Method, system, readable storage medium and computer for rotation-invariant point cloud analysis
Patent Information
- Application Number
- CN202611273731.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-09-22
AI Technical Summary
早期的不变性方法主要依赖距离和角度等手工几何特征,虽然计算高效,但因丢失了相对方向信息而导致判别力下降
[0024]本发明当中的旋转不变点云分析方法、系统、可读存储介质及计算机,通过构建局部参考帧,消除了群中任意旋转对特征提取的干扰,解决了传统网络存在的姿态模糊问题,通过采用单流式架构,在保持同等甚至更高精度的前提下,减少了模型参数量,提高计算效率,极大地降低了算力成本,利用几何感知Transformer通过条件位置编码和几何先验注意力,克服了标准Transformer对绝对坐标的依赖,能够在不破坏旋转不变性的基础上,有效捕捉点云的长距离拓扑结构和细粒度几何特征,显著提升了在复杂场景下的识别准确率。
Smart Images

Figure CN122799126A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to a rotation-invariant point cloud analysis method, system, readable storage medium, and computer. Background Technology
[0002] With the widespread application of LiDAR and RGB-D depth cameras, 3D point clouds have become a key data modality in fields such as autonomous driving, robot navigation, and augmented reality. However, point cloud data inherently possesses sparsity and disorder, posing significant challenges to deep learning models. Although PointNet and its variants have achieved remarkable progress in shape classification and segmentation tasks under ideal laboratory conditions, in real-world open-world scenarios (such as robot grasping or autonomous driving), the acquired point clouds are often located in an unknown global coordinate system and are subject to... Interference from arbitrary rotations within the group.
[0003] Traditional standard deep learning networks typically rely heavily on the absolute coordinates of points. Feature extraction is performed. When the input point cloud undergoes global rotation, the data input distribution changes drastically, leading to a disastrous decline in model performance. To address this issue, existing research mainly focuses on three aspects: data augmentation, normalized pose alignment, and rotation-invariant / isovariant feature learning. However, all of these have significant limitations. 1. Data augmentation method: This is the most intuitive strategy, but it forces the network to consume a lot of capacity to "memorize" various possible rotation poses instead of learning the essential geometry of the object. This results in the model having poor generalization ability when faced with unseen viewpoints.
[0004] 2. Standardized Pose Alignment Methods: These methods attempt to calibrate point clouds to a uniform pose using Principal Component Analysis (PCA) or Spatial Transformation Networks (STN). However, defining a unique "standardized pose" is extremely difficult for objects with symmetry or occlusion, easily leading to "pose ambiguity" problems, which amplify alignment errors in subsequent networks.
[0005] 3. Rotation-Invariant / Isovariant Feature Learning Methods: This is a cutting-edge technology. Early invariant methods mainly relied on handcrafted geometric features such as distance and angle. While computationally efficient, they suffered from decreased discriminative power due to the loss of relative orientation information. Recent rotation-invariant networks (such as TFN and VN-PointNet) have achieved excellent performance by designing convolutional kernels that satisfy group equivariance. However, these methods typically involve complex spherical harmonic transformations or high-dimensional vector operations, resulting in extremely high computational complexity. For example, the existing high-performance network LocoTrans employs a complex bi-branch structure (an invariant feature branch and a pose recovery branch), leading to high computational and memory costs and making it difficult to deploy on resource-constrained edge devices.
[0006] Furthermore, while the Transformer architecture excels at capturing global contextual information, its standard positional encoding mechanism relies on absolute coordinates, which makes it inherently lacking in rotational invariance, thus limiting its application in robust 3D perception tasks.
[0007] Therefore, there is an urgent need in the existing technology for a point cloud analysis method that can eliminate rotational interference from the geometric level, efficiently capture long-distance contextual dependencies, and maintain low parameter count and low computational complexity. Summary of the Invention
[0008] Based on this, the purpose of the present invention is to provide a rotation-invariant point cloud analysis method, system, readable storage medium, and computer to at least address the shortcomings of the aforementioned technologies.
[0009] This invention proposes a rotation-invariant point cloud analysis method, comprising: The three-dimensional point cloud data of the target object is acquired using a three-dimensional acquisition device, and the three-dimensional point cloud data is then augmented and multi-view generated. The processed point cloud data is input into a rotation-invariant edge convolution model. By constructing a local reference frame, the point cloud is projected onto a local coordinate system to eliminate global rotation interference and output rotation-invariant geometric feature vectors. The rotation-invariant geometric feature vector is input into the geometry-aware Transformer model. Through conditional position encoding and geometric prior attention mechanism, the long-range contextual dependencies of the point cloud are captured while maintaining rotation invariance, so as to obtain geometry-aware features. A multi-scale pyramid pooling model is constructed to process the geometric perception features. Through multi-level pooling, upsampling, and feature fusion, a final feature representation that combines local geometric details and global semantic information is generated. The final feature representation is input into a preset task prediction model to generate the final 3D point cloud analysis result.
[0010] Furthermore, the processed point cloud data is input into a rotation-invariant edge convolutional model. By constructing a local reference frame to project the point cloud onto a local coordinate system, global rotation interference is eliminated, and the rotation-invariant geometric feature vector is output. The steps include: Using the k-nearest neighbor algorithm to analyze any point in the processed point cloud data Search neighborhood point set And calculate the local centroid. Covariance Matrix ; For the covariance matrix Eigenvalue decomposition is performed to obtain the eigenvector matrix, and a unique local reference frame is determined by the sign disambiguation rule; Relative to the local centroid, each point in the neighborhood is considered. The relative coordinates are projected onto the local reference frame to obtain rotation-invariant coordinates. ; With the rotation-invariant coordinates As input, geometric features are extracted through edge convolution, and the rotation-invariant coordinates are... Its difference relative to the center point The features are concatenated, nonlinearly mapped using a multilayer perceptron, and then aggregated using max pooling to output a rotation-invariant geometric feature vector. .
[0011] Furthermore, the local centroid The calculation formula is:
[0012] In the formula, Represents points within the neighborhood. Indicates the number of neighboring points; The covariance matrix The calculation formula is:
[0013] in, Indicates matrix transpose; The rotation-invariant coordinates The calculation formula is:
[0014] In the formula, Indicates a local reference frame; The rotation-invariant geometric eigenvectors The calculation formula is:
[0015] In the formula, This indicates a splicing operation. This represents a multilayer perceptron.
[0016] Furthermore, the steps of inputting the rotation-invariant geometric feature vector into a geometry-aware Transformer model, and capturing long-range contextual dependencies of the point cloud while maintaining rotation invariance through conditional position encoding and geometric prior attention mechanisms to obtain geometry-aware features include: The rotation-invariant geometric feature vector is input into the standard edge convolutional layer of the geometry-aware Transformer model to increase the feature dimension and obtain enhanced features; The rotation-invariant coordinates are used to generate a position code, which is then superimposed on the enhanced feature to obtain a feature that fuses position information. In the self-attention mechanism, a geometric bias term driven by rotation-invariant coordinates is introduced, attention weights are calculated, the value vector is weighted and summed according to the attention weights, and the position encoding is incorporated to output aggregated features with local geometric and global contextual information.
[0017] Furthermore, the steps of constructing a multi-scale pyramid pooling model to process the geometrically perceived features, and generating a final feature representation that combines local geometric details and global semantic information through multi-level pooling, upsampling, and feature fusion include: Multi-scale adaptive average pooling and upsampling are performed on the geometrically perceived features to obtain a multi-scale feature set; Global max pooling is performed on the geometrically perceived features to obtain global features, local features of the geometrically perceived features are retained, and the local features, the global features, and the multi-scale feature set are concatenated by channels. The concatenated features are then fed into a fusion multilayer perceptron for dimensionality reduction and feature recombination to obtain the final feature representation.
[0018] This invention also proposes a rotation-invariant point cloud analysis system, comprising: The data acquisition module is used to acquire three-dimensional point cloud data of the target object using a three-dimensional acquisition device, and to perform data enhancement and multi-view generation on the three-dimensional point cloud data; The interference cancellation module is used to input the processed point cloud data into the rotation-invariant edge convolution model, project the point cloud onto the local coordinate system by constructing a local reference frame, eliminate global rotation interference, and output rotation-invariant geometric feature vectors. The geometry perception module is used to input the rotation-invariant geometric feature vector into the geometry perception Transformer model. Through conditional position encoding and geometric prior attention mechanism, it captures the long-distance contextual dependencies of the point cloud while maintaining rotation invariance, so as to obtain geometry perception features. The feature fusion module is used to construct a multi-scale pyramid pooling model, process the geometric perception features, and generate a final feature representation that combines local geometric details and global semantic information through multi-level pooling, upsampling and feature fusion. The point cloud analysis module is used to input the final feature representation into a preset task prediction model to generate the final three-dimensional point cloud analysis result.
[0019] Furthermore, the interference cancellation module is specifically used for: Using the k-nearest neighbor algorithm to analyze any point in the processed point cloud data Search neighborhood point set And calculate the local centroid. Covariance Matrix ; For the covariance matrix Eigenvalue decomposition is performed to obtain the eigenvector matrix, and a unique local reference frame is determined by the sign disambiguation rule; Relative to the local centroid, each point in the neighborhood is considered. The relative coordinates are projected onto the local reference frame to obtain rotation-invariant coordinates. ; With the rotation-invariant coordinates As input, geometric features are extracted through edge convolution, and the rotation-invariant coordinates are... Its difference relative to the center point The features are concatenated, nonlinearly mapped using a multilayer perceptron, and then aggregated using max pooling to output a rotation-invariant geometric feature vector. .
[0020] Furthermore, the geometric sensing module is specifically used for: The rotation-invariant geometric feature vector is input into the standard edge convolutional layer of the geometry-aware Transformer model to increase the feature dimension and obtain enhanced features; The rotation-invariant coordinates are used to generate a position code, which is then superimposed on the enhanced feature to obtain a feature that fuses position information. In the self-attention mechanism, a geometric bias term driven by rotation-invariant coordinates is introduced, attention weights are calculated, the value vector is weighted and summed according to the attention weights, and the position encoding is incorporated to output aggregated features with local geometric and global contextual information.
[0021] Furthermore, the feature fusion module is specifically used for: Multi-scale adaptive average pooling and upsampling are performed on the geometrically perceived features to obtain a multi-scale feature set; Global max pooling is performed on the geometrically perceived features to obtain global features, local features of the geometrically perceived features are retained, and the local features, the global features, and the multi-scale feature set are concatenated by channels. The concatenated features are then fed into a fusion multilayer perceptron for dimensionality reduction and feature recombination to obtain the final feature representation.
[0022] The present invention also proposes a storage medium on which a computer program is stored, which, when executed by a processor, implements the above-described rotation-invariant point cloud analysis method.
[0023] The present invention also proposes a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described rotation-invariant point cloud analysis method.
[0024] The rotation-invariant point cloud analysis method, system, readable storage medium, and computer of this invention eliminate [the problem] by constructing a local reference frame. The interference of arbitrary rotations within the group on feature extraction solves the pose ambiguity problem existing in traditional networks. By adopting a single-stream architecture, the number of model parameters is reduced while maintaining the same or even higher accuracy, improving computational efficiency and greatly reducing computing costs. By using the geometrically aware Transformer through conditional position encoding and geometric prior attention, it overcomes the dependence of the standard Transformer on absolute coordinates. It can effectively capture the long-distance topological structure and fine-grained geometric features of point clouds without destroying rotation invariance, significantly improving the recognition accuracy in complex scenes. Attached Figure Description
[0025] Figure 1 This is a flowchart of the rotation-invariant point cloud analysis method in the first embodiment of the present invention; Figure 2 This is a structural block diagram of the rotation-invariant point cloud analysis system in the second embodiment of the present invention; Figure 3 This is a structural block diagram of the computer in the third embodiment of the present invention.
[0026] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation
[0027] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.
[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.
[0029] Example 1 Please see Figure 1 The figure shows a rotation-invariant point cloud analysis method in the first embodiment of the present invention, the method specifically including steps S101 to S105: S101, use a 3D acquisition device to acquire 3D point cloud data of the target object, and perform data enhancement and multi-view generation on the 3D point cloud data; In practice, three-dimensional acquisition devices (such as LiDAR or RGB-D depth cameras) are used to acquire three-dimensional point cloud data of the target object. During the acquisition process, it is important to ensure that the sensor maintains an appropriate distance from the target object to ensure that the acquired point cloud data has sufficient density and coverage to clearly characterize the geometry of the object. Furthermore, in order to improve the model's generalization ability to different rotational poses and support subsequent three-view simulations... Figure 1 For consistency training, after acquiring the raw point cloud data, specific data augmentation and multi-view generation operations are required. S102, the processed point cloud data is input into the rotation-invariant edge convolution model, the point cloud is projected onto the local coordinate system by constructing a local reference frame, the global rotation interference is eliminated, and the rotation-invariant geometric feature vector is output. Furthermore, step S102 specifically includes steps S1021 to S1024: S1021, Use the k-nearest neighbor algorithm to process any point in the processed point cloud data. Search neighborhood point set And calculate the local centroid. Covariance Matrix ; S1022, regarding the covariance matrix Eigenvalue decomposition is performed to obtain the eigenvector matrix, and a unique local reference frame is determined by the sign disambiguation rule; S1023, relative to the local centroid, each point in the neighborhood. The relative coordinates are projected onto the local reference frame to obtain rotation-invariant coordinates. ; S1024, with the rotation-invariant coordinates As input, geometric features are extracted through edge convolution, and the rotation-invariant coordinates are... Its difference relative to the center point The features are concatenated, nonlinearly mapped using a multilayer perceptron, and then aggregated using max pooling to output a rotation-invariant geometric feature vector. .
[0030] In practical implementation, the processed point cloud data obtained above is input into a rotation-invariant edge convolution model (in this embodiment, the model chosen is RI-EdgeConv). This model employs a Local Reference Frame (LRF) calibration strategy based on Principal Component Analysis (PCA) to eliminate global rotation interference from a geometrical level. Specifically, the k-nearest neighbor algorithm (K-NN algorithm) is used to process any point in the processed point cloud data. Search neighborhood point set And calculate the local centroid. Covariance Matrix :
[0031]
[0032] In the formula, Indicates matrix transpose. Represents points within the neighborhood. Indicates the number of neighboring points; For covariance matrix Perform eigenvalue decomposition to satisfy .in, It contains three feature vectors, forming a rotationally equivalent basis for the local reference frame. When the input point cloud is rotated... hour, The corresponding transformation is To address the potential sign flipping issue in the PCA spindle, [the following is a separate, unrelated sentence:] ... By applying sign disambiguation rules, deterministic local reference frames are forced to be generated by aligning each axis direction with consistent neighborhood statistics. These constructed local reference frames are then used... Points within the neighborhood Relative to the local centroid The relative coordinates are projected onto this local reference frame to obtain rotation-invariant coordinates. :
[0033] in, This represents the matrix transpose, due to the rotation matrix. orthogonality (i.e.) And the relative subtraction of the centroid, this method guarantees the coordinates in the mathematical sense of the input level. For any Rotational invariance.
[0034] Furthermore, the obtained rotation-invariant transformed neighborhood point coordinates are... As input, an edge convolution (EdgeConv) operation is performed to extract geometric features. The transformed neighborhood point coordinates are then used. Its difference relative to the center point The features are concatenated, nonlinearly mapped using a multilayer perceptron (MLP), and then aggregated using max pooling to obtain the final rotation-invariant features that are rotation-robust. :
[0035] In the formula, This indicates a splicing operation. This represents a multilayer perceptron.
[0036] S103, The rotation-invariant geometric feature vector is input into the geometry-aware Transformer model. Through conditional position encoding and geometric prior attention mechanism, the long-distance contextual dependency of the point cloud is captured while maintaining rotation invariance, so as to obtain geometry-aware features. Furthermore, step S103 specifically includes steps S1031 to S1033: S1031, The rotation-invariant geometric feature vector is input into the standard edge convolutional layer of the geometry-aware Transformer model to increase the feature dimension and obtain enhanced features; S1032, use the rotation-invariant coordinates to generate a position code, and superimpose it onto the enhanced feature to obtain a feature with fused position information; S1033, In the self-attention mechanism, a geometric deviation term driven by rotation-invariant coordinates is introduced, attention weights are calculated, the value vector is weighted and summed according to the attention weights, and the position encoding is incorporated to output aggregated features with local geometric and global contextual information.
[0037] In practical implementation, the rotation-invariant characteristics obtained above will be applied. The input is fed into a standard EdgeConv layer. By constructing a dynamic graph in the feature space and aggregating neighborhood features, the feature dimension is increased from C1 to C2, resulting in an enhanced feature vector. This provides the Transformer module with inputs that have stronger semantic expressive power.
[0038] Furthermore, in order to preserve position information in the Transformer without violating rotation invariance, the rotation-invariant local coordinates obtained above are used. Generate positional codes instead of using absolute coordinates Conditional position encoding The calculation formula is as follows:
[0039] in, This represents a multilayer perceptron (MLP) used to generate positional codes, which are then directly superimposed onto the input features. The above yields features derived from the fused location information. This enables the network to perceive local geometry.
[0040] Furthermore, geometric bias is introduced into the self-attention mechanism to enhance the perception of point cloud topology. For query points... and key points First, the query vector is calculated using a linear mapping. Key vector Sum value vector Calculate attention weights When adding locally invariant coordinates Driven geometric deviation term The formula is as follows:
[0041] in, Indicates at key points The Softmax normalization function executed above, Indicates matrix transpose. For feature dimension, This is the learnable geometric bias function obtained through a small MLP. This represents the coordinates of the center point after the transformation.
[0042] Based on the calculated attention weights The value vectors are weighted and summed, and positional encoding information is incorporated again to output aggregated features. :
[0043] The final output features of the GA-Transformer are obtained through a feedforward neural network (FFN) with residual connections and layer normalization (LayerNorm). This feature combines local geometric precision with global contextual information.
[0044] S104, Construct a multi-scale pyramid pooling model to process the geometric perception features. Through multi-level pooling, upsampling and feature fusion, generate a final feature representation that combines local geometric details and global semantic information. Furthermore, step S104 specifically includes steps S1041 to S1043: S1041, Perform multi-scale adaptive average pooling and upsampling on the geometric perception features to obtain a multi-scale feature set; S1042, Perform global max pooling on the geometric perception features to obtain global features, retain the local features of the geometric perception features, and concatenate the local features, the global features, and the multi-scale feature set into channels; S1043, the spliced features are input into the fusion multilayer perceptron for dimensionality reduction and feature recombination to obtain the final feature representation.
[0045] In practical implementation, to enhance the model's ability to perceive targets of different scales in complex scenes and eliminate semantic ambiguity caused by rotation, a multi-scale pyramid pooling module (PPM) is constructed to process the geometric perception features obtained above. Specifically: Assuming the aggregated sensing features are , Representing feature dimension, defined Different pooling scales , for features Adaptive average pooling at the corresponding scale is performed to obtain context features at different resolutions. Then, upsampling is used to restore these features to the original point cloud size. To obtain a multi-scale feature set The formula is expressed as:
[0046] In the formula, Indicates an upsampling operation. This represents adaptive average pooling, with an output scale of . , Indicated in scale The captured contextual features.
[0047] To comprehensively utilize various information from point clouds, a fused feature vector is constructed. First, the input features... Perform global max pooling to obtain the global feature vector. And copy and extend it to First, preserve the input features; second, retain the input features. As a local feature Finally, the local features, global features, and the multi-scale features generated above are concatenated along the channel dimension:
[0048] Indicated in scale The captured context features will be concatenated into the final features. Input to fusion multilayer perceptron ( The dimensionality reduction and feature recombination are performed to output the final feature representation. :
[0049] This feature It integrates rotation-invariant geometric priors, long-distance dependencies, and multi-scale semantic context.
[0050] S105, the final feature representation is input into a preset task prediction model to generate the final three-dimensional point cloud analysis result.
[0051] In specific implementation, this feature The data is transmitted to a task head designed for a specific task for processing to generate the final 3D point cloud analysis results. Depending on the specific application scenario (such as 3D shape classification or semantic segmentation), the task head performs non-linear mapping on the features using a multilayer perceptron (MLP) and calculates the probability of each category using a softmax function, ultimately outputting the global category label or point-by-point semantic label of the point cloud.
[0052] The proposed Geometry-Aware Transformer (GA-Transformer) overcomes the standard Transformer's dependence on absolute coordinates through conditional position encoding and geometric prior attention. It effectively captures the long-range topological structure and fine-grained geometric features of point clouds without violating rotation invariance, significantly improving recognition accuracy in complex scenes (such as indoor scene segmentation). Thanks to the stable construction strategy of local reference frames and the effective aggregation of global context features, this method exhibits strong robustness to noise interference and density variations in the input point cloud. Experiments demonstrate that even with extremely sparse input points (e.g., 600 points) or the presence of Gaussian noise, the model performance does not show significant degradation.
[0053] In summary, the rotation-invariant point cloud analysis method in the above embodiments of the present invention eliminates [the problem] by constructing a local reference frame. The interference of arbitrary rotations within the group on feature extraction solves the pose ambiguity problem existing in traditional networks. By adopting a single-stream architecture, the number of model parameters is reduced while maintaining the same or even higher accuracy, improving computational efficiency and greatly reducing computing costs. By using the geometrically aware Transformer through conditional position encoding and geometric prior attention, it overcomes the dependence of the standard Transformer on absolute coordinates. It can effectively capture the long-distance topological structure and fine-grained geometric features of point clouds without destroying rotation invariance, significantly improving the recognition accuracy in complex scenes.
[0054] Example 2 In another aspect, this invention also proposes a rotation-invariant point cloud analysis system, please refer to [link / reference needed]. Figure 2 The figure shows a rotation-invariant point cloud analysis system according to a second embodiment of the present invention, the system comprising: The data acquisition module 11 is used to acquire three-dimensional point cloud data of the target object using a three-dimensional acquisition device, and to perform data enhancement and multi-view generation on the three-dimensional point cloud data. The interference cancellation module 12 is used to input the processed point cloud data into the rotation-invariant edge convolution model, project the point cloud onto the local coordinate system by constructing a local reference frame, eliminate global rotation interference, and output rotation-invariant geometric feature vectors. Furthermore, the interference cancellation module 12 is specifically used for: Using the k-nearest neighbor algorithm to analyze any point in the processed point cloud data Search neighborhood point set And calculate the local centroid. Covariance Matrix ; For the covariance matrix Eigenvalue decomposition is performed to obtain the eigenvector matrix, and a unique local reference frame is determined by the sign disambiguation rule; Relative to the local centroid, each point in the neighborhood is considered. The relative coordinates are projected onto the local reference frame to obtain rotation-invariant coordinates. ; With the rotation-invariant coordinates As input, geometric features are extracted through edge convolution, and the rotation-invariant coordinates are... Its difference relative to the center point The features are concatenated, nonlinearly mapped using a multilayer perceptron, and then aggregated using max pooling to output a rotation-invariant geometric feature vector. .
[0055] The geometry perception module 13 is used to input the rotation-invariant geometric feature vector into the geometry perception Transformer model. Through conditional position encoding and geometric prior attention mechanism, it captures the long-distance contextual dependencies of the point cloud while maintaining rotation invariance, so as to obtain geometry perception features. Furthermore, the geometric sensing module 13 is specifically used for: The rotation-invariant geometric feature vector is input into the standard edge convolutional layer of the geometry-aware Transformer model to increase the feature dimension and obtain enhanced features; The rotation-invariant coordinates are used to generate a position code, which is then superimposed on the enhanced feature to obtain a feature that fuses position information. In the self-attention mechanism, a geometric bias term driven by rotation-invariant coordinates is introduced, attention weights are calculated, the value vector is weighted and summed according to the attention weights, and the position encoding is incorporated to output aggregated features with local geometric and global contextual information.
[0056] The feature fusion module 14 is used to construct a multi-scale pyramid pooling model, process the geometric perception features, and generate a final feature representation that combines local geometric details and global semantic information through multi-level pooling, upsampling and feature fusion. Furthermore, the feature fusion module 14 is specifically used for: Multi-scale adaptive average pooling and upsampling are performed on the geometrically perceived features to obtain a multi-scale feature set; Global max pooling is performed on the geometrically perceived features to obtain global features, local features of the geometrically perceived features are retained, and the local features, the global features, and the multi-scale feature set are concatenated by channels. The concatenated features are then fed into a fusion multilayer perceptron for dimensionality reduction and feature recombination to obtain the final feature representation.
[0057] The point cloud analysis module 15 is used to input the final feature representation into a preset task prediction model to generate the final three-dimensional point cloud analysis result.
[0058] The functions or operation steps implemented by the above modules and units are largely the same as those in the above method embodiments, and will not be repeated here.
[0059] The rotation-invariant point cloud analysis system provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment.
[0060] Example 3 This invention also proposes a computer, please refer to [link / reference]. Figure 3The computer shown in the third embodiment of the present invention includes a memory 10, a processor 20, and a computer program 30 stored in the memory 10 and executable on the processor 20. When the processor 20 executes the computer program 30, it implements the above-described rotation-invariant point cloud analysis method.
[0061] The memory 10 includes at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 10 can be an internal storage unit of a computer, such as the computer's hard disk. In other embodiments, the memory 10 can be an external storage device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Furthermore, the memory 10 can include both internal and external storage units of the computer. The memory 10 can be used not only to store application software and various types of data installed on the computer, but also to temporarily store data that has been output or will be output.
[0062] In some embodiments, the processor 20 may be an electronic control unit (ECU), a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip, used to run program code stored in the memory 10 or process data, such as executing access restriction programs.
[0063] It should be pointed out that, Figure 3 The structure shown does not constitute a limitation on the computer. In other embodiments, the computer may include fewer or more components than shown, or combine certain components, or have different component arrangements.
[0064] This invention also proposes a storage medium storing a computer program that, when executed by a processor, implements the rotation-invariant point cloud analysis method described above.
[0065] Those skilled in the art will understand that the logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.
[0066] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0067] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0068] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0069] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A rotation-invariant point cloud analysis method, characterized in that, include: The three-dimensional point cloud data of the target object is acquired using a three-dimensional acquisition device, and the three-dimensional point cloud data is then augmented and multi-view generated. The processed point cloud data is input into a rotation-invariant edge convolution model. By constructing a local reference frame, the point cloud is projected onto a local coordinate system to eliminate global rotation interference and output rotation-invariant geometric feature vectors. The rotation-invariant geometric feature vector is input into the geometry-aware Transformer model. Through conditional position encoding and geometric prior attention mechanism, the long-range contextual dependencies of the point cloud are captured while maintaining rotation invariance, so as to obtain geometry-aware features. A multi-scale pyramid pooling model is constructed to process the geometric perception features. Through multi-level pooling, upsampling, and feature fusion, a final feature representation that combines local geometric details and global semantic information is generated. The final feature representation is input into a preset task prediction model to generate the final 3D point cloud analysis result.
2. The rotation-invariant point cloud analysis method according to claim 1, characterized in that, The steps of inputting the processed point cloud data into a rotation-invariant edge convolution model, projecting the point cloud onto a local coordinate system by constructing a local reference frame to eliminate global rotation interference, and outputting rotation-invariant geometric feature vectors include: Using the k-nearest neighbor algorithm to analyze any point in the processed point cloud data Search neighborhood point set And calculate the local centroid. Covariance Matrix ; For the covariance matrix Eigenvalue decomposition is performed to obtain the eigenvector matrix, and a unique local reference frame is determined by the sign disambiguation rule; Relative to the local centroid, each point in the neighborhood is considered. The relative coordinates are projected onto the local reference frame to obtain rotation-invariant coordinates. ; With the rotation-invariant coordinates As input, geometric features are extracted through edge convolution, and the rotation-invariant coordinates are... Its difference relative to the center point The features are concatenated, nonlinearly mapped using a multilayer perceptron, and then aggregated using max pooling to output a rotation-invariant geometric feature vector. .
3. The rotation-invariant point cloud analysis method according to claim 2, characterized in that, The local centroid The calculation formula is: In the formula, Represents points within the neighborhood. Indicates the number of neighboring points; The covariance matrix The calculation formula is: in, Indicates matrix transpose; The rotation-invariant coordinates The calculation formula is: In the formula, Indicates a local reference frame; The rotation-invariant geometric eigenvectors The calculation formula is: In the formula, This indicates a splicing operation. This represents a multilayer perceptron.
4. The rotation-invariant point cloud analysis method according to claim 2, characterized in that, The steps of inputting the rotation-invariant geometric feature vector into a geometry-aware Transformer model, and capturing long-range contextual dependencies of the point cloud while maintaining rotation invariance through conditional position encoding and geometric prior attention mechanisms to obtain geometry-aware features include: The rotation-invariant geometric feature vector is input into the standard edge convolutional layer of the geometry-aware Transformer model to increase the feature dimension and obtain enhanced features; The rotation-invariant coordinates are used to generate a position code, which is then superimposed on the enhanced feature to obtain a feature that fuses position information. In the self-attention mechanism, a geometric bias term driven by rotation-invariant coordinates is introduced, attention weights are calculated, the value vector is weighted and summed according to the attention weights, and the position encoding is incorporated to output aggregated features with local geometric and global contextual information.
5. The rotation-invariant point cloud analysis method according to claim 1, characterized in that, The steps of constructing a multi-scale pyramid pooling model to process the geometrically perceived features, and generating a final feature representation that combines local geometric details and global semantic information through multi-level pooling, upsampling, and feature fusion include: Multi-scale adaptive average pooling and upsampling are performed on the geometrically perceived features to obtain a multi-scale feature set; Global max pooling is performed on the geometrically perceived features to obtain global features, local features of the geometrically perceived features are retained, and the local features, the global features, and the multi-scale feature set are concatenated by channels. The concatenated features are then fed into a fusion multilayer perceptron for dimensionality reduction and feature recombination to obtain the final feature representation.
6. A rotation-invariant point cloud analysis system, characterized in that, include: The data acquisition module is used to acquire three-dimensional point cloud data of the target object using a three-dimensional acquisition device, and to perform data enhancement and multi-view generation on the three-dimensional point cloud data; The interference cancellation module is used to input the processed point cloud data into the rotation-invariant edge convolution model, project the point cloud onto the local coordinate system by constructing a local reference frame, eliminate global rotation interference, and output rotation-invariant geometric feature vectors. The geometry perception module is used to input the rotation-invariant geometric feature vector into the geometry perception Transformer model. Through conditional position encoding and geometric prior attention mechanism, it captures the long-distance contextual dependencies of the point cloud while maintaining rotation invariance, so as to obtain geometry perception features. The feature fusion module is used to construct a multi-scale pyramid pooling model, process the geometric perception features, and generate a final feature representation that combines local geometric details and global semantic information through multi-level pooling, upsampling and feature fusion. The point cloud analysis module is used to input the final feature representation into a preset task prediction model to generate the final three-dimensional point cloud analysis result.
7. The rotation-invariant point cloud analysis system according to claim 6, characterized in that, The interference cancellation module is specifically used for: Using the k-nearest neighbor algorithm to analyze any point in the processed point cloud data Search neighborhood point set And calculate the local centroid. Covariance Matrix ; For the covariance matrix Eigenvalue decomposition is performed to obtain the eigenvector matrix, and a unique local reference frame is determined by the sign disambiguation rule; Relative to the local centroid, each point in the neighborhood is considered. The relative coordinates are projected onto the local reference frame to obtain rotation-invariant coordinates. ; With the rotation-invariant coordinates As input, geometric features are extracted through edge convolution, and the rotation-invariant coordinates are... Its difference relative to the center point The features are concatenated, nonlinearly mapped using a multilayer perceptron, and then aggregated using max pooling to output a rotation-invariant geometric feature vector. .
8. The rotation-invariant point cloud analysis system according to claim 7, characterized in that, The geometry sensing module is specifically used for: The rotation-invariant geometric feature vector is input into the standard edge convolutional layer of the geometry-aware Transformer model to increase the feature dimension and obtain enhanced features; The rotation-invariant coordinates are used to generate a position code, which is then superimposed on the enhanced feature to obtain a feature that fuses position information. In the self-attention mechanism, a geometric bias term driven by rotation-invariant coordinates is introduced, attention weights are calculated, the value vector is weighted and summed according to the attention weights, and the position encoding is incorporated to output aggregated features with local geometric and global contextual information.
9. A readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the rotation-invariant point cloud analysis method as described in any one of claims 1 to 5.
10. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the rotation-invariant point cloud analysis method as described in any one of claims 1 to 5.
