Key point cloud segmentation method and system for navigation of oil sample collection robot

Abstract

The application discloses a kind of key position point cloud segmentation methods and systems for oil sample collection robot navigation, method includes: through adaptive data enhancer, sample perception enhancement is carried out to input point cloud;Local geometric features and global context information are fused by CB-Transformer module in encoder through two-stage attention mechanism;Weighted residual clustering and cross-level cascade are carried out to multi-scale features using C-VLAD module, and comprehensive global descriptor is generated;Finally, the resolution of point cloud is restored by decoder, and the segmentation result is output.The application effectively solves the problems of sparse point cloud annotation data, large equipment scale difference and complex geometric structure in industrial scene, significantly improves the segmentation accuracy and robustness of key parts such as valve and flange, reduces the data annotation cost, and provides reliable technical support for accurate navigation and operation of oil sample collection robot.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of point cloud segmentation, and particularly relates to a key part point cloud segmentation method and system for navigation of an oil sample collection robot. BACKGROUND

[0002] Point cloud segmentation technology is a core link in the fields of machine vision and robot environment perception, and its goal is to divide three-dimensional point cloud data into components with specific semantics. In the industrial automation scene, especially in the autonomous navigation of an oil sample collection robot and substation equipment inspection, this technology is crucial for the robot to accurately identify the operation target (such as a valve or a flange) and achieve obstacle avoidance and path planning.

[0003] Currently, point cloud segmentation methods are mainly divided into the following two categories:

[0004] Traditional geometry feature-based methods:

[0005] Such methods rely on artificially designed low-level geometric features (such as normal vectors, curvatures, and densities) for region growing or clustering (e.g., DBSCAN algorithm). Although they are computationally efficient, their performance is heavily dependent on the completeness of feature design. In complex industrial environments such as substations, there are dense equipment, diverse structures, and a large amount of noise and occlusions. Traditional methods have difficulty in robustly extracting discriminative features, resulting in low segmentation accuracy and poor robustness, which cannot meet the needs of high-precision navigation and operation.

[0006] Deep learning-based methods:

[0007] With the development of deep learning, point cloud processing methods based on neural networks such as PointNet and PointNet++ have become mainstream. These methods can automatically learn more robust semantic features through hierarchical feature extraction and nonlinear transformation. However, when directly applied to the navigation scene of an oil sample collection robot, they still face many serious challenges:

[0008] Annotation data sparsity problem: Point cloud data in industrial scenarios requires expert fine annotation, which is extremely costly, resulting in very scarce annotated data available for supervised training. General deep learning models are prone to overfitting under sparse annotation conditions, and their performance significantly degrades.

[0009] Multi-scale structure challenge: In the substation environment, there are both large equipment (such as transformer housings) and small key components (such as valve handwheels and bolt holes), with a large difference in scale. The down-sampling strategy of existing models (such as RandLA-Net) easily loses the detailed information of small-scale targets, resulting in insufficient segmentation accuracy.

[0010] Insufficient global context modeling: Existing methods focus more on the extraction of local features, and lack effective modeling of the overall layout of the scene and the spatial relationship between components (such as the connection relationship between pipes and valves), which limits the model's deep semantic understanding of complex industrial environments.

[0011] Insufficient use of geometric prior knowledge: Industrial equipment components have obvious geometric priors (such as pipes being cylindrical and flanges being ring-shaped planes). Existing general network structures fail to effectively incorporate these prior knowledge to enhance the perception ability of specific structures, resulting in blurred segmentation boundaries and high component misidentification rate.

[0012] In summary, existing point cloud segmentation techniques either lack robustness due to simple models or are difficult to achieve high-precision and high-robust key part segmentation in substation environments with sparse annotations, complex structures, and multiple scales coexisting due to not being optimized for the special nature of industrial scenes. This has become a technical bottleneck restricting the intelligent level of oil sample collection robots. SUMMARY

[0013] The present application provides a key part point cloud segmentation method and system for oil sample collection robot navigation, which solves the technical problem of being difficult to achieve high-precision and high-robust key part segmentation in substation environments with sparse annotations, complex structures, and multiple scales coexisting.

[0014] In a first aspect, the present application provides a key part point cloud segmentation method for oil sample collection robot navigation, comprising:

[0015] Obtain the original point cloud, perform sample perception enhancement on the original point cloud according to a preset adaptive data enhancer, generate an enhanced point cloud, and input the enhanced point cloud into a preset encoder, wherein the encoder extracts multi-scale local features, and the encoder includes a multi-layer SetAbstraction module and a CB-Transformer module;

[0016] Input the multi-scale local features into a preset C-VLAD module, and the C-VLAD module performs weighted residual clustering and cross-level cascade, to generate a comprehensive global descriptor tensor;

[0017] Cross-layer jump connection is performed between the comprehensive global descriptor tensor and the up-sampling features in the decoder, and each target point cloud is gradually recovered by inverse distance weighted interpolation, and the semantic labels of the target point clouds are output, to complete the segmentation of key parts.

[0018] In a second aspect, the present application provides a key part point cloud segmentation system for oil sample collection robot navigation, comprising:

[0019] The enhancement module is configured to obtain an original point cloud, perform sample-aware enhancement on the original point cloud according to a preset adaptive data enhancer, generate an enhanced point cloud, and input the enhanced point cloud into a preset encoder, and the encoder extracts a multi-scale local feature, wherein the encoder comprises a multi-layer SetAbstraction module and a CB-Transformer module.

[0020] The generation module is configured to input the multi-scale local feature into a preset C-VLAD module, the C-VLAD module performs weighted residual clustering and cross-level cascade, and generates a comprehensive global descriptor tensor.

[0021] The segmentation module is configured to perform cross-layer jump connection between the comprehensive global descriptor tensor and up-sampling features in a decoder, gradually recover each target point cloud through inverse distance weighted interpolation, and output semantic labels of the each target point cloud, and complete segmentation of key parts.

[0022] In a third aspect, an electronic device is provided, which includes at least one processor and a memory connected to the at least one processor in communication, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the steps of the key part point cloud segmentation method for oil sample collection robot navigation according to any one of the embodiments of the present application.

[0023] In a fourth aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, and the program instructions are executed by a processor to enable the processor to perform the steps of the key part point cloud segmentation method for oil sample collection robot navigation according to any one of the embodiments of the present application.

[0024] The key point cloud segmentation method and system for robot navigation of oil sample collection provided by the application, through hierarchical downsampling and feature enhancement module to extract multi-scale semantic information, the decoder restores the spatial details based on the feature propagation and cross-layer fusion mechanism, and finally completes the point cloud segmentation. Compared with the traditional method, through the cooperative optimization of adaptive data enhancement, center attention transformer and multi-scale global descriptor, the segmentation performance in the sparse labeling scene is significantly improved. In the encoding stage, the original point cloud is first subjected to multi-layer perception (MLP) to increase the feature dimension, and the three-dimensional coordinates are mapped to a high-dimensional feature space to enhance the representation ability. Then, through the cascaded SetAbstraction module, iterative farthest point sampling (FPS) and K nearest neighbor grouping (KNN) are performed to gradually reduce the point cloud density and aggregate local geometric features. To overcome the feature learning degradation problem caused by limited labeling, the CB-Transformer module is introduced, which is based on a two-stage attention mechanism. First, the global feature guided neighborhood weight is generated, and then the center semantic is propagated to the local area through dynamic weight injection, thereby establishing a robust local-global feature correlation under the condition of sparse labeling. Further, the C-VLAD module performs weighted residual clustering on the multi-scale features output by each level of the encoder, and generates a comprehensive global descriptor tensor through cross-resolution feature fusion, effectively solving the problem of fine-grained information loss caused by traditional single-scale aggregation. The decoder uses a progressive upsampling strategy to restore the point cloud spatial resolution. Specifically, the inverse distance weighted interpolation algorithm is used to upsample the deep abstract features, and the point cloud density is gradually restored to the original base. To enhance the semantic consistency in the feature propagation process, a cross-layer skip connection (SkipConnection) mechanism is designed to concatenate the high-dimensional features output by the CB-Transformer module with the decoder upsampled features in the channel dimension, realizing the deep fusion of local geometric details and global semantic context, and finally completing the point cloud semantic segmentation task. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0026] Figure 1 A flowchart of a key point cloud segmentation method for robot navigation of oil sample collection provided by an embodiment of the application;

[0027] Figure 2 A flowchart of a key point cloud segmentation method for a specific embodiment provided by an embodiment of the application;

[0028] Figure 3 FIG. 1 shows a schematic diagram of a two-stage attention mechanism according to an embodiment of the present application;

[0029] Figure 4 FIG. 2 shows a flowchart of a method for generating a comprehensive global description according to an embodiment of the present application;

[0030] Figure 5 FIG. 3 shows a structural diagram of a key point cloud segmentation system for navigation of an oil sample collection robot according to an embodiment of the present application;

[0031] Figure 6 FIG. 4 shows a structural diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0032] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0033] Please refer to Figure 1 , which shows a flowchart of a key point cloud segmentation method for navigation of an oil sample collection robot according to an embodiment of the present application.

[0034] As shown in Figure 1 , the key point cloud segmentation method specifically includes the following steps:

[0035] In step S101, an original point cloud is acquired, and sample-aware enhancement is performed on the original point cloud according to a preset adaptive data enhancer to generate an enhanced point cloud. The enhanced point cloud is input to a preset encoder, and multi-scale local features are extracted by the encoder. The encoder includes a multi-layer SetAbstraction module and a CB-Transformer module.

[0036] In this step, point-by-point features and overall shape features of the original point cloud are extracted.

[0037] A noise vector is generated based on a Gaussian distribution, the noise vector is spliced with the overall shape features, and a transformation matrix is calculated according to a multi-layer perceptron layer.

[0038] N copies of the overall shape features are spliced with the point-by-point features, a Gaussian noise matrix is introduced, and a displacement matrix is calculated by a multi-layer perceptron layer.

[0039] According to the displacement matrix and the transformation matrix, the original point cloud is enhanced to obtain an enhanced point cloud, and the enhanced point cloud is limited according to a preset enhancement loss function to obtain a final enhanced point cloud, wherein an expression of the enhanced point cloud is:

[0040] ,

[0041] In the formula, is the enhanced point cloud, is the original point cloud, is the transformation matrix, is the displacement matrix;

[0042] An expression of the enhancement loss function is:

[0043] ,

[0044] ,

[0045] ,

[0046] ,

[0047] ,

[0048] In the formula, is the enhancement loss function, is a cross-entropy loss of the enhanced point cloud, is a natural exponential function, is a cross-entropy loss of the point cloud P, is a total number of categories, is a label of a real category, is a probability that the point cloud P belongs to the c-th category, is an enhancement amplitude, is an upper limit of enhancement, is a transition variable in a calculation process.

[0049] In the embodiment, the introduced enhancement device has sample perception ability, and can generate a specific enhancement function according to each input sample. In order to simplify the representation, P represents a training sample of the input enhancement device A, and the enhanced point cloud is represented as The enhanced sample generated by the enhancer. First, the point-wise features F are extracted by a series of shared multi-layer perceptron (MLP), and then the overall shape feature vector G is obtained by max-pooling. On this basis, the enhancement function related to the input sample is regressed by shape regression and point-wise regression respectively. Among them, the transformation matrix M can be generated by shape regression. The specific operation is to generate a C-dimensional noise vector based on Gaussian distribution, and then splice it with G, and then calculate M through MLP. This matrix contains the rotation and scaling operations of the point cloud. The introduction of the noise vector provides randomness for the enhancer, so as to explore more diversified transformation matrices. The displacement matrix D can be generated by point-wise regression. The specific operation is to splice N copies of G with F, while introducing a Gaussian noise matrix, and finally calculate D through MLP. This matrix includes point-wise translation and jitter.

[0050] It should be noted that the specific structure of the constructed Center-Bridged Attention Transformer (CB Transformer) module is as shown in Figure 2 , in which a Central-based Attention (CA) is embedded between the multi-layer perceptron (MLP) layers, a Dropout module is added to suppress the overfitting phenomenon, and a residual connection is adopted to realize efficient transmission of feature flow. With the help of the two-stage embedding strategy, the central-based attention mechanism can effectively solve the difficulties in feature learning under the condition of sparse annotation, and significantly improve the learning efficiency and generalization ability of the model in the limited annotation data scene.

[0051] First, define the input point set as and its corresponding feature set , where is the feature vector of the point cloud, and N is the base of the point cloud. For each query point , local representation learning is performed on its K neighborhood point set and the corresponding feature . At this time is defined as the neighborhood center point , and is its adjacent point. Specifically, this method first extracts an initial representation with a dimension of NxC from the center point feature, and obtains an Nxl-dimensional center weight coefficient after dimension reduction by linear transformation. The global feature (denoted as Figure 3 ) is constructed for the neighborhood points, which is realized by aggregating the node features arranged in spatial order within each neighborhood. As shown in , the weight coefficient of the query point is obtained by calculating through a one-dimensional linear layer , and the K neighboring nodes are obtained by using the k-nearest neighbor algorithm, and the point set of the K neighboring nodes iswith the feature set together constitute a neighborhood representation space.

[0052] In the first embedding stage, the global feature is constructed by fusing the center weight and the neighborhood point feature, and the neighborhood weight of the second embedding is generated after the softmax function σ normalization. It is worth noting that, although effectively improves the representation ability of unlabelled neighborhood points in weakly supervised point clouds, the capture efficiency of independent neighborhood local features and the guiding role of the center node are still insufficient, resulting in optimization space in weight calculation accuracy compared with classical methods. Therefore, the second embedding mechanism is needed to accurately inject the center node information into the corresponding neighborhood points under the guidance of global semantics. As a dynamic weight parameter, the center node feature is directionally propagated to each neighborhood space through matrix operation. The second embedding feature generated in this process can ensure accurate injection of center node information into corresponding neighborhood points under the guidance of global semantics.

[0053] To construct a composite attention mechanism that fuses global features and center weights, this study innovatively integrates spatial encoding technology to enhance geometric features. In specific implementation, the global feature is mapped through a linear layer , and then fused with the feature through learnable parameters α, β. This strategy significantly enhances the expression ability of local features while maintaining the integrity of global semantics. In addition, by introducing position encoding to establish spatial geometric correlation, it effectively supplements the topological features based on pose relationship. Finally, the enhanced attention weight normalized by softmax is multiplied with the value feature map, and the weighted sum strategy is used to eliminate point sequence sensitivity, thereby outputting the optimized feature representation.

[0054] This application adopts a position encoding mechanism that fuses multiple modal geometric features. By introducing multi-dimensional geometric descriptors such as height angle , azimuth angle , and Euclidean distance , and through multi-layer perceptron MLP and learnable parameter λ weighted fusion, an enhanced position encoding is constructed.

[0055] Step S102, input the multi-scale local feature into the preset C-VLAD module, and the C-VLAD module performs weighted residual clustering and cross-level cascade, generating a comprehensive global descriptor tensor.

[0056] In this step, to aggregate the local features into global vectors, the present application introduces a Comprehensive Vector of Locally Aggregated Descriptors (C-VLAD) module to realize global context modeling across scales and resolutions through a multi-level feature fusion mechanism, as shown in Figure 4 . This module significantly improves the feature representation capability of the network for the multi-scale heterogeneous structure in the transformer site cloud through a multi-resolution feature joint modeling strategy, and provides an effective solution to the feature capture problem of coexistence of large-scale transformers and small-scale bushings in the power scene.

[0057] First, a multi-scale local feature set is extracted from the output of each level of the encoder , wherein is the feature matrix generated by the th layer of the encoding layer. The C-VLAD module can generate a set of level clustering centers . For each clustering center , a corresponding feature descriptor vector is calculated through weighted residual aggregation.

[0058] ,

[0059] wherein is the feature descriptor vector, is the natural exponential function, is the feature matrix generated by the th layer, is the clustering center, is the bias term, is the weight vector;

[0060] After the local features of each level are encoded, the global descriptor vector can be represented as . Through cross-level simultaneous operation, the global descriptor vectors output by each layer are dimensionally expanded and feature spliced to finally build a comprehensive global descriptor tensor containing multi-scale information. In the formula, is the comprehensive global descriptor tensor, is the th feature descriptor vector, is the vector splicing operation.

[0061] In this embodiment, the C-VLAD module can effectively alleviate the problem of loss of fine-grained information caused by single feature level in traditional methods, and provide more discriminative global context prior for subsequent segmentation tasks.

[0062] Step S103, the integrated global descriptor tensor is cross-layer jump connected with the up-sampling features in the decoder, each target point cloud is recovered step by step by inverse distance weighted interpolation, and the semantic labels of the target point clouds are output, and the segmentation of the key parts is completed.

[0063] In summary, the method of the present application first introduces an adaptive data enhancer, which can dynamically generate the optimal enhancement strategy according to the sample characteristics (such as density, noise level, component scale) of the input point cloud, greatly improving the robustness and generalization ability of the model in complex industrial environments, effectively overcoming the overfitting and performance fluctuation problems caused by traditional fixed enhancement strategies or lack of enhancement. Secondly, the core module CB-Transformer significantly enhances the perception ability of the geometric structure (such as cylindrical surface, plane) of the device components through its unique two-stage attention mechanism (local geometric perception + global context aggregation), and establishes a strong global context understanding, so that the model can accurately identify key parts (such as valves, flanges) and clearly define their boundaries, even under partial occlusion or noise interference, it can maintain high-precision segmentation. At the same time, the innovative C-VLAD module efficiently fuses feature information of different scales through the weighted residual clustering and cross-layer cascading mechanism, effectively solving the multi-scale device segmentation problem, ensuring that large-scale device main bodies and small-scale key operation parts can be accurately segmented. In addition, the entire scheme design focuses on efficiency, combining random down-sampling and optimized feature extraction process, ensuring high precision while achieving fast processing speed, meeting the needs of real-time navigation of robots. Most importantly, the dependence of this method on labeled data is significantly reduced, and it can still achieve segmentation performance close to full labeling under sparse labeling conditions, greatly reducing the data labeling cost of industrial applications and accelerating the application of technology. Ultimately, these effects work together to enable the oil sample collection robot to more reliably and accurately perceive the environment and identify operation targets, thereby significantly improving its success rate and safety in autonomous navigation, obstacle avoidance, and oil sample collection operations.

[0064] In one specific embodiment, the method of the present application is applied to a certain type of transformer substation oil sample collection robot. The laser radar carried by the robot collects the scene point cloud of the transformer substation at a resolution of 0.3m, and the goal is to accurately segment the key parts for navigation and operation from it: valves, flanges, and pipeline connection points. The implementation process is as follows: first, the raw point cloud is input into the adaptive data enhancer, which analyzes the point cloud density distribution and noise level, identifies areas containing potential valves (small scale, high curvature) and flanges (ring structure), and applies stronger rotation, scaling, and noise injection enhancement to these areas, while applying weaker enhancement to the area of the large transformer shell (large scale, flat). The enhanced point cloud enters the encoder, which uses a hierarchical downsampling strategy (such as random downsampling), and each layer contains a CB-Transformer module. In the CB-Transformer, first, the local geometric perception attention is used to calculate the similarity weight of geometric features such as normal vector and curvature in the point neighborhood, highlighting structures such as planes and cylindrical surfaces; then, the global context aggregation attention uses position encoding (fusing elevation angle, azimuth angle, and Euclidean distance) to calculate the correlation weight between all points, capturing the spatial layout relationship between devices (such as pipeline connecting to valve). The multi-scale feature map output by the encoder is input into the C-VLAD module, which performs weighted residual clustering on the features of each scale (the number of cluster centers is adaptively set according to the complexity of the component), generating a descriptor vector that describes the local structure of that scale; then the descriptor vectors of different scales are concatenated across levels to form a comprehensive descriptor tensor that contains global and local information. This tensor is cross-layer connected with the upsampled features obtained by inverse distance interpolation in the decoder, fusing detailed information. Finally, the decoder outputs the classification label of each point (background, valve, flange, pipeline connection point).

[0065] Referring to Figure 5 , a structure block diagram of a key part point cloud segmentation system for navigation of an oil sample collection robot of the present application is shown.

[0066] As Figure 5 shown, the key part point cloud segmentation system 200 includes an enhancement module 210, a generation module 220, and a segmentation module 230.

[0067] The enhancement module 210 is configured to obtain an original point cloud, perform sample-aware enhancement on the original point cloud according to a preset adaptive data enhancer, generate an enhanced point cloud, and input the enhanced point cloud into a preset encoder, and the encoder extracts multi-scale local features, wherein the encoder includes a multi-layer SetAbstraction module and a CB-Transformer module; the generation module 220 is configured to input the multi-scale local features into a preset C-VLAD module, and the C-VLAD module performs weighted residual clustering and cross-layer concatenation, and generates a comprehensive global descriptor tensor; and the segmentation module 230 is configured to perform cross-layer jump connection between the comprehensive global descriptor tensor and up-sampling features in a decoder, gradually restores each target point cloud through inverse distance weighted interpolation, and outputs semantic labels of the each target point cloud, thereby completing segmentation of key parts.

[0068] It should be understood that Figure 5 the modules described in the above Figure 1 correspond to the respective steps in the methods described in the above Figure 5 the operations and features described above for the methods and the corresponding technical effects are equally applicable to the modules in the

[0069] In some other embodiments, the present application also provides a computer readable storage medium having a computer program stored thereon, and the program instructions are executed by a processor to make the processor execute the key part point cloud segmentation method for oil sample collection robot navigation in any method embodiment described above.

[0070] As an implementation form, the computer readable storage medium of the present application stores computer executable instructions, and the computer executable instructions are configured to:

[0071] obtain an original point cloud, perform sample-aware enhancement on the original point cloud according to a preset adaptive data enhancer, generate an enhanced point cloud, and input the enhanced point cloud into a preset encoder, and the encoder extracts multi-scale local features, wherein the encoder includes a multi-layer SetAbstraction module and a CB-Transformer module;

[0072] input the multi-scale local features into a preset C-VLAD module, and the C-VLAD module performs weighted residual clustering and cross-layer concatenation, and generates a comprehensive global descriptor tensor;

[0073] perform cross-layer jump connection between the comprehensive global descriptor tensor and up-sampling features in a decoder, gradually restore each target point cloud through inverse distance weighted interpolation, and output semantic labels of the each target point cloud, thereby completing segmentation of key parts.

[0074] Computer-readable storage media may include a program storage area and a data storage area. The program storage area may store an operating system and an application program required for at least one function. The data storage area may store data created based on the use of a point cloud segmentation system for navigating key parts of an oil sample collection robot. Furthermore, the computer-readable storage medium may include high-speed random access memory and may also include memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include memory remotely configured relative to a processor, which can be connected via a network to the point cloud segmentation system for navigating key parts of an oil sample collection robot. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0075] Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present invention, such as... Figure 6 As shown, the device includes a processor 310 and a memory 320. The electronic device may also include an input device 330 and an output device 340. The processor 310, memory 320, input device 330, and output device 340 can be connected via a bus or other means. Figure 6 Taking a bus connection as an example, the memory 320 is the computer-readable storage medium described above. The processor 310 executes various server functions and data processing by running non-volatile software programs, instructions, and modules stored in the memory 320, thereby implementing the key part point cloud segmentation method for oil sample collection robot navigation described in the above embodiment. The input device 330 can receive input digital or character information and generate key signal inputs related to user settings and function control of the key part point cloud segmentation system for oil sample collection robot navigation. The output device 340 may include a display screen or other display device.

[0076] The aforementioned electronic device can execute the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0077] In one implementation, the above-described electronic device is applied to a point cloud segmentation system for key parts of an oil sample collection robot for navigation. As a client, it includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to:

[0078] The original point cloud is acquired, the original point cloud is sample-aware enhanced according to a preset adaptive data enhancer, an enhanced point cloud is generated, and the enhanced point cloud is input into a preset encoder, and the encoder extracts a multi-scale local feature, wherein the encoder comprises a multi-layer SetAbstraction module and a CB-Transformer module.

[0079] The multi-scale local feature is input into a preset C-VLAD module, the C-VLAD module performs weighted residual clustering and cross-level cascade, and generates a comprehensive global descriptor tensor;

[0080] The comprehensive global descriptor tensor is cross-layer jump connected with an up-sampling feature in a decoder, each target point cloud is gradually recovered through inverse distance weighted interpolation, and semantic labels of the target point clouds are output, and segmentation of key parts is completed.

[0081] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and necessary general hardware platforms, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of the embodiments or some parts of the embodiments.

[0082] Finally, it should be noted that: the above examples are used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A point cloud segmentation method for key parts of an oil sample collection robot for navigation, characterized in that, include: The process involves acquiring an original point cloud, performing sample-aware augmentation on the original point cloud using a preset adaptive data augmenter to generate an augmented point cloud, and inputting the augmented point cloud into a preset encoder. The encoder extracts multi-scale local features. The encoder includes a multi-layer SetAbstraction module and a CB-Transformer module. The process of performing sample-aware augmentation on the original point cloud using the preset adaptive data augmenter to generate the augmented point cloud includes: Extract the point-by-point features and overall shape features of the original point cloud; A noise vector is generated based on a Gaussian distribution, and the noise vector is concatenated with the overall shape feature. The transformation matrix is ​​then calculated based on the multilayer perceptron layers. The N copies of the overall shape features are concatenated with the point-by-point features, and a Gaussian noise matrix is ​​introduced to calculate the displacement matrix using a multilayer perceptron. The original point cloud is enhanced based on the displacement matrix and the transformation matrix to obtain an enhanced point cloud. The enhanced point cloud is then constrained according to a preset enhancement loss function to obtain the final enhanced point cloud. The expression for calculating the enhanced point cloud is as follows: ， In the formula, For the enhanced point cloud, The original point cloud, Let be the transformation matrix. It is the displacement matrix; The expression for the enhancement loss function is: ， ， ， ， ， In the formula, To enhance the loss function, To enhance the cross-entropy loss of the point cloud, It is a natural exponential function. Let the cross-entropy loss be the value of point cloud P. The total number of categories, Labels for the real category, Let P be the probability that point cloud P belongs to class c. To enhance the amplitude, To increase the upper limit, These are transitional variables during the calculation process; The multi-scale local features are input into a preset C-VLAD module, which performs weighted residual clustering and cross-level suffixation to generate a comprehensive global descriptor tensor. The comprehensive global descriptor tensor is connected across layers to the upsampled features in the decoder, and each target point cloud is gradually recovered through inverse distance weighted interpolation. The semantic labels of each target point cloud are then output to complete the segmentation of key parts.

2. The point cloud segmentation method for key parts of an oil sample collection robot for navigation according to claim 1, characterized in that, The CB-Transformer module enhances the features of the original point cloud through a two-stage attention mechanism, which includes a first-stage attention mechanism and a second-stage attention mechanism. The first stage attention mechanism specifically includes: constructing global features by fusing center weights and neighborhood point features, and generating neighborhood weights for the second stage attention mechanism after normalization by the softmax function; The second-stage attention mechanism specifically includes: using global features as dynamic weight parameters, and propagating the central node features to each neighborhood space through matrix operations.

3. The point cloud segmentation method for key parts of an oil sample collection robot for navigation according to claim 1, characterized in that, The step of inputting the multi-scale local features into a preset C-VLAD module, wherein the C-VLAD module performs weighted residual clustering and cross-level syntax to generate a comprehensive global descriptor tensor includes: Weighted residual clustering is performed on the multi-scale local features output by each level of the encoder; The feature descriptor vector for each cluster center is calculated as follows: ， In the formula, For feature descriptor vectors, It is a natural exponential function. For the first The feature matrix generated by the layer, As cluster center, For bias terms, This is the weight vector; The global descriptor vectors at each level are expanded in dimension and their features are concatenated to construct a comprehensive global descriptor tensor, expressed as: ， In the formula, To synthesize the global descriptor tensor, For the first Each feature descriptor vector This is a vector concatenation operation.

4. A point cloud segmentation system for key components in an oil sample collection robot navigation system, characterized in that, include: An enhancement module is configured to acquire the original point cloud, perform sample-aware enhancement on the original point cloud according to a preset adaptive data augmenter to generate an enhanced point cloud, and input the enhanced point cloud into a preset encoder. The encoder extracts multi-scale local features, wherein the encoder includes a multi-layer SetAbstraction module and a CB-Transformer module. The step of performing sample-aware enhancement on the original point cloud according to the preset adaptive data augmenter to generate the enhanced point cloud includes: Extract the point-by-point features and overall shape features of the original point cloud; A noise vector is generated based on a Gaussian distribution, and the noise vector is concatenated with the overall shape feature. The transformation matrix is ​​then calculated based on the multilayer perceptron layers. The N copies of the overall shape features are concatenated with the point-by-point features, and a Gaussian noise matrix is ​​introduced to calculate the displacement matrix using a multilayer perceptron. The original point cloud is enhanced based on the displacement matrix and the transformation matrix to obtain an enhanced point cloud. The enhanced point cloud is then constrained according to a preset enhancement loss function to obtain the final enhanced point cloud. The expression for calculating the enhanced point cloud is as follows: ， In the formula, For the enhanced point cloud, The original point cloud, Let be the transformation matrix. It is the displacement matrix; The expression for the enhancement loss function is: ， ， ， ， ， In the formula, To enhance the loss function, To enhance the cross-entropy loss of the point cloud, It is a natural exponential function. Let the cross-entropy loss be the value of point cloud P. The total number of categories, Labels for the real category, Let P be the probability that point cloud P belongs to class c. To enhance the amplitude, To increase the upper limit, These are transitional variables during the calculation process; The generation module is configured to input the multi-scale local features into a preset C-VLAD module, which performs weighted residual clustering and cross-level syntax to generate a comprehensive global descriptor tensor. The segmentation module is configured to perform cross-layer skip connections between the comprehensive global descriptor tensor and the upsampled features in the decoder, gradually recover each target point cloud through inverse distance weighted interpolation, and output the semantic labels of each target point cloud to complete the segmentation of key parts.

Images