A method and system for fusing granular particle generation with curve-guided point cloud representation

Abstract

This invention discloses a method and system for point cloud representation that integrates sphere-based and curve-guided methods. The method includes: acquiring raw point cloud data; dividing the raw point cloud data into spheres and sampling boundary points for each generated sphere to obtain sphere boundary sampling points; performing multiple downsampling operations on the sphere boundary sampling points; performing point cloud weaving on the compressed point cloud data at each scale; extracting local point cloud features and curve features from the compressed point cloud data and weaving lines at each scale; and then fusing the local and curve features at each scale to obtain point-line fused features at different scales, forming multi-scale point cloud features. This invention's method can achieve lightweight, high-precision, and high-efficiency 3D point cloud representation while maintaining the integrity of the point cloud's geometric structure.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and 3D data processing, and specifically relates to a point cloud representation method and system that integrates particle spheres and curve guidance. Background Technology

[0002] With the development of 3D sensing technologies such as LiDAR, structured light, and depth cameras, point cloud data has become a core foundational data source for fields such as robot navigation, autonomous driving, 3D reconstruction, and intelligent terminal perception. However, raw point clouds are typically massive in quantity, sparsely and unevenly distributed, and contain noise. Directly inputting them into deep neural networks leads to excessive computational load, large inference latency, and severe resource consumption. Therefore, point cloud downsampling and lightweight representation techniques have become an important research direction for both industry and academia.

[0003] Currently, mainstream point cloud sampling methods include random sampling, voxel grid sampling, farthest point sampling (FPS), cluster sampling (K-means, etc.), and granular ball sampling (GBS), which has been proposed in recent years. While random and voxel sampling methods are fast, they cannot effectively preserve key geometric structures such as surface boundaries. FPS can maintain uniformity but has high computational complexity, making it unsuitable for real-time applications. GBS possesses adaptive region partitioning and near-linear time complexity, and can generate many small spheres in boundary regions, thus exhibiting some robustness to noise and boundaries. However, existing GBS methods are mainly used for sample selection in data classification, rather than for optimizing 3D geometry, and have not achieved synergy with multi-scale feature extraction in point cloud neural networks. On the other hand, existing deep point cloud networks (such as PointNet++, DGCNN, and curve-guided cross-scale weaving methods) can effectively learn local geometry, but they rely on a large number of raw point inputs, still incurring high computational costs, being sensitive to noise, and lacking a front-end adaptive sampling mechanism. Therefore, the main challenge facing current technology is how to construct a new technical solution that can efficiently downsample while maintaining the key geometric structure and can directly collaborate with lightweight point cloud feature fusion networks.

[0004] Therefore, there is an urgent need for a unified technical solution that combines particle-sphere adaptive structure downsampling with cross-scale point cloud weaving networks to simultaneously achieve efficient, robust, lightweight, and structure-preserving 3D point cloud representations. Summary of the Invention

[0005] This invention aims to overcome the problems of insufficient geometric preservation capability of existing point cloud sampling methods, large computational load and sensitivity to noise in deep point cloud networks, insufficient adaptability of particle sphere sampling in 3D geometric processing, and lack of front-end data structure optimization in cross-scale feature fusion networks. It proposes an efficient 3D point cloud representation method that combines particle sphere adaptive downsampling with curve-guided cross-scale point cloud weaving network, which realizes structured compression of the original point cloud, extraction of geometric key points, and lightweight high-precision feature modeling, thereby significantly improving the efficiency and robustness of point cloud processing.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A point cloud representation method integrating particle sphere compilation and curve guidance includes:

[0008] Step 1: Obtain raw point cloud data;

[0009] Step 2: Divide the original point cloud data into spheres and sample the boundary points of each generated sphere to obtain the sphere boundary sampling points;

[0010] Step 3: Perform multiple downsampling operations on the particle boundary sampling points. Represent the particle boundary sampling points and the downsampled data obtained from each downsampling operation as point cloud compressed data at different scales. Perform the following process on the point cloud compressed data at each scale:

[0011] S31: Weave point cloud compressed data based on the RRT-P algorithm to generate weave lines. ;

[0012] S32: Compress point cloud data and braided lines Feature extraction is performed separately to generate local features of the point cloud. and curve characteristics ;

[0013] S33: Based on the point-line attention mechanism, feature fusion is performed on the local features and curve features of the point cloud to generate point-line fused features at the specified scale;

[0014] Step 4: The point and line fusion features of different scales form the multi-scale features of the point cloud.

[0015] Furthermore, step 2 specifically includes:

[0016] S21: Randomly select two points from the original point cloud data as center points. Based on the Euclidean distance between the points in the original point cloud data and these two center points, split the original point cloud data into two sub-spheres.

[0017] S22: For each sub-sphere, determine whether the splitting termination condition is met. If it is met, the sub-sphere stops splitting and is placed into the sphere set. If the splitting termination condition is not met, randomly select two points from the sub-sphere as new center points, and split the sub-sphere into two new sub-spheres according to the Euclidean distance between the points in the sub-sphere and the two new center points.

[0018] S23: Repeat S22 until all sub-spheres meet the splitting termination condition, thereby generating the final sphere set;

[0019] S24: For each particle in the final particle set, determine the intersection point of the particle with the positive and negative directions of each coordinate axis in the point cloud space dimension, and find the actual point in each particle that is closest to the intersection point as the sampling point, thereby obtaining the boundary sampling point of each particle.

[0020] Furthermore, step S31 specifically includes:

[0021] S311: Use the FPS algorithm to determine K sampling points from the compressed point cloud data;

[0022] S312: Each pair of sampling points in the K sampling points forms a sampling point pair, and the Euclidean distance between the two sampling points in each sampling point pair is determined;

[0023] S313: Select L pairs of sampling points whose Euclidean distance satisfies the distance threshold to form the starting point set. and the final set ,in, Indicates the first that satisfies the distance threshold Group sampling point pairs, ;

[0024] S314: Perform point cloud weaving on the L sets of sampling point pairs to obtain L weaving lines, wherein, for each set of sampling point pairs Perform the following knitting process:

[0025] S3141: Initialize the node set Empty;

[0026] S3142: with Begin weaving from the starting point, using the starting point... Encapsulate it as the initial node, denoted as a triple. ,in, The coordinates of the starting point. The cost representing the starting point. The parent node representing the starting point is added to the node set. ;

[0027] S3143: Randomly sample a point from the compressed point cloud data. From the set of nodes Select with nearest point ;

[0028] S3144: Calculation and The direction vector between them is denoted as the current direction. ,Will and The direction vector between them is denoted as the global direction. Determine if the angle between the current direction and the global direction meets an angle threshold. If it does, proceed along the current direction. Generate new points If the angle threshold is not met, then the current direction will be changed. Adjusted to And generate new points along the adjusted current direction. ;

[0029] S3145: Find and match points in compressed point cloud data. The closest point is the projection point. ;

[0030] S3146: Traversing the Node Set Among all nodes, select the node with the minimum total cost as the projection point. The parent node is denoted as The total cost is the cost from the starting point to the parent node and the cost between the parent node and the parent node. The sum of the distance values ​​between them is used to record the total cost as the projection point. The cost Thus, a new node is obtained. , A pointer to the parent node of the new node;

[0031] S3147: Move the node Add to node set middle;

[0032] S3148: Repeat S3143 to S3147 above until the termination condition is met, and finally reach the endpoint. Encapsulate as a node and add to the node collection middle;

[0033] S3149: In the set of nodes By backtracking from the end point to the parent node pointer of the starting point, the node coordinate sequence is extracted, thus obtaining the result from... arrive The braided yarn.

[0034] Furthermore, step S32 specifically includes:

[0035] The Set Abstraction module in PointNet++ is used to extract features from compressed point cloud data to obtain local features of the point cloud. ;

[0036] The curve features are obtained by extracting the features of each braided thread using a local aggregation coding method. ,in,

[0037] , This indicates max pooling. Indicates a fully connected transformation. Indicates searching within a given radius neighborhood points, This indicates braided yarn, where L represents the number of braided yarns. Indicates the number of dots on the weaving thread. Indicates the number of feature channels.

[0038] Furthermore, step S33 specifically includes:

[0039] S331: For braided yarn Feature fusion is performed along both the weaving line dimension and the in-line point dimension to obtain the line-level structure features. and features within the line ;

[0040] S332: Employs a dual-branch structure. In the first branch, local point cloud features are used as queries, with line-level structural features... As keys and values, attention-weighted fusion is performed to obtain the first point-line fusion feature. In the second branch, local features of the point cloud are used as queries, with in-line features as the query. As keys and values, attention-weighted fusion is performed to obtain the second point-line fusion feature. ;

[0041] S333: Merge the first point and line feature Second point line fusion feature Perform feature splicing;

[0042] S334: Perform a linear transformation on the feature concatenation result and combine it with the local features of the point cloud. Residual connections are performed to obtain the final point-line fusion features corresponding to the scale. .

[0043] Furthermore, the multi-scale features of point clouds include Layer features, the The scale of the layer features decreases sequentially from top to bottom. The method further includes: Step 5: Generating fused features by further cross-scale fusion of the multi-scale features of the point cloud, wherein, for the current layer... Perform the following process:

[0044] S51: Perform self-attention calculation on the current layer features to obtain the attention score corresponding to each feature point in the current layer features;

[0045] S52: Highest score in attention extraction 10 feature points are used as key points, and key point features are extracted;

[0046] S53: Aggregate the key point features with the lower layer features using PointTransformer to obtain the initial aggregated features corresponding to the current layer;

[0047] S54: Concatenate the initial aggregated features with the lower-level features along the channel dimension to generate the final fused features corresponding to the current layer. .

[0048] Furthermore, for point cloud segmentation tasks, from the current layer Initially, the lower layer is Each layer executes steps S51 to S54 to obtain the corresponding final aggregated features. Then Perform feature propagation operations, use the features as lower-level features, and... Decrease by 1, continue executing S51 to S54, repeat the process until the final fusion feature corresponding to the top layer is obtained. The result of feature propagation of the final fusion feature corresponding to the top layer is used to generate point cloud segmentation results through MLP.

[0049] Furthermore, for point cloud classification tasks, Take 1, that is, use the first layer and the second layer. Layer features are fused across scales to generate fused features. The fused features are then passed through MLP and then through max pooling and average pooling respectively. The results of max pooling and average pooling are multiplied and then passed through MLP again to obtain the point cloud classification result.

[0050] This invention also provides a point cloud representation system that integrates granular and curve-guided methods, the system comprising:

[0051] The data acquisition module is used to acquire raw point cloud data;

[0052] The sphere generation and adaptive downsampling module is used to divide the original point cloud data into spheres and sample the boundary points of each generated sphere to obtain the sphere boundary sampling points.

[0053] The multi-scale point cloud weaving and fusion module includes a downsampling unit, a particle-sphere point cloud weaving unit, a multi-scale feature extraction unit, a point-line fusion unit, and a cross-scale iterative fusion unit.

[0054] The downsampling unit performs multiple downsampling operations on the particle boundary sampling points to obtain downsampled data. The particle boundary sampling points and the downsampled data obtained from each downsampling operation represent point cloud compressed data at different scales.

[0055] The granular point cloud weaving unit uses the RRT-P algorithm to weave compressed point cloud data at each scale, generating weaving lines corresponding to each scale. ;

[0056] The multi-scale feature extraction unit extracts the point cloud compressed data and the braided lines corresponding to each scale. Feature extraction is performed separately to generate local point cloud features corresponding to each scale. and curve characteristics ;

[0057] The point-line fusion unit performs feature fusion on the local features and curve features of the point cloud corresponding to each scale according to the point-line attention mechanism, and generates point-line fusion features corresponding to each scale. The point-line fusion features of different scales constitute the multi-scale features of the point cloud.

[0058] The cross-scale iterative fusion unit generates fused features by fusing multi-scale features of point clouds across scales.

[0059] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the point cloud representation method of fusing and compiling spheres and curve-guided methods as described above.

[0060] The beneficial effects of this invention are:

[0061] 1. This invention uses the collaborative design of a particle-sphere adaptive downsampling structure and a curve-guided cross-scale braided fusion network to enable the original point cloud to maintain its key geometric structure while significantly compressing its size. Moreover, the particle-sphere hierarchical splitting mechanism reduces the interference of noise points, making the downsampling results more stable and reliable.

[0062] 2. The point cloud weaving method of the present invention guides the interaction of cross-scale features through the direction of geometric curves, so that the global information of the coarse scale and the local geometry of the fine scale work together, thereby enabling the lightweight point cloud to still have good geometric reconstruction performance and strong expressive ability.

[0063] 3. The reduced number of input points eliminates redundant points and unnecessary local feature calculations, resulting in a significant reduction in overall model performance in terms of storage, computing power, and memory usage. The front-end sampling structure reduces data processing scale, while the back-end weaving and fusion reduces the number of convolutional layers and feature channels, further reducing overall energy consumption and making it possible to adapt to low-power hardware platforms. This not only enables the model to run in real-time on mobile devices but also provides lower costs and higher operating efficiency for large-scale deployment.

[0064] 4. The cross-scale weaving fusion method used in this invention does not rely on the traditional local Euclidean neighborhood during feature propagation, but is guided by the direction of the geometric curve, making it difficult for noise points to spread during cross-scale updates. This significantly improves the overall system's noise resistance. Whether it is sparse point cloud, sensor data with uneven density, or complex scenes containing outliers and local interference, this invention can maintain stable structural representation and high recognition or reconstruction accuracy.

[0065] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0067] Figure 1 This is a schematic flowchart of a point cloud representation method that integrates particle and curve guidance proposed in this invention;

[0068] Figure 2 This is a schematic structural block diagram of a point cloud representation system that integrates particle and curve guidance proposed in this invention;

[0069] Figure 3(a) is a diagram of the overall architecture of point cloud classification in this invention;

[0070] Figure 3(b) is a diagram of the overall architecture of point cloud segmentation of the present invention;

[0071] Figure 3(c) is a CFF structure diagram;

[0072] Figure 3(d) is a diagram of the dot-line attention structure;

[0073] Figure 3(e) is a structural diagram of IMFF;

[0074] Figure 4 This is a schematic flowchart of particle generation and adaptive downsampling;

[0075] Figure 5 It is a schematic flowchart illustrating the overall process of point cloud weaving;

[0076] Figure 6 It is a schematic flowchart illustrating the generation of each braided thread;

[0077] Figure 7(a) is a visualization of the point cloud weaving curve;

[0078] Figure 7(b) is a visualization of the direction vectors of the starting and ending points;

[0079] Figure 7(c) is a visualization of the similarity between global and local directions;

[0080] Figure 7(d) is a visualization of the generation of new nodes;

[0081] Figure 7(e) is a visualization of the braided lines obtained by tracing back the parent node;

[0082] Figure 8 It is a schematic flowchart illustrating the fusion of points and lines;

[0083] Figure 9 This is a schematic flowchart illustrating cross-scale fusion of multi-scale features;

[0084] Figure 10 It is a visualization of point cloud weaving and segmentation effect;

[0085] Figures 11(a)-11(d) show the comparison results of the method of the present invention with other methods. Detailed Implementation

[0086] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the preferred embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0087] Figure 1 This is a schematic flowchart illustrating a point cloud representation method that integrates particle-based and curve-guided methods, such as... Figure 1 As shown, the method includes the following steps:

[0088] Step 1: Obtain raw point cloud data. The raw point cloud data can be 3D point cloud data from devices such as LiDAR and depth cameras. The raw point cloud data has a large amount of data, and the computational cost is very high when processed directly. Therefore, it is necessary to compress the data.

[0089] Step 2: Divide the original point cloud data into spheres and sample the boundary points of each generated sphere to obtain the sphere boundary sampling points;

[0090] Step 3: Perform multiple downsampling operations on the particle boundary sampling points. Represent the particle boundary sampling points and the downsampled data obtained from each downsampling operation as point cloud compressed data at different scales. Perform the following process on the point cloud compressed data at each scale:

[0091] S31: Weave point cloud compressed data based on the RRT-P algorithm to generate weave lines. ;

[0092] S32: Compress point cloud data and braided lines Feature extraction is performed separately to generate local features of the point cloud. and curve characteristics ;

[0093] S33: Based on the point-line attention mechanism, feature fusion is performed on the local features and curve features of the point cloud to generate point-line fused features at the specified scale;

[0094] Step 4: The point and line fusion features of different scales form the multi-scale features of the point cloud.

[0095] This invention also proposes a point cloud representation system that integrates granular and curve-guided methods. Figure 2 This is a schematic structural diagram of the system. For example... Figure 2 As shown, the system may include a data acquisition module, a sphere generation and adaptive downsampling module (GBS module), and a multi-scale point cloud weaving and fusion module (KPC module).

[0096] The data acquisition module is used to acquire raw point cloud data;

[0097] The sphere generation and adaptive downsampling module is used to divide the original point cloud data into spheres and sample the boundary points of each generated sphere to obtain the sphere boundary sampling points.

[0098] The multi-scale point cloud weaving and fusion module may include a downsampling unit, a particle-sphere point cloud weaving unit, a multi-scale feature extraction unit, a point-line fusion unit, and a cross-scale iterative fusion unit, among which...

[0099] The downsampling unit performs multiple downsampling operations on the particle boundary sampling points to obtain downsampled data. The particle boundary sampling points and the downsampled data obtained from each downsampling operation represent point cloud compressed data at different scales.

[0100] The granular point cloud weaving unit uses the RRT-P algorithm to weave compressed point cloud data at each scale, generating weaving lines corresponding to each scale. ;

[0101] The multi-scale feature extraction unit extracts the point cloud compressed data and the braided lines corresponding to each scale. Feature extraction is performed separately to generate local point cloud features corresponding to each scale. and curve characteristics ;

[0102] The point-line fusion unit performs feature fusion on the local features and curve features of the point cloud corresponding to each scale according to the point-line attention mechanism, and generates point-line fusion features corresponding to each scale. The point-line fusion features of different scales constitute the multi-scale features of the point cloud.

[0103] The cross-scale iterative fusion unit generates fused features by fusing multi-scale features of point clouds across scales.

[0104] Figure 4 This is a schematic flowchart of particle generation and adaptive downsampling, as shown below. Figure 4 As shown, particle generation and adaptive downsampling specifically include:

[0105] S21: Randomly select two points from the original point cloud data as center points, and split the original point cloud data into two sub-spheres based on the Euclidean distance between the points in the original point cloud data and these two center points;

[0106] S22: For each sub-sphere, determine whether the splitting termination condition is met. If it is met, the sub-sphere stops splitting and is placed into the sphere set. If the splitting termination condition is not met, randomly select two points from the sub-sphere as new center points, and split the sub-sphere into two new sub-spheres according to the Euclidean distance between the points in the sub-sphere and these two new center points.

[0107] The splitting termination condition can be that the radius of the obtained sphere is less than a preset value (e.g., 0.1) or the number of points in the obtained sphere is less than a preset minimum size (e.g., 2d points, where d represents the spatial dimension of the point cloud). When either of these conditions is met, it indicates that the splitting of the sphere has ended.

[0108] Among them, the radius of the sphere , Indicates the number of points in a pellet. Indicates the first in the granules One point, Indicates the center of the grain.

[0109] The above particle splitting method is for unlabeled point cloud data. For labeled point cloud data, the above method can be used, or the purity of each particle ball can be calculated directly based on the label corresponding to the point cloud (i.e., the ratio of the number of points of the most common category in the particle ball to the total number of points in the particle ball). When the purity of the particle ball is greater than the purity threshold or the number of points in the particle ball is less than the preset minimum size, the particle ball does not need to be split.

[0110] S23: Repeat S22 until all sub-spheres meet the splitting termination condition, thereby generating the final sphere set;

[0111] S24: For each grain in the final grain set, determine the intersection points of the grain with each coordinate axis in the point cloud spatial dimension in both positive and negative directions. Find the nearest actual point to each intersection point within each grain as a sampling point, thus obtaining the boundary sampling points for each grain. For example, for 3D point cloud data, each grain will have two intersection points in both positive and negative directions with each of the three coordinate axes in the 3D point cloud coordinate system, meaning each grain will have six intersection points with the three coordinate axes. Then, find the nearest actual point to each of these six intersection points within the grain, and use these six actual points as the boundary sampling points for each grain. For ease of description, in this invention, the total number of boundary sampling points corresponding to each grain generated after grain division is called the grain boundary sampling points. In this way, the geometric boundaries and structure of the target to be segmented or classified can be effectively preserved.

[0112] This sphere partitioning method can produce spheres of different sizes, and each sphere has 6 boundary sampling points, thus achieving effective compression of the original point cloud data while preserving the geometric boundaries.

[0113] After obtaining the particle boundary sampling points, in order to improve the ability to express three-dimensional geometric features, this invention also performs sampling at multiple scales, multi-scale feature extraction, and multi-scale feature fusion on the particle boundary sampling points.

[0114] Specifically, the sampling points at the particle boundary are downsampled multiple times, and the sampling points at the particle boundary (i.e., the unsampled data) and the downsampled data obtained from each downsampling are represented as point cloud compressed data at different scales. This process can be completed by the downsampling unit.

[0115] Next, the granular point cloud weaving unit can weave the compressed point cloud data at each scale based on the RRT-P algorithm (i.e., Point Rapidly-exploring Random Tree), generating weaving lines corresponding to each scale. (i.e., step S31). This method models local and global shapes by explicitly "weaving" curves within the grains. Curves generated by traditional random sampling suffer from redundancy, high randomness, and a tendency to lose spatial information. This method borrows from path planning, treating each point in the point cloud as a node in a graph, and preserving spatial structure by constructing curves that approximate straight lines. For a detailed explanation of the point cloud weaving process, please refer to [link to relevant documentation]. Figure 5 , Figure 6 And related descriptions.

[0116] Figure 5 This is the overall process of point cloud weaving, such as Figure 5 As shown, point cloud weaving can include the following steps for point cloud compressed data at each scale:

[0117] S311: Use the FPS algorithm to determine K sampling points from the compressed point cloud data;

[0118] S312: Each pair of sampling points in the K sampling points forms a sampling point pair, and the Euclidean distance between the two sampling points in each sampling point pair is determined;

[0119] S313: Select L groups (e.g., 20 groups) of sampling point pairs whose Euclidean distance satisfies the distance threshold to form the starting point set. and the final set ,in, Indicates the first that satisfies the distance threshold Group sampling point pairs, ;

[0120] S314: Perform point cloud weaving on the L sets of sampling point pairs to obtain L weaving lines.

[0121] In this invention, the weaving process uses each starting point as an initial node and gradually expands on the point cloud surface using an improved RRT-P strategy. Specifically, each step first randomly samples directions in space and generates candidate nodes, which are then projected onto the point cloud surface to obtain new weaving points. Subsequently, in the existing node set, considering the cumulative path cost from the starting point to the candidate parent node and the spatial distance from the parent node to the new node, the node with the minimum total cost is selected as the parent node and the connection is completed. By continuously repeating the process of "sampling—projection—parent node selection—connection," a continuous, smooth, and geometrically conforming weaving curve is gradually generated on the point cloud, ultimately achieving an effective expression of the main geometric trend of the target point cloud. Finally, by tracing back from the endpoint to the starting point's parent node, RRT-P can gradually generate "weaving curves" with clear geometric directions in the point cloud, thereby explicitly capturing the spatial structural features of the point cloud.

[0122] Figure 6 This describes the specific process of weaving any pair of sampling points to generate weaving lines. Figures 7(a) to 7(e) are schematic diagrams of the point cloud weaving process, as shown below. Figure 6 As shown in Figures 7(a) to 7(e), for sampling point pairs As shown in Figure 7(a), the green square represents the starting point and the purple pentagram represents the ending point. The following knitting process is performed on the starting point and the ending point:

[0123] S3141: Initialize the node set Empty;

[0124] S3142: with Begin weaving from the starting point, using the starting point... Encapsulate it as the initial node, denoted as a triple. ,in, The coordinates of the starting point (i.e.) (coordinates) The cost representing the starting point. The parent node of the starting point is represented by the initial node. For the initial node, its cost is 0, and its parent node is empty (represented by None). The initial node is then added to the node set. ;

[0125] S3143: Randomly sample a point from the compressed point cloud data. (As shown by the red dots in Figure 7(d)), from the set of nodes Select with nearest point ;

[0126] S3144: Calculation and The direction vector between them is denoted as the current direction. ,Will and The direction vector between them is denoted as the global direction. (As shown in Figure 7(b), where a and b represent the vectors corresponding to these two points respectively), determine whether the angle between the current direction (i.e., the local direction) and the global direction (as shown in Figure 7(c), where a and b represent the global direction vector and the current direction vector respectively, and the cosine of their angle represents the similarity between the two directions) satisfies the angle threshold. If the angle threshold is satisfied, proceed along the current direction. Generate new points (As shown by the pink dot in Figure 7(d)); if the angle threshold is not met, the current direction will be changed. Adjusted to Direction, and generate new points along the adjusted current direction. The new point can be represented as:

[0127] ,

[0128] in, This indicates the maximum step size value set (for example, it can be 0.2). express and The Euclidean distance between them;

[0129] S3145: Find and match points in compressed point cloud data. The closest point is the projection point. (As shown by the blue dot in Figure 7(d)), expressed by the formula: ,in, This represents a compressed point cloud dataset.

[0130] S3146: Traversing the Node Set Among all nodes, select the node with the minimum total cost as the projection point. The parent node is denoted as The total cost is the cost from the starting point to the parent node and the cost between the parent node and the parent node. The sum of the distance values ​​between them is used to record the total cost as the projection point. The cost Thus, a new node is obtained. , A pointer to the parent node of the new node;

[0131] S3147: Move the node Add to node set middle;

[0132] S3148: Repeat S3143 to S3147 above until the termination condition is met, and finally reach the endpoint. Encapsulate as a node and add to the node collection middle;

[0133] In some embodiments, the termination condition may be to find a node whose coordinates are less than the Euclidean distance between the endpoint and the endpoint, which indicates that the node is very close to the endpoint, and the node is then taken as the parent node of the endpoint.

[0134] In some embodiments, the termination condition may be that the number of iterations meets the requirement. For example, S3143 to S3147 may be executed a preset number of times (e.g., 100 times). Among the new points generated during the preset number of iterations, the new point with the closest Euclidean distance to the endpoint is found, and the node corresponding to the new point is taken as the parent node of the endpoint.

[0135] S3149: In the set of nodes The process involves backtracking from the endpoint to the parent node pointer of the starting point (as shown in Figure 7(e), where the blue line segment represents the random exploration tree constructed using the RRT-P algorithm, and the red line represents the curve formed by backtracking from the endpoint to the starting point with the minimum cost), extracting the node coordinate sequence, and thus obtaining the result from... arrive The braided yarn.

[0136] After determining the weaving lines, feature extraction can then be performed. For point cloud compressed data at multiple scales, the above-described point cloud weaving process yields point cloud weaving results at multiple scales, i.e., weaving lines at multiple scales.

[0137] In this invention, feature extraction (i.e., step S32) is required for each scale. As shown in Figure 3, this step can specifically include:

[0138] The Set Abstraction module in PointNet++ (i.e., the point set abstraction module, whose specific structure is the same as that in PointNet++, can abstract local to global geometric features from the point cloud layer by layer, see the point set abstraction module in Figure 3(a) and Figure 3(b)) is used to extract features from the compressed point cloud data to obtain the local features of the point cloud. ;

[0139] The curve features are obtained by extracting the features of each braided thread using a local aggregation coding method. ,in,

[0140] , This indicates max pooling. Indicates a fully connected transformation. Indicates searching within a given radius neighborhood points, This indicates braided yarn, where L represents the number of braided yarns. Indicates the number of dots on the weaving thread. B represents the number of feature channels, and B represents the batch size during data processing.

[0141] For the features extracted at each scale, feature fusion is then performed. As shown in Figure 3, the point-line fusion unit (i.e., Curve Feature Fusion, CFF, see CFF in Figure 3(c)) performs feature fusion on the local features and curve features of the point cloud corresponding to each scale according to the point-line attention mechanism (as shown in the point-line attention mechanism in Figure 3(d)), generating the point-line fusion features corresponding to each scale (i.e., step S33). The point-line fusion features of different scales constitute the multi-scale features of the point cloud.

[0142] Specifically, as shown in Figure 3(c), the local features of the point cloud (i.e., the "point cloud features" in Figure 3(c), whose dimensions are (B, C, P)) are processed by a one-dimensional convolution (i.e., Conv1d) and normalized, and then input into the point-line attention mechanism module with the curve features (whose dimensions are (B, C, L, n)) to perform point-line attention calculation.

[0143] Referring to the point-line attention mechanism block diagram in Figure 3(d), step S33 specifically includes:

[0144] S331: For braided yarn Feature fusion is performed along both the weaving line dimension and the in-line point dimension (i.e., first through 2D convolution, then through average pooling) to obtain line-level structure features. and features within the line ;

[0145] S332: Employs a dual-branch structure. In the first branch, local point cloud features are used as the query (specifically, the result of linear transformation of the position-encoded local point cloud features is used as the query), with line-level structural features. As the key and value (i.e., the result of the linear transformation of this feature), attention-weighted fusion is performed to obtain the first point-line fused feature. In the second branch, local point cloud features are used as the query (specifically, the query is obtained by linearly transforming the sum of local point cloud features and their positional codes), with in-line features as the query. Using the key and value (i.e., the result of the feature after linear transformation) as the key and value, attention-weighted fusion is performed to obtain the second point-line fused feature. ;

[0146] S333: Merge the first point and line feature Second point line fusion feature Perform feature splicing;

[0147] S334: Perform a linear transformation on the feature concatenation result and combine it with the local features of the point cloud. Residual connections are performed to obtain the final point-line fusion features corresponding to the scale. ,Right now ,in, Indicates normalization, Represents linear transformation, This indicates feature splicing (i.e., "stacking").

[0148] By fusing the local features of the point cloud at each scale with the corresponding curve features, the final fused feature for that scale can be obtained. The fused features at multiple scales constitute the multi-scale features of the point cloud.

[0149] In some embodiments of the present invention, the system further includes a cross-scale iterative fusion unit (i.e., IMFF (Iterative Multi-level Feature Fusion) in Figure 3(e), which generates fused features (i.e., feature decoding) by fusing multi-scale features of the point cloud across scales. Cross-scale fusion can enhance semantic associations within a multi-scale, cross-level framework, thereby more effectively fusing global and local information in the encoder.

[0150] The multi-scale features of point clouds include Layer features (as shown in Figure 3(a) and Figure 3(b)) The value can be 4, meaning there are 4 layers of features in total. The scale of the layer features decreases from top to bottom (i.e., the amount of data corresponding to the feature gradually decreases, but the number of feature channels increases). The features of each layer are denoted as follows: , .

[0151] Multi-scale features of point clouds are further fused across scales to generate fused features, where, for the current layer Its characteristics can be expressed as B represents the batch size, C represents the number of feature channels in the current layer, and N represents the number of feature points in the current layer. This is related to the feature size in the next layer. Perform cross-scale fusion. Indicates the number of feature channels in the lower layer. This represents the number of feature points in the lower-level features, since the two have different dimensions ( Therefore, the following procedure can be performed:

[0152] S51: Perform self-attention calculation on the current layer features to obtain the attention score corresponding to each feature point in the current layer features. This can be expressed as a formula: ,in, and This indicates a linear transformation, with the superscript T indicating transpose and d indicating the feature dimension.

[0153] S52: Highest score in attention extraction Each point is used as a key point, and key point features are extracted. For example, the Top-K strategy can be used to select the highest-scoring candidate. Each point is used as a key point, and its corresponding features are extracted, expressed by the formula: ;

[0154] S53: Perform PointTransformer aggregation on the keypoint features and the features of the lower layer (i.e., the lower layer) to obtain the initial aggregated features corresponding to the current layer, expressed by the formula: ,in, , , and It also represents linear transformation processing. express Location encoding;

[0155] S54: Concatenate the initial aggregated features with the lower-layer features along the channel dimension (i.e., the stacking operation in Figure 3(e)) to generate the final fused features corresponding to the current layer. This can be expressed as a formula: =concat( , ).

[0156] For point cloud segmentation tasks, as shown in Figure 3(b), from the current layer Initially, the lower layer is Layer, execute S51 to S54, and obtain the result with the first... The final aggregated features corresponding to the layer Then, The features propagated by the feature propagation module are used as the lower-level features, and Decrease by 1, continue executing S51 to S54, repeat the process until the final fusion feature corresponding to the top layer is obtained. The result of feature propagation of the final fusion feature corresponding to the top layer is used to generate point cloud segmentation results through MLP.

[0157] For the point cloud classification task, as shown in Figure 3(a), Take 1, Pick That is, using the first layer and the second layer Layer features are fused across scales to generate fused features. The fused features are then passed through MLP and then through max pooling and average pooling respectively. The results of max pooling and average pooling are multiplied and then passed through MLP again to obtain the point cloud classification result.

[0158] The present invention also provides a computer program that, when executed by a processor, implements the point cloud representation method described above, which integrates particle spheres and curve-guided representations.

[0159] The advantages of the present invention will be illustrated below through specific embodiments.

[0160] This embodiment uses the ModelNet40 and ShapeNet datasets as training and testing data sources. For each 3D model, 2048 points are first extracted using a uniform random sampling method, and the point cloud is normalized to a unit sphere space, with its centroid translated to the origin and its maximum radius scaled to 1. To enhance the robustness of the model, this embodiment employs various geometric perturbation methods: including random rotation (rotation angle uniformly sampled from [0°, 45°]), random translation (translation amount sampled from [-0.3, 0.3]), and random occlusion (cropping approximately 30%–40% of the points using a random directional half-space), to simulate missing points and viewpoint changes in real-world applications.

[0161] The SGD optimizer was used with an initial learning rate of 0.001, a cosine annealing strategy, a batch size of 64, and a training epoch of 500. The cross-entropy loss function was employed. Training hardware consisted of an NVIDIA RTX 4090D GPU (24GB) and an Intel Core i9-14900KF CPU.

[0162] Figure 10 This is the result of point cloud segmentation using the method of this invention on several existing labeled point cloud datasets. From Figure 10 It can be seen that the point cloud processing method of the present invention can achieve very good segmentation results.

[0163] Figures 11(a) to 11(d) show the simulation results of the present invention (represented by KPC) and existing models PointNet++, RS-CNN, CurveNet, and RepSurf-U on the existing ModelNet40 point cloud dataset. Figure 11(a) shows the classification results of different models with varying point cloud quantity. As can be seen from Figure 11(a), the method of the present invention achieves the highest classification accuracy with a smaller point cloud quantity. Figure 11(b) shows the sensitivity of different models to the amount of noise in the point cloud. As can be seen from Figure 11(b), the method of the present invention is not sensitive to noise. Figure 11(c) shows the segmentation results of different models under high density conditions. As can be seen from Figure 11(c), the present invention achieves the highest mean accuracy (mAP) under high density conditions, where the "pixel-level classification" is treated as a multi-class classification problem, the accuracy is calculated for each class, and then the average is taken for all classes. Figure 11(d) shows the cosine similarity of channel features in the encoding stage for different models. Cosine similarity is an indicator that measures how similar two vectors are. In neural networks, the output of each layer is usually a multi-channel feature tensor, with different channels corresponding to different feature patterns extracted by the model from the input data. Channel cosine similarity is used to determine whether the features represented by these channels are redundant or highly correlated by calculating the cosine angle between the feature vectors of different channels. As can be seen from Figure 11(d), compared with PointNet++ / CurveNet / PointNeXt, the method of this invention can learn more decorrelational and less redundant feature representations in the mid-level encoding stage.

[0164] Table 1 presents the comparative experimental metrics of the method in this invention and existing models on the ShapeNet dataset. Here, KPC is an abbreviation for this invention, and KPC* represents a model with a larger number of parameters, increasing the number of feature channels for point cloud features in the network. In Table 1, Cls.mIoU measures the segmentation balance between different semantic categories, while Ins.mIoU measures the segmentation accuracy of different instances within the same semantic category.

[0165] Table 1. Comparison of experimental indicators between the method of the present invention and existing models.

[0166]

[0167] As shown in Table 1, the semantic segmentation experiments on the ShapeNet dataset demonstrate that, compared to current mainstream methods, the proposed method achieves the best results in both the instance mean intersection-over-union (IoU) and class mean IoU evaluation metrics. This result indicates that the proposed method not only achieves more accurate point-level prediction at the overall level but also maintains more balanced and stable segmentation performance across different semantic categories, reflecting its comprehensive advantages in complex 3D structural feature modeling and multi-class discrimination.

[0168] In summary, this invention, through the collaborative design of a sphere-adaptive downsampling structure and a curve-guided cross-scale weaving fusion network, enables the original point cloud to retain key geometric structures while significantly compressing its size, thus demonstrating a clear advantage over existing technologies in overall accuracy and expressive power. Sphere sampling characterizes local regions through a center-radius structure and generates feature points along the boundary direction, preserving key morphological information such as edges, curvature abrupt changes, and detailed protrusions while drastically reducing the number of points. The weaving point cloud method guides the interaction of cross-scale features through geometric curve directions, allowing coarse-scale global information and fine-scale local geometry to work together, resulting in lightweight point clouds that still possess excellent geometric reconstruction performance and strong expressive power.

[0169] Thanks to the near-linear computational complexity of the particle-sphere algorithm, this invention maintains a high sampling rate when processing point clouds of hundreds of thousands, reducing the overall system's computational cost. Compared to traditional methods based on farthest point sampling (FPS) or voxelization, this invention reduces the number of points by 40% to 80% in the front-end data processing stage, significantly reducing the computational cost of subsequent kNN neighborhood search and multi-scale convolution in deep networks. Due to the reduced number of input points, the memory usage and computational power consumption during model inference are also reduced, enabling this invention to run lightweight on GPUs, NPUs, or mobile chips, making it particularly suitable for real-time or low-power environments such as robotics, autonomous driving, and smart terminals.

[0170] In terms of robustness, this invention utilizes particle purity determination and a stepwise splitting mechanism to reduce the interference of noise points, making the downsampling results more stable and reliable. Simultaneously, the cross-scale woven fusion network does not rely on traditional local Euclidean neighborhoods during feature propagation, but is guided by the direction of geometric curves, making it difficult for noise points to spread during cross-scale updates, thus significantly improving the overall system's noise resistance. Whether dealing with sparse point clouds, unevenly dense sensor data, or complex scenes containing outliers and local interference, this invention maintains stable structural representation and high recognition or reconstruction accuracy. Figures 11(a)-11(d) show the comparison results with other methods and experimental tables.

[0171] In terms of resource conservation and energy consumption, this invention significantly reduces the overall model's storage, computing power, and memory usage by reducing redundant points and unnecessary local feature calculations. The front-end sampling structure reduces the scale of data processing, and the back-end weaving and fusion reduces the number of convolutional layers and feature channels, further reducing overall energy consumption and making it possible to adapt to low-power hardware platforms. This not only enables the model to run in real-time on mobile devices but also provides lower costs and higher operating efficiency for large-scale deployment.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A point cloud representation method integrating particle sphere and curve-guided methods, characterized in that, include: Step 1: Obtain raw point cloud data; Step 2: Divide the original point cloud data into spheres and sample the boundary points of each generated sphere to obtain the sphere boundary sampling points; Step 3: Perform multiple downsampling operations on the particle boundary sampling points. Represent the particle boundary sampling points and the downsampled data obtained from each downsampling operation as point cloud compressed data at different scales. Perform the following process on the point cloud compressed data at each scale: S31: Weave point cloud compressed data based on the RRT-P algorithm to generate weave lines. ; S32: Compress point cloud data and braided lines Feature extraction is performed separately to generate local features of the point cloud. and curve characteristics ; S33: Based on the point-line attention mechanism, feature fusion is performed on the local features and curve features of the point cloud to generate point-line fused features at the specified scale; Step 4: The point and line fusion features of different scales form the multi-scale features of the point cloud.

2. The point cloud representation method integrating particle and curve guidance as described in claim 1, characterized in that, Step 2 specifically includes: S21: Randomly select two points from the original point cloud data as center points. Based on the Euclidean distance between the points in the original point cloud data and these two center points, split the original point cloud data into two sub-spheres. S22: For each sub-sphere, determine whether the splitting termination condition is met. If it is met, the sub-sphere stops splitting and is placed into the sphere set. If the splitting termination condition is not met, randomly select two points from the sub-sphere as new center points, and split the sub-sphere into two new sub-spheres according to the Euclidean distance between the points in the sub-sphere and the two new center points. S23: Repeat S22 until all sub-spheres meet the splitting termination condition, thereby generating the final sphere set; S24: For each particle in the final particle set, determine the intersection point of the particle with the positive and negative directions of each coordinate axis in the point cloud space dimension, and find the actual point in each particle that is closest to the intersection point as the sampling point, thereby obtaining the boundary sampling point of each particle.

3. The point cloud representation method integrating particle and curve guidance according to claim 1, characterized in that, Step S31 specifically includes: S311: Use the FPS algorithm to determine K sampling points from the compressed point cloud data; S312: Each pair of sampling points in the K sampling points forms a sampling point pair, and the Euclidean distance between the two sampling points in each sampling point pair is determined; S313: Select L pairs of sampling points whose Euclidean distance satisfies the distance threshold to form the starting point set. and the final set ,in, Indicates the first that satisfies the distance threshold Group sampling point pairs, ; S314: Perform point cloud weaving on the L sets of sampling point pairs to obtain L weaving lines, wherein, for each set of sampling point pairs Perform the following knitting process: S3141: Initialize the node set Empty; S3142: with Begin weaving from the starting point, using the starting point... Encapsulate it as the initial node, denoted as a triple. ,in, The coordinates of the starting point. The cost representing the starting point. The parent node representing the starting point is added to the node set. ; S3143: Randomly sample a point from the compressed point cloud data. From the set of nodes Select with nearest point ; S3144: Calculation and The direction vector between them is denoted as the current direction. ,Will and The direction vector between them is denoted as the global direction. Determine if the angle between the current direction and the global direction meets an angle threshold. If it does, proceed along the current direction. Generate new points If the angle threshold is not met, then the current direction will be changed. Adjusted to And generate new points along the adjusted current direction. ; S3145: Find and match points in compressed point cloud data. The closest point is the projection point. ; S3146: Traversing the Node Set Among all nodes, select the node with the minimum total cost as the projection point. The parent node is denoted as The total cost is the cost from the starting point to the parent node and the cost between the parent node and the parent node. The sum of the distance values ​​between them is used to record the total cost as the projection point. The cost Thus, a new node is obtained. , A pointer to the parent node of the new node; S3147: Move the node Add to node set middle; S3148: Repeat S3143 to S3147 above until the termination condition is met, and finally reach the endpoint. Encapsulate as a node and add to the node collection middle; S3149: In the set of nodes By backtracking from the end point to the parent node pointer of the starting point, the node coordinate sequence is extracted, thus obtaining the result from... arrive The braided yarn.

4. The point cloud representation method integrating particle and curve guidance according to claim 1, characterized in that, Step S32 specifically includes: The Set Abstraction module in PointNet++ is used to extract features from compressed point cloud data to obtain local features of the point cloud. ; The curve features are obtained by extracting the features of each braided thread using a local aggregation coding method. ,in, , This indicates max pooling. Indicates a fully connected transformation. Indicates searching within a given radius neighborhood points, This indicates braided yarn, where L represents the number of braided yarns. Indicates the number of dots on the weaving thread. Indicates the number of feature channels.

5. The point cloud representation method integrating particle and curve guidance according to claim 1, characterized in that, Step S33 specifically includes: S331: For braided yarn Feature fusion is performed along both the weaving line dimension and the in-line point dimension to obtain the line-level structure features. and features within the line ; S332: Employs a dual-branch structure. In the first branch, local point cloud features are used as queries, with line-level structural features... As keys and values, attention-weighted fusion is performed to obtain the first point-line fusion feature. In the second branch, local features of the point cloud are used as queries, with in-line features as the query. As keys and values, attention-weighted fusion is performed to obtain the second point-line fusion feature. ; S333: Merge the first point and line feature Second point line fusion feature Perform feature splicing; S334: Perform a linear transformation on the feature concatenation result and combine it with the local features of the point cloud. Residual connections are performed to obtain the final point-line fusion features corresponding to the scale. .

6. The point cloud representation method integrating particle spheres and curve guidance according to claim 1, characterized in that, The multi-scale features of point clouds include Layer features, the The scale of the layer features decreases sequentially from top to bottom. The method further includes: Step 5: Generating fused features by further cross-scale fusion of the multi-scale features of the point cloud, wherein, for the current layer... Perform the following process: S51: Perform self-attention calculation on the current layer features to obtain the attention score corresponding to each feature point in the current layer features; S52: Highest score in attention extraction 10 feature points are used as key points, and key point features are extracted; S53: Aggregate the key point features with the lower layer features using PointTransformer to obtain the initial aggregated features corresponding to the current layer; S54: Concatenate the initial aggregated features with the lower-level features along the channel dimension to generate the final fused features corresponding to the current layer. .

7. The point cloud representation method integrating particle and curve guidance according to claim 6, characterized in that, For point cloud segmentation tasks, starting from the current layer Initially, the lower layer is Each layer executes steps S51 to S54 to obtain the corresponding final aggregated features. Then The result of the feature propagation operation is used as the lower-level feature, and... Decrease by 1, continue executing S51 to S54, repeat the process until the final fusion feature corresponding to the top layer is obtained. The result of feature propagation of the final fusion feature corresponding to the top layer is used to generate point cloud segmentation results through MLP.

8. The point cloud representation method integrating particle and curve guidance according to claim 6, characterized in that, For point cloud classification tasks Take 1, that is, use the first layer and the second layer. Layer features are fused across scales to generate fused features. The fused features are then passed through MLP and then through max pooling and average pooling respectively. The results of max pooling and average pooling are multiplied and then passed through MLP again to obtain the point cloud classification result.

9. A point cloud representation system integrating particle sphere and curve-guided methods, characterized in that, include The data acquisition module is used to acquire raw point cloud data; The sphere generation and adaptive downsampling module is used to divide the original point cloud data into spheres and sample the boundary points of each generated sphere to obtain the sphere boundary sampling points. The multi-scale point cloud weaving and fusion module includes a downsampling unit, a particle-sphere point cloud weaving unit, a multi-scale feature extraction unit, a point-line fusion unit, and a cross-scale iterative fusion unit. The downsampling unit performs multiple downsampling operations on the particle boundary sampling points to obtain downsampled data. The particle boundary sampling points and the downsampled data obtained from each downsampling operation represent point cloud compressed data at different scales. The granular point cloud weaving unit uses the RRT-P algorithm to weave compressed point cloud data at each scale, generating weaving lines corresponding to each scale. ; The multi-scale feature extraction unit extracts the point cloud compressed data and the braided lines corresponding to each scale. Feature extraction is performed separately to generate local point cloud features corresponding to each scale. and curve characteristics ; The point-line fusion unit performs feature fusion on the local features and curve features of the point cloud corresponding to each scale according to the point-line attention mechanism, and generates point-line fusion features corresponding to each scale. The point-line fusion features of different scales constitute the multi-scale features of the point cloud. The cross-scale iterative fusion unit generates fused features by fusing multi-scale features of point clouds across scales.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the point cloud representation method that integrates particle and curve guidance as described in any one of claims 1 to 8.

Images