordinate axis is divided by the voxel side length to obtain the global 3D mesh size of the substation. .

[0042] Next, the voxel index coordinates of each data point in the 3D spatial data are calculated. Using the origin of the 3D voxel mesh as a reference, for each data point, the voxel index coordinates corresponding to that data point in the voxel mesh are calculated based on the operational relationship between its spatial coordinates and the voxel side length v. The voxel index coordinates are a triplet, denoted as i, j, and k, representing the voxel cell number to which the data point belongs in the x-axis, y-axis, and z-axis directions, respectively. These index coordinates uniquely determine which voxel cell the data point is assigned to; that is, each data point obtains a voxel ID.

[0043] Then, data points with the same voxel index coordinates are merged into the same voxel unit. Points with the same index belong to the same voxel. By grouping all data points with the same voxel index coordinates i, j, and k into the same voxel unit, the aggregation and compression of the original data points are achieved. During the merging process, the nearest neighbor method can be used, where each data point is uniquely assigned to a voxel unit based on its voxel index coordinates. Alternatively, a soft voting method can be used, where each data point contributes a weighted value to multiple voxel units based on its distance to the centers of the nearest multiple voxel units; the smaller the distance, the greater the contribution. Specifically, the weighted allocation can be based on the distance of each data point to the centers of the eight nearest voxel units, thus distributing a data point to multiple voxel units with different weights. The soft voting method can better preserve the spatial distribution characteristics of the data and reduce quantization errors caused by voxelization.

[0044] After merging, each voxel unit summarizes information from all data points belonging to that unit. The information that a voxel unit can store includes point density, surface normals, and RGB color information, providing a data foundation for subsequent 3D convolutional neural network feature extraction. For large-scale point clouds, hash tables or octree storage structures can be used to store only non-empty voxel units, further reducing storage and computational overhead. After the above processing, the original 3D spatial data is transformed into a regular 3D voxel mesh. This voxel mesh retains the spatial structure information of the scene and represents the distribution of various equipment in the substation in 3D space in a compact form.

[0045] S103. Based on the voxel grid, a three-dimensional convolutional neural network is used to detect device objects in the three-dimensional spatial data to obtain at least one device object's category information and corresponding three-dimensional bounding box.

[0046] A 3D convolutional neural network (CNN) is a deep learning model specifically designed for processing 3D data. It performs sliding window operations in 3D space using 3D convolutional kernels, effectively extracting local spatial features and global semantic features from voxel meshes. Compared to convolutional neural networks that process 2D images, 3D CNNs retain high-dimensional information, making them suitable for the structured data of 3D voxel meshes used in this application. Object detection refers to identifying device objects in a scene and determining their spatial location. The detection result includes the device object's category information and its corresponding 3D bounding box. The 3D bounding box is a cuboid defined by the 3D spatial coordinates of its upper left and lower right corners, used to define the approximate spatial range of the device within the scene.

[0047] In one possible implementation, feature encoding is performed first. For each non-empty voxel cell in the 3D voxel mesh (i.e., a voxel cell containing at least one data point), the attribute information of all data points contained in that voxel cell is extracted. The attribute information may include at least one of the following: 3D coordinates of the data point, radar reflection intensity, RGB color information, and surface normal information. The arithmetic mean of these attribute information is calculated, and the average value is used as the voxel feature of that voxel cell. Through the averaging operation, each non-empty voxel cell is compressed into a feature vector. This feature vector eliminates the disorder of data points within the voxel while preserving the statistical characteristics of the data points, providing a regular input for the subsequent neural network.

[0048] Next, 3D feature extraction is performed. The voxel features of each non-empty voxel unit are arranged at their corresponding positions in the 3D voxel grid to form a complete 3D feature tensor. This 3D feature tensor preserves the spatial topology of the voxel grid, meaning that the spatial positional relationships of adjacent voxel units are fully preserved. This 3D feature tensor is input into a 3D convolutional neural network, which performs convolution operations on the 3D feature tensor level by level through stacked 3D convolutional layers. The convolution kernels in the 3D convolutional layers slide in 3D space, extracting spatial features within the local neighborhood at each convolution operation. Through the stacking of multiple convolutional layers, the receptive field of the network gradually expands. Lower-level convolutional layers extract local geometric features such as edges and corners, while higher-level convolutional layers extract global semantic features such as device components and overall structure. The network also includes batch normalization layers and activation function layers. Batch normalization layers are used to accelerate training convergence and improve network stability, while activation function layers are used to introduce nonlinear transformations to enhance the network's expressive power.

[0049] Then, feature fusion and decoding are performed. The 3D convolutional neural network aggregates voxel features of different scales through multi-scale feature fusion. Specifically, the network aggregates voxel features with different resolution scales output from different convolutional layers onto key points through voxel set abstraction, thereby simultaneously capturing local detail information and global context information, improving detection accuracy and robustness. The multi-scale fused features are input into the detection head network, which regresses to generate candidate 3D bounding boxes. Each candidate 3D bounding box contains center point coordinates, 3D dimensions, and orientation angle.

[0050] After the detection task is completed, the processing begins. The candidate 3D bounding boxes output by the 3D convolutional neural network exhibit overlap and redundancy, meaning that the same device object may be covered by multiple candidate boxes. This embodiment of the application performs non-maximum suppression (NMS) processing on the candidate 3D bounding boxes. NMS is a commonly used post-processing algorithm in object detection. Its core idea is to retain only the candidate box with the highest confidence level among overlapping detection boxes of the same object, and delete other candidate boxes whose overlap exceeds a preset threshold, thereby removing redundant detection boxes and obtaining the final 3D bounding boxes.

[0051] A 3D convolutional neural network (CNN) requires training before use. During training, a training set is constructed, containing a large number of voxel mesh samples labeled with device object categories and 3D bounding box labels. These voxel mesh samples are input into the CNN, and predictions are obtained through forward propagation. A loss function is calculated between the predictions and the labels, comprising classification and regression losses. The classification loss determines whether the anchor box contains a target object, while the regression loss adjusts the position and size of the anchor boxes to approximate the labeled 3D bounding boxes. The network parameters are updated using backpropagation based on the loss function values, iterating until the network converges. After training, the CNN can perform object detection on the input voxel mesh during the inference phase. During inference, the network outputs multiple candidate device object detection results, each including the device object's type name, target confidence probability, and 3D bounding box. For example, it may include the type name cls_name of the device object, the target confidence probability prob, and the bounding box of the object in three-dimensional space box=(x0,y0,z0,x1,y1,z1), where (x0,y0,z0) represents the coordinates of the upper left corner of the cuboid and (x1,y1,z1) represents the coordinates of the lower right corner of the cuboid.

[0052] After obtaining the category information of at least one device object and its corresponding three-dimensional bounding box, this application embodiment also provides a manual verification step.

[0053] See Figure 2 , Figure 2This is a flowchart illustrating the three-dimensional device object detection process provided in an embodiment of this application. Figure 2 As shown, the 3D equipment object target detection process provided in this application embodiment includes the following steps. First, input the 3D LiDAR point cloud information or 3D Gaussian sphere information of the substation as the data basis for the entire detection process. Second, perform 3D voxelization on the 3D spatial point cloud or Gaussian sphere, mapping the disordered and sparse point cloud onto a regular 3D voxel mesh. Third, run the 3D convolutional neural network target detection network for inference to achieve the detection output of specific equipment objects in the scene. This process includes feature encoding, 3D feature extraction, feature fusion and decoding, and post-processing of the detection task. Fourth, output the type name cls_name, confidence probability prob, and spatial bounding box box=(x0,y0,z0,x1,y1,z1) of the equipment object. The specific implementation methods of voxelization processing, 3D convolutional neural network detection, manual confirmation and binding involved in the above process have been described in detail in the preceding text of the specification and will not be repeated here. The following describes the fifth step, which involves manually deleting or confirming candidate equipment objects, assisting in adding missed equipment, and entering the binding equipment identifier and information for equipment object management in the substation.

[0054] Specifically, candidate device objects corresponding to 3D bounding boxes are visually highlighted. This means that data points within the bounding box are highlighted in the 3D scene display interface, allowing annotators to intuitively check the correctness of each detection result. Annotators check the detection results, and in response to manual confirmation instructions, retain correct detection results or delete the 3D bounding boxes corresponding to incorrectly detected candidate device objects; in response to manual supplement instructions, they manually add 3D bounding boxes for missed device objects. Through this manual confirmation and correction, the accuracy and completeness of the detection results can be further improved.

[0055] See below for the subsequent process Figure 3 . Figure 3 This is a schematic diagram illustrating the process of 3D device object pose estimation and point cloud segmentation provided in an embodiment of this application. This process can... Figure 2 The execution is based on the output structure of the 3D device object target detection shown, or it can be executed independently after manually selecting the bounding box. That is, the input is the name of the device object (cls_name) and the bounding box information (box=(x0,y0,z0,x1,y1,z1)). This input can come from... Figure 2 The output of the 3D device object detection shown can also be obtained by manually selecting objects directly in the 3D scene. For devices with simple bounding boxes, if precise alignment of the device pose is required, relying solely on manual adjustment is not only time-consuming and laborious, but also makes it difficult to guarantee accuracy. Figure 3The automatic alignment process shown can quickly and easily complete device pose estimation and point cloud segmentation. The process specifically includes the following steps.

[0056] S104. Based on the three-dimensional bounding box, cut out the corresponding instance point cloud from the three-dimensional spatial data.

[0057] After obtaining the 3D bounding box of a device object using a 3D convolutional neural network, it is necessary to extract the corresponding point cloud data from the original 3D spatial data for subsequent accurate pose estimation and point cloud segmentation. The 3D bounding box is a spatial range defined by the 3D coordinates of the top-left and bottom-right corners of a cuboid, which defines the approximate area of ​​the device object in the scene. Clipping refers to extracting all data points located inside the cuboid from the complete 3D scene spatial data according to the spatial range of the 3D bounding box, forming the instance point cloud of the device object.

[0058] Since 3D bounding boxes are obtained through 3D convolutional neural networks based on voxel meshes, their accuracy is limited by the voxel resolution. Therefore, the bounding box may contain a small number of background data points that do not belong to the device object. For example, in the process of detecting a circuit breaker, the 3D bounding box may include the air area around the circuit breaker or part of the point cloud of neighboring devices. However, these mixed background points do not affect the accuracy of subsequent pose estimation because the subsequent iterative nearest-point registration step can effectively suppress the influence of background noise through precise alignment between point clouds. Instance point clouds are the extracted results of the original data of a single device object, preserving the actual 3D coordinate information of the device in the scene, and are the basic data for subsequent registration with standard template point clouds.

[0059] In one possible implementation, all data points in the 3D spatial data are traversed, and each data point's 3D coordinates are checked one by one to determine whether they fall within the bounded box of the 3D bounding box. For each data point, its x-coordinate is compared to the minimum and maximum values ​​of the bounding box's x-axis, y-coordinate, and z-coordinate. If a data point satisfies these conditions in all three coordinate axes, it is considered an instance point cloud belonging to the device object and retained; otherwise, it is discarded. After the traversal is complete, all retained data points constitute the instance point cloud of the device object.

[0060] In another possible implementation, for scenarios using 3D Gaussian splashed Gaussian sphere data as 3D spatial data, since the Gaussian sphere has a spatial range defined by its spatial mean and covariance matrix, the clipping method can be based on whether the center position of the Gaussian sphere is within the 3D bounding box, or based on the degree of overlap between the spatial coverage of the Gaussian sphere and the 3D bounding box. Gaussian spheres whose centers are within the bounding box or intersect with the bounding box are included in the instance data of the device object. The resulting set of data points or Gaussian spheres after clipping constitutes the instance point cloud of the device object, denoted as det_obj, where each data point is represented by its 3D coordinates. This instance point cloud preserves the spatial distribution and geometric shape of the device object in the actual scene, providing target data for subsequent registration with the standard template point cloud.

[0061] S105. Based on the category information of the device object, obtain the corresponding standard template point cloud from the preset template library.

[0062] After obtaining the instance point cloud of the device object through cropping, it is necessary to acquire reference baseline data that matches the device object type for subsequent precise registration. This application embodiment pre-constructs a standard template library containing various device types, storing standard template point clouds for each type of device. The standard template point cloud is 3D point cloud data saved after fine annotation of typical device models. Its characteristics include high point cloud quality, low noise, and standardized pose, making it a geometric reference for this type of device under ideal conditions. Given the category information of the device object, the corresponding standard template point cloud is read from the preset template library according to the category name.

[0063] The standard template point cloud is constructed as follows. Substations contain a large number of devices with identical models and appearances. For example, A-phase bushings, B-phase bushings, and C-phase bushings belong to the same type of bushing equipment, with highly consistent external structures. Similarly, many duplicate devices of the same model exist among insulator strings, circuit breakers, disconnect switches, and other equipment. Based on this characteristic, a representative typical device from a certain type of equipment in the substation scenario can be selected, and its point cloud can be finely annotated. Fine annotation refers to accurately extracting the point cloud from the outer surface of the device, removing attached debris and background point clouds, and filtering noise points in the point cloud to obtain a clean and complete device point cloud. Subsequently, the point cloud of this typical device is adjusted to a preset standard spatial position and standard rotation posture, such as placing the device's center at the origin and aligning the device's main axis to the coordinate axis direction. After adjustment, the point cloud is saved as a reference template for this device type. Using the above method, corresponding standard template point clouds are constructed for various types of substation equipment such as transformers, circuit breakers, disconnect switches, bushings, and insulators. All templates are indexed and organized according to equipment type names to form a preset template library.

[0064] For example, given a device type name cls_name, the point cloud of the corresponding reference template in the template library is read, denoted as ref_obj={(xi,yi,zi)}. The device object instance det_obj={(xi,yi,zi)} and the object reference template ref_obj={(xi,yi,zi)} are then automatically aligned using the ICP (Iterative Closest Point) algorithm. The core idea is that, given a reference point set P and a point set Q to be registered, based on the initial rotation matrix R and displacement vector t, the nearest point pairs (point-to-point or point-to-surface) are continuously searched, minimizing the sum of squared Euclidean distances between them. New R and t are then solved using SVD (Singular Value Decomposition) until convergence.

[0065] In one possible implementation, after the 3D convolutional neural network detects and outputs the category name of a device object, it uses that category name as a query keyword to search a pre-defined template library. If a standard template point cloud corresponding to that category name exists in the template library, the point cloud data is read. For example, if the category of a device object is detected as a circuit breaker, the pre-saved standard template point cloud of a circuit breaker is read from the template library. If a template for that category does not yet exist in the template library, a new standard template point cloud can be created for that category through manual fine-grained annotation and added to the template library, continuously enriching the coverage of the template library.

[0066] The standard template point cloud corresponds to the instance point cloud obtained in step S104: the instance point cloud is the actual data of the device object in the substation scenario, containing the device's position and orientation in the actual space, but mixed with some noise and background artifacts; the standard template point cloud is the pure geometric data of this type of device under ideal conditions, with a regular orientation and clear contour, but it does not have the pose information of the instance object in the actual scene. The two complement each other. The instance point cloud provides a reference for the actual spatial pose, while the standard template point cloud provides a reference for the precise geometric contour. By combining the two in the subsequent registration step, pose estimation and accurate segmentation can be achieved.

[0067] S106. Perform iterative nearest point registration on the standard template point cloud and the instance point cloud, and solve for the transformation matrix used to transform the standard template point cloud to the spatial pose of the instance point cloud.

[0068] After obtaining the instance point cloud and the corresponding standard template point cloud of the device object, precise spatial alignment between the two is required. Iterative nearest-point registration is a classic algorithm for precisely aligning two sets of point clouds. Its core idea is to iteratively find the optimal rotation and translation transformation between the two sets of point clouds to achieve the best spatial match. In this embodiment, the standard template point cloud is used as the source point cloud, and the instance point cloud is used as the target point cloud. Iterative nearest-point registration is used to solve for the transformation matrix that transforms the source point cloud to the spatial pose of the target point cloud. This transformation matrix includes a rotation matrix and a translation vector. The rotation matrix describes the orientation of the device in three-dimensional space, and the translation vector describes the position offset of the device in three-dimensional space.

[0069] Formalistically, given a source point cloud Target point cloud ICP iterative solution for a rigid transformation (rotation) Peaceful relocation ), minimize ,in Point At the nearest point after transformation. The ICP of a point to a plane minimizes the sum of squared distances from the source point to the tangent plane. ,in Let P be the normal to the tangent plane of the target point, Q be the set of reference points, R be the initial rotation matrix, and t be the displacement vector.

[0070] In one possible implementation, the specific process of iterative nearest-point registration begins with initializing the transformation parameters. The initial transformation parameters can be set as a unit rotation matrix and a zero translation vector, meaning the source point cloud is initially placed in its original position. For scenarios where the device type is known and the installation orientation is relatively fixed, an initial rotation matrix and an initial translation vector can also be preset based on the device type to accelerate the convergence speed of subsequent iterations.

[0071] In each iteration, the source point cloud is transformed to the target point cloud space based on the current transformation parameters. Specifically, for each point in the source point cloud, the current rotation matrix and translation vector are applied to transform the point from its original coordinate system to the coordinate system of the target point cloud, resulting in the transformed source point cloud. Subsequently, the nearest corresponding point is searched in the target point cloud for each point in the transformed source point cloud, establishing a nearest-point correspondence. The nearest-point search can employ spatial index data structures such as kd-trees to accelerate the search process and improve matching efficiency. Through the above nearest-point search, each point in the source point cloud has a corresponding nearest point in the target point cloud, forming a point pair, and all point pairs together constitute the nearest-point correspondence.

[0072] Based on the established nearest neighbor correspondence, the updated transformation parameters are solved using singular value decomposition (SVD). Specifically, according to the current point pair correspondence, a rigid transformation that minimizes the distance metric between the source and target point clouds is calculated. This optimization problem can be solved analytically using SVD: first, the centroids of the source and target point clouds are calculated, and both sets of point clouds are decentered. Then, a covariance matrix is ​​constructed and SVD is performed on the covariance matrix. Based on the decomposition results, the optimal rotation matrix is ​​obtained, and then the optimal translation vector is calculated. The obtained updated transformation parameters minimize the registration error in the current iteration. Subsequently, the current transformation parameters are iteratively updated using these updated transformation parameters, and the next iteration begins.

[0073] In each iteration, the distance metric between the source point cloud and the target point cloud needs to be calculated as the objective of the iterative optimization. The choice of distance metric directly affects the registration accuracy and convergence speed. This application provides two distance metric methods in its embodiments.

[0074] In one possible implementation, the distance metric is the sum of squared point-to-point distances. In each iteration, for each point in the transformed source point cloud, the Euclidean distance between that point and its nearest corresponding point in the target point cloud is calculated. The squares of the Euclidean distances of all point pairs are then summed to obtain the sum of squared point-to-point distances. The iteration aims to minimize this sum of squared distances. The calculation of the point-to-point distance metric is relatively simple and suitable for scenarios where the point cloud surface features are relatively uniform and the noise level is low.

[0075] In another possible implementation, the distance metric is the sum of squared distances from a point to a plane. In each iteration, for each point in the transformed source point cloud, the nearest corresponding point in the target point cloud is first determined, and then the normal direction of the tangent plane containing that corresponding point is obtained. The normal to each point in the target point cloud can be estimated using principal component analysis of its neighboring point clouds. Next, the distance from a point in the source point cloud along this normal direction to the tangent plane containing its corresponding point is calculated; this distance is the perpendicular distance from the source point to the tangent plane of the target point. The above perpendicular distances are calculated for all point pairs, and the sum of squared distances is taken. Minimizing this sum of squared distances is used as the iteration objective. The point-to-plane distance metric introduces geometric constraints on the target point cloud surface, typically converges faster than the point-to-point distance metric, and has higher registration accuracy, making it suitable for scenarios where the target point cloud surface is continuous and the normal estimation is accurate.

[0076] During the iteration process, iteration stops when a preset convergence condition is met. The preset convergence condition may include the following: the number of iterations reaches a preset maximum number of iterations; the update amount of the transformation parameters is less than a preset threshold, meaning the change in the rotation matrix and the change in the translation vector are both less than their respective thresholds, indicating that registration has converged; the change in the distance metric is less than a preset threshold, meaning the error change between two adjacent iterations is sufficiently small, indicating that further iterations are unlikely to significantly improve registration accuracy. When any of the above convergence conditions are met, iteration stops and the current transformation parameters are output as the final transformation matrix.

[0077] For example, the output template object is transformed into the relative pose of the instance object, and the rigid transformation matrix is ​​used. It contains relative translation information. Spatial rotation information If the shape of the device object remains constant but its size can vary, scale information can also be added. ,at this time .

[0078] The output transformation matrix is ​​a 4x4 homogeneous transformation matrix that contains relative translation and spatial rotation information. The rotation matrix describes the rotation angles around the three coordinate axes required to convert the standard template point cloud into an instance point cloud, i.e., the roll, pitch, and yaw attitude of the device in 3D space. The translation vector describes the displacement along the three coordinate axes required to convert the standard template point cloud into an instance point cloud, i.e., the position coordinates of the device in 3D space. When there is a difference between the instance size of the device object and the standard template size, the transformation matrix can also add scale information, expressed as a scaling factor multiplied by the rotation matrix, to adapt to size differences of the same model of equipment caused by different voltage levels or different batches of production.

[0079] Through the above iterative nearest point registration, the embodiments of this application achieve accurate spatial alignment between the standard template point cloud and the instance point cloud. The solved transformation matrix accurately reflects the six-degree-of-freedom pose information of the device object in three-dimensional space, providing a reliable transformation benchmark for subsequent accurate segmentation based on the template contour.

[0080] S107. Based on the transformation matrix, the instance point cloud is segmented using the contour of the standard template point cloud to obtain the target point cloud of the device object.

[0081] After obtaining the transformation matrix through iterative nearest-point registration, the standard template point cloud is spatially transformed according to the transformation matrix to match the contour of the standard template point cloud with that of the instance point cloud in space. Specifically, the transformation matrix is ​​applied to each point in the standard template point cloud ref_obj to perform coordinate transformation, so that the standard template point cloud is transformed from the standard spatial pose in the template library to the actual spatial pose of the instance point cloud. After the transformation, the 3D contour of the standard template point cloud corresponds spatially to the outer surface of the device in the instance point cloud.

[0082] The standard template point cloud is a finely annotated point cloud of a typical device, and its outline is a precise description of the geometry of the device's outer surface. Using the outline of the transformed standard template point cloud as the segmentation boundary, points located inside the outline boundary are extracted from the instance point cloud `det_obj` and used as the target point cloud for the device object. In the specific implementation, for each point in the instance point cloud, it is determined whether the point's spatial coordinates lie within the spatial region enclosed by the outline of the transformed standard template point cloud. If the point is inside the outline boundary, it is determined to belong to the device object and is retained; if the point is outside the outline boundary, it is determined to be background noise or a point adjacent to other devices and is discarded.

[0083] After the above segmentation operation, points located inside the outline boundary of the standard template point cloud in the instance point cloud are preserved, while background points and noise located outside the outline boundary are removed. The preserved points constitute the target point cloud of the device object. The target point cloud is the pure point cloud data in the instance point cloud that precisely belongs to the device object, preserving both the actual spatial pose of the device in the scene and the outline accuracy at the standard template point cloud level.

[0084] Thus, through steps S101 to S107, this embodiment of the application has completed the entire process of detection, pose estimation, and accurate point cloud segmentation for a single device object in a substation scenario. For multiple device objects in a substation scenario, steps S103 to S107 can be repeated to complete the segmentation process for each device object one by one, ultimately obtaining the category information, six-degree-of-freedom pose information, and accurately segmented target point cloud for all device objects in the entire substation scenario.

[0085] The above are some specific implementations of the segmentation method provided in the embodiments of this application. Based on this, this application also provides a corresponding segmentation system. The system provided in the embodiments of this application will be described below from the perspective of functional modularity. The system includes: The first acquisition unit is used to acquire three-dimensional spatial data of the substation scene; The generation unit is used to perform spatial discretization processing on the three-dimensional spatial data, assign each data point in the three-dimensional spatial data to the corresponding voxel unit, and generate a three-dimensional voxel mesh. The detection unit is used to detect device objects in the three-dimensional spatial data based on the voxel grid using a three-dimensional convolutional neural network, and obtain the category information of at least one device object and its corresponding three-dimensional bounding box. The cropping unit is used to crop out the corresponding instance point cloud from the three-dimensional spatial data based on the three-dimensional bounding box; The second acquisition unit is used to acquire the corresponding standard template point cloud from the preset template library according to the category information of the device object; The solving unit is used to perform iterative nearest point registration between the standard template point cloud and the instance point cloud, and to solve for the transformation matrix used to transform the standard template point cloud to the spatial pose of the instance point cloud. The segmentation unit is used to segment the instance point cloud according to the transformation matrix and the contour of the standard template point cloud to obtain the target point cloud of the device object.

[0086] This application also provides corresponding devices and computer storage media for implementing the segmentation scheme provided in this application.

[0087] The device includes a memory and a processor. The memory is used to store instructions or code, and the processor is used to execute the instructions or code to cause the device to perform the segmentation method described in any embodiment of this application.

[0088] The computer storage medium stores code, and when the code is run, the device running the code implements the segmentation method described in any embodiment of this application.

[0089] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

[0090] It should be understood that in this application, "at least one" refers to one or more items, and "more" refers to two or more items. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one" or similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, "at least one" of a, b, or c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0091] It should be understood that the terms center, longitudinal, transverse, up, down, front, back, left, right, vertical, horizontal, top, bottom, inside, outside, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0092] It should be noted that, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0093] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the statement "including a…" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0094] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0095] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A segmentation method, characterized in that, include: Acquire three-dimensional spatial data of the substation scene; The three-dimensional spatial data is spatially discretized, and each data point in the three-dimensional spatial data is assigned to a corresponding voxel unit to generate a three-dimensional voxel mesh. Based on the voxel grid, a three-dimensional convolutional neural network is used to detect device objects in the three-dimensional spatial data to obtain at least one device object's category information and corresponding three-dimensional bounding box. Based on the three-dimensional bounding box, the corresponding instance point cloud is cropped from the three-dimensional spatial data; Based on the category information of the device object, obtain the corresponding standard template point cloud from the preset template library; Iterative nearest point registration is performed between the standard template point cloud and the instance point cloud to solve for the transformation matrix used to transform the standard template point cloud to the spatial pose of the instance point cloud; Based on the transformation matrix, the instance point cloud is segmented using the contour of the standard template point cloud to obtain the target point cloud of the device object.

2. The method according to claim 1, characterized in that, The step of spatial discretizing the three-dimensional spatial data, assigning each data point in the three-dimensional spatial data to a corresponding voxel unit, and generating a three-dimensional voxel mesh includes: Based on the spatial range of the three-dimensional spatial data and the preset voxel side length, determine the number of grids in each coordinate axis direction of the three-dimensional voxel mesh; A three-dimensional voxel grid is constructed according to the number of grids, and the voxel index coordinates corresponding to each data point in the three-dimensional spatial data are calculated with the origin of the three-dimensional voxel grid as the reference. Data points with the same voxel index coordinates are merged into the same voxel cell to generate a 3D voxel mesh.

3. The method according to claim 1, characterized in that, The method of detecting device objects in the three-dimensional spatial data using a three-dimensional convolutional neural network based on the voxel grid includes: The mean value of the attribute information of the data points contained in each non-empty voxel unit is used as the voxel feature of that voxel unit. The voxel features of each non-empty voxel unit are arranged in the corresponding positions in the three-dimensional voxel grid to form a three-dimensional feature tensor, which is then input into a three-dimensional convolutional neural network. The three-dimensional convolutional neural network performs convolution operations on the three-dimensional feature tensor through stacked three-dimensional convolutional layers, gradually extracts semantic features at different levels of abstraction, and outputs candidate three-dimensional bounding boxes through multi-scale feature fusion. Non-maximum suppression is performed on the candidate 3D bounding boxes to remove overlapping candidate boxes, thus obtaining the 3D bounding boxes.

4. The method according to claim 1, characterized in that, The step of segmenting the instance point cloud using the contour of the standard template point cloud includes: The standard template point cloud is spatially transformed according to the transformation matrix so that the contour of the standard template point cloud matches the instance point cloud in space. Using the contour of the transformed standard template point cloud as the segmentation boundary, points located inside the contour boundary are extracted from the instance point cloud and used as the target point cloud of the device object.

5. The method according to claim 3, characterized in that, The method for determining the three-dimensional convolutional neural network includes: Construct a training set, which contains voxel mesh samples labeled with device object categories and 3D bounding box labels; The voxel grid samples are input into a three-dimensional convolutional neural network, and the prediction results are obtained through forward propagation. Calculate the loss function value between the prediction result and the label. The loss function includes classification loss and regression loss. The classification loss is used to determine whether the anchor box contains the target object. The regression loss is used to adjust the position and size of the anchor box to make it approximate the labeled 3D bounding box. The network parameters of the three-dimensional convolutional neural network are updated through backpropagation based on the loss function value, iterating until convergence.

6. The method according to claim 1, characterized in that, After obtaining the category information and corresponding 3D bounding box of the at least one device object, the method further includes: The candidate device objects corresponding to the three-dimensional bounding box are visualized and highlighted. In response to a manual confirmation command, retain or delete the 3D bounding box corresponding to the candidate device object; in response to a manual supplement command, add a 3D bounding box for the missed device object.

7. The method according to claim 1, characterized in that, The iterative nearest-point registration of the standard template point cloud and the instance point cloud, and the solution of the transformation matrix used to transform the standard template point cloud to the spatial pose of the instance point cloud, include: Initialize the transformation parameters using the standard template point cloud as the source point cloud and the instance point cloud as the target point cloud; In each iteration, the source point cloud is transformed to the space where the target point cloud is located according to the current transformation parameters. The nearest corresponding point to each point in the transformed source point cloud is searched in the target point cloud, and the correspondence between the nearest points is established. Based on the nearest point correspondence, the update transformation parameters that minimize the distance metric between the source point cloud and the target point cloud are solved by singular value decomposition, and the transformation parameters are iteratively updated with the update transformation parameters. The iteration stops when the preset convergence condition is met, and the current transformation parameters are output as the transformation matrix.

8. The method according to claim 7, characterized in that, The method further includes: In each iteration, for each point in the transformed source point cloud, the sum of squared Euclidean distances between the point and the nearest corresponding point in the target point cloud is calculated, and minimizing the sum of squared Euclidean distances is taken as the iteration objective.

9. The method according to claim 7, characterized in that, The method further includes: In each iteration, for each point in the transformed source point cloud, the nearest corresponding point in the target point cloud is determined; the normal of the tangent plane containing the corresponding point is obtained, and the distance from the point along the normal direction to the tangent plane is calculated. The sum of the squares of the distances is minimized as the iteration objective.

10. A segmentation system, characterized in that, include: The first acquisition unit is used to acquire three-dimensional spatial data of the substation scene; The generation unit is used to perform spatial discretization processing on the three-dimensional spatial data, assign each data point in the three-dimensional spatial data to the corresponding voxel unit, and generate a three-dimensional voxel mesh. The detection unit is used to detect device objects in the three-dimensional spatial data based on the voxel grid using a three-dimensional convolutional neural network, and obtain the category information of at least one device object and its corresponding three-dimensional bounding box. The cropping unit is used to crop out the corresponding instance point cloud from the three-dimensional spatial data based on the three-dimensional bounding box; The second acquisition unit is used to acquire the corresponding standard template point cloud from the preset template library according to the category information of the device object; The solving unit is used to perform iterative nearest point registration between the standard template point cloud and the instance point cloud, and to solve for the transformation matrix used to transform the standard template point cloud to the spatial pose of the instance point cloud. The segmentation unit is used to segment the instance point cloud according to the transformation matrix and the contour of the standard template point cloud to obtain the target point cloud of the device object.

Images