Infinite-class identity detection method, system and electronic device

Abstract

The application discloses an infinite-class identity detection method and system and electronic equipment, and relates to the field of identity detection. The method comprises the following steps: performing multi-resolution semantic discrimination motion segmentation on two motion data streams with identity information respectively to extract first and second semantic action data, wherein the motion data stream is data collected by an inertial sensor; inputting the first and second semantic action data into an identity recognition model to obtain an identity recognition result; the identity recognition result is the same identity or different identity; the identity recognition model is used for extracting the hidden features, the explainable features and the prior class features of the two semantic action data respectively, and the features extracted from each semantic action data are optimized and aggregated to obtain first and second aggregated features, and the first and second aggregated features are discriminated to obtain the identity recognition result. The application realizes infinite-class identity detection.

Description

Technical Field

[0001] This invention relates to the field of identity detection technology, and in particular to an infinite number of identity detection methods, systems, and electronic devices. Background Technology

[0002] Human movement, especially writing, contains a wealth of semantic information. This semantic information can be divided into two aspects: the content of the writing and the identity information reflected by the writer's movement characteristics. In fact, the writer's identity information and content information are independent of each other. Regardless of whether the writing content is English letters, Chinese characters, or any other language, the writer's identity information can be reflected through their movement habits. This is the basic principle of handwriting identification. Handwriting identification can be seen as a special type of identity verification technology; however, its applicability is very limited. Traditional handwriting relies on paired writing tools and a writing surface, such as pen and paper. Furthermore, handwriting requires sufficient light and stable movement—demanding conditions that make accurate writing difficult while walking or running. Therefore, handwriting-based identity verification is only suitable for extremely limited scenarios such as document signing. For diverse identity verification scenarios such as unlocking electronic devices, controlling important devices, and logging into special software, handwriting-based identity verification technology clearly cannot meet the complex writing environments and conditions in these diverse scenarios. Summary of the Invention

[0003] The purpose of this invention is to provide an infinite category identity detection method, system, and electronic device, which realizes infinite category identity detection and improves the accuracy of identity detection.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] An infinite-category identity detection method includes:

[0006] Motion segmentation with multi-resolution semantic discrimination is performed on the motion data stream of the first target to be detected to extract the first semantic action data; the motion data stream of the first target to be detected is acquired by an inertial sensor carried by the first target to be detected.

[0007] The motion data stream of the second target to be detected is segmented by multi-resolution semantic discrimination to extract the second semantic action data; the motion data stream of the second target to be detected is acquired by an inertial sensor carried by the second target to be detected.

[0008] The first semantic action data and the second semantic action data are input into the identity recognition model to obtain the identity recognition result; the identity recognition result is that the first target to be detected and the second target to be detected have the same identity, or the first target to be detected and the second target to be detected have different identities.

[0009] The identity recognition model includes a residual neural network, an interpretable feature extraction module, a priori category feature extraction module, a converter, a feature aggregation module, and a discriminator;

[0010] The residual neural network is used to extract features from the first semantic action data to obtain a first latent feature, and to extract features from the second semantic action data to obtain a second latent feature; the interpretability feature extraction module is used to extract preset interpretability features from the first semantic action data to obtain a first interpretable feature, and to extract preset interpretable features from the second semantic action data to obtain a second interpretable feature; the prior category feature extraction module is used to obtain a first prior category feature based on the matching degree between the first semantic action data and a set number of known identity categories, and to obtain a second prior category feature based on the matching degree between the second semantic action data and a set number of known identity categories. The converter is used to optimize the first latent feature, the first interpretability feature, the first prior category feature, the second latent feature, the second interpretability feature, and the second prior category feature, respectively. The feature aggregation module is used to aggregate the optimized first latent feature, the first interpretability feature, and the first prior category feature to obtain a first aggregated feature, and to aggregate the optimized second latent feature, the second interpretability feature, and the second prior category feature to obtain a second aggregated feature. The discriminator is used to obtain the identity recognition result based on the first aggregated feature and the second aggregated feature.

[0011] Optionally, motion segmentation with multi-resolution semantic discrimination is performed on the motion data stream of the first target to be detected to extract the first semantic action data, specifically including:

[0012] Set α windows of different lengths, and slide each window sequentially on the motion data stream of the first target to be detected. Input the window data of the current window at each slide into the first random forest model to perform a binary classification task of semantic action data and non-semantic action data, and obtain the probability value of each window sliding at each slide being semantic action data; the first random forest model is a trained random forest model.

[0013] Cubic spline interpolation is used to convert the probability value corresponding to each window into a probability curve;

[0014] The point in each probability curve with a probability value of 0.5 is defined as a potential split point;

[0015] Integrate the potential segmentation points in group α to obtain the positions of the final segmentation points;

[0016] Semantic action data is extracted from the motion data stream of the target to be detected based on the position of the final segmentation point.

[0017] Optionally, the process of integrating the α group of potential segmentation points to obtain the location of the final segmentation point specifically includes:

[0018] For each group of potential segmentation points in group α, sort them in chronological order. If there are multiple consecutive potential segmentation points in the current group, and the time between any two consecutive potential segmentation points is less than a time threshold, delete the multiple consecutive potential segmentation points if the number of such points is even. If the number of such points is odd, use the average coordinate of the multiple consecutive potential segmentation points as a segmentation point. The remaining potential segmentation points in the current group are retained as segmentation points, thus obtaining group α of segmentation points.

[0019] The mode of the number of segmentation points in group α is taken as the final number of segmentation points;

[0020] Select a β group of segmentation points from the α group of segmentation points, with the same number of segmentation points as the final number of segmentation points;

[0021] For a set of β segmentation points, the average time of the β i-th segmentation points is taken as the position of the i-th final segmentation point, where the value of i ranges from 1 to the number of final segmentation points, thus obtaining a set of final segmentation points.

[0022] Optionally, the process for determining the predefined interpretability features includes:

[0023] Obtain a sample semantic action data set; the sample semantic action data in the sample semantic action data set is determined by motion segmentation based on motion data streams collected when multiple identities perform different types of actions, using multi-resolution semantic discrimination.

[0024] A first set number of samples are randomly selected from the sample semantic action data set to obtain multiple feature data for each sample; the feature data includes time domain data and frequency domain data, and both the time domain data and the frequency domain data include mean, variance, slope, number of peaks and correlation coefficient;

[0025] Each feature data point takes values ​​from a first set number of samples, which constitute a feature vector.

[0026] The k-means clustering algorithm is used to cluster multiple feature vectors to obtain a second set number of feature clusters;

[0027] For each feature cluster:

[0028] Initialize the set of important features to empty;

[0029] Initialize the feature cluster identity recognition accuracy;

[0030] The remaining feature vectors are traversed sequentially from the current feature cluster. The current second random forest model is trained using the currently acquired feature vectors and the feature vectors in the current important feature set. The identity recognition accuracy of the trained current second random forest model is obtained. The second random forest model is used to perform identity recognition on the input features.

[0031] If the remaining feature vectors in the current feature cluster have been traversed to completion, then the identity recognition accuracy corresponding to each feature vector in the current feature cluster is obtained;

[0032] If the highest identity recognition accuracy of each feature vector in the current feature cluster is greater than the current feature cluster identity recognition accuracy, then the current feature cluster identity recognition accuracy is updated to the highest current identity recognition accuracy. The feature vector with the highest identity recognition accuracy in the current feature cluster is extracted and placed into the important feature set. The process of traversing the remaining feature vectors in the current feature cluster sequentially, using the currently obtained feature vectors and the feature vectors in the current important feature set to train the current second random forest model, and obtaining the identity recognition accuracy of the trained current second random forest model continues until the highest identity recognition accuracy in the current feature cluster is less than or equal to the current feature cluster identity recognition accuracy.

[0033] The important feature sets of each feature cluster are merged to obtain the merged important features, which are then denoted as merged important features.

[0034] Return to the step of randomly extracting a first set number of samples from the sample semantic action data set, and obtaining multiple feature data of each sample, until m merged important features are obtained;

[0035] The intersection of m important features is taken as the set of most valuable features;

[0036] The most valuable feature set is input into the first multilayer perceptron to obtain the first interpretable feature with the same dimension as the first latent feature.

[0037] Optionally, in obtaining the first prior category feature based on the matching degree between the first semantic action data and multiple identity categories, the prior category feature extraction module is specifically used for:

[0038] Obtain the matching degree between the first semantic action data and the feature codes of a third set number of known identity categories;

[0039] Each matching degree is used as a weight and weighted and summed with the feature codes of a third set number of known identity categories to obtain the first prior category feature of the first semantic action data.

[0040] Optionally, the identity recognition model is a trained identity recognition network; the training process of the identity recognition network includes a first training phase and a second training phase.

[0041] The first training phase involves training the residual neural network, specifically including:

[0042] Initialize the action content classifier; the action content classifier is used to classify action categories based on the features output by the residual neural network and the features output by the interpretable feature extraction module;

[0043] The training objective is to reduce the prediction error of the action content classifier. The residual neural network and the action content classifier are trained using a first training set and combined with the early stopping method to obtain a trained action content classifier and a residual neural network after the first training. The first training set includes multiple first sample data pairs. Each first sample data pair includes first input data and first label data. The first input data is a motion sample data stream and the first label data is an action category.

[0044] The parameters of the trained action content classifier are fixed. A gradient inversion layer is added between the residual neural network and the action content classifier after the first training. The motion identity classifier is initialized and connected in parallel with the action content classifier. The motion identity classifier is used to classify the identity categories of the features output by the residual neural network and the features output by the interpretable feature extraction module.

[0045] Using a gradient inversion layer, the training objectives are to increase the prediction error of the action content classifier and reduce the prediction error of the motion identity classifier. A second training set is used, combined with an early stopping method, to train the residual neural network and the motion identity classifier after the first training, resulting in a trained residual neural network and a trained motion identity classifier. The second training set includes multiple second sample data pairs, each of which includes second input data and second label data. The second input data is a motion sample data stream, and the second label data includes action category and identity category.

[0046] The matching degree between the first semantic action data and the feature codes of a third set number of known identity categories is the output feature vector obtained by inputting the first semantic action data into a trained motion identity classifier.

[0047] Optionally, the prior category feature extraction module includes an identity category encoding matrix, which is used to store feature codes of a third set number of known identity categories;

[0048] The second training phase involves training the identity category encoding matrix, the converter, the feature aggregation module, and the discriminator, specifically including:

[0049] The training objective is to reduce the prediction error of the regularization term and the discriminator. The identity category coding matrix, the converter, the feature aggregation module and the discriminator are trained using the third training set and combined with the early stopping method, so as to obtain the trained identity category coding matrix, the converter, the feature aggregation module and the discriminator.

[0050] The third training set includes multiple third sample data pairs, each of which includes third input data and third label data. The third input data is a first motion sample data stream and a second motion sample data stream, and the third label data is the identity recognition result.

[0051] The regular expression is represented as follows:

[0052] Among them, R S S represents the regularization term. α (E) represents the α-order Renyi information entropy of the identity category coding matrix E.

[0053] Optionally, the motion data stream of the first target to be detected is the motion data stream of the first person to be detected when writing; the motion data stream of the second target to be detected is the motion data stream of the second person to be detected when writing.

[0054] This invention also discloses an infinite category identity detection system, comprising:

[0055] The first semantic action data extraction module is used to perform multi-resolution semantic discrimination motion segmentation on the motion data stream of the first target to be detected, and extract the first semantic action data; the motion data stream of the first target to be detected is collected by an inertial sensor carried by the first target to be detected;

[0056] The second semantic action data extraction module is used to perform multi-resolution semantic discrimination motion segmentation on the motion data stream of the second target to be detected, and extract the second semantic action data; the motion data stream of the second target to be detected is collected by an inertial sensor carried by the second target to be detected;

[0057] An identity recognition module is used to input the first semantic action data and the second semantic action data into an identity recognition model to obtain an identity recognition result; the identity recognition result is that the first target to be detected and the second target to be detected have the same identity, or the first target to be detected and the second target to be detected have different identities.

[0058] The identity recognition model includes a residual neural network, an interpretable feature extraction module, a priori category feature extraction module, a converter, a feature aggregation module, and a discriminator;

[0059] The residual neural network is used to extract features from the first semantic action data to obtain a first latent feature, and to extract features from the second semantic action data to obtain a second latent feature; the interpretability feature extraction module is used to extract preset interpretability features from the first semantic action data to obtain a first interpretable feature, and to extract preset interpretable features from the second semantic action data to obtain a second interpretable feature; the prior category feature extraction module is used to obtain a first prior category feature based on the matching degree between the first semantic action data and a set number of known identity categories, and to obtain a second prior category feature based on the matching degree between the second semantic action data and a set number of known identity categories. The converter is used to optimize the first latent feature, the first interpretability feature, the first prior category feature, the second latent feature, the second interpretability feature, and the second prior category feature, respectively. The feature aggregation module is used to aggregate the optimized first latent feature, the first interpretability feature, and the first prior category feature to obtain a first aggregated feature, and to aggregate the optimized second latent feature, the second interpretability feature, and the second prior category feature to obtain a second aggregated feature. The discriminator is used to obtain the identity recognition result based on the first aggregated feature and the second aggregated feature.

[0060] The present invention also discloses an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the infinite category identity detection method.

[0061] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0062] This invention first extracts semantic action data from the data stream based on motion segmentation with multi-resolution semantic discrimination. Then, it extracts the latent features, interpretability features, and prior category features of the semantic action data, performing multi-faceted feature extraction on the semantic action data. This enables identity detection of infinite categories and improves the accuracy of identity detection. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This invention provides a schematic flowchart of an infinite category identity detection method.

[0065] Figure 2 A schematic diagram illustrating the principle of motion segmentation for multi-resolution semantic discrimination is provided for embodiments of the present invention;

[0066] Figure 3 A schematic diagram of the data processing flow of the identity recognition model is provided for embodiments of the present invention;

[0067] Figure 4 This invention provides a schematic diagram of an infinite category identity detection system. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] The purpose of this invention is to provide an infinite number of identity detection methods, systems, and electronic devices, thereby improving the accuracy of identity detection.

[0070] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0071] Example 1

[0072] like Figure 1 As shown, this embodiment provides a method for detecting identities with infinite categories, including the following steps:

[0073] Step 101: Perform multi-resolution semantic discrimination motion segmentation on the motion data stream of the first target to be detected, and extract the first semantic action data; the motion data stream of the first target to be detected is collected by the inertial sensor carried by the first target to be detected.

[0074] Inertial sensors are characterized by their small size, ease of wear, and low cost, and can be easily integrated into mobile phones, watches, bracelets, and even rings. Therefore, this invention enables identity detection by wearing these devices and performing semantic actions, thereby achieving downstream tasks such as remote control and human-computer interaction.

[0075] Step 101 specifically includes:

[0076] Set up α windows of different lengths, and sequentially slide each window over the motion data stream of the first target to be detected. Input the window data of each sliding step into a first random forest model to perform a binary classification task of semantic action data and non-semantic action data, and obtain the probability value of each sliding step being semantic action data. The first random forest model is a pre-trained random forest model. Define each probability value and its corresponding time coordinate as a test point.

[0077] Cubic spline interpolation is used to convert the probability value corresponding to each window into a probability curve;

[0078] The point in each probability curve with a probability value of 0.5 is defined as a potential split point;

[0079] Integrate the potential segmentation points in group α to obtain the positions of the final segmentation points;

[0080] Semantic action data is extracted from the motion data stream of the target to be detected based on the position of the final segmentation point.

[0081] When the window resolution is high, the window length is small, and the data length within the window is naturally short as well. Therefore, the classification accuracy based on this data fluctuates greatly, which leads to a large number of potential split points clustering together.

[0082] Specifically, the process of integrating the α group of potential segmentation points to obtain the location of the final segmentation point includes:

[0083] For each group of potential segmentation points in group α, sort them in chronological order. If there are multiple consecutive potential segmentation points in the current group, and the time between any two consecutive potential segmentation points is less than a time threshold, delete the multiple consecutive potential segmentation points if the number of such points is even. If the number of such points is odd, use the average coordinate of such points as a segmentation point. The remaining potential segmentation points in the current group are retained as segmentation points, thus obtaining group α of segmentation points.

[0084] The mode of the number of segmentation points in group α is taken as the final number of segmentation points.

[0085] Select β groups of segmentation points from the α group of segmentation points, with the same number of segmentation points as the final number of segmentation points.

[0086] For a set of β segmentation points, the average time of the β i-th segmentation points is taken as the position of the i-th final segmentation point, where the value of i ranges from 1 to the number of final segmentation points, thus obtaining a set of final segmentation points.

[0087] In summary, for d potential split points t i In the sequence (i = 1, 2, 3, ..., d), the split points that are too close to each other are integrated to obtain d′ (d′ ≤ d) split points t′. i (i = 1, 2, 3…d′).

[0088] Choosing different resolution windows has a significant impact on the segmentation results of the above motion segmentation scheme. Using a high-resolution window results in high accuracy of segmentation points, while using a low-resolution window is less susceptible to noise interference in the data stream and therefore has better robustness.

[0089] Inertial measurement units (IMUs) can collect acceleration and angular velocity data during motion, unaffected by external environmental factors such as light, obstructions, and noise. The identity detection method proposed in this invention can identify whether the identity information corresponding to any two sets of IMU motion data matches, thereby completing identity detection and recognition. Furthermore, due to the advantages of IMUs such as small size, easy wearability, low power consumption, low cost, and ease of mass production, the motion capture process based on IMUs is almost unrestricted by any external environment. When capturing writing, a writing panel is not required, thus the writing area is not limited to a small piece of paper. In summary, IMU-based identity detection technology has broad application prospects.

[0090] Step 102: Perform multi-resolution semantic discrimination motion segmentation on the motion data stream of the second target to be detected, and extract the second semantic action data; the motion data stream of the second target to be detected is collected by the inertial sensor carried by the second target to be detected.

[0091] The specific process and principle of step 102 are the same as those of step 101.

[0092] Step 103: Input the first semantic action data and the second semantic action data into the identity recognition model to obtain the identity recognition result; the identity recognition result is that the first target to be detected and the second target to be detected have the same identity, or the first target to be detected and the second target to be detected have different identities.

[0093] The identity recognition model includes a residual neural network, an interpretable feature extraction module, a prior category feature extraction module, a transformer, a feature aggregation module, and a discriminator.

[0094] The residual neural network is specifically a one-dimensional residual neural network (1D-ResNet).

[0095] The residual neural network is used to extract features from the first semantic action data to obtain a first latent feature, and to extract features from the second semantic action data to obtain a second latent feature; the interpretability feature extraction module is used to extract preset interpretability features from the first semantic action data to obtain a first interpretable feature, and to extract preset interpretable features from the second semantic action data to obtain a second interpretable feature; the prior category feature extraction module is used to obtain a first prior category feature based on the matching degree between the first semantic action data and a set number of known identity categories, and to obtain a second prior category feature based on the matching degree between the second semantic action data and a set number of known identity categories. The converter is used to optimize the first latent feature, the first interpretability feature, the first prior category feature, the second latent feature, the second interpretability feature, and the second prior category feature, respectively. The feature aggregation module is used to aggregate the optimized first latent feature, the first interpretability feature, and the first prior category feature to obtain a first aggregated feature, and to aggregate the optimized second latent feature, the second interpretability feature, and the second prior category feature to obtain a second aggregated feature. The discriminator is used to obtain the identity recognition result based on the first aggregated feature and the second aggregated feature.

[0096] Each sample (either the first semantic action data or the second semantic action data) includes 824 feature data that contain the aforementioned time-domain and frequency-domain data. These 824 feature data reveal interpretable characteristics of the statistical, morphological, and dynamic properties of the input data.

[0097] In the identity recognition model, the feature aggregation module is a second multilayer perceptron, and the discriminator is a third multilayer perceptron.

[0098] Using all of these features as explanatory features would make the subsequent network structure too large, consuming a lot of memory and computing power. More importantly, not all of these 824 features will contribute to the target task, i.e., identity detection; some redundant features may even negatively impact the classification process. Therefore, these features need to be selected.

[0099] In step 103, the process of determining the preset interpretability features includes:

[0100] Obtain a sample semantic action data set; the sample semantic action data in the sample semantic action data set is determined by motion segmentation based on motion data streams collected when multiple identities perform different types of actions, using multi-resolution semantic discrimination.

[0101] A first predetermined number of samples are randomly selected from the sample semantic action data set to obtain multiple feature data for each sample; the feature data includes time-domain data and frequency-domain data, and both the time-domain data and the frequency-domain data include mean, variance, slope, number of peaks, and correlation coefficient. The first predetermined number is 100.

[0102] Each feature data point takes a value from a first set number of samples, which forms a feature vector. Each discrete feature can be represented as a 1×100 vector (feature vector), resulting in 824 feature vectors.

[0103] The k-means clustering algorithm is used to cluster multiple feature vectors, resulting in a second predetermined number of feature clusters. This second predetermined number is 80, meaning 80 feature clusters are obtained. Each cluster of features reflects similar signal characteristics. Let represent the i-th feature cluster. Therefore, we only need to select the most valuable feature (MVF) from each feature cluster as the final selected interpretability feature. The MVF selection process is as follows:

[0104] For each feature cluster

[0105] Initialize the set of important features to empty;

[0106] Initialize the feature cluster identity recognition accuracy;

[0107] The remaining feature vectors are traversed sequentially from the current feature cluster. The current second random forest model is trained using the currently acquired feature vectors and the feature vectors in the current important feature set. The identity recognition accuracy of the trained current second random forest model is obtained. The second random forest model is used to perform identity recognition on the input features.

[0108] If the remaining feature vectors in the current feature cluster have been traversed to completion, then the identity recognition accuracy corresponding to each feature vector in the current feature cluster is obtained;

[0109] If the highest identity recognition accuracy of each feature vector in the current feature cluster is greater than the current feature cluster identity recognition accuracy, then the current feature cluster identity recognition accuracy is updated to the highest current identity recognition accuracy. The feature vector with the highest identity recognition accuracy in the current feature cluster is extracted and placed into the important feature set. The process of traversing the remaining feature vectors in the current feature cluster sequentially, using the currently obtained feature vectors and the feature vectors in the current important feature set to train the current second random forest model, and obtaining the identity recognition accuracy of the trained current second random forest model continues until the highest identity recognition accuracy in the current feature cluster is less than or equal to the current feature cluster identity recognition accuracy.

[0110] The important feature sets of each feature cluster are merged to obtain the merged important features, which are then denoted as merged important features.

[0111] Return to the step of randomly extracting a first set number of samples from the sample semantic action data set, and obtaining multiple feature data of each sample, until m merged important features are obtained;

[0112] The intersection of m important features is taken as the set of most valuable features;

[0113] The most valuable feature set is input into the first multilayer perceptron to obtain the first interpretable feature with the same dimension as the first latent feature.

[0114] In step 103, regarding obtaining the first prior category feature based on the matching degree between the first semantic action data and multiple identity categories, the prior category feature extraction module is specifically used for:

[0115] Obtain the matching degree between the first semantic action data and the feature codes of a third set number of known identity categories;

[0116] Each matching degree is used as a weight and weighted and summed with the feature codes of a third set number of known identity categories to obtain the first prior category feature of the first semantic action data.

[0117] The prior category feature extraction module includes an identity category encoding matrix, which is used to store the feature codes of a third set number of known identity categories.

[0118] The training process of the identity recognition network includes a first training phase and a second training phase.

[0119] The first training phase is the training of the residual neural network;

[0120] Based on the target-focusing mechanism, the training of the residual neural network specifically includes:

[0121] Initialize the action content classifier; the action content classifier is used to classify action categories based on the features output by the residual neural network and the features output by the interpretable feature extraction module;

[0122] The training objective is to reduce the prediction error of the action content classifier. The residual neural network and the action content classifier are trained using a first training set and combined with the early stopping method to obtain a trained action content classifier and a residual neural network after the first training. The first training set includes multiple first sample data pairs. Each first sample data pair includes first input data and first label data. The first input data is a motion sample data stream, and the first label data is an action category.

[0123] Fix the trained action content classifier (the parameters in the trained action content classifier will no longer change), initialize the motion identity classifier, and connect the action content classifier and the motion identity classifier in parallel; the motion identity classifier is used to classify the identity categories of the features output by the residual neural network and the features output by the interpretability feature extraction module.

[0124] In order to remove action content information and focus on identity information from the features extracted by the residual neural network, that is, under this feature, the motion identity classifier has high accuracy and the action content classifier has low accuracy.

[0125] However, in general, when the classifier and the residual neural network are directly connected, the gradient update of the residual neural network by the classifier's loss function will increase its accuracy. To decrease its accuracy, there are two methods.

[0126] The first method requires changing the sign of the gradient update. Specifically, a gradient inversion layer (GRL) is added between the residual neural network and the action content classifier after the first training. This reverses the update of the residual neural network by the action content classifier loss function, thereby achieving the goal of ignoring action content information and focusing on identity information.

[0127] The second approach: Modify the loss function of the action content classifier to: 1 / sqrt(L content ).

[0128] Among them, L content This represents the prediction error of the trained action content classifier.

[0129] By using a gradient reversal layer or modifying the loss function of the action content classifier, the training objectives are to increase the prediction error of the action content classifier and reduce the prediction error of the motion identity classifier. A second training set is used, combined with early stopping, to train the residual neural network and the motion identity classifier after the first training, resulting in a trained residual neural network and a trained motion identity classifier. The second training set includes multiple second sample data pairs, each of which includes second input data and second label data. The second input data is a motion sample data stream, and the second label data is the action category and the identity category.

[0130] The matching degree between the first semantic action data and the feature codes of a third set number of known identity categories is the output feature vector obtained by inputting the first semantic action data into a trained motion identity classifier.

[0131] The second training phase involves training the identity category encoding matrix, the converter, the feature aggregation module, and the discriminator, specifically including:

[0132] The training objective is to reduce the prediction error of the regularization term and the discriminator. The identity category coding matrix, the converter, the feature aggregation module and the discriminator are trained using the third training set and combined with the early stopping method, so as to obtain the trained identity category coding matrix, the converter, the feature aggregation module and the discriminator.

[0133] The third training set includes multiple third sample data pairs, each of which includes third input data and third label data. The third input data is a first motion sample data stream and a second motion sample data stream, and the third label data is the identity recognition result.

[0134] The regular expression is represented as follows:

[0135] Among them, R S S represents the regularization term. α (E) represents the α-order Renyi information entropy of the identity category coding matrix E.

[0136] The motion data stream of the first target to be detected is the motion data stream of the first person to be detected when writing; the motion data stream of the second target to be detected is the motion data stream of the second person to be detected when writing.

[0137] The identity detection method based on inertial sensors proposed in this invention requires only a small amount of training data (up to 10 people) to achieve identity detection for any individual in the world. This means that it is necessary to achieve accurate classification of nearly infinite categories based on training samples with extremely limited categories. To achieve this, this invention proposes a Draw Inference Learning framework (identity recognition model). This framework enables machine learning methods to summarize, deduce, and inductively represent data like humans. It can then learn the knowledge and experience required for identity detection from limited data samples, perform feature representation on samples of unknown categories, and ultimately achieve identity detection for any sample.

[0138] Example 2

[0139] This embodiment provides a method for detecting identities with infinite categories, including the following steps:

[0140] Step 1: Motion segmentation based on multi-resolution semantic discrimination.

[0141] Motion data streams include acceleration data streams and angular velocity data streams. The semantic action data upon which identity detection tasks rely is hidden within this continuous data stream, requiring automatic identification and extraction of the data segments corresponding to the semantic actions. To address this, this invention proposes a data extraction, or motion segmentation, scheme.

[0142] like Figure 2 As shown, the motion segmentation scheme generally includes:

[0143] We obtain a human motion data stream (the motion data stream of the target to be detected) collected by an IMU within a certain set time period, denoted as X, where the data length of X is N. To accurately identify the data corresponding to semantic actions in the data stream, we first set up a sliding window of size L on the data stream, along with a pre-trained random forest model. Each time the sliding window slides L across the data stream, the window data is input into the random forest model to perform a binary classification task between semantic action data and general action data, obtaining the probability p that each window data point in the data stream is semantic action data. i (p i ∈[0,1],i=1,2,3…n), where n=N%L is the quotient obtained by dividing the length N of the data stream X by the sliding window size L.

[0144] Through cubic spline interpolation, these discrete probability values ​​are transformed into a continuous probability curve. Points on this curve with higher probability values ​​(test points) correspond to semantic action data, while points with lower probability values ​​correspond to general action data. Specifically, on the probability curve, points with a probability greater than 0.5 are considered semantic action data, points with a probability less than 0.5 are considered general action data, and points with a probability equal to 0.5 are considered potential action segmentation points t. i ∈[1,N](i=1,2,3…d), where d is the number of split points. The types of split points can be divided into two categories: semantic action start points and semantic action end points.

[0145] When the window resolution is high, the window length is small, and the data length within the window is naturally short. Therefore, the classification accuracy based on this data fluctuates significantly, leading to a large clustering of potential split points. When an even number of potential split points cluster together, the action types corresponding to the data before and after these potential split points are consistent; therefore, all of these potential split points are deleted. Conversely, when an odd number of potential split points cluster together, the action types corresponding to the data before and after these potential split points are inconsistent; the average coordinates of these potential split points are taken as the final split point. In summary, for d potential split points t... i In the sequence (i = 1, 2, 3, ..., d), the split points that are too close to each other are integrated to obtain d′ (d′ ≤ d) split points t′. i (i = 1, 2, 3…d′).

[0146] Choosing different resolution windows has a significant impact on the segmentation results of the above motion segmentation scheme. Using a high-resolution window results in high accuracy of segmentation points, while using a low-resolution window is less susceptible to noise interference in the data stream and therefore has better robustness.

[0147] To balance the advantages of both high-resolution and low-resolution windows, α resolution windows were selected within a window size range of 0.4s to 2s. The segmentation results from these α resolution windows were then fused to complete the motion segmentation of the data samples. Specifically, for a data stream, α sets of motion segmentation points T were obtained using these α resolution windows according to the algorithm described above. i s (s=1,2,3…α; i=1,2,3…d s ), where d s It is the number of segmentation points in the s-th motion segmentation point group. Then, the number of these α segmentation points, d... s The mode d of (s = 1, 2, 3…α) mode As the final motion segmentation point T * The number of dividing points, and assuming that the number of dividing points in this α group is equal to the mode d. modeThere are β (β≤α) sets of identical split points. Use these β sets of segmentation points to calculate the final segmentation point T. * Location of each dividing point The average value of the segmentation points at the same position in these β (β≤α) groups of segmentation points is taken as the final segmentation point, as shown in the following formula:

[0148]

[0149] Based on the final segmentation point T * Extract semantic action data from the data stream.

[0150] Step 2: Based on the extracted semantic action data, perform infinite category identity detection based on Draw Inference Learning.

[0151] Most machine learning models can only accurately classify samples of known categories, and they cannot cope with a large number of new category samples. This is not difficult for humans. Thanks to their superior deductive and inductive abilities, humans can easily and accurately classify an infinite number of unknown categories based on experience. For example, when faced with a large number of unfamiliar facial images, humans can easily classify them according to their identities.

[0152] To address this, this invention proposes a "deductive-inductive machine learning" framework. This framework simulates the human learning process, deeply mining and memorizing knowledge from labeled samples of a finite number of categories, thereby achieving accurate classification of an infinite number of unfamiliar category samples. The structure of the "deductive-inductive machine learning" framework is as follows: Figure 3 As shown.

[0153] The overall idea of ​​this framework is to train several feature extractors based on a finite number of labeled samples to achieve comprehensive feature representation of any sample. Then, based on the feature representation results of the sample, it is determined whether different samples belong to the same category, thereby completing the classification of infinitely many unfamiliar category samples.

[0154] It can be seen that comprehensive feature representation of samples is the focus of this technology. Therefore, feature extraction and representation of input data are performed from three aspects: latent features, interpretable features, and prior category features. The extraction methods for each feature are described below. This invention sets up a one-dimensional residual neural network to extract n1 channel features from the input data, with a feature dimension of d. Residual neural networks have excellent feature extraction capabilities, and the extracted features contain the potential patterns of the input data; therefore, the n1 d-dimensional features extracted by the neural network are called latent features. However, the latent features extracted by the one-dimensional residual neural network do not have practical mathematical or physical meaning for humans; this lack of interpretability is also a drawback of this type of feature. Therefore, this invention constructs an interpretable feature extraction module.

[0155] 1. Interpretable feature extraction.

[0156] This embodiment selects 824 interpretable features that can reveal the statistical, morphological, and dynamic properties of the input data, including features such as mean, variance, slope, number of peaks, and correlation coefficient in the time and frequency domains.

[0157] However, using all these features as interpretable features would make the subsequent network structure too large, consuming a lot of memory and computing power. More importantly, not all of these 824 features will contribute to the target task, i.e., identity detection; some redundant features may even negatively impact the classification process. Therefore, these features need to be filtered. First, considering that interpretable features are all discrete scalar features, each discrete interpretable feature needs to be vectorized. Specifically, 100 samples were randomly selected from the dataset, so the value of each discrete feature under these 100 samples can be obtained, and each discrete feature can be represented as a 1×100 vector. After obtaining the vector representations of the 824 features, the k-means clustering algorithm was used to cluster all features, resulting in 80 feature clusters. Each feature cluster reflects similar signal characteristics. Therefore, only the Most Valuable Feature (MVF) needs to be selected from each feature cluster as the final interpretable feature. The MVF selection process is as follows:

[0158] For each feature cluster When performing feature filtering, a set of important features is preset. When selecting features to add to the set, iterate through all remaining features in the feature cluster. The second random forest model is then trained sequentially with existing features from the important feature set to complete one round of identity detection classification. Clearly, each candidate feature will obtain a classification result. Higher classification accuracy means a greater performance improvement of that feature compared to existing features in the important feature set, thus making that feature more valuable for the identity detection task. Therefore, the feature c that achieves the highest classification accuracy in this round of iteration is selected. best Add to the set of important features Then, the next round of feature selection begins, which involves selecting the features from the remaining features in this feature cluster that provide the greatest improvement to the existing features and adding them to the important feature set. When all remaining features can no longer improve the classification accuracy of the important feature set, the set is considered important. It contains all the features in this feature cluster that are most valuable for identity detection.

[0159] Similarly, for each feature cluster Each can obtain a corresponding set of important features. The 80 feature clusters correspond to the 80 important feature sets. By merging, all important features of the sample can be obtained. The training of the random forest model is random, and the features selected each time will have some differences. To obtain features that are more robust to identity detection, the experiment is repeated m times, and the resulting set of m important features is analyzed. By calculating the intersection, we can ultimately obtain the most valuable features of the sample in the identity detection task.

[0160] Furthermore, the feature sets selected based on different samples may differ, and these differences in feature selection may stem from variations in the action content acquired during the sample collection process. Since the identity detection task does not rely on any action content information, features reflecting differences in action content should be removed. Therefore, for the N feature sets obtained from N samples... By further calculating their intersection, we can ultimately obtain the core set of interpretable features that are most valuable for the identity detection task. The interpretable features in the set form a vector of length m. In order to aggregate it while maintaining the same feature dimension as the latent features, it is input into the MLP to obtain a vector of length d, which is the interpretable feature of the sample.

[0161] 2. Target-Focused Mechanism

[0162] The latent feature and interpretable feature representation modules mine and represent the semantic information of the input samples from two perspectives. However, since this semantic information includes not only the identity of the action performer but also the content information of the action, this content information is also included in the latent and interpretable features. The goal of this invention is to achieve the detection and classification of unfamiliar identities based on any action, without relying on any special action. Obviously, under this condition, the content information of the action not only does not help with identity recognition but also introduces interference. Therefore, this invention designs a "target focusing mechanism" that can separate and remove information related to action content from the sample representation, thereby enhancing the ability to represent identity features. Specifically, firstly, an action content classifier is pre-trained, which can accurately identify the characters contained in the input motion data, i.e., the content information.

[0163] Subsequently, the action content classifier is frozen (its weights are no longer updated) and connected in parallel with the identity classifier as a fixed structure. Both classifiers simultaneously perform motion identity detection and motion content detection based on latent features and interpretability features. Let the error of the action content classifier be L. content Then 1 / sqrt(L) content The loss function of this action content classifier is used as the implicit feature and interpretable feature. As the loss function decreases, the frozen action content classifier gradually becomes unable to identify and classify content information based on implicit and interpretable features, which means that the content information in the implicit and interpretable features is removed. Therefore, the generation process of implicit and interpretable features can focus on identity information, and the identity recognition task is no longer interfered with by content information. Regardless of what characters the test subject writes or what gestures they perform, this identity detection system can achieve accurate identity detection.

[0164] 3. Prior category information representation.

[0165] Manually extracted interpretable features and deep learning-based latent features effectively represent the information in the input data samples. Especially under the "target-focusing mechanism," almost all features irrelevant to identity information are removed, while all identity-related features are retained. However, a large amount of prior information remains untapped, such as category information. Different categories themselves contain a wealth of prior information; it is by sorting and summarizing this information that humans can reason and classify a large number of samples with unknown categories.

[0166] Clearly, humans' superior deductive and inductive abilities are based on summarizing and memorizing prior categories. Therefore, it is necessary to encode, represent, and store the categories present in the training set. Based on the representations of a finite number of categories in the training set, the model can further infer the category vector representation of any unknown category sample. Naturally, the category vector corresponding to the sample will provide crucial information for sample classification.

[0167] To obtain the category features of samples of unknown categories, we first need to obtain the feature representations of 10 known categories. For this purpose, a learnable matrix E (size d×10) is initialized. This matrix is ​​multiplied by the classification result vector Y of the known samples (size 10×1). Since samples of known categories are easy to classify, the classification result vector Y is a sparse vector, meaning that the elements in this vector corresponding to the true category have values ​​extremely close to 1, while the values ​​of the remaining elements are close to 0. Therefore, multiplying matrix E by vector Y means that one column of elements in matrix E is extracted, and the remaining columns, after being multiplied by elements close to 0, also become close to 0, thus unable to participate in subsequent calculations or be updated in this training. This means that each column of matrix E corresponds to an identity category. Therefore, matrix E is called the "identity category encoding matrix," and each column of this matrix is ​​a vector representation of a known category. These 10 vectors are called the category feature vectors.

[0168] While the identity category coding matrix E can provide a vector representation for each known category, an ideal vector representation should have the following two advantages:

[0169] (1) The category feature vector should reflect the differences between different patterns as much as possible, therefore the feature vectors of different categories should be as orthogonal as possible;

[0170] (2) Category feature vectors should store as much information as possible. Although three feature vectors in the form of [1,0,0], [0,1,0], [0,0,1] are orthogonal to each other, they cannot fully represent the semantic information of each category.

[0171] To ensure that the categorical feature vectors meet the above two requirements, a plug-and-play Entropy Increase Representation Learning (EIRL) structure was designed. This structure can make the feature vectors of different categories as orthogonal as possible while increasing the amount of information stored in each feature. Specific operations are as follows: Figure 3 As shown.

[0172] Figure 3 The identity classification is a motion identity classifier, and the gesture classification is an action content classifier.

[0173] Since the identity category encoding matrix E stores the feature vectors of all categories, the α-order Renyi information entropy S of the matrix is ​​used first. α The information content of matrix E can be measured by the following equation:

[0174]

[0175]

[0176] G[i][j]= <E (i) E (j) >;

[0177] Where G represents the Gram matrix of matrix E, and the element G[i][j] in the i-th row and j-th column of G is equal to the inner product of the i-th column and the j-th column of matrix E. For the trace normalization of G, i.e. In similar problems, taking α to be 2 generally yields better results.

[0178] S α (E) measures the information content of the identity category coding matrix E, S α The larger (E) is, the greater the information content of matrix E, and the greater the information content of each category feature vector. Meanwhile, from S... α It is not difficult to find in the calculation process of (E) that, as S α As (E) increases, the Gram matrix G of the identity category encoding matrix E decreases, meaning the smaller the inner product of the feature vectors of different categories, the greater the degree of transaction. It can be observed that both requirements increase with S. α (E) is satisfied simultaneously. Therefore, a regularization term R is set. S This regularization term can be added to the loss function of a deep learning model. As this regularization term decreases with training, S... α (E) gradually increases, thereby obtaining high-quality feature encoding for each identity category.

[0179]

[0180] Although the identity category encoding matrix only stores the feature codes of the 10 categories in the training set, the feature codes of unknown categories can be obtained through a linear combination of the features of known categories. As mentioned earlier, when a known category sample is input into the classifier, the classification result is a sparse vector, which means that the known sample has a very high matching degree with a certain category, while the matching degree with other categories is almost 0. When an unknown category sample is input into the trained classifier, the matching degree distribution vector between the sample and the 10 known categories can be obtained. By using this as a weight and weighting it with the feature vectors of the 10 categories, the category vector representation of the sample at that location can be obtained, thus realizing the representation of the unknown category features based on the existing category features.

[0181] 4. Multi-source feature aggregation and classification.

[0182] At this point, for any input sample, three feature representations based on that sample can be obtained: latent features, interpretability features, and categorical features. To further explore the correlation information among these three different source features and to achieve interaction between them, they are input into a Transformer, which possesses strong multi-source information fusion and global information capture capabilities. Therefore, the vector set output by the Transformer has stronger feature representation capabilities and is more valuable for identity detection tasks. To aggregate this vector set, it is input into a Multilayer Perceptron (MLP) to obtain the final vector representation of the sample.

[0183] The objective of this invention is to determine whether any two input data samples belong to the same identity category, thereby achieving accurate classification of samples across an infinite number of categories. Therefore, the identity recognition model takes two data samples as input each time, generates unique feature vectors for each sample as described above, and inputs these vectors into an MLP classifier. This classifier performs a binary classification task, where 1 indicates that the input samples belong to the same identity category, and 0 indicates that the input samples belong to different identity categories. Because the information of the input data samples has been deeply mined and comprehensively represented, this binary classifier can achieve ultra-high accuracy. In summary, the model of this invention achieves identity detection for an infinite number of categories.

[0184] Step 3: Dataset creation.

[0185] Inertial sensor-based identity detection is influenced not only by individual movement habits but also by physiological differences such as height, weight, age, and gender. To ensure the method is applicable to almost all individuals, the dataset (including the first and second training sets) should include individuals with diverse physiological characteristics. Therefore, the dataset created for this invention covers heights from 1.2m to 1.9m, weights from 30kg to 100kg, and ages from 12 to 60 years old, including both genders. Furthermore, to avoid the influence of motion content information on identity detection, the identity recognition model needs to be guided during training to learn to avoid motion content information in the data. Therefore, the dataset should contain rich motion content information for training. To this end, 62 gesture actions (10 Arabic numerals and 52 uppercase and lowercase English letters) containing content information were included in the first and second training sets.

[0186] This invention proposes an infinite category identity detection method based on inertial sensors, which can be used in diverse identity detection scenarios such as unlocking electronic devices, controlling important devices, and logging into special software.

[0187] Similar to biometric technologies such as fingerprint recognition, facial recognition, and voiceprint recognition, this invention proposes a biometric technology based on motion information. In fact, no research has ever proposed or attempted to use inertial sensors to collect motion data and extract biometric features from it.

[0188] This invention proposes a machine learning framework with memory, reasoning, and deductive-inductive capabilities: "Deductive-Inductive Machine Learning". It can simulate the human learning process, generate unique feature representations for samples under an infinite number of categories based on extremely limited data samples, and complete downstream tasks such as recognition and prediction based on these feature representations.

[0189] This invention proposes a "target-focused mechanism" to guide the feature extraction direction of deep learning models, thereby enabling deep learning models to focus on the features required by the target task to achieve deeper learning effects and ultimately significantly improve the performance of the target task.

[0190] This invention proposes a "category feature representation method". In classification tasks, it can generate a representation vector for each category to describe and characterize the main features of that category, thereby improving the accuracy of the model in classification tasks.

[0191] This invention proposes a novel identity detection task: biometrics based on motion information. In fact, no research has ever attempted to use inertial sensors to collect motion features and present it as a biometric technology on par with fingerprint recognition, facial recognition, and voiceprint recognition.

[0192] Example 3

[0193] like Figure 4 As shown, this embodiment provides an infinite category identity detection system, including:

[0194] The first semantic action data extraction module 201 is used to perform multi-resolution semantic discrimination motion segmentation on the motion data stream of the first target to be detected, and extract the first semantic action data; the motion data stream of the first target to be detected is collected by an inertial sensor carried by the first target to be detected.

[0195] The second semantic action data extraction module 202 is used to perform multi-resolution semantic discrimination motion segmentation on the motion data stream of the second target to be detected, and extract the second semantic action data; the motion data stream of the second target to be detected is collected by the inertial sensor carried by the second target to be detected.

[0196] The identity recognition module 203 is used to input the first semantic action data and the second semantic action data into the identity recognition model to obtain the identity recognition result; the identity recognition result is that the first target to be detected and the second target to be detected have the same identity, or the first target to be detected and the second target to be detected have different identities.

[0197] The identity recognition model includes a residual neural network, an interpretable feature extraction module, a priori category feature extraction module, a converter, a feature aggregation module, and a discriminator;

[0198] The residual neural network is used to extract features from the first semantic action data to obtain a first latent feature, and to extract features from the second semantic action data to obtain a second latent feature; the interpretability feature extraction module is used to extract preset interpretability features from the first semantic action data to obtain a first interpretable feature, and to extract preset interpretable features from the second semantic action data to obtain a second interpretable feature; the prior category feature extraction module is used to obtain a first prior category feature based on the matching degree between the first semantic action data and a set number of known identity categories, and to obtain a second prior category feature based on the matching degree between the second semantic action data and a set number of known identity categories. The converter is used to optimize the first latent feature, the first interpretability feature, the first prior category feature, the second latent feature, the second interpretability feature, and the second prior category feature, respectively. The feature aggregation module is used to aggregate the optimized first latent feature, the first interpretability feature, and the first prior category feature to obtain a first aggregated feature, and to aggregate the optimized second latent feature, the second interpretability feature, and the second prior category feature to obtain a second aggregated feature. The discriminator is used to obtain the identity recognition result based on the first aggregated feature and the second aggregated feature.

[0199] Example 4

[0200] This embodiment provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the infinite category identity detection method described in Embodiment 1.

[0201] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0202] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting infinite class identity, characterized by, The method comprises the following steps: performing motion segmentation on a motion data stream of a first target to be detected by multi-resolution semantic discrimination to extract first semantic action data; the motion data stream of the first target to be detected is collected by an inertial sensor carried by the first target to be detected; performing motion segmentation on a motion data stream of a second target to be detected by multi-resolution semantic discrimination to extract second semantic action data; the motion data stream of the second target to be detected is collected by an inertial sensor carried by the second target to be detected; inputting the first semantic action data and the second semantic action data into an identity recognition model to obtain an identity recognition result; the identity recognition result is that the first target to be detected and the second target to be detected have the same identity or that the first target to be detected and the second target to be detected have different identities; the identity recognition model comprises a residual neural network, an explainable feature extraction module, a prior category feature extraction module, a converter, a feature aggregation module and a discriminator; the residual neural network is used for feature extraction on the first semantic action data to obtain first hidden features and feature extraction on the second semantic action data to obtain second hidden features; the explainable feature extraction module is used for extracting preset explainable features in the first semantic action data to obtain first explainable features and extracting preset explainable features in the second semantic action data to obtain second explainable features; the prior category feature extraction module is used for obtaining first prior category features based on matching degrees of the first semantic action data and a set number of known identity categories and obtaining second prior category features based on matching degrees of the second semantic action data and the set number of known identity categories; the converter is used for feature optimization on the first hidden features, the first explainable features, the first prior category features, the second hidden features, the second explainable features and the second prior category features respectively; the feature aggregation module is used for aggregating the first hidden features, the first explainable features and the first prior category features after feature optimization to obtain first aggregated features and aggregating the second hidden features, the second explainable features and the second prior category features after feature optimization to obtain second aggregated features; and the discriminator is used for obtaining an identity recognition result according to the first aggregated features and the second aggregated features. performing motion segmentation on a motion data stream of a first target to be detected by multi-resolution semantic discrimination to extract first semantic action data, specifically comprising:

2. The method of claim 1, wherein, setting alpha windows of different lengths, sliding each window on the motion data stream of the first target to be detected in turn, inputting window data of each sliding window into a first random forest model to perform a binary classification task of semantic action data and non-semantic action data, and obtaining a probability value of each window data of each sliding window being semantic action data; the first random forest model is a trained random forest model; converting the probability value corresponding to each window into a probability curve by using cubic spline interpolation; ​ A point in each of the probability curves where the probability value is equal to 0.5 is defined as a potential segmentation point; The alpha group of potential segmentation points are integrated to obtain the position of the final segmentation point; According to the position of the final segmentation point, semantic action data is extracted from the motion data stream of the target to be detected.

3. The method of claim 2, wherein, The alpha group of potential segmentation points are integrated to obtain the position of the final segmentation point, specifically including: For each group of potential segmentation points in the alpha group of potential segmentation points, sort them in time sequence. If there are multiple consecutive potential segmentation points in the current group of potential segmentation points, the time between each two consecutive potential segmentation points satisfies the condition that the time is less than a time threshold. When the number of the multiple consecutive potential segmentation points is even, the multiple consecutive potential segmentation points are deleted. When the number of the multiple consecutive potential segmentation points is odd, the average coordinates of the multiple consecutive potential segmentation points are taken as a segmentation point. The remaining potential segmentation points in the current group are retained as segmentation points, and finally the alpha group of segmentation points are obtained. The mode of the number of the alpha group of segmentation points is taken as the final number of segmentation points. From the alpha group of segmentation points, a beta group of segmentation points with the same number of segmentation points as the final number of segmentation points are selected. For the beta group of segmentation points, the time average of the i-th segmentation point in the beta group is taken as the position of the i-th final segmentation point, i ranges from 1 to the final number of segmentation points, and a group of final segmentation points are obtained.

4. The method of claim 1, wherein, The determination process of the pre-set explainable feature includes: A sample semantic action data set is obtained. Each sample semantic action data in the sample semantic action data set is determined by multi-resolution semantic discrimination of motion segmentation based on motion data streams collected when different types of actions are performed by multiple identities. A first set number of samples are randomly extracted from the sample semantic action data set, and a plurality of feature data of each sample are obtained. The feature data includes time domain data and frequency domain data, and the time domain data and the frequency domain data each include mean, variance, skewness, peak number, and correlation coefficient. The value of each feature data under the first set number of samples constitutes a feature vector. A k-means clustering algorithm is used to cluster the plurality of feature vectors to obtain a second set number of feature clusters. For each feature cluster: An important feature set is initialized as empty; The feature cluster identity recognition accuracy is initialized; The remaining feature vectors in the current feature cluster are iteratively traversed, a current second random forest model is trained using the current obtained feature vector and the feature vectors in the current important feature set, and the identity recognition accuracy of the trained current second random forest model is obtained. The second random forest model is used for identity recognition of input features; If the traversal of the remaining feature vectors in the current feature cluster is completed, the identity recognition accuracy corresponding to each feature vector in the current feature cluster is obtained. If the highest identity recognition accuracy of the feature vectors in the current feature cluster is greater than the identity recognition accuracy of the current feature cluster, the identity recognition accuracy of the current feature cluster is updated to the highest identity recognition accuracy, the feature vector with the highest identity recognition accuracy in the current feature cluster is extracted and put into the important feature set, and the steps of sequentially traversing the remaining feature vectors in the current feature cluster, training the current second random forest model with the current obtained feature vector and the feature vectors in the current important feature set, and obtaining the identity recognition accuracy of the trained current second random forest model are returned until the corresponding highest identity recognition accuracy in the current feature cluster is less than or equal to the identity recognition accuracy of the current feature cluster. The important feature sets of each feature cluster are merged to obtain merged important features, and the merged important features are denoted as merged important features. Return to the step of randomly extracting a first set number of samples from the sample semantic action data set to obtain a plurality of feature data of each sample until m merged important features are obtained. The intersection of the m merged important features is taken as a most valuable feature set. The most valuable feature set is input into a first multilayer perceptron to obtain a first explainability feature with the same dimension as the first implicit feature.

5. The method of claim 1, wherein, In terms of obtaining a first prior category feature based on the matching degree of the first semantic action data and a plurality of identity categories, the prior category feature extraction module is specifically configured to: Obtain the matching degree of the first semantic action data and the feature encoding of the third set number of known identity categories. Sum the matching degrees as weights and the feature encodings of the third set number of known identity categories to obtain the first prior category feature of the first semantic action data.

6. The method of claim 5, wherein, The identity recognition model is a trained identity recognition network; the training process of the identity recognition network includes a first training phase and a second training phase; The first training phase is the training of a residual neural network, specifically including: Initializing an action content classifier; the action content classifier is used to classify the features output by the residual neural network and the features output by the explainability feature extraction module into action categories; Taking the reduction of the prediction error of the action content classifier as the training target, training the residual neural network and the action content classifier using a first training set combined with an early stopping method to obtain a trained action content classifier and a residual neural network after the first training; the first training set includes a plurality of first sample data pairs, each first sample data pair includes first input data and first label data, the first input data is a motion sample data stream, and the first label data is an action category; Fixing the parameters in the trained action content classifier, adding a gradient reversal layer between the residual neural network after the first training and the action content classifier, initializing a motion identity classifier, and connecting the action content classifier and the motion identity classifier in parallel; the motion identity classifier is used to classify the features output by the residual neural network and the features output by the explainability feature extraction module into identity categories. The gradient inversion layer is used to take the prediction error of the action content classifier and the prediction error of the motion identity classifier as training targets, and the residual neural network and the motion identity classifier after the first training are trained using a second training set and in combination with an early stopping method to obtain a trained residual neural network and a trained motion identity classifier; the second training set includes a plurality of second sample data pairs, each of the second sample data pairs including second input data and second label data, the second input data being a motion sample data stream, and the second label data including an action category and an identity category; The matching degree of the first semantic action data and the feature encoding of the third set number of known identity categories is an output feature vector obtained by inputting the first semantic action data into the trained motion identity classifier.

7. The method of claim 6, wherein, The prior category feature extraction module includes an identity category encoding matrix, and the identity category encoding matrix is used to store feature encoding of a third set number of known identity categories. The second training stage is training of the identity category encoding matrix, the converter, the feature aggregation module, and the discriminator, and specifically includes: Taking the reduction regular term and the prediction error of the discriminator as training targets, the identity category encoding matrix, the converter, the feature aggregation module, and the discriminator are trained using a third training set and in combination with an early stopping method to obtain the trained identity category encoding matrix, the converter, the feature aggregation module, and the discriminator. The third training set includes a plurality of third sample data pairs, each of the third sample data pairs including third input data and third label data, the third input data being a first motion sample data stream and a second motion sample data stream, and the third label data being an identity recognition result. The regularizer is expressed as: where R S represents the regularization term, S α (E) represents the αth-order Renyi information entropy of the identity category coding matrix E.

8. The method of claim 1, wherein, The motion data stream of the first to-be-detected target is a motion data stream of a first to-be-detected person when writing, and the motion data stream of the second to-be-detected target is a motion data stream of a second to-be-detected person when writing.

9. A system for detecting infinite class identity, the system comprising: It includes: A first semantic action data extraction module is configured to perform multi-resolution semantic discrimination motion segmentation on a motion data stream of a first to-be-detected target to extract first semantic action data; the motion data stream of the first to-be-detected target is collected by an inertial sensor carried by the first to-be-detected target; A second semantic action data extraction module is configured to perform multi-resolution semantic discrimination motion segmentation on a motion data stream of a second to-be-detected target to extract second semantic action data; the motion data stream of the second to-be-detected target is collected by an inertial sensor carried by the second to-be-detected target; An identity recognition module is configured to input the first semantic action data and the second semantic action data into an identity recognition model to obtain an identity recognition result; the identity recognition result is that the first to-be-detected target and the second to-be-detected target have the same identity or that the first to-be-detected target and the second to-be-detected target have different identities; The identity recognition model includes a residual neural network, an interpretable feature extraction module, a prior category feature extraction module, a converter, a feature aggregation module, and a discriminator. The residual neural network is configured to extract first hidden features from the first semantic action data and extract second hidden features from the second semantic action data. The interpretable feature extraction module is configured to extract first interpretable features from the first semantic action data and extract second interpretable features from the second semantic action data. The prior category feature extraction module is configured to obtain first prior category features based on matching degrees between the first semantic action data and a set number of known identity categories and obtain second prior category features based on matching degrees between the second semantic action data and the set number of known identity categories. The transformer is configured to perform feature optimization on the first hidden features, the first interpretable features, the first prior category features, the second hidden features, the second interpretable features, and the second prior category features, respectively. The feature aggregation module is configured to aggregate the first hidden features, the first interpretable features, and the first prior category features after feature optimization to obtain first aggregated features, and aggregate the second hidden features, the second interpretable features, and the second prior category features after feature optimization to obtain second aggregated features.

10. An electronic device, comprising: The discriminator is configured to obtain an identity recognition result based on the first aggregated features and the second aggregated features. The electronic device includes a memory and a processor. The memory is configured to store a computer program. The processor is configured to run the computer program to enable the electronic device to perform the infinite-category identity detection method according to any one of claims 1 to 8.

Images