WO2023185887A1 - Model acquisition system, gesture recognition method and apparatus, device, and storage medium - Google Patents

Abstract

The present application relates to the field of electrophysiologic signal recognition, and discloses a model acquisition system, a gesture recognition method and apparatus, a device, and a storage medium. The model acquisition system comprises: an input unit 101, which is used to acquire current electromyographic data and a gesture label of a current individual; and a processing unit 102, which is configured to: acquire an initial gesture recognition model when the current individual does not have a personal gesture recognition model, wherein the initial gesture recognition model is generated by carrying out meta-learning training according to a plurality of meta-learning tasks generated on the basis of historical electromyographic data and gesture labels of a plurality of individuals, and each meta-learning task is generated on the basis of historical electromyographic data and a gesture label of the same individual, or each meta-learning task is generated in a hybrid manner on the basis of historical electromyographic data and gesture labels of similar individuals; and fine-tune the initial gesture recognition model according to the current electromyographic data and the gesture label of the current individual to obtain a current gesture recognition model. The current gesture recognition model is accurately obtained on the basis of individual data.

Description

Model acquisition system, gesture recognition method, device, equipment and storage medium

cross reference

This application is filed based on a Chinese patent application with application number “2022103417552” and a filing date of March 29, 2022, and claims the priority of the Chinese patent application. The entire content of the Chinese patent application is hereby incorporated by reference. Apply.

Technical field

Embodiments of the present application relate to the field of electrophysiological signal recognition, and in particular to a model acquisition system, gesture recognition method, device, equipment and storage medium.

Background technique

Electromyography (EMG) is the recording of electrophysiological signals obtained by discharging muscle motor neurons through electromyographic electrodes. It contains rich neural information that can be decoded into many limb-related activity signals. The collection of EMG is harmless to the human body, easy to obtain and easy to operate. It has good application prospects in the field of gesture recognition and classification, especially in medical treatment, entertainment and other industries involving machine control.

Traditional myoelectric gesture recognition uses a large amount of pre-collected data from a single individual to train the corresponding pattern recognition classifier. Commonly used classifier models include support vector machines, random forests and other models of traditional machine learning, as well as convolutional neural networks based on deep learning. , recurrent neural network and other methods to build models. At present, there are many inherent problems in the application of electromyographic signals in gesture recognition. Due to the specificity of individual muscle nerve distribution and the influence of skin impedance, muscle structure and other factors that are highly dependent on individual characteristics, there are significant individual differences in electromyographic signals. Differential,classification accuracy relies on training on a large,amount of data from a single individual. In addition, affected by factors such as electrode displacement and environment, the characteristics of myoelectric signals change very rapidly. As a result, even if a model is trained on a large amount of individual data, its accuracy will gradually decrease as the use time increases. For the above reasons, models trained based on a large amount of individual data often lead to problems such as overfitting and unstable model performance due to factors such as limited data collection and waste of prior knowledge.

In order to deal with the problem of waste of prior knowledge, the transfer learning method provides a solution, but transfer learning will face the problem of catastrophic forgetting, and in order to ensure the accuracy of the model, a large amount of individual data is still needed for training, model training and model application. Scene switching is still relatively complex, and training and adjustment efficiency is low.

Contents of the invention

The embodiments of this application aim to provide a model acquisition system, gesture recognition method, device, equipment and storage medium, aiming to use meta-learning to accurately obtain the current gesture recognition model that adapts to the current individual characteristics through a small amount of individual data. , improve the efficiency of gesture recognition model training and adjustment, and then complete the accurate recognition of the current individual gesture.

In order to solve one or more of the above problems and achieve the above objects, embodiments of the present application provide a model acquisition system, including: an input unit for acquiring the current myoelectric data and gesture tags of the current individual. The gesture tags Containing gesture information; the processing unit is configured to: obtain an initial gesture recognition model when the current individual does not have a personal gesture recognition model, wherein the initial gesture recognition model is based on historical myoelectric data and gestures based on multiple individuals. Several meta-learning tasks for label generation are generated through meta-learning training, and each of the meta-learning tasks is based on the historical electromyography data and gesture label generation of the same individual, or each of the meta-learning tasks is based on the historical electromyography of similar individuals. The data and gesture labels are mixed and generated; according to the current electromyographic data and gesture labels of the current individual, the initial gesture recognition model is fine-tuned to obtain the current gesture recognition model of the current individual.

In order to solve one or more of the above problems and achieve the above objects, embodiments of the present application provide a model acquisition method, which includes: acquiring the current myoelectric data and gesture tags of the current individual, where the gesture tags include gesture information; When the current individual does not have a personal gesture recognition model, an initial gesture recognition model is obtained, wherein the initial gesture recognition model is meta-learned based on several meta-learning tasks generated based on historical myoelectric data and gesture labels of multiple individuals. Learning and training are generated, and each of the meta-learning tasks is generated based on the historical myoelectric data and gesture labels of the same individual, or each of the meta-learning tasks is generated based on a mixture of historical myoelectric data and gesture labels of similar individuals; according to the Using the current electromyographic data and gesture tags of the current individual, fine-tune the initial gesture recognition model to obtain the current gesture recognition model of the current individual.

In order to solve one or more of the above problems and achieve the above objects, embodiments of the present application provide a gesture recognition method, including: obtaining the current gesture recognition model of the current individual through the above model acquisition system; obtaining the current individual's current gesture recognition model. Real-time electromyographic data; through the current gesture recognition model, the gesture of the current individual is obtained according to the real-time electromyographic data.

In order to solve one or more of the above problems and achieve the above objects, embodiments of the present application also provide a gesture recognition device, including: a first acquisition module for acquiring the current gesture recognition of the current individual through the above model acquisition system. model; a second acquisition module, used to acquire the real-time electromyographic data of the current individual; an identification module, used to pass the The current gesture recognition model acquires the gesture of the current individual based on the real-time electromyographic data.

In order to solve one or more of the above problems and achieve the above objects, embodiments of the present application also provide an electronic device, including: at least one processor; and a memory communicatively connected with the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor, so that the at least one processor can perform the above-mentioned gesture recognition method, or execute the above-mentioned model. Get method.

In order to solve one or more of the above problems and achieve the above objects, embodiments of the present application also provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the above gesture recognition method is implemented, or Execute the above model acquisition method.

The model acquisition system proposed in this application, when the current individual does not have a personal gesture recognition model, first obtains the current electromyographic data and gesture information of the current individual through the input unit, and based on the current individual's current electromyographic data and gesture tags, The initial gesture recognition model is fine-tuned to obtain the current gesture recognition model suitable for the current individual, and the initial gesture recognition model is generated based on meta-learning training based on several meta-learning tasks generated based on the historical electromyographic data and gesture labels of multiple individuals, and Each meta-learning task is generated based on the historical EMG data and gesture labels of the same individual or based on a mixture of historical EMG data and gesture labels of similar individuals. First, several meta-learning tasks are first formed based on the historical myoelectric data and gesture labels of multiple individuals, which are generated based on the historical myoelectric data and gesture labels of the same individual or based on a mixture of historical myoelectric data and gesture labels of similar individuals. Combined with the meta-learning tasks, The initial gesture recognition model obtained through learning and training is suitable for the crowd and has good generalization ability. When applied to a specific current individual, the initial gesture recognition model is evaluated based on the current myoelectric data and gesture labels of the current individual. Fine-tuning, in this way, with the support of a small amount of individual data, the initial gesture recognition model can be fine-tuned accurately and efficiently to obtain the current gesture recognition model that adapts to the current individual, thereby ensuring the accuracy of gesture recognition.

Description of drawings

One or more embodiments are exemplified by the pictures in the corresponding drawings. These illustrative illustrations do not constitute limitations to the embodiments. Elements with the same reference numerals in the drawings are represented as similar elements. Unless otherwise stated, the figures in the drawings are not intended to be limited to scale.

Figure 1 is a schematic structural diagram of a model acquisition system in an embodiment of the present application;

Figure 2 is a schematic structural diagram of a meta-learning task in the embodiment of the present application;

Figure 3 is a schematic diagram of a meta-learning training process in an embodiment of the present application;

Figure 4 is a flow chart of a model acquisition method in another embodiment of the present application;

Figure 5 is a flow chart of a gesture recognition method in another embodiment of the present application;

Figure 6 is a schematic structural diagram of a gesture recognition device in another embodiment of the present application;

Figure 7 is a schematic structural diagram of an electronic device in another embodiment of the present application.

Detailed ways

It can be seen from the background technology that traditional model training methods have problems of over-fitting and unstable model performance, while models obtained through transfer learning have low training and adjustment efficiency and need to rely on a large amount of individual data to ensure recognition accuracy. Therefore, how to provide a model acquisition method that can meet the needs of efficient switching, training and adjustment in different application scenarios to ensure accurate and efficient recognition of gestures is an urgent problem that needs to be solved.

In order to solve the above problems, embodiments of the present application provide a model acquisition system, including: an input unit, used to acquire the current electromyographic data and gesture tags of the current individual, where the gesture tags contain gesture information; a processing unit, configured as: When the current individual does not have a personal gesture recognition model, an initial gesture recognition model is obtained. The initial gesture recognition model is generated based on meta-learning training based on several meta-learning tasks generated based on historical myoelectric data and gesture labels of multiple individuals. , and each meta-learning task is generated based on the historical EMG data and gesture labels of the same individual, or each meta-learning task is generated based on a mixture of historical EMG data and gesture labels of similar individuals; based on the current EMG data and gesture labels of the current individual Label, fine-tune the initial gesture recognition model to obtain the current gesture recognition model of the current individual.

The model acquisition system proposed in this application, before acquiring the current gesture recognition model of the current individual, when the current individual does not have a personal gesture recognition model, acquires the historical myoelectric data and gesture tags of multiple individuals, based on each individual The historical EMG data and gesture labels form a meta-learning task, or the historical EMG data and gesture labels of similar individuals are mixed to form a meta-learning task, and a more general initial gesture recognition model is generated through meta-learning training. Then, when obtaining the current gesture recognition model of the current individual, first obtain the current electromyographic data and gesture labels of the current individual through the input unit, and fine-tune the initial gesture recognition model based on the obtained current electromyographic data and gesture labels, and obtain Current gesture recognition model applicable to the current individual. First, several meta-learning tasks are obtained based on the historical EMG data and gesture labels of multiple individuals. Each meta-learning task is generated based on the historical EMG data and gesture labels of the same individual, or each meta-learning task is based on the historical EMG data and gesture labels of similar individuals. Electrical data and gesture labels are mixed and generated, combined with the meta-learning training method to obtain an initial gesture recognition model that is suitable for the crowd and has good generalization ability. When applied to a specific current individual, it is based on the current EMG of the current individual. Data and gesture labels to fine-tune the initial gesture recognition model. In this way, with the support of a small amount of individual data, the initial gesture recognition model can be fine-tuned accurately and efficiently to obtain the current gesture recognition model that adapts to the current individual, thereby ensuring the accuracy of gesture recognition.

In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, each embodiment of the present application will be described in detail below with reference to the accompanying drawings. However, those of ordinary skill in the art can understand that in each embodiment of the present application, many technical details are provided to enable readers to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solution claimed in this application can also be implemented. The division of the following embodiments is for convenience of description and should not constitute any limitation on the specific implementation of the present application. They can be combined and referenced with each other under the premise of shielding.

The implementation details of the model acquisition system recorded in this application will be described in detail below with reference to specific embodiments. The following contents are only implementation details provided for convenience of understanding and are not necessary for the real-time solution.

The first aspect of the embodiment of the present application provides a model acquisition system. As shown in Figure 1, the model acquisition system includes an input unit 101 and a processing unit 102:

The input unit 101 is used to obtain the current electromyographic data and gesture tags of the current individual. The gesture tags include gesture information.

Specifically, when obtaining the current gesture recognition model with better recognition effect for the current individual, the input unit 101 is first used to correspond to the electromyographic data of the current individual within a certain period of time in the current state and the gesture correspondence of the individual when the electromyographic data is collected. Collect and record the gesture tags to obtain the current myoelectric data and gesture tags of the current individual. Among them, the gesture tag is a data tag created for gesture storage. Different gesture tags correspond to different gestures. For example, gesture tag 1 corresponds to the individual's raised index finger, gesture tag 2 corresponds to the individual's raised middle finger, gesture tag 3 corresponds to the individual's fist, etc. In specific applications, the meaning of each tag can be set as needed. This embodiment does not limit the specific setting and meaning of gesture tags.

For example, when collecting the current electromyographic data of the current individual, the input unit 101 automatically prompts the current individual through voice or text to make gestures corresponding to several preset gesture tags in a certain order within a specified time, and The electromyographic signals generated by muscle motor neurons when the current individual makes different gestures are recorded through electromyographic electrodes, and the electromyographic signals and gesture labels are encoded and stored according to the time interval corresponding to each gesture. By obtaining the current electromyographic data and gesture labels of the current individual in the current state, it is convenient to fine-tune the initial gesture recognition model in the future, and then efficiently and accurately obtain the current gesture recognition model for the current individual's current state.

In some other alternative embodiments, gesture labels may be obtained in an unsupervised or self-supervised manner. In this collection method, the collection system or staff do not need to set a specific "label meaning" for each gesture. Instead, the labels are handed over to the statistical model for automatic classification. This embodiment has no special restrictions on processing classification methods, such as dimensionality reduction, clustering, autoencoders and other machine learning methods. Exemplarily, the processing classification method is principal component analysis PCA, K-Means clustering method.

The processing unit 102 is configured to: obtain an initial gesture recognition model when the current individual does not have a personal gesture recognition model, wherein the initial gesture recognition model is generated based on several historical myoelectric data and gesture tags of multiple individuals. Meta-learning tasks are generated through meta-learning training, and each meta-learning task is generated based on the historical electromyographic data and gesture labels of the same individual, or each meta-learning task is generated based on a mixture of historical electromyographic data and gesture labels of similar individuals; based on the current The individual's current electromyographic data and gesture labels are used to fine-tune the initial gesture recognition model to obtain the current gesture recognition model of the current individual.

Specifically, the processing unit 102 obtains the current electromyographic data and gestures of the current individual through the input unit 101 Before or after the tag, according to the pre-configured program, first detect whether the current individual's personal gesture recognition model is pre-stored based on the current individual's identity. If it is detected that the current individual does not have a personal gesture recognition model, the processing unit 102 obtains the initial Gesture recognition model. The initial gesture recognition model is generated based on meta-learning training based on several meta-learning tasks generated based on the historical electromyographic data and gesture labels of multiple individuals, and each meta-learning task is generated based on the historical electromyographic data and gesture labels of the same individual, or Each meta-learning task is generated based on a mixture of historical EMG data and gesture labels of similar individuals. When acquiring the current gesture recognition model for the current individual, the processing unit 102 may read the pre-trained initial gesture recognition model from the storage address where the initial gesture recognition model is stored through communication. The processing unit 102 can also read the historical myoelectric data and gesture tags of several individuals from the designated storage address, first generate several meta-learning tasks based on the historical myoelectric data and gesture tags of each individual or similar individuals, and then perform a The trained meta-learner undergoes meta-learning training to generate an initial gesture recognition model. The processing unit 102 can also directly read several meta-learning tasks pre-generated based on historical myoelectric data and gesture labels of different individuals to train an untrained meta-learner to generate an initial gesture recognition model. This embodiment does not limit the specific acquisition method of the initial gesture recognition model. This embodiment uses a similarity measurement method to determine similar individuals, which can increase the sample size. This embodiment has no special restrictions on the specific method of similarity measurement. For example, Euclidean distance and Chebyshev distance are used to determine similarity by obtaining the "distance" between two objects.

After acquiring the current electromyographic data and gesture tags of the current individual and the initial gesture recognition model, the processing unit 102 fine-tunes the initial gesture recognition model for the current individual according to the current electromyographic data and gesture tags of the current individual, for example, by The current myoelectric data is used as input information, and the gesture corresponding to the gesture tag is used as supervision information. Supervised learning is performed on the initial gesture recognition model, so that the initial gesture recognition model has better pertinence and adaptability to the current individual. The fine-tuned initial gesture recognition model is used as the input information. The gesture recognition model serves as the current gesture recognition model of the current individual. By fine-tuning the initial gesture recognition model for the current individual based on the current myoelectric data and gesture labels of the current individual, the pertinence of the obtained current gesture recognition model is improved and the accuracy of gesture recognition for the current individual is ensured.

Furthermore, the initial gesture recognition model can be trained and generated in the following way: based on the historical myoelectric data and gesture labels of multiple individuals, train a network model with gradient backpropagation to obtain the original gesture recognition model; Historical EMG data and gesture labels are used to generate several meta-learning tasks, and each meta-learning task is based on the historical EMG data and gesture labels of the same individual, or each meta-learning task is based on the historical EMG data and gestures of similar individuals. Label hybrid generation; use each meta-learning task to perform meta-learning training on the meta-learner based on the original gesture recognition model to obtain the initial gesture recognition model.

Specifically, when the processing unit 102 or other model training equipment trains the initial gesture recognition model based on the historical electromyographic data and gesture tags of multiple individuals, the processing unit 102 or other model training equipment may first train the initial gesture recognition model based on the historical electromyographic data and gesture tags of multiple individuals. The network model with gradient backpropagation is trained. The training methods include supervised training, unsupervised training, etc., and the original A gesture recognition model is started, and then several meta-learning tasks are formed from the EMG data set consisting of historical EMG data and gesture labels of multiple individuals according to preset constraints, and the data included in each meta-learning task comes from the same individual. , or from similar individuals. For example, an EMG data set composed of historical EMG data and gesture labels of multiple individuals is randomly sampled to extract historical EMG data and gesture labels of multiple individuals. Then one or more meta-learning tasks are generated based on the historical myoelectric data and gesture labels of each individual obtained through random sampling. In this way, each meta-learning task among several tasks contains data from the same individual. Alternatively, after randomly sampling and extracting the historical EMG data and gesture labels of multiple individuals, the similarity measurement method is used to determine whether the individuals are similar, and then based on the historical EMG data and gesture labels of similar individuals obtained through random sampling, a generated One or more meta-learning tasks. Then, according to several generated meta-learning tasks, each meta-learning task is used to perform meta-learning training on the meta-learner based on the original gesture recognition model to optimize the meta-learner parameters, and then obtain the initial gesture recognition model. Here, the acquisition of the original gesture recognition model and the formation of the meta-learning task can also be carried out at the same time, or the meta-learning task can be formed first and then the original gesture recognition model is obtained. By obtaining the electromyographic data and gesture tags corresponding to the electromyographic data of multiple individuals, and based on the preset constraints, the electromyographic data and gesture tags of multiple individuals are randomly selected, and then based on the electromyographic data of each individual or similar individuals and gesture labels, generate one or more meta-learning tasks, and use the generated meta-learning tasks to perform meta-learning training on the meta-learner based on the original gesture recognition model, so that the trained initial gesture recognition model can target different individuals and new Individuals have good generalization ability.

It is worth mentioning that the network model with gradient backpropagation is any of the convolutional neural network CNN model, the long-short-term neural network LSTM model and the recurrent neural network RNN model.

For example, a neural network model optimized by an optimization algorithm based on stochastic gradient descent can be used as the basic model, such as a 3-layer convolutional neural network model using the Adam optimizer as the basic model, combined with a preset loss function, such as cross-entropy loss. function to perform model training. The historical EMG data in the EMG data set are used as input, and the gestures to which the historical EMG data belong are used as supervision signals. The loss value is calculated using the preset loss function, and the model parameters are optimized based on the gradient update of the loss value to complete the training of the model. Get the original gesture recognition model.

In addition, after each collection of the current EMG data and gesture labels of the current individual, the EMG data set containing the historical EMG data and gesture labels of several individuals can be expanded, and then based on the expanded EMG data Set, retrain the original gesture model to improve the generalization effect of the original gesture recognition model. Retraining may be performed regularly according to a preset period, or may be performed after the myoelectric data set has been expanded to a certain extent. This embodiment does not limit this.

In another example, the meta-learning task includes a support set (Support Set) and a query set (Query Set). Each meta-learning task is used to perform meta-learning training on a meta-learner based on the original gesture recognition model, including: for each element For learning tasks, first assign the parameters of the meta-learner to the base learner, and then train the base learner based on the EMG data in the support set and the gesture labels corresponding to the EMG data to optimize the parameters of the base learner; according to the query set The EMG data and the gesture label corresponding to the EMG data are obtained to obtain the prediction error of the gesture label prediction result of the parameter-optimized base learner; Based on the prediction error, gradient updates are performed on the meta-learner to optimize the parameters of the meta-learner.

Specifically, the meta-learning task includes a support set and a query set. Among them, the support set has an important setting of N-way K-shot, that is, there are N types of samples in the support set, and K samples of each type are labeled. Data, and the query set can contain n types of samples, each type of sample has k labeled data, where N, K, n and k are all positive integers. For example, the support set contains 6 gesture tags, and each gesture tag samples three pieces of electromyographic data. The query set contains two gesture tags, and each gesture tag samples one piece of myoelectric data. As an example, this embodiment The meta-learning task generated based on the historical electromyographic data and gesture labels of each individual can be referred to Figure 2. The labeled data represents the acquired electromyographic data, and the samples are various types of gesture labels.

In the process of meta-learning training of the meta-learner based on the original gesture recognition model according to each meta-learning task, for each meta-learning task, the parameters in the meta-learner are first copied through the base learner, and then the support is read Each concentrated EMG data and the gesture labels corresponding to the EMG data are used as input. The base learner is used to predict the gesture labels, and the loss value between the prediction results and the gesture labels corresponding to the EMG data is used. Gradient updates to optimize parameters in the base learner. Then read the electromyographic data in the query set and the gesture tags corresponding to the electromyographic data, use the electromyographic data corresponding to each gesture tag in the query set as input, and use the parameter-optimized base learner to predict the gesture tags and predict the gesture tags according to the query The gesture labels corresponding to the electromyographic data are concentrated to obtain the prediction error of the gesture label prediction result output by the parameter-optimized base learner. Then, based on the obtained prediction error, the gradient of the meta-learner is updated to obtain the optimized new parameters, and the new parameters are used as parameters of the meta-learner. In the next round of training, the optimized parameters of the meta-learner are copied back to the base learner. Multiple rounds of training are carried out in this way until the prediction error of the gesture label no longer decreases, then the training is stopped, and the meta-learner with optimized parameters at this time is used as the initial gesture recognition model. By performing double-layer loop training and parameter optimization based on the support set and query set in the meta-learning task, the required initial gesture recognition model is obtained accurately and efficiently based on the original gesture recognition model. Moreover, the obtained initial gesture recognition model only requires a small number of gradient update steps and a meta-learning task related to the specific task for fine-tuning to adapt to the specific task.

In addition, the preset constraints can also include that when generating meta-learning tasks, the electromyographic data and gesture labels used to generate the support set can be taken from the data collected relatively early in the individual electromyographic data collected, and The electromyographic data and gesture labels used to generate the query set are taken from the relatively later collection time of the individual electromyographic data collected. By using the individual's earlier EMG data to generate the support set and the later EMG data to generate the query set, the trained meta-learner is based on the earlier EMG data versus the later EMG data. The prediction model can then improve the consistency between the trained model and the actual gesture recognition process, and further improve the accuracy of the model's gesture prediction. Moreover, because the original gesture recognition model will learn to a certain extent the data distribution shift caused by differences in acquisition time (i.e. different muscle/surface electromyographic states), it can better adapt to subsequent actual usage scenarios.

In another example, base learning is performed based on the EMG data in the support set and the gesture labels corresponding to the EMG data. The machine is trained to optimize the parameters of the base learner, including: using the EMG data in the support set as input, the gesture label corresponding to the EMG data as a supervision signal, and obtaining the loss value of the gesture label prediction result of the base learner; according to the loss value , perform gradient updates on the base learner to obtain new parameters of the base learner.

Specifically, when training the base learner, the electromyographic data is used as the input signal, and the gesture label corresponding to the centrally recorded electromyographic data is used as the supervision signal during the training process. The base learner uses the input electromyoelectric data to The data predicts the gesture label, and calculates the loss value of the gesture label prediction result through a preset loss function, such as a cross-entropy loss function, a pairwise loss function or an exponential loss function, the gesture label prediction result and the supervision signal, and then calculates the loss value of the gesture label prediction result according to The loss value performs a gradient update on the base learner to obtain the parameters optimized by the base learner. Through supervised training, the parameter optimization of the base learner is accurately completed to ensure the effect of meta-learning training.

In another example, obtaining the prediction error of the gesture label prediction result of the parameter-optimized base learner includes: obtaining the gesture label prediction result of the electromyographic data based on the parameter-optimized base learner and the electromyographic data in the query set; Through the preset loss function, gesture label prediction results, and gesture labels corresponding to the electromyographic data in the query set, the loss value corresponding to the gesture label prediction result is obtained, and the prediction error is obtained based on the loss values corresponding to all myoelectric data in the query set.

Specifically, the electromyographic data in the query set are used as input, a base learner with optimized parameters is used to predict gesture labels, and the gesture labels corresponding to the electromyographic data in the query set are used as supervision signals. Combining the preset loss function, such as cross-entropy loss function, pairwise loss function or exponential loss function, the gesture label predicted by the parameter-optimized base learner, and the gesture label corresponding to the electromyographic data in the query set, the parameter optimization is obtained The loss value of the gesture label prediction result of the base learner is obtained, and the prediction error is obtained based on the loss value corresponding to all the electromyographic data in the query set, so as to facilitate the subsequent gradient update of the meta-learner based on the prediction error, and then complete the meta-learner. Learner parameter optimization. The prediction error of the prediction result is accurately obtained through the preset loss function, which facilitates accurate parameter adjustment of the meta-learner and completes model optimization.

Further, obtaining the prediction error based on the loss values corresponding to all the electromyographic data in the query set includes: obtaining the average value of the loss values corresponding to all the electromyographic data in the query set, and using the average value as the prediction error. The average here can be any one of arithmetic mean, weighted mean, geometric mean, root mean square mean and harmonic mean. Preferably, the average value is an arithmetic mean. Specifically, in the process of obtaining the prediction error of the parameter-optimized base learner, the loss values corresponding to each EMG data in the query set can be obtained one by one, and then the loss values corresponding to each EMG data are arithmetic averaged to obtain all muscle parameters. The arithmetic mean of the loss values corresponding to the electrical data is used as the prediction error. By performing an arithmetic average of the loss values corresponding to all electromyographic data in the query set, and using the arithmetic average as the prediction error, we can obtain the prediction error as accurately as possible and avoid the impact of accidental factors in a single prediction on the prediction result.

Furthermore, during the meta-learning training process of the meta-learner based on the original gesture recognition model using each meta-learning task, the gradient of the base learner is updated based on the electromyographic data in the support set and the gesture labels corresponding to the electromyographic data. The learning rate when is smaller than the learning rate when updating the gradient of the meta-learner based on the EMG data in the query set and the gesture labels corresponding to the EMG data. By limiting the learning rate of the inner loop training based on the support set and the outer loop training based on the query set, and combining the high learning rate inner loop with the low learning rate outer loop, the meta-learner can converge as quickly as possible and improve the meta-learner. Learner training efficiency.

In summary, in the meta-learner training based on the original gesture recognition model, the process of conducting a round of meta-learning training based on a meta-learning task can be referred to Figure 3, including:

Step 301: Obtain the meta-learning task used in the current round of training.

Step 302: Assign the parameters of the meta-learner to the base learner.

Step 303: Train the base learner based on the support set data of the meta-learning task used in the current round of training, and obtain the optimized parameters of the base learner after the gradient update.

Step 304: Based on the parameter-optimized base learner and the query set data of the meta-learning task used in the current round of training, calculate the loss value of the meta-learning task as the prediction error, and calculate the corresponding gradient.

Step 305: Update the parameters of the meta-learner according to the calculated gradient to complete the current round of meta-learning training.

In another alternative example, the initial gesture recognition model can also be trained and generated in the following way: generating several meta-learning tasks based on the historical electromyographic data and gesture labels of multiple individuals, and each meta-learning task is based on the same individual or Historical EMG data and gesture labels of similar individuals are generated; each meta-learning task is used to perform meta-learning training on a meta-learner based on a network model with gradient backpropagation to obtain an initial gesture recognition model.

Specifically, when acquiring the initial gesture recognition model, the historical EMG data and gesture labels of multiple individuals can be randomly obtained, and then one or more meta-learning tasks are generated based on the historical EMG data and gesture labels of each individual. , multiple meta-learning tasks are finally obtained, and each meta-learning task is generated based on the historical electromyographic data and gesture labels of the same individual, or based on the historical electromyographic data and gesture labels of similar individuals. Then a grid model with gradient backpropagation is taken as a meta-learner, and each meta-learning task is used to perform meta-learning training on the network model to obtain the initial gesture recognition model. The parameters of the grid model with gradient backpropagation are randomly obtained at the initial stage of training. The meta-task training method in meta-learning drives the model to optimize parameters in a direction with stronger generalization and improve the accuracy of the model. By directly obtaining a grid model with random parameters and gradient backpropagation, and obtaining the initial gesture recognition model through meta-learning training, the acquisition process of the initial gesture recognition model is simplified, the model acquisition efficiency is improved, and only a small amount of gradients are required. The number of steps is updated, and a meta-learning task related to the specific task can be fine-tuned to adapt to the specific task.

In another example, fine-tuning the initial gesture recognition model based on the current electromyographic data and gesture labels of the current individual includes: supervising and training the initial gesture recognition model based on the current electromyographic data and gesture labels of the current individual; Preset the number of gradient update steps, and perform gradient updates on the initial gesture recognition model to obtain new parameters to form the current Current gesture recognition model of ex-individuals.

Specifically, when fine-tuning the initial gesture recognition model based on the current EMG data and gesture labels of the current individual, it can be achieved through supervised training, that is, the current EMG data of the current individual is used as input data, and the current EMG data is used as the input data. The gesture label corresponding to the electromyographic data is used as a supervision signal for supervised training. During the gradient update process, the initial gesture recognition model is gradient updated according to the preset gradient update steps to obtain optimized parameters and form the current individual's current Gesture recognition model. Through supervised training and preset gradient update steps, the initial gesture recognition model can be accurately fine-tuned to ensure the recognition accuracy of the current gesture recognition model.

In another example, the processing unit 102 is further configured to: generate a current meta-learning task for the current individual based on the current electromyographic data and gesture labels of the current individual; and perform meta-learning based on the initial gesture recognition model based on the current meta-learning task. The machine performs meta-learning training to generate the current individual's personal gesture recognition model.

Specifically, after acquiring the current myoelectric data and gesture labels of the current individual, the processing unit 102 will also generate the current meta-learning task of the current individual based on the current myoelectric data and gesture labels of the current individual. Then, according to the current meta-learning task of the current individual, the meta-learner based on the initial gesture recognition model is trained for the current individual, so that the meta-learner with optimized parameters has better pertinence and adaptability to the current individual. The meta-learner after completing meta-learning training is used as the current individual's personal gesture recognition model. For example, collect myoelectric data when an individual makes gestures corresponding to different gesture tags within 2 minutes in the current state. Based on the collected current myoelectric data and gesture tags, the initial gesture recognition model is fine-tuned and the preset settings are also used. As a constraint, from the 2 minutes of data collected, the electromyographic data collected first and the corresponding gesture tags are taken as the support set, and the electromyographic data collected later and the corresponding gesture tags are used as the query set to generate the current individual's The current meta-learning training task is performed, and the meta-learner based on the initial gesture recognition model is meta-learning trained according to the current meta-learning task to obtain the current individual's personal gesture recognition model. Under the condition of ensuring the data of the same individual, the method of constructing the support set and query set generation meta-task according to the time sequence can enable the personal gesture recognition model to have the ability to learn the overall distribution drift of myoelectric data over time.

In addition, when generating the current meta-learning task based on the current individual's current EMG data and gesture labels, the EMG data when the individual makes gestures corresponding to different gesture labels within 2 minutes of the current state can also be collected. The electromyographic data and the gesture labels corresponding to the electromyographic data generate a meta-learning training task, and the meta-learning verification task is generated based on the electromyographic data and the gesture labels corresponding to the electromyographic data within the next minute. Obviously, it is also possible to generate a meta-learning training task based on the EMG data and the gesture labels corresponding to the EMG data in the first minute and a half, and to generate a meta-learning verification task based on the EMG data and the gesture labels corresponding to the EMG data in the second half minute. Task. Of course, other time ratios can also be used to divide time to obtain training tasks and verification tasks, and this embodiment is not limited to this. In addition, the proportion of time before and after segmentation can be determined according to the proportion of K-WAY-N-SHOT, which can ensure the balance of the sample. After meta-learning training is performed on the initial gesture recognition model according to the meta-learning training task, the gesture prediction accuracy of the trained initial gesture recognition model is obtained according to the meta-learning verification task. Sum detection, when the gesture prediction accuracy reaches the preset threshold, and the loss value of the gesture label prediction result no longer decreases (or its fluctuation does not decrease significantly within several training rounds), it is determined that the meta-learning training is completed, no need Then make adjustments and use the resulting initial gesture recognition model as the current individual's personal gesture recognition model. For example, setting the threshold of prediction accuracy to 0.9 means that the initial gesture recognition model after meta-learning training has a probability of 90% or above to correctly predict the current individual's gesture, and the loss value of the gesture label prediction result will no longer decrease. In the case of , it is judged that the training is completed. When the gesture prediction accuracy does not reach the preset threshold and/or the loss value of the gesture label prediction result is still declining, it is determined that the training is not completed, and the initial gesture recognition model continues to be repeatedly trained and parameter fine-tuned according to the meta-learning training task, or Collect the electromyographic data of the current individual over a longer period of time and the gesture labels corresponding to the electromyographic data, generate a new meta-learning task, retrain the initial gesture recognition model and fine-tune parameters until the gesture prediction accuracy reaches the preset threshold, and the gesture labels The loss value of the prediction result no longer decreases. By using the verification task to detect the prediction accuracy, the accuracy of the obtained personal gesture recognition model is guaranteed.

It is worth mentioning that the preset duration and prediction accuracy thresholds can be set as needed. When generating meta-learning tasks, training tasks can be generated based on data collected earlier, and verification tasks can be generated based on data collected later. Training tasks can be generated based on data collected later, and verification tasks can be generated based on data collected earlier. All collected data can be used for meta-learning task generation, or only part of the data can be selected for meta-learning task generation. In this embodiment There are no restrictions on this.

Further, the model acquisition system also includes: a storage unit; a storage unit for storing a personal gesture recognition model; a processing unit, also configured to: when the current individual has a personal gesture recognition model, according to the current individual's current muscle According to the electromyographic data and gesture labels of the current individual, the personal gesture recognition model is fine-tuned to obtain the current gesture recognition model of the current individual; based on the current individual's electromyographic data and gesture labels, the current meta-learning task of the current individual is generated, and the current meta-learning task is generated based on the current meta-learning task. The personal gesture recognition model continues to undergo meta-learning training to generate a parameter-optimized personal gesture recognition model.

Specifically, after detecting whether the current individual's personal gesture recognition model is pre-stored according to the current individual's identity, the processing unit 102 reads the current individual's personal gesture recognition model if it is stored in the storage unit. The individual's personal gesture recognition model is then fine-tuned based on the current individual's current myoelectric data and gesture tags to obtain the current individual's current gesture recognition model. That is, when it is detected that the current individual is not a new user, the pre-stored personal gesture recognition model of the current individual is read and used as the basis for obtaining the current gesture recognition model. By storing and reusing individual personal gesture recognition models, the waste of prior experience caused by the generation of personal gesture recognition models in the current gesture recognition model generation process is avoided.

In addition, the processing unit 102 is also configured to generate a current meta-learning task for the current individual based on the current electromyographic data and gesture labels of the current individual, and then perform a personal gesture recognition model of the current individual based on the current meta-learning task for the current individual. Meta-learning training of the individual's current state enables the personal gesture recognition model to have better performance for the current individual. Targeted and adaptable, the parameters of the personal gesture recognition model are optimized through meta-learning training, and the optimized personal gesture recognition model is stored in the storage unit.

This embodiment has no particular limitation on the type of processing unit. The processing unit can be hardware that performs logical operations, such as a microcontroller, a microprocessor, a programmable logic controller (PLC) or a field-programmable logic gate array (FPGA), or in a Software programs, functional modules, functions, object libraries (Object Libraries) or dynamic link libraries (Dynamic-Link Libraries) that implement the above functions based on hardware. Or, a combination of both.

Another aspect of the embodiments of the present application provides a method for generating a gesture model. The process of the gesture model generation method can be referred to Figure 4, including the following steps:

Step 401: Obtain the current electromyographic data and gesture tag of the current individual. The gesture tag contains gesture information.

Step 402: When the current individual does not have a personal gesture recognition model, obtain an initial gesture recognition model. The initial gesture recognition model is meta-learned based on several meta-learning tasks generated based on historical myoelectric data and gesture labels of multiple individuals. Learning and training generation, each meta-learning task is generated based on the historical electromyographic data and gesture labels of the same individual, or each meta-learning task is generated based on the historical electromyographic data and gesture labels of similar individuals.

Step 403: Fine-tune the initial gesture recognition model based on the current electromyographic data and gesture labels of the current individual to obtain the current gesture recognition model of the current individual.

It is not difficult to find that this embodiment is a method embodiment corresponding to the system embodiment, and this embodiment can be implemented in cooperation with the system embodiment. The relevant technical details mentioned in the system embodiment are still valid in this embodiment. In order to reduce duplication, they will not be described again here. Correspondingly, the relevant technical details mentioned in this embodiment can also be applied to the system embodiment.

Another aspect of the embodiment of the present application provides a gesture recognition method. The flow of the gesture recognition method can be referred to Figure 5, which includes the following steps:

Step 501: Obtain the current gesture recognition model of the current individual. Specifically, the terminal device that recognizes the gesture of the current individual obtains the current gesture recognition model obtained by fine-tuning the initial gesture recognition model for the current individual through the above model acquisition system.

Step 502: Obtain real-time electromyographic data of the current individual. Specifically, when performing gesture recognition, the electromyographic data of the current individual is collected in real time through electromyographic electrodes, and the collected real-time electromyographic data is input into the current gesture recognition model. Exemplarily, during data collection (for example, during model training or real-time inference), the method of cumulative sharding of data sampling points is used to determine the size of the data slices input to the statistical model. Fragmentation methods include but are not limited to sliding window segmentation methods and endpoint detection segmentation methods (endpoint detection based on RMS or other methods).

Step 503: Obtain the current individual's gesture based on the real-time myoelectric data through the current gesture recognition model. Specifically, after obtaining the real-time electromyographic data of the current individual, the current gesture recognition model performs Predict the gestures of the current individual and output the predicted gestures of the current individual.

In one example, after obtaining the current gesture recognition model of the current individual through the above model acquisition system, it also includes: obtaining the prediction accuracy of the current gesture recognition model; when the prediction accuracy does not meet the preset threshold, re-obtaining a new prediction Set the current electromyographic data and gesture tags of the current individual within the time period; re-fine-tune the current gesture recognition model based on the re-obtained current electromyographic data and gesture tags.

Specifically, due to the characteristics of electromyographic data, the prediction accuracy of the current gesture recognition model for the current individual will gradually decrease as the use time increases. Therefore, after obtaining the current gesture recognition model of the current individual, the current gesture recognition model Monitor the prediction accuracy of the current gesture recognition model. When the prediction accuracy of the current gesture recognition model is insufficient, reacquire the gestures corresponding to different gesture labels and the electromyographic data when making different gestures within the preset time period of the current individual's current state, and based on The current myoelectric data and gesture labels are re-acquired, and the current gesture recognition model is fine-tuned again, so that the current gesture recognition model can better adapt to the current status of the current individual and ensure the accuracy of gesture recognition. Taking into account the easy-to-change nature of electromyographic signals, the gesture recognition model is updated and iterated periodically to ensure the accuracy of the gesture recognition results obtained by the gesture recognition model.

Another aspect of the embodiment of the present application also provides a gesture recognition device. Referring to Figure 6, it includes:

The first acquisition module 601 is used to acquire the current gesture recognition model of the current individual through the above-mentioned model acquisition system.

The second acquisition module 602 is used to acquire the real-time electromyographic data of the current individual;

The recognition module 603 is used to obtain the current individual's gesture according to the real-time myoelectric data through the current gesture recognition model.

In one example, the gesture recognition device further includes a third acquisition model; the third acquisition module is used to acquire the prediction accuracy of the current gesture recognition model; when the prediction accuracy does not meet the preset threshold, re-obtain a new preset duration. The current myoelectric data and gesture tags of the current individual in the system are retrieved; the current gesture recognition model is re-fine-tuned based on the re-obtained current myoelectric data and gesture tags.

It is not difficult to find that this embodiment is a device embodiment corresponding to the method embodiment, and this embodiment can be implemented in cooperation with the method embodiment. The relevant technical details mentioned in the method embodiment are still valid in this embodiment. In order to reduce duplication, they will not be described again here. Correspondingly, the relevant technical details mentioned in this embodiment can also be applied to the method embodiment.

Another aspect of the embodiment of the present application also provides an electronic device, with reference to Figure 7, including: at least one processor 701; and a memory 702 communicatively connected to the at least one processor 701; wherein the memory 702 stores data that can be Instructions executed by at least one processor 701, the instructions are executed by at least one processor 701, so that at least one processor 701 can execute the gesture recognition method as described above, or the model acquisition method as described above.

The memory 702 and the processor 701 are connected using a bus. The bus may include any number of interconnected buses and bridges. The bus connects various circuits of one or more processors 701 and the memory 702 together. The bus is ok To connect various other circuits together, such as peripherals, voltage regulators, and power management circuits, etc., these are all well known in the art and, therefore, will not be described further herein. The bus interface provides the interface between the bus and the transceiver. A transceiver may be one element or may be multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. The data processed by the processor 701 is transmitted on the wireless medium through the antenna. Furthermore, the antenna also receives the data and transmits the data to the processor 701 .

Processor 701 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. The memory 702 may be used to store data used by the processor 701 when performing operations.

Another aspect of the embodiments of the present application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by the processor, the gesture recognition method as described above, or the model acquisition method as described above, is implemented.

That is, those skilled in the art can understand that all or part of the steps in the methods of the above embodiments can be completed by instructing relevant hardware through a program. The program is stored in a storage medium and includes several instructions to cause a device ( It may be a microcontroller, a chip, etc.) or a processor (processor) that executes all or part of the steps of the methods described in various embodiments of this application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program code. .

Those of ordinary skill in the art can understand that the above-mentioned embodiments are specific embodiments for realizing the present invention, and in practical applications, various changes can be made in form and details without departing from the spirit and spirit of the present invention. scope.

Claims (20)

1. A model acquisition system is characterized by including:

   An input unit, used to obtain the current electromyographic data and gesture tags of the current individual, where the gesture tags include gesture information;

   Processing unit, configured as:

   When the current individual does not have a personal gesture recognition model, an initial gesture recognition model is obtained, wherein the initial gesture recognition model is performed based on several meta-learning tasks generated based on historical myoelectric data and gesture labels of multiple individuals. Meta-learning training is generated, and each of the meta-learning tasks is generated based on the historical electromyographic data and gesture labels of the same individual, or each of the meta-learning tasks is generated based on a mixture of historical electromyographic data and gesture labels of similar individuals;

   According to the current electromyographic data and gesture tags of the current individual, the initial gesture recognition model is fine-tuned to obtain the current gesture recognition model of the current individual.
2. The model acquisition system according to claim 1, wherein the initial gesture recognition model can be trained and generated in the following manner:

   Based on the historical electromyographic data and gesture labels of multiple individuals, the network model with gradient backpropagation is trained to obtain the original gesture recognition model;

   Generate several of the meta-learning tasks based on the historical myoelectric data and gesture labels of multiple individuals, and each meta-learning task is generated based on the historical myoelectric data and gesture labels of the same individual, or each of the meta-learning tasks The task is generated based on a mixture of historical EMG data and gesture labels of similar individuals;

   Each of the meta-learning tasks is used to perform meta-learning training on a meta-learner based on the original gesture recognition model to obtain the initial gesture recognition model.
3. The model acquisition system according to claim 2, wherein the network model with gradient backpropagation is any one of a convolutional neural network CNN model, a long-short-term neural network LSTM model, and a recurrent neural network RNN model.
4. The model acquisition system according to claim 2, wherein the meta-learning task includes a support set and a query set, and each of the meta-learning tasks is used to meta-learn a meta-learner based on the original gesture recognition model. Learning and training, including:

   For each meta-learning task, the parameters of the meta-learner are first assigned to the base learner, and then the base learner is evaluated based on the electromyographic data in the support set and the gesture labels corresponding to the electromyographic data. training to optimize the parameters of said base learner;

   According to the electromyographic data in the query set and the gesture labels corresponding to the electromyographic data, obtain the prediction error of the gesture label prediction result of the base learner after parameter optimization;

   Based on the prediction error, the meta-learner is gradient updated to optimize the parameters of the meta-learner.
5. The model acquisition system according to claim 4, wherein the base learner is trained according to the electromyographic data in the support set and the gesture tag corresponding to the electromyographic data to optimize the base learner. parameters, including:

   Using the electromyographic data in the support set as input and the gesture label corresponding to the electromyographic data as a supervision signal, the loss value of the gesture label prediction result of the base learner is obtained;

   According to the loss value, the gradient of the base learner is updated to obtain new parameters of the base learner.
6. The model acquisition system according to claim 4, wherein the prediction error of the gesture label prediction result of the base learner after the parameters are optimized includes:

   According to the parameter-optimized base learner and the electromyographic data in the query set, obtain the gesture label prediction result of the electromyographic data;

   Through the preset loss function, the gesture label prediction result and the gesture label corresponding to the myoelectric data in the query set, the loss value corresponding to the gesture label prediction result is obtained, based on all the myoelectric data in the query set The corresponding loss value is used to obtain the prediction error.
7. The model acquisition system according to claim 6, wherein obtaining the prediction error based on the loss values corresponding to all myoelectric data in the query set includes:

   Obtain the average value of the loss values corresponding to all the electromyographic data in the query set, and use the average value as the prediction error.
8. The model acquisition system according to claim 4, characterized in that, in the process of performing meta-learning training on the meta-learner based on the original gesture recognition model using each of the meta-learning tasks, according to the support set The learning rate when updating the gradient of the base learner based on the electromyographic data and the gesture tags corresponding to the electromyographic data is smaller than the myoelectric data and the gesture tags corresponding to the electromyographic data in the query set, and the learning rate for the meta-learning The learning rate when the gradient of the controller is updated.
9. The model acquisition system according to claim 1, characterized in that the initial gesture recognition model can also be trained and generated in the following manner:

   Generate several of the meta-learning tasks based on the historical myoelectric data and gesture labels of multiple individuals, and each meta-learning task is generated based on the historical myoelectric data and gesture labels of the same individual, or each of the meta-learning tasks The task is generated based on a mixture of historical EMG data and gesture labels of similar individuals;

   Each of the meta-learning tasks is used to perform meta-learning training on a meta-learner based on a network model with gradient backpropagation to obtain the initial gesture recognition model.
10. The model acquisition system according to claim 1, characterized in that, fine-tuning the initial gesture recognition model according to the current electromyographic data and gesture tags of the current individual includes:

    Supervise and train the initial gesture recognition model according to the current electromyographic data and gesture tags of the current individual;

    According to the preset number of gradient update steps, gradient update is performed on the initial gesture recognition model to obtain new parameters to form the current gesture recognition model of the current individual.
11. The model acquisition system according to any one of claims 1 to 10, wherein the processing unit is further configured to:

    Generate a current meta-learning task for the current individual based on the current electromyographic data and gesture labels of the current individual;

    According to the current meta-learning task, meta-learning training is performed on a meta-learner based on the initial gesture recognition model to generate the personal gesture recognition model of the current individual.
12. The model acquisition system according to claim 11, characterized in that the model acquisition system further includes: a storage unit;

    The storage unit is used to store the personal gesture recognition model;

    The processing unit is also configured to:

    When the personal gesture recognition model exists for the current individual, fine-tune the personal gesture recognition model according to the current electromyographic data and gesture tags of the current individual to obtain the current gesture of the current individual. identification model;

    Generate the current meta-learning task of the current individual based on the electromyographic data and gesture tags of the current individual, and continue to perform meta-learning training on the personal gesture recognition model based on the current meta-learning task to generate parameter optimization The latter personal gesture recognition model.
13. The model acquisition system according to claim 1, characterized in that:

    The meta-learning task includes a support set and a query set. In the process of generating the meta-learning task, the data with a relatively early collection time among the current individual's current myoelectric data and gesture tags is used as the support set. , use the data collected relatively later in the current individual's current electromyographic data and gesture tags as the query set.
14. A model acquisition method is characterized by including:

    Obtain the current electromyographic data and gesture tags of the current individual, where the gesture tags include gesture information;

    When the current individual does not have a personal gesture recognition model, an initial gesture recognition model is obtained, wherein the initial gesture recognition model is performed based on several meta-learning tasks generated based on historical myoelectric data and gesture labels of multiple individuals. Meta-learning training is generated, and each of the meta-learning tasks is generated based on the historical electromyographic data and gesture labels of the same individual, or each of the meta-learning tasks is generated based on a mixture of historical electromyographic data and gesture labels of similar individuals;

    According to the current electromyographic data and gesture tags of the current individual, the initial gesture recognition model is fine-tuned to obtain the current gesture recognition model of the current individual.
15. A gesture recognition method, characterized in that it includes: acquiring the current gesture recognition model of the current individual through the model acquisition system according to any one of claims 1 to 13;

    Obtain real-time electromyographic data of the current individual;

    Through the current gesture recognition model, the gesture of the current individual is obtained according to the real-time myoelectric data.
16. The gesture recognition method according to claim 15, characterized in that, after obtaining the current gesture recognition model through the model acquisition system according to any one of claims 1 to 12, it further includes:

    Obtain the prediction accuracy of the current gesture recognition model;

    If the prediction accuracy does not meet the preset threshold, reacquire the current electromyographic data and gesture tags of the current individual within a new preset time period;

    The current gesture recognition model is fine-tuned again based on the reacquired current electromyographic data and gesture tags.
17. A gesture recognition device, characterized by including:

    A first acquisition module, configured to acquire the current gesture recognition model of the current individual through the model acquisition system according to any one of claims 1 to 13;

    The second acquisition module is used to acquire the real-time electromyographic data of the current individual;

    A recognition module, configured to obtain the gesture of the current individual according to the real-time myoelectric data through the current gesture recognition model.
18. The gesture recognition device according to claim 17, further comprising: a third acquisition model;

    The third acquisition module is used to acquire the prediction accuracy of the current gesture recognition model;

    When the prediction accuracy does not meet the preset threshold, re-obtain the current electromyographic data and gesture tags of the current individual within a new preset time period;

    The current gesture recognition model is fine-tuned again based on the reacquired current electromyographic data and gesture tags.
19. An electronic device, characterized by including:

    at least one processor; and,

    a memory communicatively connected to the at least one processor; wherein,

    The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the model acquisition method as claimed in claim 14, or The gesture recognition method according to claim 15 or 16.
20. A computer-readable storage medium storing a computer program, characterized in that when the computer program is executed by a processor, it implements the model acquisition method as claimed in claim 14, or the gesture recognition method as claimed in claim 15 or 16. .