A hierarchical lightweight point cloud semantic segmentation method and system fusing geometric residual error

Abstract

The application provides a hierarchical lightweight point cloud semantic segmentation method and system fusing geometric residual, and the method comprises the following steps: selecting a point cloud dataset, preprocessing the point cloud dataset to obtain a point cloud training sample set and a test sample set; inputting the training sample set into a kernel point convolution network to form feature representation data with geometric perception ability; inputting the feature representation data into an encoder layer, explicitly modeling hierarchical context information from local details to global semantics through a parallel multi-branch structure and an adaptive fusion mechanism; connecting the context information and the encoder layer output information in residual connection to output point cloud semantic geometric feature information; inputting the point cloud semantic geometric feature information into a decoder layer to output a semantic prediction result, inputting the semantic prediction result into a point cloud test sample set to be segmented, automatically segmenting the test data and evaluating the precision. The application realizes high-precision automatic semantic segmentation of indoor scene three-dimensional point cloud data.

Description

Technical Field

[0001] This invention relates to the field of 3D point cloud data processing technology for indoor scenes, and more specifically, to a hierarchical lightweight point cloud semantic segmentation method and system that integrates geometric residuals. Background Technology

[0002] With the widespread use of smart homes and indoor robots, and the growing demand for location services, point cloud semantic segmentation technology is receiving increasing attention and has become a crucial component of scene perception. Currently, 3D point cloud semantic segmentation methods are mainly implemented through traditional machine learning and deep learning methods. Traditional point cloud semantic segmentation primarily relies on geometric constraints and statistical rules, followed by manually designed features to obtain the target segmentation result. Based on their working principles, they can be categorized into five types: edge-information-based segmentation methods, region-growing-based segmentation methods, model-fitting-based segmentation methods, attribute-based segmentation methods, and graph optimization-based segmentation methods. Edge-based segmentation methods segment regions by finding edge information of varying brightness. While computationally efficient, these methods are highly sensitive to noise and lack precision. Region-growing methods merge non-adjacent points to form a seed region based on certain criteria, but this process is slow and lacks robustness. Model-fitting methods use cylindrical or spherical models to match corresponding points, offering noise-insensitive and fast inference speeds, but accuracy is difficult without precise prior knowledge. Attribute-based segmentation methods use feature vectors with similar characteristics from clusters to segment point cloud data. While offering some noise resistance, these methods suffer from low processing efficiency and computational overhead. Graph optimization-based segmentation methods transform point clouds into graph-structured data and use graph convolutional networks to synthesize node-related features. This approach maintains some stability under spatial transformations but struggles to represent the relationships between points.

[0003] In summary, traditional point cloud semantic segmentation methods are effective in simple scenarios, but they have poor generalization ability and poor anti-interference ability in complex environments, so they are not suitable for segmentation requirements in complex indoor environments. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a hierarchical lightweight point cloud semantic segmentation method and system that integrates geometric residuals. It aims to solve the problems of edge missegmentation and large number of model parameters in point cloud indoor scene segmentation.

[0005] According to a first aspect of the present invention, a hierarchical lightweight point cloud semantic segmentation method incorporating geometric residuals is provided, comprising the following steps: Select a point cloud dataset, preprocess the point cloud dataset to obtain a point cloud training sample set and a test sample set; The point cloud training sample set is input into a kernel convolutional network to form feature representation data with geometric perception capabilities; The feature representation data with geometric perception capability is input into the encoder layer, and through a parallel multi-branch structure and adaptive fusion mechanism, hierarchical contextual information from local details to global semantics is explicitly modeled. The hierarchical context information is residually concatenated with the encoder layer output information to output point cloud semantic geometric feature information. The semantic geometric feature information of the point cloud is input into the decoder layer, and the semantic prediction result is output. The semantic prediction result is input into the test sample set of the point cloud to be segmented, and the test data is automatically segmented into point clouds and the accuracy is evaluated.

[0006] Based on the above technical solution, the present invention can also be improved as follows.

[0007] Optionally, the preprocessing of the point cloud dataset includes: Perform parameter initialization and data loading, and complete point cloud downsampling and spatial index construction; Establish a label index mapping system and sampling point generation mechanism, dynamically select the center point of the local area based on the set sampling strategy, extract the corresponding point cloud area according to the geometric morphology parameters, implement composite data augmentation processing including geometric transformation and color adjustment, and perform feature selection and normalization operations. By combining point clouds from multiple regions to construct training batches, multi-level pyramid feature representations are generated, and automatic calibration and optimization of batch parameters and the number of neighbors are achieved through statistical analysis.

[0008] Optionally, the step of inputting the point cloud training sample set into a kernel-point convolutional network to form feature representation data with geometric awareness includes: Input point cloud data and generate query points for large-scale point cloud data using the k-nearest neighbor algorithm; A group attention mechanism is introduced to generate modulation coefficients in the kernel dimension by performing multilayer perceptron transformation on the center point features or neighborhood pooling features. A depthwise separable convolutional architecture is adopted to perform channel-wise convolution operations on the modulated weights and neighborhood features.

[0009] Optionally, the input of geometrically perceptive feature representation data into the encoder layer includes: Multiple parallel feature extraction branches are designed to capture contextual information of point cloud feature data at different spatial scales. A learnable branch weight vector is introduced, and attention weights are generated through Softmax normalization. The weighted multi-scale features are concatenated along the channel dimension, and the features are integrated and the dimensions are restored through a fusion layer. Finally, the features are fused with the original input features through scalable residual connections.

[0010] Optionally, the multiple parallel feature extraction branches include: a local branch focusing on fine-grained geometric structures, a mesoscale branch extracting conventional neighborhood features, a large receptive field branch capturing macroscopic regional patterns, and a global branch obtaining scene-level semantic representations through global average pooling and linear transformation, with the number of output channels of each branch dynamically adjusted.

[0011] Optionally, the hierarchical context information and the encoder layer output information are residually connected, and the output point cloud semantic geometric feature information adopts a dual-path structure. The main path compresses feature channels through dimensionality reduction transformation and performs multi-scale neighborhood feature aggregation through configurable geometric convolutional layers. The auxiliary path maintains feature dimension alignment through projection transformation, thereby improving the accurate segmentation of features of objects of different sizes in the point cloud data.

[0012] Optionally, the step of performing a residual connection between the hierarchical context information and the encoder layer output information to output point cloud semantic geometric feature information includes: The number of intermediate channels is optimized to one-quarter of the number of output channels, and a dimensionality-reducing MLP is used to compress the input features to the intermediate dimension; Based on attention grouping configuration, standard convolution or attention-enhanced convolution is dynamically selected to complete adaptive modeling of local neighborhood geometry. The intermediate features are restored to the target output dimension by using an upscaling MLP. At the same time, the auxiliary path performs local max pooling on the input features according to the step size sampling requirements. The feature dimensions are aligned by projection transformation. The outputs of the main path and the auxiliary path are added together and residual connections are formed by activation functions to output high-precision point cloud semantic geometric feature information.

[0013] Optionally, the point cloud semantic geometric feature information is input to the decoder layer, and the output semantic prediction result is used to complete the deep analysis of the point cloud semantic features through multi-level feature fusion, thereby realizing semantic segmentation of point cloud data.

[0014] Optionally, the step of inputting the point cloud semantic geometric feature information output by the encoder layer into the decoder layer and outputting semantic prediction results includes: The high-dimensional features output by the decoder are mapped to the semantic label space of each point. The encoder is responsible for multi-level feature extraction and abstraction, and the decoder is responsible for feature reconstruction and resolution restoration, outputting the semantic prediction result of each point.

[0015] According to a second aspect of the present invention, a hierarchical lightweight point cloud semantic segmentation system incorporating geometric residuals is provided, comprising: The data input and preprocessing module is used to select point cloud datasets, preprocess the point cloud datasets, and obtain point cloud training sample sets and test sample sets. The feature extraction module is used to input the point cloud training sample set into the kernel convolutional network to form feature representation data with geometric perception capabilities; The hierarchical multi-scale context fusion module is used to input geometrically perceptive feature representation data into the encoder layer. Through a parallel multi-branch structure and adaptive fusion mechanism, it explicitly models hierarchical context information from local details to global semantics. The geometric residual module is used to perform residual connection between hierarchical context information and encoder layer output information to output point cloud semantic geometric feature information; The data output module inputs the semantic feature information of the point cloud set into the decoder layer, outputs the semantic prediction result, inputs the semantic prediction result into the test sample set of the point cloud to be segmented, and automatically segments the test data into points and evaluates the accuracy.

[0016] The technical effects and advantages of this invention are as follows: This invention provides a hierarchical lightweight point cloud semantic segmentation method and system that integrates geometric residuals. Through the collaborative design of a deformable kernel point mechanism and attention modulation, it achieves efficient computation while maintaining geometric perception capabilities, providing strong support for point cloud deep feature extraction and enhancing the subsequent feature learning ability of the kernel point convolutional network on point cloud data. By employing a parallel multi-branch structure and adaptive fusion mechanism, it explicitly models hierarchical contextual information from local details to global semantics, improving the ability to recognize and judge global and local features of point cloud data. Through explicit multi-scale feature extraction and adaptive fusion mechanism, it significantly enhances the contextual perception and expressive power of point cloud features while maintaining consistency in input and output dimensions. The dual-path structure improves the accurate segmentation of features of objects of different sizes in point cloud data, further enhancing the accuracy of point cloud semantic segmentation. This invention achieves high-precision automatic semantic segmentation of 3D point cloud data for indoor scenes. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall network structure provided in an embodiment of the present invention; Figure 2 A schematic diagram of the hierarchical context aggregation module provided in an embodiment of the present invention; Figure 3 A schematic diagram of the geometric residual unit provided in an embodiment of the present invention; Figure 4 The figure shows a comparison of the segmentation results of the S3DIS dataset provided in the embodiments of the present invention. In the figure, (a) is the true label image and (b) is the classification result image of the present method. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In recent years, thanks to the significant advancements achieved in tasks such as image classification and object detection using deep learning, more and more researchers are applying deep learning-based methods to the field of 3D point cloud segmentation. Currently, research on deep learning-based 3D point cloud segmentation methods has become a new research hotspot.

[0020] Projection-based point cloud semantic segmentation leverages 2D convolutional neural networks through dimensionality reduction mapping, achieving a balance between computational efficiency and feature representation, but is limited by geometric distortion and viewport dependence. Multi-view Convolutional Neural Networks (MVCNNs) aggregate global shapes through multi-view projection, significantly reducing the burden of 3D convolution; Projection-Based Point Convolution (PPConv) achieves real-time inference on mobile devices with a lightweight 2D convolution-MLP hybrid architecture, but is prone to losing scene edge details due to point cloud coordinate projection distortion.

[0021] Voxel-based methods regularize non-Euclidean point clouds into 3D meshes, enabling parallel computation while expanding the receptive field for semantic segmentation. Early 3D Convolutional Neural Networks (3D-CNNs) employed uniform voxelization and trilinear interpolation to preserve symmetrical details, but their computation time wasted up to 90% of the spacetime voxel count, resulting in significant GPU memory waste. The Fast Point Cloud Segmentation Using Permutohedral Lattices (LatticeNet) method introduces non-uniform voxels from permutohedral lattices, aggregating features across scales using deformable slicing modules. Combined with CRF semantic propagation, this reduces the number of parameters while preserving geometric details.

[0022] Point-based segmentation methods skip regular preprocessing such as voxelization or projection, directly modeling on unordered point clouds, thus avoiding the loss of geometric details and additional computational overhead, and have become the focus of current research. These methods can be summarized into three main lines: models based on multilayer perceptrons and convolutions extract point-wise features using shared weights; PointNet (Deep Learning On Point Sets For 3D Classification and Segmentation) achieves permutation invariance through max pooling but lacks local structure; PointNet++ (Deep Hierarchical Feature Learning On Point Sets In a Metric Space) introduces hierarchical sampling and ball query aggregation to enhance neighborhood awareness but is density-sensitive; and graph-based models explicitly model point clouds as dynamic graphs. Dynamic Graph CNN for Learning On Point Clouds (DGCNN) constructs a k-NN edge point set in real time during training and updates nodes with EdgeConv, taking into account both local geometry and global topology, and is particularly effective for fine indoor boundaries. Generative Adversarial-Driven Cross-Aware Network (GACNet) further embeds graph attention adaptive weighted neighborhoods to improve segmentation sharpness, but its memory usage increases sharply at the scale of millions of points. Attention-based models capture long-range dependencies through self-attention. PointTransformer (Learning On Point Clouds with Self-Attention) integrates position encoding into high-dimensional similarity to achieve context awareness, but its quadratic complexity limits the number of points.

[0023] However, the semantic prediction results mentioned above only focus on exploring local geometric structures and ignore the global structure of indoor scenes. They lack multi-scale feature processing and fusion methods, making it difficult to dynamically adjust according to the differences in object scale, which leads to edge missegmentation. In addition, most models have a large number of parameters, require a lot of computing power, and have high memory consumption due to time-consuming inference.

[0024] Understandably, given the deficiencies in the background technology, this invention proposes a hierarchical lightweight point cloud semantic segmentation method that integrates geometric residuals, specifically as follows: Figure 1 As shown, it includes the following steps: Step S1: Select a point cloud dataset, preprocess the point cloud dataset to obtain a point cloud training sample set and a test sample set; The preprocessing of the point cloud dataset to obtain the point cloud training sample set and test sample set includes: First, parameter initialization and data loading are performed to complete point cloud downsampling and spatial index construction. Subsequently, a label index mapping system and a sampling point generation mechanism are established; the center point of a local region is dynamically selected based on the set sampling strategy, and the corresponding point cloud region is extracted based on the geometric morphology parameters; Then, a composite data augmentation process involving geometric transformation and color adjustment is performed, followed by feature selection and normalization operations. Then, the point clouds from multiple regions are combined to construct training batches and generate multi-level pyramid feature representations; Finally, through statistical analysis, automatic calibration and optimization of batch parameters and the number of neighbors are achieved, realizing the complete processing of 3D point cloud preprocessing data.

[0025] It should be noted that the point cloud dataset mentioned includes, but is not limited to, the S3DIS dataset. The S3DIS dataset is labeled with thirteen semantic categories: ceiling (ceiling, c_l), floor (floor, f_r), wall (wall, w_l), beam (beam, b_m), column (column, c_n), window (window, w_n), door (door, d_r), table (table, t_b), chair (chair, c_r), sofa (sofa, s_f), bookshelf (bookcase, b_c), and device (clutter, c_t).

[0026] For the preprocessing of the S3DIS dataset, this invention employs a hierarchical sampling strategy to adapt to the semantic segmentation task of large-scale indoor scene point clouds. The specific steps for preprocessing the S3DIS dataset are as follows: Step 1.1: Data Loading and Feature Extraction; Using the S3DIS dataset as the input source, which includes the 3D coordinates (XYZ), color features (RGB), and semantic labels for each point. In preprocessing, we use the original point cloud coordinates and color features as the basic feature vector for network input; Step 1.2: Scene Segmentation and Memory Optimization; Due to the limited GPU memory capacity, complete architectural scenes cannot be directly input into the network. Therefore, a spatial cube partitioning strategy is adopted, dividing each large scene into multiple regular 3D block regions. During the training phase, input regions are selected from these blocks through random sampling. Each region contains a limited number of points, ensuring that the computational load of each forward propagation is controllable. Step 1.3: Data Augmentation and Sampling during Training; To improve model robustness and prevent overfitting, various data augmentation operations are implemented during training: random rotation, scaling, and flipping to simulate different viewing angles and scene scale changes; coordinate jitter and color perturbation to enhance robustness to noise; height normalization and floor centering to eliminate height deviations between scenes; random point discarding (with a 12.5% ​​probability) and random color loss to simulate sensor loss. Step 1.4: Full-scene coverage during the testing phase; During the testing phase, a rule-based sliding window strategy is used to perform dense coverage sampling across the entire scene. Each sampling region is evenly distributed in a grid pattern to ensure that every point in the scene is contained within at least one sampling region. The final full-scene semantic segmentation is achieved through a voting mechanism based on the prediction results of multiple regions. Step 1.5: Ensuring Region Integrity; To avoid small fragmented regions at scene edges affecting training results, a region merging strategy is implemented during region partitioning. When a region has too few points, it is merged into a larger, adjacent region to ensure that each input region contains sufficient geometric and semantic information.

[0027] Step 2: Input the point cloud training sample set into the kernel convolutional network to form feature representation data with geometric awareness; In this embodiment, the point cloud training sample set is input into the feature extraction layer of the kernel point convolutional network. The feature extraction layer constructs a deep feature extraction network to form a feature representation with geometric perception capabilities, thereby enhancing the kernel point convolutional network's ability to learn features from point cloud data. The specific steps for inputting the point cloud training sample set into a kernel-point convolutional network to form feature representation data with geometric awareness are as follows: Step 2.1: Construct the kernel point neighborhood; input point cloud data, and generate query points for large-scale point cloud data using the k-nearest neighbor algorithm. For each query point ( First, collect support points within its spherical neighborhood. Construct a local graph structure. Initialize multi-scale shell-based kernel points and calculate the relative spatial relationship between each kernel point and its neighbors. Dynamically assign the influence weights of kernel points on their neighbors using a linear or Gaussian influence mode to form a geometrically adaptive convolutional receptive field. This process can be represented as: (1) in, The weight of the influence of the core point on its neighboring points. The location of the kth core point. For the influence function, To affect the radius, For the j-th neighborhood support point, For the spherical neighborhood of the query point, is the Euclidean distance between the neighboring point and the core point.

[0028] Step 2.2: Perform attention modulation and weight adaptation; introduce a grouped attention mechanism to enhance feature adaptability. By performing a multilayer perceptron transformation on the center point features (or neighborhood pooling features), modulation coefficients in the kernel dimension are generated. These coefficients are normalized by activation functions (sigmoid / softmax / tanh) and then multiplied element-wise with the basic convolutional weights to achieve dynamic weight adjustment. (2) in, For attention modulation coefficients, For multilayer perceptron mapping function The feature vector of the query point, The kernel weights are after attention modulation. Based on the convolution weights, This represents the attention modulation coefficient.

[0029] Step 2.3: Depthwise Separable Convolution and Feature Aggregation; A depthwise separable convolution architecture is adopted, and the modulated weights and neighborhood features are convolved channel by channel. Through a nearest neighbor kernel matching strategy, each neighborhood point is only affected by its nearest kernel point, significantly reducing computational complexity. Finally, neighborhood information is aggregated through weighted summation: (3) in, For the first Query points The output characteristics, For query point The set of neighborhood point indices The geometric influence weight of a neighboring point relative to its nearest kernel point. For the first The modulated convolution weights corresponding to each kernel point For neighborhood points The nearest kernel index, As a support point Input feature vector This represents channel-level multiplication, where ∑ represents the weighted summation of all support points in the neighborhood.

[0030] This step, through the collaborative design of kernel point deformability mechanism and attention modulation, achieves efficient computation while maintaining geometric perception capabilities, providing strong support for point cloud depth feature extraction and enhancing the kernel point convolutional network's ability to subsequently learn features from point cloud data.

[0031] Step 3: Input the feature representation data with geometric perception capabilities into the encoder layer. Through parallel multi-branch structure and adaptive fusion mechanism, explicitly model the hierarchical context information from local details to global semantics, and improve the ability to recognize and judge global and local features of point cloud data. To address the need for multi-scale geometric pattern perception in point cloud scenarios, this invention designs a hierarchical multi-scale context fusion module. This module explicitly models hierarchical contextual information from local details to global semantics through a parallel multi-branch structure and an adaptive fusion mechanism.

[0032] The hierarchical context aggregation module achieves efficient modeling of contextual information from different receptive fields through a parallel multi-branch structure. Specifically, for example... Figure 2 As shown, the hierarchical context aggregation module comprises three local feature branches (corresponding to small, medium, and large receptive fields, respectively) and a global pooling branch. Each branch extracts hierarchical feature representations through an independent feedforward network. The local branches capture features ranging from fine-grained geometric details to macroscopic structures using receptive fields of different scales, while the global branch obtains scene-level semantic information through average pooling and linear transformation. All branch outputs are adaptively fused using learnable weights, then integrated through feature concatenation and a fusion layer. Finally, they are added to the original input features via scalable residual connections, significantly enhancing the multi-scale expressive power of point cloud features while maintaining consistency in input and output dimensions.

[0033] The specific steps for inputting geometrically perceptual feature representation data into the encoder layer are as follows: Step 3.1: Construction of Multi-Scale Feature Extraction Branches; Four parallel feature extraction branches are designed to capture contextual information of point cloud feature data at different spatial scales. Three local scale branches employ different receptive field configurations: the local branch (scale=0.5) focuses on fine-grained geometric structures, the mid-scale branch (scale=1.0) extracts regular neighborhood features, and the large receptive field branch (scale=2.0) captures macroscopic region patterns. The global branch obtains scene-level semantic representations through global average pooling and linear transformation. The number of output channels for each branch is dynamically adjusted and ensured to be a multiple of 16 to optimize computational efficiency.

[0034] Step 3.2: Adaptive weighted fusion mechanism; to balance the contribution of features at different scales, a learnable branch weight vector is introduced. Attention weights are generated using Softmax normalization: (4) in, For attention weights, These are the weights of the original vector.

[0035] The weighted multi-scale features are represented as follows: (5) in, This represents the weighted multi-scale features. These are the outputs of the local scale branches. For global branch output, , , , These are represented as four learnable branch weight vectors.

[0036] Step 3.3: Feature Fusion and Residual Enhancement; The weighted multi-scale features are concatenated along the channel dimension, and feature integration and dimensionality restoration are performed through a fusion layer (UnaryBlock). Finally, the features are fused with the original input features through scalable residual connections. (6) in, Output features for the module. For learnable residual scaling factors, This represents the original input characteristics of the module. This represents a multi-layer perceptron in the fusion layer. The fusion features are obtained by splicing after multi-scale weighting.

[0037] This step significantly enhances the context awareness and expressiveness of point cloud features by using explicit multi-scale feature extraction and adaptive fusion mechanisms, while maintaining consistency in input and output dimensions. It is suitable for semantic understanding tasks in complex indoor and outdoor scenes.

[0038] Step 4: Perform residual connection between the hierarchical context information and the encoder layer output information to output point cloud semantic geometric feature information.

[0039] In this embodiment, hierarchical context information is residually concatenated with the encoder layer output information to output high-precision point cloud semantic geometric feature information. This step adopts a dual-path structure. The main path compresses feature channels through dimensionality reduction transformation and performs multi-scale neighborhood feature aggregation through configurable geometric convolutional layers. The auxiliary path maintains feature dimension alignment through projection transformation, improving the accurate segmentation of features of objects of different sizes in point cloud data and further improving the semantic segmentation accuracy of point clouds.

[0040] To address the lightweight requirements in point cloud scenarios, this invention designs a geometric residual unit for residual connections, further enhancing the learning ability of point cloud geometric features. The geometric residual unit is an optimized residual module specifically designed for learning point cloud geometric features. This unit employs a dual-path structure: the main path achieves geometrically perceptual feature transformation through a "dimensionality reduction-convolution-dimensionality increase" process, while the auxiliary path maintains feature dimension alignment through projection transformation.

[0041] Specifically, such as Figure 3 As shown, the geometric residual unit adopts a dual-path structure: the main path first compresses feature channels through dimensionality reduction transformation, and then performs multi-scale neighborhood feature aggregation through configurable geometric convolutional layers. The attention-enhanced version balances computational overhead and feature representation capability through a dynamically optimized grouping mechanism. The convolutional output, after passing through a merged normalization-activation layer, is then subjected to dimensionality increase transformation to restore the channel dimension. The auxiliary path processes input features through learnable projection transformation and supports local pooling operations during stride sampling to ensure alignment with the feature dimension of the main path. Finally, the outputs of the two paths are fused through residual connections, enhancing deep feature representation while preserving the original geometric information, achieving an optimized balance between computational efficiency and geometric perception capability.

[0042] The specific steps for performing residual concatenation between the hierarchical context information and the encoder layer output information to output point cloud semantic geometric feature information are as follows: Step 4.1: Intermediate channel optimization and dimensionality reduction; To avoid information bottlenecks, the number of intermediate channels is optimized to one-quarter of the number of output channels and not less than 32. Based on the difference between input and output dimensions, dimensionality reduction MLP is selectively used to compress input features to intermediate dimensions to reduce subsequent computational overhead. Step 4.2: Adaptive Convolution Selection and Feature Aggregation; Based on attention grouping configuration, dynamically select standard convolution (KPConvD) or attention-enhanced convolution (KPConvX). The latter balances computational complexity and feature interaction capability through an optimized grouping mechanism, achieving adaptive modeling of local neighborhood geometry. After the convolution output, the computation process is simplified through a merged normalization-activation layer; Step 4.3: Feature Upscaling and Residual Connection; The intermediate features are restored to the target output dimension using an upscaling MLP. Simultaneously, the auxiliary path performs local max pooling on the input features according to the step size sampling requirements and aligns the feature dimensions through projection transformation. Finally, the outputs of the two paths are added together, and a residual connection is formed through an activation function. This enhances feature representation while preserving the original geometric information, outputting high-precision point cloud semantic geometric feature information.

[0043] Step 5: Input the point cloud semantic geometric feature information output by the encoder layer into the decoder layer and output the semantic prediction result; This step achieves deep parsing of point cloud semantic features through multi-level feature fusion, thereby realizing semantic segmentation of point cloud data. This invention employs a symmetrical encoder-decoder architecture. The encoder progressively downsamples through stacked geometric residual units, with each unit using a dual-path structure to extract and fuse local geometric features. A hierarchical context aggregation module is integrated into the deep features, explicitly modeling multi-scale spatial dependencies from local details to global semantics through a multi-branch parallel structure, constructing a feature pyramid with rich contextual information. The decoder progressively restores the point cloud resolution through nearest-neighbor upsampling and fuses feature representations from each level of the encoder through skip connections, achieving progressive aggregation and enhancement of multi-scale features. At the network's end, a lightweight fully connected layer is used for category prediction, outputting an accurate semantic label for each point.

[0044] The specific steps for inputting the point cloud semantic geometric feature information output by the encoder layer into the decoder layer and outputting the semantic prediction result are as follows: Step 5.1: Decoder reconstructs and fuses hierarchical context information; the decoder adopts a symmetrical structure and gradually restores the spatial resolution of the point cloud through nearest neighbor upsampling operations; Step 5.2: In each upsampling stage, the feature representations of the corresponding level of the encoder are fused through skip connections, combining deep semantic information with shallow geometric details; Step 5.3: The fused features are subjected to dimension alignment and feature enhancement through a linear transformation layer. Optionally, a decoder convolutional layer is added to further refine the feature representation, realizing the effective aggregation and reconstruction of multi-level features, and outputting a semantic prediction result file. Step 5.4: End-to-end semantic prediction architecture; a lightweight fully connected layer is used at the end of the network for category prediction, mapping the high-dimensional features output by the decoder to the semantic label space of each point. The entire network adopts an end-to-end training approach, with the encoder responsible for multi-level feature extraction and abstraction, and the decoder responsible for feature reconstruction and resolution restoration, ultimately outputting the semantic prediction result for each point. This symmetric architecture ensures a balance between maintaining geometric structure and semantic information transmission, while improving the expressive power of features and the generalization performance of the model through residual connections and multi-scale fusion mechanisms.

[0045] Step 6: Input the semantic prediction results into the point cloud test sample set to be segmented, and automatically segment the test data into point clouds and evaluate the accuracy.

[0046] Examples of the results of automatic semantic segmentation of 3D point clouds in indoor scenes according to this invention are as follows: Figure 4As shown in the figure, (a) is the real label (left) and (b) is the classification result using this method (right). It can be seen that the hierarchical lightweight point cloud semantic segmentation method with geometric residual fusion described in this embodiment of the invention can stably and accurately segment indoor scene point clouds with semantic segmentation targets of different types, quantities and sizes. In addition, this invention provides a hierarchical lightweight point cloud semantic segmentation system that integrates geometric residuals, comprising: The data input and preprocessing module is used to select point cloud datasets, preprocess the point cloud datasets, and obtain point cloud training sample sets and test sample sets. The feature extraction module is used to input the point cloud training sample set into the kernel convolutional network to form feature representation data with geometric perception capabilities; The hierarchical multi-scale context fusion module is used to input geometrically perceptive feature representation data into the encoder layer. Through a parallel multi-branch structure and adaptive fusion mechanism, it explicitly models hierarchical context information from local details to global semantics. The geometric residual module is used to perform residual connection between hierarchical context information and encoder layer output information to output point cloud semantic geometric feature information; The data output module inputs the semantic geometric feature information of the point cloud into the decoder layer, outputs the semantic prediction result, inputs the semantic prediction result into the test sample set of the point cloud to be segmented, performs automatic point cloud segmentation on the test data and evaluates the accuracy.

[0047] It is understood that the hierarchical lightweight point cloud semantic segmentation system with geometric residual fusion provided by the present invention corresponds to the hierarchical lightweight point cloud semantic segmentation method with geometric residual fusion provided in the foregoing embodiments. The relevant technical features of the hierarchical lightweight point cloud semantic segmentation system with geometric residual fusion can be referred to the relevant technical features of the hierarchical lightweight point cloud semantic segmentation method with geometric residual fusion, and will not be repeated here.

[0048] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0049] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0050] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A hierarchical lightweight point cloud semantic segmentation method incorporating geometric residuals, characterized in that, Includes the following steps: Select a point cloud dataset, preprocess the point cloud dataset to obtain a point cloud training sample set and a test sample set; The point cloud training sample set is input into a kernel convolutional network to form feature representation data with geometric perception capabilities; The feature representation data with geometric perception capability is input into the encoder layer, and through a parallel multi-branch structure and adaptive fusion mechanism, hierarchical contextual information from local details to global semantics is explicitly modeled. The hierarchical context information is residually concatenated with the encoder layer output information to output point cloud semantic geometric feature information. The semantic geometric feature information of the point cloud is input into the decoder layer, and the semantic prediction result is output. The semantic prediction result is input into the test sample set of the point cloud to be segmented, and the test data is automatically segmented into point clouds and the accuracy is evaluated.

2. The hierarchical lightweight point cloud semantic segmentation method based on geometric residuals as described in claim 1, characterized in that, The preprocessing of the point cloud dataset includes: Perform parameter initialization and data loading, and complete point cloud downsampling and spatial index construction; Establish a label index mapping system and sampling point generation mechanism, dynamically select the center point of the local area based on the set sampling strategy, and extract the corresponding point cloud area according to the geometric morphology parameters; implement composite data augmentation processing including geometric transformation and color adjustment, and perform feature selection and normalization operations. By combining point clouds from multiple regions to construct training batches, multi-level pyramid feature representations are generated, and automatic calibration and optimization of batch parameters and the number of neighbors are achieved through statistical analysis.

3. The hierarchical lightweight point cloud semantic segmentation method based on geometric residuals as described in claim 1, characterized in that, The step of inputting the point cloud training sample set into a kernel-point convolutional network to form feature representation data with geometric awareness includes: Input point cloud data and generate query points for large-scale point cloud data using the k-nearest neighbor algorithm; A group attention mechanism is introduced to generate modulation coefficients in the kernel dimension by performing multilayer perceptron transformation on the center point features or neighborhood pooling features. A depthwise separable convolutional architecture is adopted to perform channel-wise convolution operations on the modulated weights and neighborhood features.

4. The hierarchical lightweight point cloud semantic segmentation method based on geometric residuals as described in claim 1, characterized in that, The step of inputting geometrically perceptual feature representation data into the encoder layer includes: Multiple parallel feature extraction branches are designed to capture contextual information of point cloud feature data at different spatial scales. A learnable branch weight vector is introduced, and attention weights are generated through Softmax normalization. The weighted multi-scale features are concatenated along the channel dimension, and the features are integrated and the dimensions are restored through a fusion layer. Finally, the features are fused with the original input features through scalable residual connections.

5. The hierarchical lightweight point cloud semantic segmentation method based on geometric residuals according to claim 4, characterized in that, The multiple parallel feature extraction branches include: a local branch focusing on fine-grained geometric structures, a mesoscale branch extracting conventional neighborhood features, a large receptive field branch capturing macroscopic regional patterns, and a global branch obtaining scene-level semantic representations through global average pooling and linear transformation. The number of output channels for each branch is dynamically adjusted.

6. The hierarchical lightweight point cloud semantic segmentation method based on geometric residuals according to claim 1, characterized in that, The hierarchical context information is residually connected with the encoder layer output information, and the output point cloud semantic geometric feature information adopts a dual-path structure. The main path compresses feature channels through dimensionality reduction transformation and performs multi-scale neighborhood feature aggregation through configurable geometric convolutional layers. The auxiliary path maintains feature dimension alignment through projection transformation, thereby improving the accurate segmentation of features of objects of different sizes in point cloud data.

7. The hierarchical lightweight point cloud semantic segmentation method based on geometric residuals according to claim 1, characterized in that, The step of performing a residual connection between the hierarchical context information and the encoder layer output information to output point cloud semantic geometric feature information includes: The number of intermediate channels is optimized to one-quarter of the number of output channels, and a dimensionality-reducing MLP is used to compress the input features to the intermediate dimension; Based on attention grouping configuration, standard convolution or attention-enhanced convolution is dynamically selected to complete adaptive modeling of local neighborhood geometry. The intermediate features are restored to the target output dimension by using an upscaling MLP. At the same time, the auxiliary path performs local max pooling on the input features according to the step size sampling requirements. The feature dimensions are aligned by projection transformation. The outputs of the main path and the auxiliary path are added together and residual connections are formed by activation functions to output the semantic geometric feature information of the point cloud.

8. The hierarchical lightweight point cloud semantic segmentation method based on geometric residuals according to claim 1, characterized in that, The semantic geometric feature information of the point cloud is input into the decoder layer, and the semantic prediction result is output. Through multi-level feature fusion, a deep analysis of the semantic features of the point cloud is completed, thereby realizing semantic segmentation of the point cloud data.

9. The hierarchical lightweight point cloud semantic segmentation method based on geometric residuals according to claim 1, characterized in that, The step of inputting the point cloud semantic geometric feature information output from the encoder layer into the decoder layer and outputting semantic prediction results includes: The high-dimensional features output by the decoder are mapped to the semantic label space of each point. The encoder is responsible for multi-level feature extraction and abstraction, and the decoder is responsible for feature reconstruction and resolution restoration, outputting the semantic prediction result of each point.

10. A hierarchical lightweight point cloud semantic segmentation system incorporating geometric residuals, characterized in that, include: The data input and preprocessing module is used to select point cloud datasets, preprocess the point cloud datasets, and obtain point cloud training sample sets and test sample sets. The feature extraction module is used to input the point cloud training sample set into the kernel convolutional network to form feature representation data with geometric perception capabilities; The hierarchical multi-scale context fusion module is used to input geometrically perceptive feature representation data into the encoder layer. Through a parallel multi-branch structure and adaptive fusion mechanism, it explicitly models hierarchical context information from local details to global semantics. The geometric residual module is used to perform residual connection between hierarchical context information and encoder layer output information to output point cloud semantic geometric feature information; The data output module inputs the semantic geometric feature information of the point cloud into the decoder layer, outputs the semantic prediction result, inputs the semantic prediction result into the test sample set of the point cloud to be segmented, performs automatic point cloud segmentation on the test data and evaluates the accuracy.

Images