Posture recognition method and system based on identity feature desensitization, medium and server

Abstract

The invention belongs to the technical field of posture recognition, and provides a posture recognition method and system based on identity feature desensitization, a medium and a server, and the method comprises the steps: carrying out the preprocessing of three-dimensional point cloud data containing human body posture information, and carrying out the feature extraction through a deep learning algorithm, and carrying out standardization processing, covariance matrix eigendecomposition and principal component analysis dimension reduction processing on the extracted eigenvector in sequence to obtain a desensitized eigenvector, inputting the desensitized eigenvector into a pre-trained support vector machine classifier, and outputting an identification result including the basic posture of the human body. According to the method, after the original point cloud is converted into the feature vector, the posture data is detected through the SVM classifier, human body joint point recognition and data output are not involved, meanwhile, identity relevance in the point cloud data is effectively eliminated through identity feature desensitization processing, and privacy protection is higher; the deep learning algorithm is adopted to perform feature extraction and posture recognition, so that the recognition accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of posture recognition technology, specifically relating to a posture recognition method, system, medium, and server based on identity feature desensitization. Background Technology

[0002] In current scenarios such as intelligent monitoring, elderly care, and campus management, attitude recognition is mainly achieved by collecting video image information through visible light cameras.

[0003] Chinese patent CN120108043A discloses a portable posture detection and gait analysis method and system, which uses two-dimensional images from a visible light camera for posture detection. Chinese patent CN119888783A discloses a multi-view vision-based high-precision human posture machine vision recognition method based on simulated data training. The main method involves training a two-dimensional posture detection model, using triangulation and fine-grained human stereo detection box segmentation strategies, combined with spatial grid and feature projection fusion, to accurately estimate the three-dimensional coordinates of human joints, thereby determining the human three-dimensional posture. Chinese patent CN119992412A discloses a gait identity information desensitization method and system based on sparse contour transitions. These methods primarily focus on processing single modalities (such as gait videos) and lack targeted solutions for richer data sources like three-dimensional point clouds.

[0004] The existing 3D point cloud data processing methods often directly use the raw data to identify human posture during the processing, and output human posture data (such as skeletal key points) through several methods. This makes it easy for personal identity information (height, information of each joint point) to be leaked. Through a series of machine learning algorithms, user information can be inferred and associated with user identity, which poses a security risk to user privacy. Summary of the Invention

[0005] The purpose of this invention is to overcome the existing defects and provide a posture recognition method, system, medium and server based on identity feature desensitization, which can accurately identify human posture while protecting personal privacy, and is suitable for scenarios that require privacy protection, such as intelligent monitoring, elderly care, and campuses.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] The first objective of this invention is to provide a pose recognition method based on identity feature desensitization, comprising:

[0008] (1) Collect three-dimensional point cloud data containing human posture information, and perform downsampling and statistical filtering noise reduction on the three-dimensional point cloud data;

[0009] (2) Deep learning algorithms are used to extract features from the preprocessed point cloud data to obtain feature vectors representing human posture information;

[0010] (3) The extracted feature vectors are subjected to standardization, covariance matrix eigenvalue decomposition and principal component analysis dimensionality reduction in sequence to eliminate sensitive information related to identity features and obtain desensitized feature vectors.

[0011] (4) Input the desensitized feature vectors into the pre-trained support vector machine classifier, output the recognition results including the basic human posture, and use temporal filtering and confidence threshold filtering to optimize the recognition results.

[0012] Furthermore, in step (1), the downsampling and statistical filtering denoising process of the 3D point cloud data includes:

[0013] Uniform downsampling: Selecting points from point cloud data at fixed intervals;

[0014] Statistical filtering for noise reduction: Outliers are removed by calculating the average distance between neighboring points and setting a threshold range.

[0015] Furthermore, in step (2), the feature extraction of the preprocessed point cloud data using a deep learning algorithm includes:

[0016] The preprocessed point cloud data is input into the PointNet++ deep learning model. Through the multi-layered structure of the PointNet++ model, the local and global features of the point cloud data are extracted layer by layer.

[0017] The multilayer perceptron is used to extract features from each point to obtain the local features of each point;

[0018] By aggregating the local features of all points through the max pooling layer, we obtain the global features that characterize the entire point cloud data.

[0019] By fusing local and global features, the final feature vector representing human posture information is obtained.

[0020] Furthermore, in step (3), the specific steps for desensitizing identity features include:

[0021] The feature vectors are standardized to ensure that all feature dimensions are within the same scale range.

[0022] Subtract the overall mean from the standardized feature vectors to make the data distribution center located at the origin;

[0023] Calculate the covariance matrix of the standardized eigenvectors and perform eigenvalue decomposition to obtain the eigenvalues ​​and corresponding eigenvectors;

[0024] Sort the eigenvalues ​​from largest to smallest, select the k smallest principal components with a cumulative variance contribution rate of 95% to construct a projection matrix, and project the original eigenvectors onto a low-dimensional space.

[0025] Furthermore, the dimensionality-reduced feature vectors are subjected to AES symmetric encryption to ensure data security during transmission and storage.

[0026] Furthermore, in step (4), the support vector machine classifier is trained using desensitized feature vectors, the kernel function is selected as RBF, the hyperparameters are optimized through grid search, and the training data contains a variety of basic pose samples.

[0027] Another object of the present invention is to provide a pose recognition system based on identity feature desensitization, comprising:

[0028] The data acquisition module is used to scan the area where people are active using a 3D scanning device to obtain 3D point cloud data containing human posture information.

[0029] The preprocessing module is used to perform downsampling and statistical filtering noise reduction on the acquired 3D point cloud data;

[0030] The feature extraction module is used to extract features from the preprocessed point cloud data using deep learning algorithms to obtain feature vectors representing human posture information.

[0031] The desensitization module is used to perform standardization, covariance matrix eigenvalue decomposition, and principal component analysis dimensionality reduction on the extracted feature vectors in sequence to eliminate sensitive information related to identity features and obtain desensitized feature vectors.

[0032] The pose recognition module is used to input the desensitized feature vectors into a pre-trained support vector machine classifier, outputting recognition results including basic human poses, and optimizing the recognition results using temporal filtering and confidence threshold filtering.

[0033] Furthermore, the feature extraction module includes:

[0034] The data preparation unit is used to input the preprocessed point cloud data into the PointNet++ deep learning model.

[0035] The local feature extraction unit is used to extract features from each point using a multilayer perceptron to obtain the local features of each point;

[0036] The global feature extraction unit is used to aggregate the local features of all points through the max pooling layer to obtain the global features that characterize the entire point cloud data.

[0037] The feature fusion unit is used to fuse local and global features to obtain the final feature vector representing human posture information.

[0038] Another object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the pose recognition method based on identity feature desensitization provided by the first object of the present invention.

[0039] Another object of the present invention is to provide a server comprising at least one processor and a memory communicatively connected to the processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the processor to cause the at least one processor to perform the pose recognition method based on identity feature desensitization provided in the first object of the present invention.

[0040] In combination with the above technical solutions, the beneficial effects of the present invention compared with the prior art are as follows:

[0041] This invention offers strong privacy protection by converting raw point clouds into feature vectors and then using an SVM classifier to detect pose data. It avoids human joint point recognition and data output, and effectively eliminates identity-related factors in the point cloud data through identity feature desensitization, thus protecting personal privacy. Compared to existing technologies that often use related human keypoint recognition models or other algorithms to obtain human joint point coordinates, perform desensitization, and then use a pose classification model to detect and classify the keypoint data, this invention provides stronger privacy protection. Pose recognition is accurate, as deep learning algorithms are used for feature extraction and pose recognition, improving accuracy. It has wide applicability, suitable for various privacy-sensitive scenarios such as intelligent monitoring, elderly care, and campuses, and has broad application prospects. Attached Figure Description

[0042] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0043] Figure 1 This is a flowchart of the pose recognition method based on identity feature desensitization provided in an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of the structure of the posture recognition system based on identity feature desensitization provided in an embodiment of the present invention. Detailed Implementation

[0045] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0046] Example 1:

[0047] like Figure 1 The image shown is an embodiment of the pose recognition method based on identity feature desensitization provided by the present invention, which specifically includes the following steps:

[0048] S1: Collect 3D point cloud data containing human posture information, and perform downsampling and statistical filtering noise reduction on the 3D point cloud data;

[0049] S2: Deep learning algorithms are used to extract features from the preprocessed point cloud data to obtain feature vectors representing human posture information;

[0050] S3: The extracted feature vectors are subjected to standardization, covariance matrix eigenvalue decomposition, and principal component analysis for dimensionality reduction to eliminate sensitive information related to identity features and obtain desensitized feature vectors.

[0051] S4: Input the desensitized feature vectors into a pre-trained support vector machine classifier, output the recognition results including basic human poses, and optimize the recognition results using temporal filtering and confidence threshold filtering.

[0052] Specifically, the technical solution and technical effects of the present invention will be further explained below with reference to specific examples.

[0053] 1. Data Collection:

[0054] Using a certain brand of 3D scanner, the human body was scanned in an intelligent monitoring scenario at a sampling frequency of 30 frames per second, resulting in 3D point cloud data containing human posture information.

[0055] 2. Pretreatment:

[0056] First, the collected point cloud data is downsampled to reduce the number of data points and improve processing speed. The specific implementation process is as follows: a uniform downsampling method is used for downsampling. Its basic principle is to select points at fixed intervals (e.g., keep 1 out of every 3 points). The calculation speed is fast and it is especially suitable for ordered point clouds collected by 3D scanners.

[0057] Then, denoising is performed using statistical filtering to remove noise points. The basic principle is that normal points are relatively uniformly distributed in space, while noise points are usually isolated or outliers far from the main point cloud. The algorithm calculates the average distance from each point to its neighbors, assuming these distances follow a Gaussian distribution, and then removes points that deviate excessively from the mean. The specific implementation process is as follows:

[0058] (1) Neighborhood distance statistics

[0059] For each point in the point cloud, calculate the average distance to its k nearest neighbors;

[0060] Construct a set of average distances for all points, and calculate the global mean μ and standard deviation σ.

[0061] (2) Outlier identification

[0062] Set the threshold range: μ±α·σ (α is a user-defined multiple of the standard deviation);

[0063] Points whose average distance exceeds this range are marked as outliers.

[0064] (3) Filtering

[0065] Remove outliers (or retain outliers) based on the labeling results;

[0066] Output filtered point cloud data.

[0067] Compared to existing preprocessing methods, this solution processes 3D point cloud data, while the comparison file CN119992412A processes 2D video image sequences. Point cloud data provides richer spatial information, but the processing complexity is higher.

[0068] This invention employs a uniform downsampling method, selecting points at fixed intervals, resulting in fast computation speed and suitability for ordered point clouds. Point cloud preprocessing retains more spatial structure information, which is beneficial for subsequent feature extraction. Statistical filtering effectively removes outliers while preserving the main point cloud structure.

[0069] The preprocessing method of this invention has low dependence on scanning equipment and is applicable to point cloud data from different sources; the parameters are adjustable (such as the number of neighboring points k and the standard deviation factor α) to adapt to different scenario requirements. The preprocessing results can be directly input into deep learning models such as PointNet++, providing a good data foundation for subsequent feature decoupling.

[0070] 3. Pose feature extraction:

[0071] The PointNet++ deep learning model is used to extract features from the preprocessed point cloud data, resulting in a 1024-dimensional feature vector that can represent human pose information. The specific implementation process is as follows:

[0072] (1) Data preparation:

[0073] The preprocessed point cloud data is then input into the deep learning model.

[0074] (2) Model selection:

[0075] The PointNet++ deep learning model is used, which can effectively process unordered point cloud data and extract key features.

[0076] (3) Feature extraction:

[0077] By leveraging the multi-layered structure of the PointNet++ model, local and global features of point cloud data are extracted layer by layer.

[0078] (4) Local feature extraction:

[0079] Feature extraction is performed on each point using a multilayer perceptron (MLP) to obtain the local features of each point.

[0080] (5) Global feature extraction:

[0081] By aggregating the local features of all points through a max pooling layer, global features representing the entire point cloud data are obtained.

[0082] (6) Feature fusion:

[0083] By fusing local and global features, a final feature vector is obtained, which can comprehensively represent human posture information.

[0084] Compared to existing posture recognition technologies, this invention does not identify / output human joint information. Instead, it extracts a 1024-dimensional high-dimensional feature vector to represent human posture information, thus avoiding the leakage of personnel information that may result from directly outputting joint information.

[0085] Existing gradient-based feature extraction methods for binary images have limited feature representation capabilities. 3D point cloud data preserves the spatial structure of the human body, avoiding the viewpoint and occlusion problems present in 2D images. The PointNet++ model can automatically learn the hierarchical features of point cloud data, extracting more discriminative pose features. Its 1024-dimensional feature vector has stronger representational capabilities than traditional methods (such as contour-based methods).

[0086] 4. Desensitization of identity features:

[0087] Principal component analysis (PCA) is used to reduce the dimensionality of the extracted feature vectors, retaining 95% of the information while removing sensitive information related to identity features.

[0088] Implementation process:

[0089] (1) Standardize the 1024-dimensional feature vectors extracted by PointNet++ by subtracting the mean of each feature dimension and dividing by the standard deviation to ensure that all feature dimensions are within the same scale range.

[0090] (2) Subtract the overall mean from the standardized feature vectors to make the data distribution center located at the origin, formula: X centered =X - mean(X);

[0091] (3) Calculate the covariance matrix of the standardized eigenvectors, perform eigenvalue decomposition on the covariance matrix, and solve the characteristic equation: |Cov-λI|=0, obtaining 1024 eigenvalues ​​and their corresponding eigenvectors. Here, Cov represents the covariance matrix, which is a 1024x1024 symmetric matrix; λ is the eigenvalue, and the equation yields 1024 λ values, each corresponding to an eigenvector; I represents the identity matrix with the same dimension as Cov.

[0092] (4) Sort the eigenvalues ​​from largest to smallest, calculate the cumulative variance contribution rate, and select the smallest k value that makes the cumulative contribution rate reach 95%. Typically, the k value is much smaller than the original dimension of 1024. The k value represents the number of principal components retained in PCA dimensionality reduction. By selecting the eigenvectors corresponding to the top k largest eigenvalues, the original high-dimensional data (1024 dimensions) is projected onto a k-dimensional subspace, achieving data compression and information preservation. Here, the selection criteria are explained: the k value must satisfy a cumulative variance contribution rate ≥ 95%, meaning that more than 95% of the original data information is retained after dimensionality reduction.

[0093] Identity feature desensitization mechanism:

[0094] Through experimental analysis, principal components related to identity information were identified. These principal components typically correspond to eigenvalues ​​of moderate size, and are neither the principal components with the largest nor the smallest variance. While retaining 95% of the information, principal components related to identity features were preferentially discarded, while those highly correlated with pose recognition were retained. Then, the first k principal components were used to construct a projection matrix W, projecting the original feature vectors onto a new low-dimensional space, as shown in the formula: Y = X. centered ×W.

[0095] This invention irreversibly eliminates identity information through mathematical transformations, making it more secure and reliable than simple noise addition or pixel modification. Unlike existing techniques that add adversarial perturbations to the pixel space, this invention achieves desensitization through linear transformations in the feature space, preserving purer pose-related principal component information, which is beneficial for improving the accuracy of subsequent pose recognition. Dimensionality reduction significantly reduces the amount of data, lowering the computational burden of subsequent encryption processing and pose recognition.

[0096] The dimensionality-reduced feature vectors are encrypted using the AES symmetric encryption algorithm to ensure data security during transmission and storage.

[0097] 5. Pose recognition:

[0098] Based on the anonymized feature vectors, a Support Vector Machine (SVM) classifier is trained for human pose recognition. The recognition results include basic poses such as standing, walking, sitting, and lying down, with an accuracy rate exceeding 95%. Implementation process:

[0099] (1) Classifier training:

[0100] The SVM classifier was trained using desensitized feature vectors, with RBF kernel function selected. Hyperparameters were optimized using grid search. The training data included samples of various postures such as standing, walking, sitting, and lying down.

[0101] (2) Real-time recognition:

[0102] After inputting new point cloud data, the same preprocessing and feature extraction process is followed, and the desensitized features are input into the trained SVM classifier, which outputs the pose category with the highest probability and the confidence level.

[0103] (3) Post-processing optimization:

[0104] Temporal filtering is used to smooth the recognition results, a confidence threshold is set to filter out low-quality detections, and the final output is optimized by combining the scene context.

[0105] Example 2:

[0106] like Figure 2 As shown, the purpose of this embodiment of the invention is to provide a pose recognition system based on identity feature desensitization, comprising:

[0107] The data acquisition module is used to scan the area where people are active using a 3D scanning device to obtain 3D point cloud data containing human posture information.

[0108] The preprocessing module is used to perform downsampling and statistical filtering noise reduction on the acquired 3D point cloud data;

[0109] The feature extraction module is used to extract features from the preprocessed point cloud data using deep learning algorithms to obtain feature vectors representing human posture information.

[0110] The desensitization module is used to perform standardization, covariance matrix eigenvalue decomposition, and principal component analysis dimensionality reduction on the extracted feature vectors in sequence to eliminate sensitive information related to identity features and obtain desensitized feature vectors.

[0111] The pose recognition module is used to input the desensitized feature vectors into a pre-trained support vector machine classifier, outputting recognition results including basic human poses, and optimizing the recognition results using temporal filtering and confidence threshold filtering.

[0112] Preferably, the feature extraction module in this embodiment of the invention includes:

[0113] The data preparation unit is used to input the preprocessed point cloud data into the PointNet++ deep learning model.

[0114] The local feature extraction unit is used to extract features from each point using a multilayer perceptron to obtain the local features of each point;

[0115] The global feature extraction unit is used to aggregate the local features of all points through the max pooling layer to obtain the global features that characterize the entire point cloud data.

[0116] The feature fusion unit is used to fuse local and global features to obtain the final feature vector representing human posture information.

[0117] Example 3: This embodiment of the invention provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the posture recognition method based on identity feature desensitization provided in Example 1 of the invention.

[0118] Example 4: This embodiment of the invention provides a server, including at least one processor and a memory communicatively connected to the processor. The memory stores instructions that can be executed by the at least one processor. The instructions are executed by the processor to cause the at least one processor to perform the pose recognition method based on identity feature desensitization provided in Example 1 of the invention.

[0119] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated in the present invention, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0120] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0121] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A pose recognition method based on identity feature desensitization, characterized in that, The method includes: (1) Collect three-dimensional point cloud data containing human posture information, and perform downsampling and statistical filtering noise reduction on the three-dimensional point cloud data; (2) Deep learning algorithms are used to extract features from the preprocessed point cloud data to obtain feature vectors representing human posture information; (3) The extracted feature vectors are subjected to standardization, covariance matrix eigenvalue decomposition and principal component analysis dimensionality reduction in sequence to eliminate sensitive information related to identity features and obtain desensitized feature vectors. (4) Input the desensitized feature vectors into the pre-trained support vector machine classifier, output the recognition results including the basic human posture, and use temporal filtering and confidence threshold filtering to optimize the recognition results.

2. The pose recognition method based on identity feature desensitization according to claim 1, characterized in that, In step (1), the downsampling and statistical filtering denoising process for the 3D point cloud data includes: Uniform downsampling: Selecting points from point cloud data at fixed intervals; Statistical filtering for noise reduction: Outliers are removed by calculating the average distance between neighboring points and setting a threshold range.

3. The pose recognition method based on identity feature desensitization according to claim 1, characterized in that, In step (2), the feature extraction of the preprocessed point cloud data using a deep learning algorithm includes: The preprocessed point cloud data is input into the PointNet++ deep learning model. Through the multi-layered structure of the PointNet++ model, the local and global features of the point cloud data are extracted layer by layer. The multilayer perceptron is used to extract features from each point to obtain the local features of each point; By aggregating the local features of all points through the max pooling layer, we obtain the global features that characterize the entire point cloud data. By fusing local and global features, the final feature vector representing human posture information is obtained.

4. The pose recognition method based on identity feature desensitization according to claim 1, characterized in that, In step (3), the specific steps for de-identifying identity features include: The feature vectors are standardized to ensure that all feature dimensions are within the same scale range. Subtract the overall mean from the standardized feature vectors to make the data distribution center located at the origin; Calculate the covariance matrix of the standardized eigenvectors and perform eigenvalue decomposition to obtain the eigenvalues ​​and corresponding eigenvectors; Sort the eigenvalues ​​from largest to smallest, select the k smallest principal components with a cumulative variance contribution rate of 95% to construct a projection matrix, and project the original eigenvectors onto a low-dimensional space.

5. The pose recognition method based on identity feature desensitization according to claim 4, characterized in that, The dimensionality-reduced feature vectors are subjected to AES symmetric encryption to ensure data security during transmission and storage.

6. The pose recognition method based on identity feature desensitization according to claim 1, characterized in that, In step (4), the support vector machine classifier is trained using desensitized feature vectors, the kernel function is selected as RBF, the hyperparameters are optimized by grid search, and the training data contains a variety of basic pose samples.

7. A gesture recognition system based on identity feature desensitization, characterized in that, The system includes: The data acquisition module is used to scan the area where people are active using a 3D scanning device to obtain 3D point cloud data containing human posture information. The preprocessing module is used to perform downsampling and statistical filtering noise reduction on the acquired 3D point cloud data; The feature extraction module is used to extract features from the preprocessed point cloud data using deep learning algorithms to obtain feature vectors representing human posture information. The desensitization module is used to perform standardization, covariance matrix eigenvalue decomposition, and principal component analysis dimensionality reduction on the extracted feature vectors in sequence to eliminate sensitive information related to identity features and obtain desensitized feature vectors. The pose recognition module is used to input the desensitized feature vectors into a pre-trained support vector machine classifier, outputting recognition results including basic human poses, and optimizing the recognition results using temporal filtering and confidence threshold filtering.

8. The pose recognition system based on identity feature desensitization according to claim 7, characterized in that, The feature extraction module includes: The data preparation unit is used to input the preprocessed point cloud data into the PointNet++ deep learning model. The local feature extraction unit is used to extract features from each point using a multilayer perceptron to obtain the local features of each point; The global feature extraction unit is used to aggregate the local features of all points through the max pooling layer to obtain the global features that characterize the entire point cloud data. The feature fusion unit is used to fuse local and global features to obtain the final feature vector representing human posture information.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the posture recognition method based on identity feature desensitization as described in any one of claims 1 to 6.

10. A server, characterized in that: It includes at least one processor and a memory communicatively connected to the processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the processor to cause the at least one processor to perform the pose recognition method based on identity feature desensitization as described in any one of claims 1-6.

Images