A Human Motion Recognition Method Based on Fabric Sensors

Abstract

This invention discloses a human motion recognition method based on fabric sensors, relating to the field of human motion recognition. The method includes: acquiring raw multi-channel temporal electrical signals of key movement parts of the limbs of a target object using a fabric sensor array; denoising and normalizing the raw multi-channel temporal electrical signals to construct a standardized temporal input; performing feature encoding on the standardized temporal input to generate a morphological primitive seed vector; and dividing the morphological primitive seed vector into a length seed and a feature seed before normalization. This invention generates morphological primitive seed vectors through a feature encoder and dynamically adjusts the time span and feature dimension of the primitives by combining the length seed and feature seed, breaking through the adaptation limitations of traditional fixed primitives and effectively solving the adaptation problem for different motion samples and individual differences.

Description

Technical Field

[0001] This invention relates to the field of human motion recognition technology, specifically a human motion recognition method based on a fabric sensor. Background Technology

[0002] Human motion recognition, as a core supporting technology in fields such as human-computer interaction, smart wearables, rehabilitation medicine, and sports and health monitoring, directly determines the effectiveness of related applications and user experience.

[0003] Currently, human motion recognition methods based on fabric sensors have become a research hotspot. However, existing technologies still have many problems that need to be solved. Traditional fabric sensor motion recognition methods mostly rely on manually designed feature extraction strategies to perform fixed-dimensional feature mining on multi-channel time-series signals collected by fabric sensors. This cannot adapt to the feature differences of different motion samples, nor can it effectively adapt to the motion habits and body feature differences of different users, resulting in weak feature extraction and easy loss of key motion details.

[0004] In the prior art, Chinese patent CN121122571A discloses "A Human Motion Recognition System Based on LSTM". This method captures human motion through a camera, extracts key points using a pose estimation algorithm, and analyzes behavioral patterns to identify the current state. However, this method mainly relies on camera acquisition, which is prone to distortion and cannot accurately extract key information about the target human body.

[0005] To address this, we propose a human motion recognition method based on fabric sensors. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a human motion recognition method based on fabric sensors, which can effectively solve the problems of the existing technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions; This invention discloses a human motion recognition method based on a fabric sensor, comprising: Step 1: Collect raw multi-channel time-series electrical signals of the target object's limb movement parts based on the fabric sensor array, preprocess the raw multi-channel time-series electrical signals, and construct a standardized time-series input; Step 2: Perform feature encoding on the standardized temporal input to generate morphological primitive seed vectors. Based on the morphological primitive seed vectors, dynamically generate multiple morphological primitives adapted to the current sample and the validity mask corresponding to each morphological primitive. After time expansion and weighted aggregation, obtain the sample adaptive morphological primitive temporal feature sequence. Step 3: Extract morphological primitive branch features based on the temporal feature sequence of adaptive morphological primitives, and extract deep branch features based on the standardized temporal input. Perform weighted fusion of morphological primitive branch features and deep branch features to generate fused temporal features. Step 4: Perform temporal phase alignment on the morphological primitive branch features, deep branch features, and fused temporal features. Based on the phase-aligned features, perform multi-granularity voting classification to generate the final human action recognition result.

[0008] Furthermore, step 2 involves feature encoding of the standardized temporal input to generate morphological primitive seed vectors, specifically including the following sub-steps: S211: Perform adaptive average pooling along the time dimension on the standardized time-series input X to obtain a pooling feature of fixed length. ; S212: The pooling feature... After being flattened into a one-dimensional vector, a linear mapping is performed to obtain a morphological primitive seed vector z. The morphological primitive seed vector z is then divided into length seeds according to a preset dimension. With characteristic seeds ; S213: For the length seed With characteristic seeds Normalization is performed to stabilize the distribution of the seed vector.

[0009] Furthermore, in step 2, multiple morphological primitives adapted to the current sample and the validity mask corresponding to each morphological primitive are dynamically generated based on the morphological primitive seed vector, including the following sub-steps: S214: Normalize the length seed The time span parameter of the i-th morphological primitive is generated through nonlinear constraint mapping. According to the time span parameter With normalized feature seeds The i-th morphological primitive vector is generated through a trainable mapping network. ,in K is the preset number of morphological primitives, and D is the feature dimension; S215: Generate a validity mask for each morphological primitive, the formula for which is as follows: ; In the formula: For validity mask; For activation functions; These are the weight parameters for the linear mapping. This is the seed mapping function corresponding to the i-th morphological primitive; These are the bias parameters for the linear mapping. and All of these are trainable parameters.

[0010] Furthermore, in step 2, the temporal feature sequence of the adaptive morphological primitives of the sample is obtained through time expansion and weighted aggregation, including the following sub-steps: S216: The i-th morphological primitive vector According to its time span parameter By performing differentiable time expansion and using continuous-time mapping and differentiable interpolation, the morphological primitive features are mapped onto a unified standard time axis, resulting in a primitive time series. ; S217: Based on Validity Mask Weighted aggregation of K primitive time series sequences yields the sample adaptive morphological primitive time series feature sequence. Its formula is: ; In the formula: For discrete-time indexing; Primary time series In the eigenvectors at time step t, ε is a minimal constant to prevent the denominator from being zero.

[0011] Furthermore, step 3 includes the following sub-steps: S311: Extract morphological primitive branch features based on the temporal feature sequence of the adaptive morphological primitives of the sample, and extract deep branch features based on the standardized temporal input; S312: Input the morphological primitive branch features and the deep branch features into the group network module for feature processing to obtain the output features corresponding to each network module; S313: Based on the morphological primitive branch features, the deep branch features, and the input features of each network module, determine the comprehensive activation weight of each network module, and normalize the comprehensive activation weight. S314: The output features of each network module are weighted and aggregated based on the normalized comprehensive activation weights, and then fused with the morphological primitive branch features and the deep branch features to obtain fused temporal features.

[0012] Furthermore, in step 312, the grouped network module includes several parallel trainable temporal feature extraction network modules, and all network modules are divided into a morphological primitive feature-specific network group and a deep feature-specific network group. Among them, the morphological primitive feature-specific network group receives and processes morphological primitive branch features, and the depth feature-specific network group receives and processes depth branch features. Different network modules correspond to different temporal modeling scales, receptive field sizes, or attention ranges.

[0013] Furthermore, step 313 specifically includes: S3131: Construct a module gating mechanism, perform global average pooling on the morphological primitive branch features and the deep branch features in the time dimension, extract the global statistical feature vectors of the two branches and concatenate them, input the concatenated features into the gating mapping function composed of a trainable linear mapping layer and a Softmax activation function, and generate the morphological primitive branch weights and the deep feature branch weights. S3132: Construct a feature quality gating mechanism, calculate the statistical variance of the input features of each network module in the time dimension, input the statistical variance into the quality mapping function composed of a trainable linear mapping layer and a Sigmoid activation function, and generate the quality weights of the corresponding network modules. S3133: Multiplicatively combine the quality weights of each network module with the branch weights of their respective branches to obtain the comprehensive activation weights of each network module; suppress network modules with comprehensive activation weights below a preset threshold, and normalize the comprehensive activation weights of the retained network modules to obtain the final normalized activation weights used for feature fusion, the formula of which is: ; In the formula: represents the normalized activation weights corresponding to the i-th network module; k is the total number of retained network modules; exp is the exponential function; The preset weight coefficients control the impact of feature energy, feature dynamic change degree, and cross-branch feature consistency on module weights, respectively. Input the average feature vector of the features in the time dimension to the i-th network module; for The L2 norm; Input the variance statistics of the features in the time dimension for the i-th network module; It is the global average feature vector of the morphological primitive branch features; This represents the global average feature vector of the deep branch features; To prevent extremely small constants with a denominator of 0; This indicates a summation operation performed on all network modules involved in the fusion computation.

[0014] Furthermore, in step 314, temporal features are fused. The calculation formula is: ; in, For feature normalization operators; The temporal characteristics output by the i-th retained network module; The morphological primitive branching characteristics; The branch weight of the morphological primitive branch feature at time step t; This is a deep branch feature; The branch weights of the deep branch features at time step t.

[0015] Furthermore, in step 4, when aligning the morphological primitive branch features, deep branch features, and fused temporal features with time phase, the process includes the following sub-steps: S411: Unify the morphological primitive branch features, deep branch features, and fused temporal features into a temporal feature sequence with consistent time length. Perform temporal convolution and nonlinear mapping on each temporal feature sequence to extract the phase structure information of each branch and generate the corresponding phase description vector. S412: Based on the phase description vector of each branch, calculate the phase matching degree of each branch feature in the time dimension, and generate the time weight sequence corresponding to each branch according to the phase matching degree; S413: Based on the time weight sequence, the time-slice weighted adjustment of each branch's time-series features is performed, and the weighted branch features are normalized and fused to obtain the fused features after time phase alignment.

[0016] Furthermore, in step 4, when performing multi-granularity voting classification based on the phase-aligned features to generate the final human action recognition result, the following sub-steps are included: S414: The fused features after time phase alignment are classified into fine-grained, medium-grained, and coarse-grained categories to obtain classification outputs of three granularities. S415: Input the classification output features of the three granularities into the linear mapping layer, transform them into voting weight vectors and normalize them to obtain the adaptive voting weights for each granularity; S416: The classification probability vectors of the three granularities are weighted and summed with their corresponding voting weights to generate the final action classification probability distribution. The category with the highest probability value is selected as the human action recognition result.

[0017] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects: 1. This invention proposes a sample adaptive morphological primitive feature extraction mechanism. It generates morphological primitive seed vectors through a feature encoder and dynamically adjusts the time span and feature dimension of primitives by combining length seeds and feature seeds. This breaks through the adaptation limitations of traditional fixed primitives and effectively solves the adaptation problem of different action samples and different individual differences.

[0018] 2. This invention constructs a group network fusion and dual-gating mechanism, which performs hierarchical processing of sample adaptive morphological primitive features and deep temporal features, further improving the discriminative ability of fused features, strengthening the expression of branch features, and providing more discriminative input for classification tasks.

[0019] 3. This invention proposes a temporal phase alignment and multi-granularity voting classification strategy. It solves the problem of temporal misalignment between fused features and features of each branch by using temporal phase alignment technology. It designs a fine, medium and coarse three-granularity classification voting mechanism and integrates the classification output through adaptive weight allocation, thereby enhancing the stability and robustness of the classification results and adapting to the recognition scenarios of complex, continuous and subtle actions.

[0020] 4. This invention designs a complete technical solution based on the signal characteristics of fabric sensors. From the standardized processing of the original multi-channel time-series signals to adaptive feature extraction, fusion and classification, it is adapted to the time-series nature, multi-channel characteristics and wearable scenario complexity of fabric sensors. It can break through the dependence of traditional action recognition methods on high-precision sensing devices, achieve deep adaptation between technology and application scenarios, and further improve the applicability of the method. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0022] Figure 1 This is a flowchart illustrating a human motion recognition method based on fabric sensors. Figure 2 This is a schematic diagram of the packet network fusion module in this invention; Figure 3 This is a schematic diagram of multi-granularity classification in this invention; Figure 4 This is a physical appearance diagram of the fabric sensor in this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] The present invention will be further described below with reference to embodiments.

[0025] Example: This embodiment presents a human motion recognition method based on a fabric sensor, such as... Figures 1-3 As shown, it includes: Step 1: Collect raw multi-channel time-series electrical signals of the target object's limb movement parts based on the fabric sensor array, preprocess the raw multi-channel time-series electrical signals, and construct a standardized time-series input; Step 2: Encode the standardized temporal input for features, generate morphological primitive seed vectors, dynamically generate multiple morphological primitives adapted to the current sample and the validity mask corresponding to each morphological primitive based on the morphological primitive seed vectors, and obtain the sample adaptive morphological primitive temporal feature sequence through time expansion and weighted aggregation. Step 2 involves feature encoding of the standardized temporal input to generate morphological primitive seed vectors, which specifically includes the following sub-steps: S211: Perform adaptive average pooling along the time dimension on the standardized time-series input X to obtain a pooling feature of fixed length. ; S212: Pooling features Flattened into a one-dimensional vector, a linear mapping is performed to obtain the morphological primitive seed vector z. The morphological primitive seed vector z is then divided into length seeds according to a preset dimension. With characteristic seeds ; The formula is as follows: ; When this formula is used, it first performs adaptive average pooling on the temporal input to obtain a fixed-length feature, then flattens it into a one-dimensional vector, and obtains a morphological primitive seed vector through linear mapping. This transforms the multi-dimensional temporal features into seed information that is adapted to the generation of morphological primitives, allowing the seed vector to accurately carry the core morphological control parameters of the sample, thus distinguishing it from the limitations of traditional fixed feature mapping. ; This formula splits the morphological primitive seed vector into length seed and feature seed according to a preset dimension, and controls the time span and feature expression of the morphological primitive respectively, so as to achieve precise two-dimensional control of primitive generation and enable the primitive to adaptively match the action temporal features of different samples. In the formula: The seed vector for morphological primitives is obtained by linear mapping and contains various control information required to describe the generation of morphological primitives for the current sample. is a linear mapping weight matrix, and is a trainable parameter whose dimension matches the flattened one-dimensional feature vector. It is used to map the one-dimensional feature vector to the morphological primitive seed space. This is a flattening operator used to flatten multidimensional temporal feature tensors. Expand them into one-dimensional feature vectors in a preset order so that they can be processed by linear mapping later. is the linear mapping bias vector, and are trainable parameters used to shift and adjust the linear mapping result. The length seed obtained by the vector segmentation operation is used to represent the span information of the morphological primitive in the time dimension; This is a vector segmentation operator used to segment the morphological primitive seed vector z into length seeds according to a preset dimension. With characteristic seeds ; S213: For length seed With characteristic seeds Normalization is performed to stabilize the distribution of the seed vector; Step 2 involves dynamically generating multiple morphological primitives adapted to the current sample and the validity mask corresponding to each morphological primitive based on the morphological primitive seed vector. This includes the following sub-steps: S214: Normalize the length seed The time span parameter of the i-th morphological primitive is generated through nonlinear constraint mapping. Based on time span parameters With normalized feature seeds The i-th morphological primitive vector is generated through a trainable mapping network. ,in K is the preset number of morphological primitives, and D is the feature dimension; S215: Generate a validity mask for each morphological primitive. The formula for the validity mask is as follows: ; In the formula: For validity mask; For activation functions; These are the weight parameters for the linear mapping. This is the seed mapping function corresponding to the i-th morphological primitive; These are the bias parameters for the linear mapping. and All of these are trainable parameters; The above formula takes the length seed as input, and after the mapping function and linear transformation, it generates an effectiveness mask through the activation function. This precisely controls the activation degree of each morphological primitive in the current sample, allowing the model to adaptively select effective primitives, avoid interference from invalid features, and improve the targeting of feature extraction. In step 2, the temporal feature sequence of adaptive morphological primitives of the samples is obtained through time expansion and weighted aggregation, including the following sub-steps: S216: The i-th morphological primitive vector According to its time span parameter By performing differentiable time expansion and using continuous-time mapping and differentiable interpolation, the morphological primitive features are mapped onto a unified standard time axis, resulting in a primitive time series. ; S217: Based on Validity Mask Weighted aggregation of K primitive time series sequences yields the sample adaptive morphological primitive time series feature sequence. Its formula is: ; In the formula: For discrete-time indexing; Primary time series In the eigenvectors at time step t, ε is a minimal constant to prevent the denominator from being zero; This formula performs weighted aggregation of the temporal sequences of each morphological primitive based on an effectiveness mask, adds a very small constant to avoid the denominator being zero, amplifies the feature contribution of effective primitives, and generates morphological primitive temporal features that are adapted to the current sample, thus solving the problem that traditional fixed primitive aggregation cannot adapt to individual action differences. Step 3: Extract morphological primitive branch features based on the temporal feature sequence of adaptive morphological primitives, and extract deep branch features based on the standardized temporal input. Perform weighted fusion of morphological primitive branch features and deep branch features to generate fused temporal features. Step 3 includes the following sub-steps: S311: Extract morphological primitive branch features based on sample adaptive morphological primitive temporal feature sequence, and extract deep branch features based on standardized temporal input; S312: Input the morphological primitive branch features and the depth branch features into the group network module for feature processing to obtain the output features of each network module; S313: Based on the morphological primitive branch features, deep branch features, and input features of each network module, determine the comprehensive activation weight of each network module and normalize the comprehensive activation weight. S314: The output features of each network module are weighted and aggregated based on the normalized integrated activation weights, and then fused with the morphological primitive branch features and deep branch features to obtain the fused temporal features. In step 312, the grouped network module includes several parallel trainable temporal feature extraction network modules, and all network modules are divided into a morphological primitive feature-specific network group and a deep feature-specific network group. Among them, the morphological primitive feature-specific network group receives and processes morphological primitive branch features, and the depth feature-specific network group receives and processes depth branch features. Different network modules correspond to different temporal modeling scales, receptive field sizes, or attention ranges. Step 313 specifically includes: S3131: Construct a module gating mechanism, perform global average pooling on the morphological primitive branch features and the deep branch features in the time dimension, extract the global statistical feature vectors of the two branches and concatenate them, input the concatenated features into the gating mapping function composed of a trainable linear mapping layer and a Softmax activation function, and generate the morphological primitive branch weights and the deep feature branch weights. S3132: Construct a feature quality gating mechanism, calculate the statistical variance of the input features of each network module in the time dimension, input the statistical variance into the quality mapping function composed of a trainable linear mapping layer and a Sigmoid activation function, and generate the quality weights of the corresponding network modules. S3133: Multiplicatively combine the quality weights of each network module with the branch weights of their respective branches to obtain the comprehensive activation weights of each network module; suppress network modules with comprehensive activation weights below a preset threshold, and normalize the comprehensive activation weights of the retained network modules to obtain the final normalized activation weights used for feature fusion, the formula of which is: ; In the formula: represents the normalized activation weights corresponding to the i-th network module; k is the total number of retained network modules; exp is the exponential function; The preset weight coefficients control the impact of feature energy, feature dynamic change degree, and cross-branch feature consistency on module weights, respectively. Input the average feature vector of the features in the time dimension to the i-th network module; for The L2 norm; Input the variance statistics of the features in the time dimension for the i-th network module; It is the global average feature vector of the morphological primitive branch features; This represents the global average feature vector of the deep branch features; To prevent extremely small constants with a denominator of 0; This indicates a summation operation performed on all network modules involved in the fusion computation. This formula normalizes the output features of the activated network module, the morphological primitive branch and the deep branch features after adaptive weighted fusion, taking into account both local action patterns and global action trends. It achieves feature complementarity through dynamic weight allocation, making the fused features more discriminative, thereby adapting to the needs of complex action recognition. In step 314, temporal features are fused. The calculation formula is: ; in, For feature normalization operators; The temporal characteristics output by the i-th retained network module; The morphological primitive branching characteristics; The branch weight of the morphological primitive branch feature at time step t; This is a deep branch feature; The branch weights of the deep branch features at time step t; Step 4: Perform temporal phase alignment on the morphological primitive branch features, deep branch features, and fused temporal features. Based on the phase-aligned features, perform multi-granularity voting classification to generate the final human action recognition result. Step 4, which involves temporal phase alignment of morphological primitive branch features, deep branch features, and fused temporal features, includes the following sub-steps: S411: Unify the morphological primitive branch features, deep branch features, and fused temporal features into a temporal feature sequence with consistent time length. Perform temporal convolution and nonlinear mapping on each temporal feature sequence to extract the phase structure information of each branch and generate the corresponding phase description vector. S412: Based on the phase description vector of each branch, calculate the phase matching degree of each branch feature in the time dimension, and generate the time weight sequence corresponding to each branch according to the phase matching degree; S413: The time-slice weighted adjustment of the temporal features of each branch is performed based on the time weight sequence, and the weighted features of each branch are normalized and fused to obtain the fused features after time phase alignment. In step 4, when performing multi-granularity voting classification based on the phase-aligned features to generate the final human action recognition result, the following sub-steps are included: S414: After the time phase is aligned, the fused features are classified at fine, medium and coarse granular levels to obtain classification outputs at three granularities. S415: Input the classification output features of the three granularities into the linear mapping layer, transform them into voting weight vectors and normalize them to obtain the adaptive voting weights for each granularity; S416: The classification probability vectors of the three granularities are weighted and summed with their corresponding voting weights to generate the final action classification probability distribution. The category with the highest probability value is selected as the human action recognition result.

[0026] In the above embodiments, the method collects multi-channel temporal signals of the limbs through fabric sensors and performs standardized processing in specific implementation scenarios. This effectively filters out noise and unifies signal amplitude. In practical scenarios such as smart wearables, rehabilitation medicine, and sports health monitoring, it can adaptively match the movement habits and body characteristics of different users, accurately capture key details of movements, effectively combine local movement features with global movement trends, and improve recognition stability through temporal alignment and multi-dimensional discrimination. It can accurately distinguish complex, continuous, and subtle movements, reduce dependence on high-precision sensing devices, adapt to daily wearable usage scenarios, and further improve the accuracy and practicality of movement recognition.

[0027] It should be noted that: The fabric sensor array used in this embodiment uses flexible conductive fabric as the substrate and acquires motion signals based on the piezoresistive strain sensing principle. The sensor unit thickness is 0.1~0.3mm, the response time is no more than 50ms, and the operating voltage is 3.3V. The sensor array is arranged according to a preset layout, covering three key motion parts of the upper limbs: shoulder, elbow, and wrist, and three key motion parts of the lower limbs: hip, knee, and ankle. 2~4 sensor units are arranged for each motion part, forming a synchronous acquisition array of 12~24 channels. The sampling rate is set to 50~200Hz, the single-channel sampling accuracy is 16bit, and the synchronous sampling error of each channel is controlled within 1ms to ensure the synchronicity and accuracy of signal acquisition from the key motion parts of the limbs.

[0028] When denoising the original multi-channel timing electrical signal, Butterworth fourth-order low-pass filtering is used first, with the cutoff frequency set to 10Hz to effectively filter out motion interference and environmental electromagnetic interference. At the same time, a median filter with a window size of 3×1 is superimposed to eliminate impulse noise in the signal. The amplitude normalization adopts the min-max normalization method, which linearly maps the amplitude of the filtered signal to the [0,1] interval. The normalization process eliminates the influence of individual sensor sensitivity differences and signal amplitude fluctuations, resulting in a stable and reliable standardized timing input.

[0029] The standardized timing input time length T is set to 100~500 time steps, corresponding to an actual sampling duration of 2~10s. The time span parameter range of the morphological primitive is set to... , The preferred number of morphological primitives, K, is 8, and the feature dimension D is consistent with the number of sensor acquisition channels. The length seed is used to generate time span parameters via a boundary-constrained Sigmoid mapping. The mapping formula is Where Wlen and blen are trainable parameters; the morphological primitive vector is generated by a trainable mapping network consisting of two fully connected perceptrons. The first layer has 2D neurons and uses the ReLU activation function, while the second layer has D neurons and uses the linear activation function to ensure that the output morphological primitive vector is suitable for subsequent feature extraction requirements.

[0030] When performing differentiable time expansion on morphological primitive vectors, a continuous-time linear mapping method is adopted to map morphological primitive features of different time spans to a unified standard time axis. This is combined with bilinear interpolation to complete feature interpolation processing, so that all morphological primitives generate primitive time series sequences consistent with the standardized time series input time length T, maintaining differentiability throughout the process and meeting the requirements of gradient backpropagation during model training.

[0031] The grouped network module consists of 6-12 parallel trainable temporal feature extraction units. All units are evenly divided into a morphological primitive feature-specific network group and a deep feature-specific network group, with each group containing 3-6 feature extraction units. The temporal feature extraction units adopt a structure combining a temporal convolutional network (TCN) and a bidirectional LSTM. The dilation coefficients of the temporal convolutional network are set to 1, 2, 4, and 8, respectively, covering temporal receptive fields at different scales: local, intermediate, and global. The number of neurons in the hidden layer of the bidirectional LSTM is D / 2, used to model long-term temporal action dependencies, enabling the two network groups to generate more complementary temporal feature representations.

[0032] The module gating mechanism generates branch weights by globally averaging and concatenating morphological primitive branch features and deep branch features, followed by a linear mapping layer and a Softmax activation function. The feature quality gating mechanism generates quality weights in the 0-1 range by calculating the time-dimensional statistical variance of the input features, followed by a linear mapping layer and a Sigmoid activation function. The branch weights and quality weights are multiplicatively combined to obtain the comprehensive activation weights. A preset threshold of 0.1 is set; network modules below this threshold are directly suppressed. The retained module weights are normalized and used for feature fusion. The weight coefficients λ1, λ2, and λ3 are fixed at 1.0, 0.5, and 0.3, respectively, to stably control the impact of feature energy, dynamic change, and cross-branch consistency on the weights.

[0033] During temporal phase alignment, the phase description vector of the fused temporal features is used as a reference. The cosine distance between the phase description vectors of the morphological primitive branch features and the deep branch features and the reference vector is calculated. The phase offset threshold is set to 0.2. Time segments with a cosine distance greater than 0.2 are considered to have large phase offsets and their weights are reduced. A temporal weight sequence is generated based on the phase matching degree. The weight coefficient for key action stages is 1.0~1.5, and the weight coefficient for phase offset segments is 0.2~0.5. Temporal phase alignment of each branch feature is achieved through weighted adjustment on a time-slice basis.

[0034] In multi-granularity voting classification, fine-granularity independently linearly maps the features at each time step to output the classification result. Medium-granularity divides the temporal features into 5-10 equal-length feature blocks, generates the classification output through global max pooling and linear mapping, and coarse-granularity extracts global features through global average pooling to complete the classification. The classification outputs of the three granularities are then normalized by linear mapping and Softmax to obtain adaptive voting weights. The weighted sum of these weights yields the final classification probability distribution, and the category with the highest probability is selected as the action recognition result.

[0035] The model training uses the cross-entropy loss function to adapt to the action classification task. The optimizer is Adam, the initial learning rate is set to 1e-3, the batch size is 16~32, and the training dataset is a public fabric-sensing action dataset containing 20 categories of daily actions such as walking, running, bending over, and raising hands. The number of training rounds is 50~100, and the validation set ratio is set to 20% to ensure the convergence and generalization ability of the model training.

[0036] In summary, the method in this embodiment uses a fabric sensor to collect multi-channel temporal signals from the limbs and performs standardized processing, effectively eliminating noise interference and signal amplitude fluctuations, improving data purity and consistency. Through sample-adaptive feature extraction, it dynamically adapts to different action samples and individual body feature differences, accurately capturing key details of the action, avoiding information loss caused by fixed feature extraction strategies, organically fusing two types of complementary features, enhancing the discriminative ability and expressive effect of the features, solving the problem of feature temporal misalignment through adaptive alignment in the time dimension, and integrating classification results through multi-granularity voting, significantly improving the stability and anti-interference ability of action recognition, adapting to the recognition needs of subtle, continuous, and complex actions, and fully conforming to the signal characteristics of the fabric sensor and the usage characteristics of wearable scenarios, reducing the dependence on high-precision sensing equipment, effectively improving the accuracy and reliability of human action recognition, and expanding the practical value of the technology in fields such as smart wearables, rehabilitation medicine, and sports health monitoring.

[0037] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for human motion recognition based on fabric sensors, characterized in that, include: Step 1: Collect raw multi-channel time-series electrical signals of the target object's limb movement parts based on the fabric sensor array, preprocess the raw multi-channel time-series electrical signals, and construct a standardized time-series input; Step 2: Perform feature encoding on the standardized temporal input to generate morphological primitive seed vectors. Based on the morphological primitive seed vectors, dynamically generate multiple morphological primitives adapted to the current sample and the validity mask corresponding to each morphological primitive. After time expansion and weighted aggregation, obtain the sample adaptive morphological primitive temporal feature sequence. Step 3: Extract morphological primitive branch features based on the temporal feature sequence of adaptive morphological primitives, and extract deep branch features based on the standardized temporal input. Perform weighted fusion of morphological primitive branch features and deep branch features to generate fused temporal features. Step 4: Perform temporal phase alignment on the morphological primitive branch features, deep branch features, and fused temporal features. Based on the phase-aligned features, perform multi-granularity voting classification to generate the final human action recognition result.

2. The human motion recognition method based on a fabric sensor according to claim 1, characterized in that, Step 2 involves feature encoding of the standardized temporal input to generate morphological primitive seed vectors, specifically including the following sub-steps: S211: Perform adaptive average pooling along the time dimension on the standardized time-series input X to obtain a pooled feature of fixed length. ; S212: The pooling feature... After being flattened into a one-dimensional vector, a linear mapping is performed to obtain a morphological primitive seed vector z. The morphological primitive seed vector z is then divided into length seeds according to a preset dimension. With characteristic seeds ; S213: For the length seed With characteristic seeds Normalization is performed to stabilize the distribution of the seed vector.

3. The human motion recognition method based on a fabric sensor according to claim 2, characterized in that, Step 2 involves dynamically generating multiple morphological primitives adapted to the current sample and a validity mask for each morphological primitive based on the morphological primitive seed vector. This includes the following sub-steps: S214: Normalize the length seed The time span parameter of the i-th morphological primitive is generated through nonlinear constraint mapping. According to the time span parameter With normalized feature seeds The i-th morphological primitive vector is generated through a trainable mapping network. ,in K is the preset number of morphological primitives, and D is the feature dimension; S215: Generate a validity mask for each morphological primitive, the formula for which is as follows: ； In the formula: For validity mask; For activation functions; These are the weight parameters for the linear mapping. is the seed mapping function corresponding to the i-th morphological primitive; These are the bias parameters for the linear mapping. and All of these are trainable parameters.

4. The human motion recognition method based on a fabric sensor according to claim 3, characterized in that, In step 2, the temporal feature sequence of adaptive morphological primitives of the samples is obtained through time expansion and weighted aggregation, including the following sub-steps: S216: The i-th morphological primitive vector According to its time span parameter By performing differentiable time expansion and using continuous-time mapping and differentiable interpolation, the morphological primitive features are mapped onto a unified standard time axis, resulting in a primitive time series. ; S217: Based on Validity Mask Weighted aggregation of K primitive time series sequences yields the sample adaptive morphological primitive time series feature sequence. Its formula is: ； In the formula: For discrete-time indexing; For primitive time series In the eigenvectors at time step t, ε is a minimal constant to prevent the denominator from being zero.

5. The human motion recognition method based on a fabric sensor according to claim 1, characterized in that, Step 3 includes the following sub-steps: S311: Extract morphological primitive branch features based on the temporal feature sequence of the adaptive morphological primitives of the sample, and extract deep branch features based on the standardized temporal input; S312: Input the morphological primitive branch features and the deep branch features into the group network module for feature processing to obtain the output features corresponding to each network module; S313: Based on the morphological primitive branch features, the deep branch features, and the input features of each network module, determine the comprehensive activation weight of each network module, and normalize the comprehensive activation weight. S314: The output features of each network module are weighted and aggregated based on the normalized comprehensive activation weights, and then fused with the morphological primitive branch features and the deep branch features to obtain fused temporal features.

6. The human motion recognition method based on a fabric sensor according to claim 5, characterized in that, In step 312, the grouped network module includes several parallel trainable temporal feature extraction network modules, and all network modules are divided into a morphological primitive feature-specific network group and a deep feature-specific network group. Among them, the morphological primitive feature-specific network group receives and processes morphological primitive branch features, and the depth feature-specific network group receives and processes depth branch features. Different network modules correspond to different temporal modeling scales, receptive field sizes, or attention ranges.

7. The human motion recognition method based on a fabric sensor according to claim 5, characterized in that, Step 313 specifically includes: S3131: Construct a module gating mechanism, perform global average pooling on the morphological primitive branch features and the deep branch features in the time dimension, extract the global statistical feature vectors of the two branches and concatenate them, input the concatenated features into the gating mapping function composed of a trainable linear mapping layer and a Softmax activation function, and generate the morphological primitive branch weights and the deep feature branch weights. S3132: Construct a feature quality gating mechanism, calculate the statistical variance of the input features of each network module in the time dimension, input the statistical variance into the quality mapping function composed of a trainable linear mapping layer and a Sigmoid activation function, and generate the quality weights of the corresponding network modules. S3133: Multiplicatively combine the quality weights of each network module with the branch weights of their respective branches to obtain the comprehensive activation weights of each network module; suppress network modules with comprehensive activation weights below a preset threshold, and normalize the comprehensive activation weights of the retained network modules to obtain the final normalized activation weights used for feature fusion, the formula of which is: ； In the formula: represents the normalized activation weights corresponding to the i-th network module; k is the total number of retained network modules; exp is the exponential function; The preset weight coefficients control the impact of feature energy, feature dynamic change degree, and cross-branch feature consistency on module weights, respectively. Input the average feature vector of the features in the time dimension to the i-th network module; for The L2 norm; Input the variance statistics of the features in the time dimension for the i-th network module; It is the global average feature vector of the morphological primitive branch features; This represents the global average feature vector of the deep branch features; To prevent extremely small constants with a denominator of 0; This indicates a summation operation performed on all network modules involved in the fusion computation.

8. The human motion recognition method based on a fabric sensor according to claim 5, characterized in that, In step 314, temporal features are fused. The calculation formula is: ； in, For feature normalization operators; The temporal characteristics output by the i-th retained network module; The branching characteristics of morphological primitives; The branch weight of the morphological primitive branch feature at time step t; This is a deep branch feature; The branch weights of the deep branch features at time step t.

9. The human motion recognition method based on a fabric sensor according to claim 1, characterized in that, Step 4, which involves temporal phase alignment of the morphological primitive branch features, deep branch features, and fused temporal features, includes the following sub-steps: S411: Unify the morphological primitive branch features, deep branch features, and fused temporal features into a temporal feature sequence with consistent time length. Perform temporal convolution and nonlinear mapping on each temporal feature sequence to extract the phase structure information of each branch and generate the corresponding phase description vector. S412: Based on the phase description vector of each branch, calculate the phase matching degree of the features of each branch in the time dimension, and generate the time weight sequence corresponding to each branch according to the phase matching degree; S413: Based on the time weight sequence, the time-slice weighted adjustment of each branch's time-series features is performed, and the weighted branch features are normalized and fused to obtain the fused features after time phase alignment.

10. The human motion recognition method based on a fabric sensor according to claim 1, characterized in that, In step 4, when performing multi-granularity voting classification based on the phase-aligned features to generate the final human action recognition result, the following sub-steps are included: S414: The fused features after time phase alignment are classified into fine-grained, medium-grained, and coarse-grained categories to obtain classification outputs of three granularities. S415: Input the classification output features of the three granularities into the linear mapping layer, transform them into voting weight vectors and normalize them to obtain the adaptive voting weights for each granularity; S416: The classification probability vectors of the three granularities are weighted and summed with their corresponding voting weights to generate the final action classification probability distribution. The category with the highest probability value is selected as the human action recognition result.

Images