An intelligent gesture intention recognition method based on multi-task learning

Abstract

The application discloses an intelligent gesture intention recognition method based on multi-task learning, comprising the following steps: 1, collecting surface electromyography and inertial data, and performing data preprocessing; 2, performing feature extraction on sEMG signal samples and IMU signal samples respectively; 3, constructing a multi-task learning network based on sEMG and IMU data fusion; 4, training the multi-task learning network to recognize forearm posture and gesture task. The application can fuse heterogeneous sensing data, utilize the correlation between gesture task and forearm posture task, realize high-precision gesture recognition and posture recognition, and further realize complex gesture intention recognition.

Description

Technical Field

[0001] This invention belongs to the field of biosignal processing, specifically a method for intelligent gesture intent recognition based on multi-task learning. Background Technology

[0002] Gestures, as an important way for humans to interact with their environment, have led to the widespread application of gesture recognition technology in fields such as automotive electronics, virtual reality games, mobile offices, and smart homes. Electromyography (EMG) wristbands can acquire high-quality sEMG signals for high-precision gesture recognition, and their high portability makes them an ideal choice for next-generation gesture interaction devices. Commercially available EMG wristbands already have mature design paradigms, typically integrating an IMU (Integrated Mutual Memory Unit) to meet signal acquisition needs at low cost.

[0003] While surface electromyography (sEMG)-based gesture recognition can achieve high accuracy under ideal experimental conditions, it still faces numerous challenges in practical applications, such as changes in limb posture. When limb posture changes, muscle fibers shift relative to the electrodes, and activation of muscles related to limb stability or movement reduces the reliability of recognition. Previous research has addressed the impact of limb posture by proposing solutions based on the fusion of IMU and sEMG data. These solutions include single-stage classifiers that concatenate IMU and sEMG feature vectors, and cascaded classifiers that first use IMU for position classification and then perform gesture recognition on a specific position classifier. Both methods effectively reduce the impact of arm position changes on gesture recognition. These multi-sensor fusion schemes for addressing limb posture changes treat limb posture as a disturbance variable and use IMU as supplementary information to improve the robustness of EMG pattern recognition.

[0004] However, treating body posture as a distractor is one-sided, because the same hand gesture can correspond to different intentions under different body postures. For example, performing the same wrist flexion gesture with the right hand is often interpreted as a leftward intention when the forearm is facing forward and the wrist is in a neutral position, while it is usually understood as a downward intention when the wrist is in an internal rotation position. Therefore, a solution is needed that simultaneously recognizes hand movements and wrist and forearm positions to achieve intelligent and accurate recognition of complex hand gesture intentions. Summary of the Invention

[0005] The present invention aims to address the shortcomings of the existing technology by proposing an intelligent gesture intention recognition method based on multi-task learning. This method combines information on different aspects of the same gesture extracted from sEMG and IMU data, and makes reasonable use of the correlation between gesture recognition and posture recognition tasks. This enables high-precision gesture and posture recognition, and thus accurately recognizes complex gesture intentions.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] The present invention provides an intelligent gesture intent recognition method based on multi-task learning, characterized by the following steps:

[0008] Step 1: Use an electromyography (EMG) wristband to collect surface EMG signals and inertial data of a user performing N hand gestures in M ​​forearm postures at the wrist.

[0009] The acquired K-channel surface electromyography signals were subjected to pairwise differences between channels to form K×K channel sEMG data; the inertial data of the C channel were linearly interpolated to make the number of sampling points equal to that of the sEMG data, thus obtaining the interpolated IMU data.

[0010] Active-segment sliding window segmentation is performed on both the sEMG data and the interpolated IMU data to obtain an sEMG signal sample set S containing D1 sEMG signal samples. raw With the IMU signal sample set I containing D1 IMU signal samples raw ;

[0011] Step 2: Extract features from the sEMG signal samples and IMU signal samples respectively to obtain the sEMG signal training set S and the IMU signal training set I, thereby constructing a training set {S, I, Y} containing multimodal signal feature map samples. g ,Y p}, where Y g Y represents the set of gesture category labels for signal feature map samples in S and I. p Let Y represent the set of pose category labels for the signal feature map samples in S and I, and let Y be the set of pose category labels for the samples in S and I. g Includes N gesture categories, Y p It includes M pose categories;

[0012] Step 3: Construct a multi-task learning network based on the fusion of sEMG and IMU data, including: a spatial module, two temporal modules, a feature cross module, an attention module, and a classification module, and process the i-th sEMG feature map sample S. i and the i-th IMU feature map sample I i The data is processed to obtain the classification probability value of the gesture task. Classification probability values ​​for pose tasks

[0013] Step 4: Based on the training database and the corresponding gesture and posture labels, the multi-task learning network is trained using gradient descent, and the total loss function L for task learning is calculated. The network parameters are updated by backpropagation until the maximum number of training iterations is reached or L converges, thereby obtaining a robust multi-task learning model for predicting gesture and posture categories from surface electromyography signals and inertial data.

[0014] The intelligent gesture intent recognition method based on multi-task learning described in this invention is also characterized in that step 2 includes:

[0015] Step 2.1: Divide each sEMG signal sample into H frames on average, and extract the sEMG signal sample set S. raw From the r temporal features of each frame of each signal sample, we obtain the sEMG signal training set S = {S} containing D1 sEMG signal feature map samples. i |i=1,2,…D1}, where S i ∈R H×K×K×r Let S represent any i-th sEMG signal feature map sample. i The gesture category label is Its posture category label is K represents the number of channels, r represents the number of temporal features, H represents the number of frames, and D1 represents the number of samples;

[0016] Step 2.2: Divide each IMU signal sample into H frames of IMU signals on average, and calculate the average absolute value of each frame of IMU signals to obtain an IMU signal training set I = {I_1} containing D1 IMU signal feature map samples. i |i=1,2,…D1}, where, I i ∈R H×C Let I represent any i-th IMU signal feature map sample. i The gesture category label is Its posture category label is C represents the number of channels.

[0017] Step 3 includes:

[0018] Step 3.1: The spatial module includes: several convolutional layers, each convolutional layer is followed by a batch normalization layer, and each convolutional layer uses b two-dimensional convolutional kernels of size e×e and a sigmoid activation function; a Flatten layer is connected after the last batch normalization layer;

[0019] The i-th sEMG feature map sample S i After being processed by the spatial module, two-dimensional sEMG spatial feature samples are obtained. b represents the number of two-dimensional convolution kernels;

[0020] Step 3.2: Each temporal module includes: a bidirectional gated recurrent unit layer and a Dropout layer, wherein the number of neurons in the bidirectional gated recurrent unit layer is J;

[0021] The i-th sEMG feature map sample and the i-th IMU feature map sample I i Each input is processed in a time series module, and the corresponding sEMG depth features are obtained. and IMU depth features

[0022] Step 3.3: The feature cross-validation module uses equations (1) and (2) to obtain the learning gesture task features. and posture task features

[0023]

[0024]

[0025] In equations (1) and (2), α g The proportionality coefficient α represents the feature of the gesture task. p The scaling factor representing the posture task characteristics is obtained from equation (3), α. c The proportionality coefficient representing shared features is a parameter to be learned, α. c initial value α c,initial It is obtained from equation (4) and updated through backpropagation:

[0026]

[0027]

[0028] In equations (3) and (4), D(·) represents the Euclidean distance; sigmoid represents the activation function.

[0029] Step 3.4: The posture task features Input the attention module and obtain the temporal attention weight map O∈R according to equations (6)-(8). H Then, the temporal attention weight map O is combined with the gesture task features. The features from each time channel are weighted and fused to obtain the gesture task fusion features.

[0030]

[0031]

[0032]

[0033] In equations (6)-(8), express The j-th feature in the h-th frame; z i,h express The average feature of the h-th frame, z i,h+1 express The average feature of the (h+1)th frame, z i,h-1 express The average features of the (h-1)th frame, express The temporal variation amplitude of the features, among which, express The similarity between the features of the h-th frame and the features of adjacent frames, W∈R H×H For a parameter to be learned;

[0034] Step 3.5: Gesture Recognition Task Features Features of forearm pose recognition task The input to the classification module is flattened by a Flatten layer, then processed by a fully connected layer and a SoftMax activation layer to obtain the classification probability value for the gesture task. Classification probability values ​​for pose tasks

[0035] The total loss function L in step 4 is established according to equations (9)-(11);

[0036]

[0037]

[0038] L = L g +L p (11)

[0039] In equations (9)-(11), L g and L p These are the cross-entropy losses for gesture tasks and posture tasks, respectively. S represents i and I i The gesture label is the label value of category n. S represents the prediction of a multi-task learning network. i and I i The probability value of the gesture label as category n; S represents i and I i The pose label is the label value of category m. S represents the prediction of a multi-task learning network. i and I i The pose label is the probability value of category m.

[0040] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the intelligent gesture intent recognition method, and the processor is configured to execute the program stored in the memory.

[0041] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program is executed by a processor to perform the steps of the intelligent gesture intent recognition method.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] 1. In response to the fact that most research on electromyography pattern recognition only focuses on hand shape recognition and ignores the different hand gesture intentions reflected under different wrist directions and forearm positions, this invention designs a multi-task learning neural network that can simultaneously realize gesture recognition and posture recognition, and has achieved a certain level of accuracy.

[0044] 2. This invention proposes a novel fusion scheme for IMU and sEMG. Although both emphasize different aspects of gesture motion characterization—IMU focusing on posture and trajectory, and sEMG focusing on delicate finger movements, hand shape, and sustained muscle exertion—this invention argues that different tasks (finger and wrist movements, hand shape recognition, and arm orientation posture recognition) are related and interdependent. Through interrelated characterization, dynamic and complex gestures can be more accurately characterized and recognized. Simultaneously recognizing hand shape and posture helps the model accurately understand the intent of the gesture, improving recognition accuracy and enhancing the model's robustness under different postures.

[0045] 3. This invention proposes an intelligent gesture intent understanding method based on multi-task learning, simultaneously recognizing gesture and posture tasks. The multi-task learning network model uses Euclidean distance and backpropagation to determine the optimal inter-task shared features and task-specific features, feeding each feature into the network corresponding to the task. Furthermore, this network model uses a cross-task attention mechanism to understand the temporal correlation between different tasks and encodes this correlation into the design of the multi-task model. By jointly learning the two tasks, information is shared between them, improving the performance of both tasks and proposing a scheme for recognizing complex gesture intentions. This scheme achieves high-precision gesture and posture recognition and can also be seen as improving the robustness of gesture recognition tasks against various forearm posture changes, overcoming the shortcomings of existing gesture recognition schemes. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the method flow provided by the present invention;

[0047] Figure 2 This is a schematic diagram of hand gestures and forearm postures provided by the present invention;

[0048] Figure 3 This is a schematic diagram illustrating the gesture recognition accuracy of the comparison method provided by the present invention;

[0049] Figure 4 This is a schematic diagram illustrating the posture recognition accuracy of the comparison method provided by the present invention. Detailed Implementation

[0050] In this example, an intelligent gesture intent recognition method based on multi-task learning is described, such as... Figure 1 As shown, the procedure is as follows:

[0051] Step 1: Use an electromyography (EMG) wristband to collect surface electromyogram (sEMG) signals and inertial measurement unit (IMU) data of a user performing N hand gestures in M ​​forearm postures at the wrist.

[0052] In this embodiment, the electromyography (EMG) wristband device is equipped with an inertial sensing unit. The EMG wristband is naturally worn on the wrist of the subject's dominant hand to collect data. The experiment designed a total of 6 different forearm postures and 7 gesture patterns. For forearm postures, three forearm positions (arm perpendicular to the ground, elbow bent at 90°, elbow bent at 45°) and two wrist directions (wrist internal rotation and wrist neutral) were considered. These three forearm positions and two wrist directions were combined in pairs to form 6 forearm postures, such as... Figure 2 As shown in section (a) of the diagram. The seven gesture patterns include extending the index finger and swiping left, extending the index finger and swiping right, flexing the wrist, extending the wrist, pinching, opening the hand, and making a fist, as shown below. Figure 2 As shown in part (b) of the document.

[0053] like Figure 2 As shown, the subjects were instructed to perform the target gesture naturally in different postures, and the pattern was repeated 15 times for each gesture.

[0054] The acquired K-channel surface electromyography signals were subjected to pairwise differences between channels to form K×K channel sEMG data; the inertial data of the C channel were linearly interpolated to make the number of sampling points equal to that of the sEMG data, thus obtaining the interpolated IMU data.

[0055] In this embodiment, the electromyography (sEMG) wristband can acquire 8 channels of surface electromyography (sEMG) signals and 9 channels of inertial data. The inertial data includes triaxial accelerometer data, triaxial gyroscope data, and Euler angle data. A single-difference operator is applied to any two channels of the 8-channel sEMG signals to form 8×8 channel sEMG data. The sampling frequency of the inertial data is 200Hz, and the sampling frequency of the sEMG signals is 600Hz. Linear interpolation is performed on the inertial data to make its sampling number equal to that of the sEMG data.

[0056] Active-segment sliding window segmentation is performed on both the sEMG data and the interpolated IMU data to obtain an sEMG signal sample set S containing D1 sEMG signal samples. raw With the IMU signal sample set I containing D1 IMU signal samples raw ;

[0057] In this embodiment, a sliding window with a window length of 200ms and a step size of 80ms is used to segment the signal. The average amplitude of each window is calculated, and the threshold for determining the active window is the average amplitude of all channels of the resting data plus three times the standard deviation of the fluctuation.

[0058] Step 2: Extract features from sEMG signal samples and IMU signal samples respectively;

[0059] Step 2.1: Divide each sEMG signal sample into H frames on average, and extract the sEMG signal sample set S. raw From the r temporal features of each frame of each signal sample, we obtain the sEMG signal training set S = {S} containing D1 sEMG signal feature map samples. i |i=1,2,…D1}, where S i ∈R H×K×K×r Let S represent any i-th sEMG signal feature map sample. i The gesture category label is Its posture category label is K represents the number of channels, r represents the number of temporal features, H represents the number of frames, and D1 represents the number of samples;

[0060] In this embodiment, each sample is divided into 10 frames on average. Four time-domain features (mean absolute value, zero crossing point, number of slope sign changes, and waveform length) are calculated for each frame of the sEMG signal sample.

[0061] Step 2.2: Divide each IMU signal sample into H frames of IMU signals on average, and calculate the average absolute value of each frame of IMU signals to obtain an IMU signal training set I = {I_1} containing D1 IMU signal feature map samples. i |i=1,2,…D1}, where, I i ∈RH×C Let I represent any i-th IMU signal feature map sample. i The gesture category label is Its posture category label is C represents the number of channels.

[0062] Step 2.3: Construct a training set {S,I,Y} containing multimodal signal feature map samples. g ,Y p}, where Y g Y represents the set of gesture category labels for signal feature map samples in S and I. p Let Y represent the set of pose category labels for the signal feature map samples in S and I, and let Y be the set of labels for the pose categories of the samples in S and I. g Includes N gesture categories, Y p It includes M pose categories.

[0063] Step 3: Construct a multi-task learning network based on the fusion of sEMG and IMU data, which includes: a spatial module, two temporal modules, a feature cross module, an attention module, and a classification module;

[0064] Step 3.1: The spatial module includes: several convolutional layers, each convolutional layer is followed by a batch normalization layer, and each convolutional layer uses b two-dimensional convolutional kernels of size e×e and a sigmoid activation function; a Flatten layer is connected after the last batch normalization layer;

[0065] The i-th sEMG feature map sample S i After being processed by the spatial module, two-dimensional sEMG spatial feature samples are obtained. b represents the number of two-dimensional convolution kernels;

[0066] In this embodiment, the spatial module contains two convolutional layers, each with eight 2×2 convolutional kernels. After operation by the spatial module, each 10×8×8×4 sEMG feature map sample is converted into a 10×512 two-dimensional sEMG spatial feature sample.

[0067] Step 3.2: Each temporal module includes: a bidirectional gated recurrent unit layer and a Dropout layer, wherein the number of neurons in the bidirectional gated recurrent unit layer is J;

[0068] The i-th sEMG feature map sample and the i-th IMU feature map sample I i Each input is processed in a time-series module, and the corresponding sEMG depth features are obtained. and IMU depth features

[0069] In this embodiment, the number of neurons in the bidirectional gated recurrent unit layer of the timing module is set to 16, and the ratio of the Dropout layer is set to 0.5. After the operation of the timing module, the size of each sEMG and IMU feature map sample is converted to 10×32.

[0070] Step 3.3: The feature cross-validation module uses equations (1) and (2) to obtain the learning gesture task features. and posture task features

[0071]

[0072]

[0073] In equations (1) and (2), α g The proportionality coefficient α represents the feature of the gesture task. p The scaling factor representing the posture task characteristics is obtained from equation (3), α. c The proportionality coefficient representing shared features is a parameter to be learned, α. c initial value α c,initial It is obtained from equation (4) and updated through backpropagation:

[0074]

[0075]

[0076] In equations (3) and (4), D(·) represents the Euclidean distance; sigmoid represents the activation function.

[0077] Step 3.4: The posture task features Input the attention module and obtain the temporal attention weight map O∈R according to equations (6)-(8). H Then, the temporal attention weight map O is combined with the gesture task features. The features from each time channel are weighted and fused to obtain the gesture task fusion features.

[0078]

[0079]

[0080]

[0081] In equations (6)-(8), express The j-th feature in the h-th frame; z i,h express The average feature of the h-th frame, zi,h+1 express The average feature of the (h+1)th frame, z i,h-1 express The average features of the (h-1)th frame, express The temporal variation amplitude of the features, among which, express The similarity between the features of the h-th frame and the features of adjacent frames, W∈R H×H This is a parameter to be learned.

[0082] Step 3.5: Gesture Recognition Task Features Features of forearm pose recognition task The input to the classification module is flattened by a Flatten layer, then processed by a fully connected layer and a SoftMax activation layer to obtain the classification probability value for the gesture task. Classification probability values ​​for pose tasks

[0083] Step 4: Training the network to recognize forearm poses and gesture tasks;

[0084] Step 4.1: Establish the total task learning loss function L according to equations (9)-(11);

[0085]

[0086]

[0087] L = L g +L p (11)

[0088] In equations (9)-(11), L g and L p These are the cross-entropy losses for gesture tasks and posture tasks, respectively. S represents i and I i The gesture label is the label value of category n. S represents the prediction of a multi-task learning network. i and I i The probability value of the gesture label as category n; S represents i and I i The pose label is the label value of category m. S represents the prediction of a multi-task learning network. i and I i The pose label is the probability value of category m.

[0089] Step 4.2: Based on the training database and the corresponding gesture and posture labels, the multi-task learning network is trained using gradient descent, and the total loss function L for task learning is calculated. The network parameters are updated by backpropagation until the maximum number of training iterations is reached or L converges, thereby obtaining a robust multi-task learning model for predicting gesture and posture categories from surface electromyography signals and inertial data.

[0090] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0091] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0092] The given test data is fed into the trained model to test the performance of the multi-task learning neural network. Furthermore, to quantify the effectiveness of this invention, a task-independently trained model and two commonly used multi-task learning models were compared with the above-described scheme of this invention. In the comparative experiments, single-task (gesture) and single-task (pose) methods used independent training methods for gesture / pose recognition tasks using sEMG and IMU fusion; hard parameter sharing was a multi-task learning method that directly shares model layers; and soft parameter sharing was a multi-task learning method based on L2 constraints. Figure 3 The gesture recognition accuracy results of the above four methods are as follows: the gesture recognition accuracy of the three comparative methods are (89.48±4.15)%, (90.91±3.62)%, and (91.09±4.11)%, respectively, and the gesture recognition accuracy of the proposed method is (92.95±4.05)%. Figure 4 The pose recognition accuracy results of the above four methods are as follows: the pose recognition accuracy of the three comparative methods are (96.04±4.10)%, (98.10±2.28)%, and (98.13±2.87)%, respectively, and the pose recognition accuracy of the proposed method is (98.51±2.50)%. Figure 3 and Figure 4 S1-S10 in the diagram represent the subjects. The results show that this invention achieves higher accuracy in gesture and posture recognition compared to other methods.

[0093] In summary, this invention integrates spatiotemporal information from deep sEMG and IMU to characterize the correlation between gesture and posture tasks. Through joint learning of these two tasks, high-precision gesture and posture recognition results are obtained, thereby achieving intelligent gesture intent recognition. The method proposed in this invention is easy to implement, highly portable, and of great significance for fields such as electromyographic gesture interaction.

Claims

1. A method for intelligent gesture intent recognition based on multi-task learning, characterized in that, The procedure is as follows: Step 1: Use an electromyography (EMG) wristband to collect surface EMG signals and inertial data of a user performing N hand gestures in M ​​forearm postures at the wrist. The acquired K-channel surface electromyography signals were subjected to pairwise differences between channels to form K×K channel sEMG data; the inertial data of the C channel were linearly interpolated to make the number of sampling points equal to that of the sEMG data, thus obtaining the interpolated IMU data. Active segment sliding window segmentation was performed on both the sEMG data and the interpolated IMU data to obtain data containing... sEMG signal sample set of sEMG signal samples With inclusion IMU signal sample set of IMU signal samples ; Step 2: Extract features from the sEMG signal samples and IMU signal samples respectively to obtain the sEMG signal training set. and IMU signal training set Thus, a training set containing multimodal signal feature map samples is constructed. },in, express The set of gesture category labels for signal feature map samples in I, express The set of pose category labels for the signal feature map samples in I, and It includes N types of gestures. It includes M pose categories; Step 3: Construct a multi-task learning network based on the fusion of sEMG and IMU data, including: a spatial module, two temporal modules, a feature cross module, an attention module, and a classification module, and process the i-th sEMG feature map sample. and the i-th IMU feature map sample The data is processed to obtain the classification probability value of the gesture task. Classification probability values ​​for pose tasks ; Step 3.1: The spatial module includes: several convolutional layers, each convolutional layer is followed by a batch normalization layer, and each convolutional layer uses b two-dimensional convolutional kernels of size e×e and a sigmoid activation function; a Flatten layer is connected after the last batch normalization layer; The i-th sEMG feature map sample After being processed by the spatial module, two-dimensional sEMG spatial feature samples are obtained. b represents the number of two-dimensional convolution kernels; Step 3.2: Each temporal module includes: a bidirectional gated recurrent unit layer and a Dropout layer, wherein the number of neurons in the bidirectional gated recurrent unit layer is J; The i-th sEMG feature map sample and the i-th IMU feature map sample Each input is processed in a time-series module, and the corresponding sEMG depth features are obtained. and IMU depth features ; Step 3.3: The feature cross-validation module uses equations (1) and (2) to obtain the learning gesture task features. and posture task features : (1) (2) In equations (1) and (2), The scaling factor representing the characteristics of the gesture task. The scaling factor represents the characteristics of the posture task and is obtained from equation (3). The proportionality coefficient representing the shared features is a parameter to be learned. initial value It is obtained from equation (4) and updated through backpropagation: = = (3) (4) In equations (3) and (4), Indicates Euclidean distance; Indicates the activation function; Step 3.4: The posture task features Input the attention module and obtain the temporal attention weight map according to equations (6)-(8). Then, the temporal attention weight map Features of gesture tasks The features from each time channel are weighted and fused to obtain the gesture task fusion features. ; (6) (7) (8) In equations (6)-(8), express The j-th feature in the h-th frame; express The average features of the h-th frame, express The average features of the (h+1)th frame, express The average features of the (h-1)th frame, ,in, express The similarity between the features of the h-th frame and the features of adjacent frames. For a parameter to be learned; Step 3.5: Gesture Recognition Task Features Features of forearm pose recognition task The input to the classification module is flattened by a Flatten layer, then processed by a fully connected layer and a SoftMax activation layer to obtain the classification probability value for the gesture task. Classification probability values ​​for pose tasks ; Step 4: Based on the training database and the corresponding gesture and posture labels, the multi-task learning network is trained using gradient descent, and the total loss function L for task learning is calculated. The network parameters are updated by backpropagation until the maximum number of training iterations is reached or L converges, thereby obtaining a robust multi-task learning model for predicting gesture and posture categories from surface electromyography signals and inertial data.

2. The intelligent gesture intent recognition method based on multi-task learning according to claim 1, characterized in that, Step 2 includes: Step 2.1: Divide each sEMG signal sample into H frames on an average basis and extract the sEMG signal sample set. The r temporal features of each frame of each signal sample are used to obtain the following: sEMG signal training set of sEMG signal feature map samples ,in, Let represent any i-th sEMG signal feature map sample, and let The gesture category label is Its posture category label is K represents the number of channels, r represents the number of temporal features, and H represents the number of frames. Indicates the number of samples; Step 2.2: Divide each IMU signal sample into H frames of IMU signals on an average basis, and calculate the average absolute value of each frame of IMU signals to obtain the total number of frames. IMU signal training set of IMU signal feature map samples ,in, Let represent any i-th IMU signal feature map sample, and let The gesture category label is Its posture category label is C represents the number of channels.

3. The intelligent gesture intent recognition method based on multi-task learning according to claim 1, characterized in that, The total loss function L in step 4 is established according to equations (9)-(11); (9) (10) (11) In equations (9)-(11), and These are the cross-entropy losses for gesture tasks and posture tasks, respectively. express and Gesture labels are categories The tag value, This indicates the prediction of a multi-task learning network. and Gesture labels are categories The probability value; express and The pose label is the category The tag value, This indicates the prediction of a multi-task learning network. and The pose label is the category The probability value.

4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the intelligent gesture intent recognition method according to any one of claims 1-3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, performs the steps of the intelligent gesture intent recognition method according to any one of claims 1-3.

Images