Method for realizing classification of three-dimensional point clouds based on NCFMama network

Abstract

The invention discloses a method for realizing three-dimensional point cloud classification based on an NCFMama network, relates to the technical field of three-dimensional data processing, and solves the problems that an existing point cloud classification method is insufficient in local feature extraction, difficult in consideration of calculation complexity and global modeling capability and the like. According to the method, three dimensions of spatial distribution, feature similarity and direction consistency of a local neighborhood of the point cloud are comprehensively analyzed through a neighborhood consistency feature modulator, adaptive feature aggregation of different geometric regions is realized, and meanwhile, a bidirectional spatial perception fusion module is provided. The module uses point cloud three-dimensional space distance information to guide feature interaction of a bidirectional state space model, information exchange between adjacent points of a point cloud space is enhanced, and high-precision and high-efficiency point cloud classification is realized. According to the method, the advantages of calculation efficiency, classification precision and robustness are combined, and a new technical path is provided for application of the Mama architecture in point cloud processing.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional data processing, and in particular to a method for classifying three-dimensional point clouds based on an NCFMamba (Neighborhood Consistency Feature Modulated Point Cloud Classification Network) network. Background Art

[0002] As a direct digital representation of the physical world, 3D point clouds are playing an increasingly important role in fields such as autonomous driving and robotic navigation. Unlike the regular pixel arrays of 2D images, point cloud data exhibits characteristics such as disorder, sparsity, and irregular distribution. These characteristics present unique technical challenges for point cloud classification. With the rapid development and widespread adoption of 3D sensor technology, achieving efficient and accurate point cloud category recognition has become a critical issue that needs to be addressed in the 3D vision field.

[0003] Although three-dimensional point clouds are currently widely used in fields such as autonomous driving and robot navigation, their characteristics such as disorder, sparsity and irregular distribution bring unique technical challenges to point cloud classification. How to achieve efficient and accurate point cloud category recognition has become a key issue that needs to be urgently solved in the field of three-dimensional vision.

[0004] Among existing point cloud processing methods, early voxel-based or multi-view methods suffer from information loss; convolutional neural network methods based on local aggregation, such as PointNet, PointNet++, and DGCNN, are limited by predefined receptive fields and have difficulty capturing long-range geometric dependencies; although global methods that introduce the Transformer architecture can capture global dependencies, the quadratic computational complexity leads to high computational costs, and existing methods find it difficult to balance computational efficiency and modeling capabilities.

[0005] The Mamba network, with its selective scanning mechanism and hardware-aware design, enables efficient global context modeling with linear computational complexity, offering new possibilities for resolving this contradiction. Research projects such as Pointmamba and Point Cloud Mamba have explored its application in point cloud tasks. These methods, which process point clouds using relevant algorithms and then input them into the Mamba module for feature extraction, demonstrate promising performance and computational complexity. However, existing Mamba point cloud methods ignore the differences in geometric properties across point cloud regions, lack an adaptive modeling mechanism for local geometric structure, and employ a simple static bidirectional scanning mechanism that makes it difficult to deeply mine three-dimensional spatial association information.

[0006] Therefore, the present invention proposes a classification method for three-dimensional point clouds based on the NCFMamba network. Summary of the Invention

[0007] In order to solve the problems faced by existing point cloud classification methods, such as insufficient local feature extraction and difficulty in balancing computational complexity and global modeling capabilities, the present invention provides a classification method for three-dimensional point clouds based on the NCFMamba network.

[0008] The classification method of 3D point cloud based on NCFMamba network is implemented by the following steps:

[0009] Step 1: Divide the point cloud into multiple local point groups and obtain G local feature sequences;

[0010] Step 2: Use the NCFMamba network model to process the local feature sequence obtained in step 1; obtain the features output by the NCFMamba network model; the specific process is:

[0011] Step 2.1: Introduce learnable classification labels [CLS] to aggregate the global information of G local feature sequences;

[0012] Step 22: Add standard learnable position encoding to the learnable classification label [CLS] and the G local feature sequences and input them into the NCFMamba network model;

[0013] Step 2 and 3: Set the NCFMamba network model to consist of T-layer stacked NCFM modules and a Bi-SSM structure based on the BSAF module; the NCFM module analyzes the multi-dimensional consistency of the local neighborhood of the local feature sequence to achieve adaptive feature aggregation for different geometric regions; the bidirectional Bi-SSM structure based on the BSAF module introduces an adaptive interaction mechanism based on three-dimensional spatial position to achieve intelligent fusion of forward and backward information flows; and finally output the fused features;

[0014] Step 3: Map the fused features output by the NCFMamba network model to the final category space to obtain the point cloud classification result.

[0015] Beneficial effects of the present invention:

[0016] 1. In response to the problems faced by point cloud classification methods, such as insufficient local feature extraction and difficulty in balancing computational complexity and global modeling capabilities, the present invention proposes a point cloud classification network based on neighborhood consistency feature modulation (NCFMamba). This method uses a neighborhood consistency feature modulator to comprehensively analyze the spatial distribution, feature similarity, and directional consistency of the local neighborhood of the point cloud, thereby achieving adaptive feature aggregation for different geometric regions. At the same time, a bidirectional spatial perception fusion module is proposed, which uses the three-dimensional spatial distance information of the point cloud to guide the feature interaction of the bidirectional state space model, enhances the information exchange between neighboring points in the point cloud space, and achieves high-precision and high-efficiency point cloud classification. Extensive experimental results on the ModelNet40 and ScanObjectNN benchmark datasets show that the proposed method has better point cloud classification accuracy than existing mainstream methods while reducing the model weight, providing a new technical solution for three-dimensional point cloud classification.

[0017] 2. The method proposed in the present invention first divides the point cloud into multiple local regions and extracts initial features through a point cloud grouping and feature encoding module. The NCFMamba encoder then performs deep feature learning through a multi-layer stacked structure. It includes two core modules: the Neighborhood Consistency Feature Modulator (NCFM) comprehensively analyzes the spatial distribution, feature similarity, and directional consistency of the local neighborhood to achieve adaptive feature aggregation for different geometric regions (smooth regions and boundary regions), effectively protecting geometric details. The Bidirectional Spatial Aware Fusion Module (BSAF) uses the three-dimensional spatial distance information of the point cloud to guide the intelligent fusion of Mamba bidirectional features, enhancing information interaction between spatially adjacent points and improving global modeling capabilities while maintaining linear computational complexity. Finally, the classification module maps the encoded features to the target category space and outputs the classification results.

[0018] 3. The NCFM module proposed in the method of the present invention fully utilizes neighborhood information for feature enhancement in smooth areas through a dynamic modulation mechanism, and relies more on center point features in boundary areas to protect geometric details, effectively solving the problem of feature confusion at geometric boundaries in traditional methods.

[0019] 4. The BSAF module in the method of the present invention introduces the 3D spatial coordinate information of the point cloud into Mamba's bidirectional feature fusion process. Through adaptive weight calculation based on Euclidean distance, it enables spatially adjacent points to have stronger bidirectional information interaction, achieving effective modeling of global spatial relationships with minimal parameter overhead.

[0020] 5. The present method has a wide range of applications: The proposed neighborhood consistency feature modulator uses multi-dimensional geometric analysis to differentiate different regions of a point cloud, enabling accurate feature extraction of the local geometric structure of the point cloud. The proposed bidirectional spatial perception fusion module optimizes the forward and backward feature interaction of the point cloud by introducing three-dimensional spatial distance information, solving the problem of adapting sequence modeling to the spatial characteristics of the point cloud. The combined advantages of the present method in terms of computational efficiency, classification accuracy, and robustness provide a new technical path for the application of the Mamba architecture in point cloud processing and offer a reliable solution for point cloud classification in practical application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a framework diagram of the NCFMamba network in the 3D point cloud classification method based on the NCFMamba network described in the present invention;

[0022] Figure 2 Schematic diagram of the neighborhood consistency feature modulator;

[0023] Figure 3 is a bidirectional space model; (a) is the bidirectional state space model framework diagram; (b) is the bidirectional space perception fusion module BSAF;

[0024] Figure 4 The experimental results are shown for different input point cloud densities;

[0025] Figure 5 The point cloud classification results under different encoder layer conditions. DETAILED DESCRIPTION

[0026] Specific implementation method 1. Combination Figures 1 to 3 This embodiment describes a classification method for three-dimensional point clouds based on the NCFMamba network. This method uses a neighborhood consistency feature modulator (NCFM module) to comprehensively analyze the spatial distribution, feature similarity, and directional consistency of the local neighborhood of the point cloud, thereby achieving adaptive feature aggregation for different geometric regions. At the same time, a bidirectional spatial model is proposed, namely, a bidirectional state space model (Bi-SSM) based on a spatial perception fusion module (BSAF). The BSAF module uses the three-dimensional spatial distance information of the point cloud to guide the feature interaction of the Bi-SSM model, thereby enhancing the information exchange between neighboring points in the point cloud space and achieving high-precision and high-efficiency point cloud classification. This provides a new technical solution for three-dimensional point cloud classification. The specific implementation process of the classification method described in this embodiment is as follows:

[0027] Step 1: Divide the point cloud into multiple local point groups through point cloud grouping and feature encoding, extract initial features, and finally obtain G local feature sequences; the specific process is as follows:

[0028] Step 1.1: Set the input original point cloud Where N represents the number of point clouds, each point p in the point cloud P i Contains 3D coordinate information. Use the Farthest Point Sampling (FPS) algorithm to select G center points from the original point cloud;

[0029] Step 1.2: For each center point c g Use the K-Nearest Neighbors (KNN) algorithm to find the K nearest neighbor points in the original point cloud to form a local point group set In this embodiment, a local grouping strategy is adopted to transform the coordinates of the points into a local coordinate system with the center point as the origin, so as to facilitate the subsequent capture of geometric structure information around the center point.

[0030] Step 1.3: Use lightweight PointNet to extract features from each local point group through a shared multi-layer perceptron (MLP) and maximum pooling operation. The formula is:

[0031]

[0032] Where, is the feature representation of the g-th local point group, and C is the feature dimension. G local feature sequences are obtained by point cloud grouping and feature encoding. Among them, the number of sequences corresponds to the number of center points;

[0033] Step 2: According to step 1, the local feature sequence of the input point cloud is obtained, and the point cloud feature sequence processing framework NCFMamba Encoder is constructed, that is, the NCFMamba network model is constructed; the NCFMamba network model consists of T-layer stacked NCFM modules and a bidirectional state space model (Bi-SSM) based on the spatial perception fusion module (BSAF);

[0034] The NCFM module achieves adaptive feature aggregation of different geometric regions (smooth regions, boundary regions) by comprehensively analyzing the spatial distribution, feature similarity and directional consistency of the local neighborhood, effectively protecting geometric details.

[0035] The BSAF module uses the 3D spatial distance information of the point cloud to guide the Bi-SSM model to enhance the information interaction between spatially adjacent points, improving the global modeling capability while maintaining linear computational complexity. After T layers of iteration, the encoder’s final output feature h T .

[0036] The specific implementation process is:

[0037] Step 2.1: The global information of the entire feature sequence is aggregated by introducing the learnable classification label [CLS].

[0038] Step 2.2: By adding a standard learnable positional encoding E to the G+1 tokens (including [CLS] and G local features) pos , which enables the NCFMamba network model to perceive the relative position relationship of different local areas in the overall point cloud;

[0039] Step 2.3: Adopt NCFM module to adaptively adjust feature aggregation strategy by comprehensively analyzing the multi-dimensional consistency of local neighborhood. Figure 2 As shown, the specific steps are:

[0040] (1) From the spatial consistency S spatial , feature consistency S feature , directional consistency S angular The neighborhood consistency is analyzed from three perspectives, and the three different scales are normalized. Finally, the neighborhood consistency features are concatenated and processed through a lightweight MLP to obtain the final neighborhood consistency score.

[0041] First, spatial consistency S spatial , by calculating the distance d from the neighborhood point to the center point i,k The standard deviation of the distance is used to measure the degree of spatial distribution dispersion, reflecting the tightness of the local geometry. In the smooth area, the neighborhood points are evenly distributed around the center point, and the distance standard deviation is small; while in the boundary area, due to geometric mutations, the neighborhood points are unevenly distributed, resulting in a larger distance standard deviation. Its calculation formula is:

[0042] d i,k =‖p i,k -c i ‖2

[0043]

[0044] In the formula, K represents the number of nearest neighbor points, c i is the center point of the point cloud, p i,k is a neighborhood point, and std(·) represents the standard deviation operator.

[0045] Secondly, the above feature consistency S feature It reflects the coherence of semantic information and is evaluated by calculating the average difference between the neighborhood features and the center point features. Within the same semantic region, the feature difference is small; while at the junction of different semantic regions, the feature difference increases significantly. Its calculation formula is:

[0046]

[0047] In the formula, K represents the number of nearest neighbor points, x i is the center point feature, x i,k is the neighborhood point feature.

[0048] In this embodiment, the directional consistency S angular It reflects the curvature change of the neighborhood by calculating the cosine value v of the angle between adjacent direction vectors i,k In flat areas, the direction vector changes slowly; in high curvature areas or at edges, the direction changes dramatically. The calculation formula is:

[0049]

[0050] Where c i is the center point of the point cloud, p i,k are neighborhood points, <·,·> represents vector inner product, and var(·) represents variance.

[0051] Then, in order to unify the consistency indicators of different scales to a similar numerical range, normalization is performed by adopting the hyperbolic tangent function:

[0052]

[0053] Among them, λ spatial ,λ feature and λ angular is a learnable scaling parameter used to adjust the sensitivity of different consistency metrics.

[0054] Finally, the three neighborhood consistency features are concatenated and processed through a lightweight MLP to obtain the final consistency score α:

[0055]

[0056] Where σ(·) is the Sigmoid activation function, which ensures that the output is between 0 and 1. α close to 1 indicates that the neighborhood is highly consistent (smooth area), while α close to 0 indicates that the neighborhood changes drastically (boundary area).

[0057] (2) Adaptive feature modulation is performed based on the calculated neighborhood consistency score α. The specific process is:

[0058] First, the neighborhood features and coordinates are normalized to eliminate the impact of scale differences:

[0059]

[0060]

[0061] Where μ x and σ xare the mean and standard deviation of the neighborhood features, σ p is the standard deviation of the spatial coordinates in the neighborhood.

[0062] Then, by splicing with the central feature to expand the feature dimension, it provides richer information for subsequent differentiation processing. In order to make the network model emphasize the central point features more in the boundary area (α is close to 0) to avoid cross-boundary feature confusion. In the smooth area (α is close to 1), the model can make full use of the neighborhood information and enhance local coherence. edge and β smooth are learnable parameters that control the weights of edge region enhancement and smooth region enhancement, respectively. Finally, an affine transformation is applied to further adjust the features:

[0063]

[0064] Where γ and δ are learnable parameters, and ⊙ represents element-wise multiplication.

[0065] (3) After feature modulation, it is necessary to aggregate the neighborhood features into a single representation. The present invention adopts an adaptive weighted aggregation strategy to assign different weights according to the importance of each neighborhood point. The specific process is:

[0066] First, the importance score of each modulated feature is calculated through a scoring network, which is:

[0067]

[0068] Then, the aggregate weight w obtained by softmax normalization of the scores of each neighborhood point is i,k Perform weighted aggregation on the modulated features. The calculation formula is:

[0069]

[0070] To further enhance the expressive power of features, the aggregated features are fed into a shared MLP for inter-channel feature interaction and nonlinear transformation:

[0071] f i =MLP mix (f i agg )

[0072] Where, MLP mix It is a two-layer fully connected network with a ReLU activation function in the middle to mix information from different channels and introduce additional nonlinearity.

[0073] Through this adaptive weighted aggregation mechanism, the network model can dynamically adjust the contribution of each neighboring point in the point cloud based on its feature quality and relevance, thereby obtaining a more discriminative local feature representation. In particular, in boundary areas, this mechanism can automatically reduce the weight of neighboring points from different semantic regions, maintaining the feature's ability to express local geometric structure.

[0074] Step 2.4, such as Figure 3 As shown in the figure, a Bi-SSM structure based on the spatial perception fusion module BSAF is proposed. Figure 3 As shown in (a); after generating the local feature representation in step 2.3, the intelligent fusion of forward and backward information flows is achieved by introducing an adaptive interaction mechanism based on three-dimensional spatial position. The specific BSAF module is as follows Figure 3 As shown in (b). The specific steps include:

[0075] (1) First, calculate the spatial distance feature. Given the 3D center coordinates of the point cloud group For each pair of center points (c i ,c j ), calculate its Euclidean distance d ij , and then construct the spatially aware position weights For the classification mark [CLS], since it does not correspond to a specific spatial position, an independent position enhancement parameter is introduced, and the calculation formula is:

[0076] d ij =‖c i -c j ‖2

[0077]

[0078] Where, d i,ref is the distance from the i-th point to the reference point, σ spatial is a learnable spatial scale parameter that controls the range of interaction, γ cls is a learnable parameter and σ(·) is the Sigmoid function.

[0079] (2) After obtaining the complete spatial position weight vector in (1), the bidirectional interaction strength β is defined based on the spatial position weight:

[0080]

[0081] β=σ(β base )·W spatial

[0082] Where, β base is the global interaction strength parameter, which is ensured to be between 0 and 1 by the Sigmoid function.

[0083] (3) The calculation formula of the enhanced bidirectional feature is:

[0084] h′ forward =h forward +β⊙h backward

[0085] h′ backward =h backward +β⊙h forward

[0086] Where h forward and h backward denote the output features of the forward and backward SSM respectively, and ⊙ denotes element-wise multiplication.

[0087] This design enables spatially close points to have stronger two-way information interaction, while the interaction between distant points is relatively weak.

[0088] (4) To further improve the effect of bidirectional information fusion, a learnable bidirectional coupling parameter θ = [θ f ,θ b ] and the adaptive weight [α f ,α b ],θ f and θ b are the forward coupling parameter and the backward coupling parameter respectively; this ensures that α f +α b =1, which ensures the scale stability of the fused features. Finally, the fused bidirectional features are obtained. The calculation formula is:

[0089] [α f ,α b ]=softmax(θ)

[0090] h fusion =α f ·h′ forward +α b ·h′ backward

[0091] Through the BSAF module, we combine Mamba's efficient sequence modeling capabilities with the three-dimensional spatial characteristics of point clouds, which only adds a small number of learnable parameters compared to the introduction of an additional attention mechanism. In addition, by introducing spatial position information, the network model can adaptively adjust the interaction strength according to the geometric distribution of the point cloud, while maintaining computational efficiency while improving the model's ability to perceive global spatial relationships, ultimately obtaining a bidirectional feature h that is fused forward and backward. fusion .

[0092] The NCFMamba encoder extracts features layer by layer through a T-layer stacked NCFM and bidirectional SSM structure. The local feature sequence of the input point cloud is effectively converted into a high-level feature representation rich in global context information and local geometric details. Specifically, its forward propagation process can be expressed as:

[0093] h0=[CLS,f1,f2,...,f G ]+E pos

[0094] h l ′=h l-1 +NCFM(LN(h l-1 ))

[0095] h l =h l ′+Bi-SSM(LN(h l ′))

[0096] In the formula, E pos The meaning is to add a standard learnable position encoding to G+1 tags (including [CLS] and G local features), h l is the output of layer l, l∈[0,T], and LN represents the layer normalization operation. The use of layer normalization helps stabilize the training process, while residual connections facilitate gradient propagation, making the training of deep networks more effective.

[0097] After T layers of iteration, the final output of the encoder is Among them, f CLS The final representation corresponding to the [CLS] tag aggregates the global semantic information of the entire point cloud; These are representations of each local grouping after deep feature extraction. These encoded features retain rich local geometric structure information and, through bidirectional state space modeling, establish effective long-range dependencies, providing a discriminative feature foundation for subsequent classification tasks. In this example, T is set to 12.

[0098] Step 3: After obtaining the features of the T-layer encoder output in step 2, map the learned point cloud representation to the final category space. The specific steps are:

[0099] (1) To preserve local geometric details, h T All local grouping features in Perform double pooling operation:

[0100]

[0101] (2) Secondly, extract the output feature f of the [CLS] tagCLS , f group and f CLS Splicing to obtain the final classification features, which are used to input the classification head to obtain the point cloud category:

[0102]

[0103] (3) The classification head uses a three-layer fully connected network with batch normalization and Dropout regularization. The final classification features are passed through the classification head to obtain the classification results of the point cloud:

[0104]

[0105] Where N c is the number of categories.

[0106] The present embodiment describes a method for classifying three-dimensional point clouds based on the NCFMamba network. This method is implemented through point cloud grouping and feature encoding, a neighborhood consistency feature modulator NCFM, a bidirectional state space model Bi-SSM based on a bidirectional spatial perception fusion module BSAF, and a classification module. According to this method, systematic training and verification were carried out on the ModelNet40 dataset and the ScanObjectNN dataset based on the PyTorch deep learning framework.

[0107] Specific implementation method 2: Figure 3 and Figure 4 This embodiment is described as a verification example of the classification method of three-dimensional point cloud based on NCFMamba network described in the first embodiment.

[0108] The model was trained on an NVIDIA TITAN RTX GPU using the PyTorch framework. The backbone network consists of T = 12 stacked encoders with a feature dimension C of 384. The grouping strategy divides each sample into 128 groups, each containing 32 points. The neighborhood size k of the local consistency feature modulator (NCFM module) is set to 6. The optimizer is AdamW, and the initial learning rate is set to 5×10 -4 , the weight decay is 0.05. The learning rate scheduling adopts the cosine annealing (CosineAnnealing withWarmRestarts) strategy, with an initial cycle length of 300 rounds, a cycle multiplication factor of 1.8, and a minimum learning rate of 1×10 -6 , the number of warm-up rounds is 20. The batch size is 32, and the gradient clipping threshold is 10. The loss function uses cross entropy loss with 0.2 label smoothing, and the DropPath rate is 0.2 to enhance regularization.

[0109] In addition, the experiment employed the Point-MAE pre-training method, which randomly masks 60% of the input point cloud blocks, reconstructs the masked points using an encoder-decoder architecture, and optimizes the distance between the predicted and true points using the Chamfer Distance loss function. Pre-training was performed on the ShapeNet dataset for 300 epochs, followed by fine-tuning for the point cloud classification task using only the pre-trained weights.

[0110] To verify the effectiveness of the proposed method, the classification performance and robustness of NCFMamba were evaluated on two standard benchmark datasets, ModelNet40 and ScanObjectNN, and compared with existing mainstream methods. Detailed ablation experiments were conducted to analyze the contribution of each module to the overall performance. At the same time, an objective evaluation indicator, namely overall accuracy (OA), was used to evaluate the proposed method. In classification, there are four possible results, including: TP (TruePositive), which represents the number of samples that are actually positive and are correctly predicted as positive; TN (TrueNegative), which represents the number of samples that are actually negative and are correctly predicted as negative; FP (FalsePositive), which represents the number of samples that are actually negative and are incorrectly predicted as positive; and FN (FalseNegative), which represents the number of samples that are actually positive and are incorrectly predicted as negative. Overall accuracy is the most commonly used evaluation indicator in point cloud classification tasks. It is defined as the ratio of correctly classified samples to the total number of samples. The calculation formula is as follows:

[0111]

[0112] In this experiment, to further improve the model's prediction accuracy and robustness, a voting mechanism (Vote) was used for test-time enhancement. Specifically, each test sample was subjected to 10 different random sampling and data augmentation operations, including random point sampling and rotation transformation. The multiple prediction results for the same sample were then probabilistically averaged, and the category with the highest average probability was selected as the final prediction. This voting strategy can effectively reduce the accidental errors caused by random sampling and transformation in a single prediction, and achieve more stable and accurate classification performance by integrating multiple prediction results. It should be noted that unless explicitly marked, the voting mechanism is not used by default. In the following experimental results analysis, "w / o vot" is used to indicate the absence of a voting mechanism, and "w / vot" is used to indicate the presence of a voting mechanism.

[0113] To verify the performance of NCFMamaba for point cloud classification, classification experiments were conducted on the synthetic ModelNet40 dataset. Several mainstream methods based on different backbone networks were also selected for comparative testing. The ModelNet40 dataset includes approximately 12,000 synthetic 3D CAD models across 40 categories. The number of ModelNet40 training samples was set to 9,843 frames, and the number of test samples was set to 2,468 frames. The experiment used N = 1,024 points as input, and applied scaling and translation for data augmentation. Table 1 shows the classification results for the ModelNet40 dataset.

[0114] Table 1

[0115]

[0116] As shown in Table 1, where P(M) represents the number of parameters and F(G) represents the model complexity (computational effort), NCFMamba achieves a classification accuracy of 94.7%. Compared to classic local feature aggregation methods, NCFMamba, with its global modeling capabilities, surpasses PointNet++ and DGCNN. Compared to Transformer-based methods, NCFMamba not only surpasses PCT and PointConT in accuracy, but also avoids the quadratic computational complexity bottleneck introduced by the self-attention mechanism.

[0117] To further validate the effectiveness and practicality of the proposed network model, we conducted experimental analysis on the more challenging ScanObjectNN dataset. The ScanObjectNN dataset contains point cloud data obtained from scans of real scenes. Compared to the synthetic data used in ModelNet40, it has more noise, occlusion, and incompleteness, and can better reflect the complexities of real-world application scenarios. Not only does the classification accuracy compare with existing methods to verify the competitiveness of the proposed method, but we also conduct an in-depth analysis of the network's performance at different input point cloud densities to evaluate the method's robustness and scale adaptability.

[0118] (1) ScanObjectNN dataset point cloud classification experiment;

[0119] The ScanObjectNN dataset contains approximately 15K objects from 15 categories. These point clouds are real-world object point clouds obtained by scanning in a cluttered background. The dataset includes three variants: OBJ_BG, OBJ_ONLY, and PB_T50_RS. OBJ_BG is the original subset containing background noise, retaining environmental interference in the real scene; OBJ_ONLY is the subset after removing the background, retaining only the pure target point cloud.

[0120] PB_T50_RS is the most challenging augmentation subset, incorporating random translation, rotation, and scaling perturbations while preserving background noise. The number of training samples for OBJ_BG and OBJ_ONLY is set to 2309 frames, and the number of testing samples is set to 581 frames; the number of training samples for PB_T50_RS is set to 11416 frames, and the number of testing samples is set to 2882 frames. During network model training, rotation is used as data augmentation, and the input point cloud size is set to 2048. The effectiveness of the proposed neighborhood consistency feature modulation strategy and spatial perception fusion module is demonstrated. Table 2 shows the classification results for the ScanObjectNN dataset.

[0121] Table 2

[0122]

[0123] As shown in Table 2, NCFMamba achieved the best performance on all three subsets. When faced with real object point cloud data, the proposed NCFMamba network model still demonstrated excellent classification performance and strong competitive advantages. First, compared with the traditional MLP backbone network method, NCFMamba's long sequence modeling capability based on the Mamba architecture enables it to better capture the global dependencies in the point cloud, especially showing stronger robustness when dealing with complex backgrounds and noise interference. Secondly, compared with other Transformer-based methods, NCFMamba can more effectively process the local geometric structure of the point cloud through the neighborhood consistency feature modulation module, which is especially important when faced with irregular sampling and local missing data in real scan data. Finally, even compared with Pointmamba and PCM, which are also based on the Mamba architecture, NCFMamba also demonstrates excellent performance, which fully verifies the effectiveness of the neighborhood consistency feature modulation strategy and spatial perception fusion module proposed in this embodiment.

[0124] (2) Point cloud classification experiments under different sparsity;

[0125] Taking into account the fact that point cloud data in practical applications often have uneven sampling density, changes in the number of points due to equipment limitations, and the requirements for computational efficiency in real-time processing, the network's ability to adapt to the sparsity of the input point cloud has become an important indicator for measuring its practicality. This implementation method designed four sets of experiments, using 2048, 1024, 512, and 256 points as input to test the classification performance of the model under different degrees of point cloud sparsity on the subset OBJ_ONLY of the ScanObjectNN dataset. This experiment aims to verify whether the proposed neighborhood consistency feature modulation mechanism can still maintain effective feature extraction and representation capabilities when the point cloud information becomes increasingly sparse, thereby ensuring the stability and reliability of the network when facing input data of different qualities.

[0126] The results are as follows Figure 2 As shown, the results show that when the number of input points drops from 2048 to 256, the classification accuracy of the model only drops from 91.91% to 89.85%, the overall performance drops by only 2.06%, and the relative performance loss is about 2.24%. This result shows that even when the point cloud information is severely sparse (only 1 / 8 of the original number of points is retained), the network can still maintain a high classification accuracy. This excellent robustness is mainly due to the design of the neighborhood consistency feature modulation module. By adaptively adjusting the feature weights, this module can more effectively utilize limited spatial information when the point cloud is sparse, ensuring that key geometric features are fully preserved and strengthened. At the same time, the long sequence modeling capability of the Mamba architecture enables the network to still capture effective global context information with fewer input points.

[0127] In order to deeply analyze the contribution of each component in the proposed method, a comprehensive ablation experiment is designed in this embodiment. The effectiveness of the two modules proposed in the present invention, the neighborhood consistency feature modulation (NCFM) module and the spatial perception fusion module, is verified by performing ablation analysis. By comparing the performance differences with and without the module, its role in improving classification accuracy is evaluated. In addition, in order to determine the impact of network depth on model performance and find the optimal architecture configuration, the performance of NCFMamaba under different stacking layers is tested. The ablation experiments are all conducted on the OBJ_ONLY subset of the ScanObjectNN dataset. As shown in Table 3, Table 3 is the effectiveness verification results of each module.

[0128] Table 3

[0129]

[0130] As shown in Table 3, where P(M) represents the number of parameters and F(G) represents the model complexity (computational effort), ablation experiments show that the proposed NCFM and BSAF modules significantly contribute to network performance. The complete model (Variant 1) achieves a classification accuracy of 91.91% on the ScanObjectNN OBJ_ONLY subset, an improvement of 2.24 percentage points over the base model (Variant 4), fully demonstrating the effectiveness of both modules.

[0131] The NCFM module (comparing Variant 1 and 3) brought a 1.89% performance improvement, demonstrating that the neighborhood consistency feature modulation mechanism can effectively distinguish features from different geometric regions, fully utilizing neighborhood information in smooth areas while preserving geometric details in boundary regions. The BSAF module (comparing Variant 1 and 2) contributed a 1.03% accuracy improvement, demonstrating that the bidirectional spatially aware fusion mechanism can enhance information exchange between spatially adjacent points by utilizing 3D spatial distance information to guide feature interaction. Notably, the synergy of the two modules yielded additional performance gains. Using NCFM (Variant 2) alone improved the baseline model by 1.21%, while using BSAF (Variant 3) alone improved by 0.35%. The total improvement of the two modules (2.24%) was greater than the simple sum of their individual contributions (1.56%), demonstrating the good complementarity and synergy between NCFM and BSAF.

[0132] Regarding the point cloud classification results under different number of encoder layers, that is, the experimental results of different network depths are as follows Figure 3 As shown, it can be seen that there is a nonlinear relationship between network performance and depth. When the depth increases from 4 layers to 12 layers, the classification accuracy steadily increases from 90.02% to 91.91%, indicating that increasing the network depth helps to improve the feature expression ability of the model. However, when the depth is further increased to 16 layers, the accuracy drops to 91.39%, and the number of parameters and computational complexity increase significantly, which indicates that an overly deep network structure will lead to overfitting, verifying that there is an optimal configuration point for the network depth. Taking into account the classification accuracy, computational efficiency and model complexity, the 12-layer network configuration achieves the best balance between accuracy, number of parameters and computational overhead, and is therefore selected as the standard architecture configuration of the present invention.

[0133] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned 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.

[0134] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A classification method for 3D point clouds based on the NCFMamba network is characterized by: The method is implemented by the following steps: Step 1: Divide the point cloud into multiple local point groups and obtain G local feature sequences; Step 2: Use the NCFMamba network model to process the local feature sequence obtained in step 1; obtain the features output by the NCFMamba network model; The specific process is: Step 21: Introduce learnable classification labels [CLS] to aggregate the global information of G local feature sequences; Step 22: Add standard learnable position encoding to the learnable classification label [CLS] and the G local feature sequences and input them into the NCFMamba network model; Step 23: Set the NCFMamba network model to consist of T-layer stacked NCFM modules and a Bi-SSM structure based on the BSAF module; the NCFM module analyzes the multi-dimensional consistency of the local neighborhood of the local feature sequence to achieve adaptive feature aggregation of different geometric regions; The bidirectional Bi-SSM structure based on the BSAF module introduces an adaptive interaction mechanism based on three-dimensional spatial position to achieve intelligent fusion of forward and backward information flows; and finally outputs the fused features; Step 3: Map the fused features output by the NCFMamba network model to the final category space to obtain the point cloud classification result.

2. The method for classifying three-dimensional point clouds based on the NCFMamba network according to claim 1, characterized in that: In step 1, the specific steps of point cloud grouping and feature encoding are as follows: Step 1: Use the farthest point sampling algorithm to select G center points from the original point cloud; Step 1 and 2: For each center point c g Use the K nearest neighbor algorithm to find K nearest neighbor points in the original point cloud to form a local point group; Step 1 and 3: Use lightweight PointNet to extract features from each local point group through a shared multi-layer perceptron (MLP) and maximum pooling operations, and obtain G local feature sequences through point cloud grouping and feature encoding.

3. The method for classifying three-dimensional point clouds based on the NCFMamba network according to claim 1, characterized in that: In steps 2 and 3, the NCFM module analyzes the multi-dimensional consistency of the local neighborhood of the local feature sequence, including spatial consistency S spatial , feature consistency S feature , directional consistency S angular Analyze neighborhood consistency from three perspectives; the specific process is: First, from the spatial consistency S spatial , feature consistency S feature , directional consistency S angular The neighborhood consistency is analyzed at three different scales and normalized. The normalized neighborhood consistency features are concatenated and processed through a lightweight MLP to obtain the final neighborhood consistency score α. Then, adaptive feature modulation is performed according to the obtained neighborhood consistency score α, and an adaptive weighted aggregation strategy is adopted to assign different weights according to the importance of each neighborhood point.

4. The method for classifying three-dimensional point clouds based on the NCFMamba network according to claim 3, characterized in that: The spatial consistency S spatial By calculating the distance d from the neighborhood point to the center point i,k The standard deviation of measures the degree of dispersion of spatial distribution; it is expressed as: d i,k =‖p i,k -c i ‖2 In the formula, K represents the number of nearest neighbor points, c i is the center point of the point cloud, p i,k is a neighborhood point, and std(·) represents the standard deviation operator.

5. The method for classifying three-dimensional point clouds based on the NCFMamba network according to claim 3, characterized in that: The characteristic consistency S feature The evaluation is performed by calculating the average difference between the neighborhood features and the center point features, which can be expressed as: In the formula, K represents the number of nearest neighbor points, x i is the center point feature, x i,k is the neighborhood point feature.

6. The method for classifying three-dimensional point clouds based on the NCFMamba network according to claim 3, characterized in that: The directional consistency S angular By calculating the cosine value v of the angle between adjacent direction vectors i,k The variance is measured and expressed as follows: Where c i is the center point of the point cloud, p i,k are neighborhood points, <·,·> represents vector inner product, and var(·) represents variance.

7. The method for classifying three-dimensional point clouds based on the NCFMamba network according to claim 3, characterized in that: The consistency indicators of different scales are unified into a similar numerical range and normalized by using the hyperbolic tangent function, which can be expressed as follows: Where λ spatial ,λ feature and λ angular is a learnable scaling parameter used to adjust the sensitivity of different consistency metrics; The three neighborhood consistency features are processed using lightweight MLP and expressed as follows: Where σ(·) is the Sigmoid activation function, which ensures that the output is between 0 and 1. α close to 1 indicates that the neighborhood is highly consistent, while α close to 0 indicates that the neighborhood changes dramatically.

8. The method for classifying three-dimensional point clouds based on the NCFMamba network according to claim 3, characterized in that: Adaptive feature modulation is performed on the neighborhood consistency score α, and the specific process is as follows: First, the neighborhood features and coordinates are normalized, and the feature dimension is expanded by concatenating them with the central features. Learnable parameters are introduced to control the weights of edge enhancement and smoothing areas respectively. Finally, an affine transformation is applied to further adjust the features. Then, an adaptive weighted aggregation strategy is adopted to assign different weights according to the importance of each neighborhood point.

9. The method for classifying three-dimensional point clouds based on the NCFMamba network according to claim 1, characterized in that: In steps 2 and 3, the implementation process of the bidirectional Bi-SSM structure is as follows: First, calculate the spatial distance feature; For each pair of center points (c i ,c j ), calculate its Euclidean distance d ij , and then construct the spatially aware position weights For classification labeling [CLS], an independent position enhancement parameter is introduced It can be expressed as: d ij =‖c i -c j ‖2 Where, d i,ref is the distance from the i-th point to the reference point, σ spatial is a learnable spatial scale parameter that controls the range of interaction, γ cls is a learnable parameter, σ(·) is the Sigmoid function; Secondly, after obtaining the complete spatial position weight W spatial Finally, the two-way interaction strength β is defined based on the spatial position weight, which is expressed as follows: β=σ(β base )·W spatial Where, β base is the global interaction strength parameter, which is ensured to be between 0 and 1 by the Sigmoid function. are the corresponding spatial position weights of 1...G respectively; Then, the enhanced bidirectional features are calculated and expressed as: h′ forward =h forward +β⊙h backward h′ backward =h backward +β⊙h forward Where h' forward and h' backward are the output features of the enhanced forward and backward SSM, h forward and h backward Denote the output features of the forward and backward SSM respectively, and ⊙ denotes element-wise multiplication; Finally, a learnable bidirectional coupling parameter θ = [θ f ,θ b ] and the adaptive weight [α f ,α b ], and finally obtain the fused bidirectional feature h fusion , which can be expressed as: [a f ,a b ]=softmax(θ) h fusion =a f ·h′ forward +a b ·h′ backward Where θ f and θ b are the forward coupling parameters and the backward coupling parameters respectively; After T layers of stacked NCFM modules and Bi-SSM modules, features are extracted layer by layer and the final output feature h is T .

10. The point cloud classification network based on neighborhood consistency feature modulation according to claim 1, characterized in that: In the step 3, the feature h output by the NCFMamba network model is T Map to the final category space to obtain the classification results of the point cloud; the specific process is: For feature h T All local grouping features in Perform double pooling operation to obtain feature f group ; Extract the output feature f of the learnable classification label [CLS] CLS , f group and f CLS Splicing to obtain the final classification feature f concat ; The classification head adopts a three-layer fully connected network, batch normalization and Dropout regularization, and finally obtains the classification results of the point cloud.

Images