Action recognition method and device based on multi-source data, electronic equipment, storage medium and program product
Through the combination of multi-source data and transfer learning methods, the problem that a single sensor cannot capture complex action information is solved, achieving higher action recognition accuracy and cross-user adaptability.
Patent Information
- Application Number
- CN202411278512.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-12
- Publication Date
- 2024-11-26
AI Technical Summary
Existing action recognition methods rely on a single type of sensor and cannot comprehensively and accurately capture complex action information of the human body, resulting in low recognition accuracy and poor performance in cross-user scenarios.
Using multi-source data action recognition method, combining electromyography sensors, force sensors, gyroscopes and acceleration sensors, the target feature data is extracted through sliding windows, and data conversion and transfer learning are performed in the regenerated core Hilbert space to train the movements Identify models to improve recognition accuracy.
It realizes more comprehensive and multi-dimensional recognition of human body movements, overcomes the problem of poor sensitivity of a single sensor, improves the accuracy of action recognition, and is suitable for a variety of scenarios and users.
Smart Images

Figure CN119014861A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of motion recognition technology, and in particular to a motion recognition method, device, electronic device, computer-readable storage medium, and computer program product based on multi-source data. Background Art
[0002] As motion recognition technology is increasingly used, it has applications in fields including health monitoring, motion analysis, and human-computer interaction. Existing motion recognition methods often rely on a single type of sensor (such as an accelerometer or gyroscope). This single data source has certain limitations in capturing complex human motion information. Moreover, since different sensors have different sensitivities to motion, a single sensor may not be able to fully and accurately capture all motion details, thus affecting the overall performance of the motion recognition system and the accuracy of motion recognition.
[0003] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention
[0004] The main purpose of this application is to provide a motion recognition method, device, electronic device, computer-readable storage medium and computer program product based on multi-source data, aiming to solve the technical problems of limitations of current motion recognition solutions and low motion recognition accuracy.
[0005] To achieve the above objectives, the present application proposes a method for action recognition based on multi-source data, the method for action recognition based on multi-source data comprising:
[0006] Extracting raw sensor data from an electromyographic sensor, a force sensor, a gyroscope, and an acceleration sensor, respectively, wherein the electromyographic sensor, the force sensor, the gyroscope, and the acceleration sensor are all worn on an object to be identified;
[0007] Based on a preset sliding window, selecting target feature data from each of the raw sensor data;
[0008] The target feature data is input into a preset target action recognition model, and the action recognition result corresponding to the object to be recognized is output through the target action recognition model.
[0009] In one embodiment, the raw sensor signal includes electromyographic signal data, force signal data, gyroscope data, and acceleration data;
[0010] After the step of respectively extracting raw sensor data from the electromyographic sensor, the force sensor, the gyroscope, and the acceleration sensor, the method further comprises:
[0011] The raw sensor data corresponding to the electromyographic sensor is rectified and integrated to obtain pre-processed electromyographic signal data;
[0012] Smoothing the raw sensor data corresponding to the force sensor, the gyroscope, and the acceleration sensor respectively through a preset filter to obtain preprocessed force signal data, gyroscope data, and acceleration data;
[0013] Based on the preprocessed electromyographic signal data, force signal data, gyroscope data and acceleration data, the step of selecting target feature data from each of the raw sensor data based on a preset sliding window is performed.
[0014] In one embodiment, before the step of inputting the target feature data into a preset target action recognition model, the method further includes:
[0015] Acquire source domain sensor data of a source domain and label data corresponding to the source domain sensor data;
[0016] The source domain sensor data and the target feature data are converted into a common feature space to obtain training sensor data and updated target feature data corresponding to the source domain sensor data, wherein the updated target feature data is used to predict the action recognition result of the object to be recognized;
[0017] Based on the training sensor data and the label data, a preset initial action recognition model is trained to obtain a target action recognition model.
[0018] In one embodiment, the common feature space is a reproducing kernel Hilbert space;
[0019] The step of converting the source domain sensor data and the target feature data into a common feature space to obtain training sensor data corresponding to the source domain sensor data comprises:
[0020] Mapping the source domain sensor data and the target feature data to a reproducing kernel Hilbert space;
[0021] With the goal of minimizing the distance between the source domain sensor data and the target feature data in the reproducing kernel Hilbert space, the source domain sensor data and the target feature data are updated, and the training sensor data corresponding to the source domain sensor data and the updated target feature data are determined.
[0022] In one embodiment, the step of converting the source domain sensor data and the target feature data into a common feature space to obtain training sensor data corresponding to the source domain sensor data includes:
[0023] Initializing a transformation matrix, transforming the source domain sensor data and the target feature data into a common feature space through the transformation matrix to obtain updated source domain sensor data and updated target domain data;
[0024] Iteratively optimizing the transformation matrix with the goal of minimizing the marginal distribution and conditional distribution between the updated source domain sensor data and the updated target domain data, and updating the updated source domain sensor data and the updated target domain data based on the optimized transformation matrix;
[0025] When the preset convergence condition is met, the iterative optimization is stopped, the current updated source domain sensor data is set as the training sensor data, and the current updated target domain data is set as the updated target feature data.
[0026] In one embodiment, before the step of inputting the target feature data into a preset target action recognition model, the method further includes:
[0027] Acquire source domain sensor data of a source domain and label data corresponding to the source domain sensor data;
[0028] Determining an optimal subspace dimension based on the source domain sensor data and the target feature data;
[0029] Constructing a geodesic streamline kernel function between the source domain sensor data and the target feature data;
[0030] Converting the source domain sensor data to the target domain where the target feature data is located according to the subspace dimension and the geodesic streamline kernel function to obtain training sensor data;
[0031] Based on the training sensor data and the label data, a preset initial action recognition model is trained to obtain a target action recognition model.
[0032] In addition, to achieve the above purpose, the present application also proposes a motion recognition device based on multi-source data, the motion recognition device based on multi-source data comprising:
[0033] A data acquisition module, used to extract raw sensor data from an electromyographic sensor, a force sensor, a gyroscope, and an acceleration sensor, respectively, wherein the electromyographic sensor, the force sensor, the gyroscope, and the acceleration sensor are all worn on the object to be identified;
[0034] A feature selection module, used to select target feature data from each of the raw sensor data based on a preset sliding window;
[0035] The action recognition module is used to input the target feature data into a preset target action recognition model, and output the current action recognition result corresponding to the object to be recognized through the target action recognition model.
[0036] In addition, to achieve the above-mentioned purpose, the present application also proposes an electronic device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the action recognition method based on multi-source data as described above.
[0037] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the action recognition method based on multi-source data as described above are implemented.
[0038] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the action recognition method based on multi-source data as described above.
[0039] The present application proposes a method for action recognition based on multi-source data, firstly extracting raw sensor data from electromyographic sensors, force sensors, gyroscopes and acceleration sensors respectively, wherein the electromyographic sensors, force sensors, gyroscopes and acceleration sensors are all worn on the object to be identified, then selecting target feature data from each of the raw sensor data based on a preset sliding window, finally inputting the target feature data into a preset target action recognition model, and outputting the action recognition result corresponding to the object to be identified through the target action recognition model. The technical solution of the present application combines sensor data corresponding to multiple sensors including electromyographic sensors, force sensors, gyroscopes and acceleration sensors, and selects the target feature data therefrom for action recognition. Compared with the traditional action recognition solution based on a single sensor, it can capture more comprehensive and multi-dimensional action details of the object to be identified, overcomes the defect that different sensors have different sensitivities to actions, and can comprehensively and accurately identify the actions of the object to be identified by integrating the data of multiple sensors, thereby improving the accuracy of action recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.
[0041] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0042] Figure 1 Schematic diagram of the process of the action recognition method based on multi-source data in an embodiment of the present application;
[0043] Figure 2 Schematic diagram of data processing flow in the training phase and prediction phase of the action recognition model in the embodiment of the present application;
[0044] Figure 3 A schematic diagram of a data processing flow for transfer learning during the training of an action recognition model in an embodiment of the present application;
[0045] Figure 4 This is a schematic diagram of the composition structure of the motion recognition device based on multi-source data in an embodiment of the present application;
[0046] Figure 5 Schematic diagram of the device structure of the hardware operating environment involved in the action recognition method based on multi-source data in the embodiment of the present application.
[0047] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0048] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.
[0049] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.
[0050] At present, motion recognition systems still face some significant problems. First, traditional methods often rely on a single type of sensor (such as accelerometers or gyroscopes), and this single data source has limitations in capturing complex motion information. Due to the different sensitivities of different sensors to motion, a single sensor may not be able to fully and accurately capture all the details of the motion, thus affecting the overall performance of the system. Secondly, the performance of existing motion recognition systems in cross-user scenarios is also insufficient. Since each user's motion patterns and physiological characteristics are different, models trained based on single user data usually perform poorly on other users. New users usually lack sufficient labeled data to train accurate recognizers, which limits the general applicability and practicality of the system. In order to overcome the above-mentioned defects, the embodiment of the present application proposes a motion recognition method that utilizes multi-sensor fusion data and transfer learning strategies.
[0051] In order to achieve the above objectives, the present invention provides a method for action recognition based on multi-source data. Figure 1 This is a flow chart of a first embodiment of a method for motion recognition based on multi-source data of the present application. The method for motion recognition based on multi-source data includes:
[0052] Step S10, extracting raw sensor data from the electromyographic sensor, the force sensor, the gyroscope, and the acceleration sensor, respectively, wherein the electromyographic sensor, the force sensor, the gyroscope, and the acceleration sensor are all worn on the object to be identified;
[0053] It should be noted that the embodiments of the present application can be applied to a wearable motion recognition system or device, which includes multiple types of sensors that can be worn on an object to be identified (such as a human body) to collect sensor data of multiple dimensions of the object to be identified, and the sensor data includes a large number of time domain features and frequency domain features to characterize the motion of the object to be identified. Exemplarily, the multiple types of sensors may include surface electromyography sensors (set on the legs or upper arms), force sensors, gyroscopes, and acceleration sensors, etc., which are distributed throughout the various positions of the object to be identified to comprehensively and accurately collect its motion information, so as to accurately identify the current action.
[0054] Step S20, selecting target feature data from each raw sensor data based on a preset sliding window;
[0055] In the process of action recognition, the collection of the above various sensor signals is a continuous process, and the data required in the process of action recognition is limited, so a sliding window can be set in advance to collect sensor data within a certain time range as target feature data. It is understandable that in actual scenes, the duration of different actions is different, and the duration of the sliding window can also be set according to the actual scene requirements. For example, the duration of detecting actions such as standing up and squatting can be 1-2s, and continuous actions such as walking and running may require a longer window time to capture complete action features.
[0056] On the other hand, when selecting feature data, the original sensor data needs to be streamlined to remove irrelevant and redundant features. Specifically, in step S20, a feature selection algorithm can be applied to optimize the feature space and select feature data with high correlation from the original sensor data as target feature data.
[0057] Among them, the feature selection algorithm may include: variance threshold method, which specifically calculates the variance of each feature and compares it with a preset threshold, removes features with variances less than the threshold to reduce the number of features. This method is suitable for removing features that vary little in the data set because they may contribute less to the model1; univariate feature selection method, which specifically evaluates the relationship between each feature and the target variable by performing statistical tests on each feature separately, such as chi-square test, F test, etc., and selects features that are significantly related to the target variable. This method determines which features should be retained based on the results of statistical tests1; recursive feature elimination (RFE): by recursively training the model, the least important features are removed each time until the predetermined number of features is reached. This method evaluates the importance of features through model training and gradually reduces the number of features1. It should be noted that when selecting target feature data from raw sensor data, the above various feature selection algorithms can be applied in one or more combinations.
[0058] Step S30, inputting the target feature data into a preset target action recognition model, and outputting the action recognition result corresponding to the object to be recognized through the target action recognition model.
[0059] Among them, the target action recognition model is a pre-trained neural network model, which is trained by labeled data (including sensor data and labeled action recognition results) in this field or similar fields. It can predict the most likely action recognition result of the object to be identified based on the input target feature data. For example, the action recognition result can be an action type, such as standing up, squatting, walking, running, jumping, etc., or it can be specific motion data, such as detailed data such as the position coordinates, movement speed, movement amplitude, etc. of the moving parts of the upper limbs, lower limbs, etc.
[0060] Based on the motion recognition results obtained above, accurate human motion data can be provided in the fields of health monitoring, motion analysis, and human-computer interaction, thereby improving the performance of the motion recognition system and bringing a better user experience.
[0061] In a feasible embodiment, the action recognition method based on multi-source data in the embodiment of the present application includes a training phase and a prediction phase. Figure 2 In the training phase, the original sensor data of the data set is first divided into segments using a sliding window, and then the feature analysis step is entered. Specifically, feature S1 can be extracted from the surface electromyography sensor of the leg, feature S2 can be extracted from the force sensor, feature S3 can be extracted from the gyroscope, and feature S4 can be extracted from the acceleration sensor. Then feature selection is performed, and action recognition training is performed after selecting features based on S1, S2, S3, and S4 to obtain a trained action recognition model. In the prediction phase, the test set is first segmented, and after completing feature selection, the target feature data obtained by action selection is input into the trained action recognition model to complete the action prediction step.
[0062] In the embodiment of the present application, a method for action recognition based on multi-source data is proposed. First, the raw sensor data are extracted from the electromyographic sensor, the force sensor, the gyroscope and the acceleration sensor, respectively, wherein the electromyographic sensor, the force sensor, the gyroscope and the acceleration sensor are all worn on the object to be identified. Then, based on a preset sliding window, target feature data is selected from each of the raw sensor data. Finally, the target feature data is input into a preset target action recognition model, and the action recognition result corresponding to the object to be identified is output through the target action recognition model. The technical solution of the present application combines the sensor data corresponding to multiple sensors including electromyographic sensors, force sensors, gyroscopes and acceleration sensors, and selects the target feature data therefrom for action recognition. Compared with the traditional action recognition solution based on a single sensor, it can capture more comprehensive and multi-dimensional action details of the object to be identified, overcome the defect that different sensors have different sensitivities to actions, and can comprehensively and accurately identify the actions of the object to be identified by integrating the data of multiple sensors, thereby improving the accuracy of action recognition.
[0063] Further, in a feasible embodiment, the original sensor signal includes electromyographic signal data, force signal data, gyroscope data and acceleration data;
[0064] After the step of respectively extracting raw sensor data from the electromyographic sensor, the force sensor, the gyroscope, and the acceleration sensor, the method further comprises:
[0065] Step S11, rectifying and integrating the raw sensor data corresponding to the electromyographic sensor to obtain pre-processed electromyographic signal data;
[0066] Step S12, smoothing the raw sensor data corresponding to the force sensor, the gyroscope and the acceleration sensor respectively through a preset filter to obtain pre-processed force signal data, gyroscope data and acceleration data;
[0067] Step S13, based on the preprocessed electromyographic signal data, force signal data, gyroscope data and acceleration data, execute the step of selecting target feature data from each raw sensor data based on the preset sliding window.
[0068] In order to capture kinematic and physiological signals, force sensors, gyroscopes, acceleration sensors and surface electromyography signal sensors are used in the embodiments of the present application. Among them, sEMG (surface electromyography) signals need to be pre-processed by rectification and integration operations, and the raw sensor data of other sensors are smoothed by filters. Finally, the processed signals are subsequently transmitted to the server for further analysis.
[0069] Exemplarily, the pre-processing of sEMG (surface electromyography) signals such as rectification and integration operations is to improve signal quality, extract more effective information, and provide a high-quality data basis for subsequent analysis. Among them, the rectification operation can include flipping the negative value part in the sEMG signal to a positive value, so that the entire signal becomes a full-wave rectified signal. This step helps to eliminate negative fluctuations in the signal, making the signal more stable and convenient for subsequent processing. Specifically, the rectification operation generally includes the following methods: full-wave rectification, multiplying all negative values in the sEMG signal by -1 to make it positive. After such processing, the signal is positive throughout the entire cycle, which is convenient for subsequent feature extraction and analysis; absolute value rectification, directly taking the absolute value of the sEMG signal, the effect is similar to full-wave rectification, but the implementation method is more direct.
[0070] In addition, the integration operation is to accumulate the rectified sEMG signal in time, so as to extract the envelope information of the signal. The integration operation can reflect the intensity and duration of human muscle activity, and is an important means to evaluate muscle activation. Specifically, the integration operation usually includes the following steps: first, select a suitable integration window, and select a suitable integration window size according to the characteristics and analysis requirements of the sEMG signal, wherein the size of the integration window will affect the smoothness and detail retention of the integration result; calculate the integral value, and integrate the rectified sEMG signal within the selected integration window, and the integral value reflects the average intensity or cumulative effect of the signal within the window. In addition, in order to obtain continuous integration results, the integration window can be slid to integrate the entire sEMG signal. The size and step size of the sliding window can be customized according to specific needs.
[0071] On the other hand, commonly used smoothing filters (such as Gaussian filters, ideal low-pass filters, Butterworth low-pass filters and Gaussian low-pass filters, etc.) can be used to smooth the force signal data, gyroscope data and acceleration data to reduce the impact of noise and improve data quality.
[0072] It should be noted that after obtaining the preprocessed electromyographic signal data, force signal data, gyroscope data and acceleration data, when executing step S20, the target feature data is selected based on the preprocessed data, which is equivalent to updating the original sensor data, and executing the step of selecting the target feature data with the updated original sensor data.
[0073] In a feasible embodiment, before the step of inputting the target feature data into a preset target action recognition model, the method further includes:
[0074] Step A10, obtaining source domain sensor data in a source domain and label data corresponding to the source domain sensor data;
[0075] Step A20, converting the source domain sensor data and the target feature data into a common feature space, obtaining training sensor data and updated target feature data corresponding to the source domain sensor data, wherein the updated target feature data is used to predict the action recognition result of the object to be recognized;
[0076] Step A30: training a preset initial action recognition model based on the training sensor data and the label data to obtain a target action recognition model.
[0077] The existing motion recognition systems also have shortcomings in their performance in cross-user scenarios. It is understandable that, since each user's motion patterns and physiological characteristics are different, models trained based on single user data usually perform poorly on other users. Moreover, new scenarios or new users usually lack sufficient labeled data to train accurate motion recognition models, which limits the general applicability and practicality of the models. Therefore, in the absence of training data and labeled data, the embodiments of the present application propose a method based on transfer learning to train motion recognition models.
[0078] Among them, the source domain sensor data of the source domain is the original sensor data and labeled data of the field close to the target field (i.e., the field to which the present action recognition method is applied), wherein the labeled data is the real action recognition result corresponding to the original sensor, and is used for supervised training to obtain the action recognition model. Since the original sensor data and labeled data in the target field are relatively scarce, the source domain sensor data and the corresponding labeled data are introduced in the embodiment of the present application to train the action recognition model, so that the action recognition result of the object to be identified in the current situation can be predicted according to the target feature data corresponding to the original sensor data in the target field, thereby overcoming the defect of insufficient support for model training due to the lack of target field data and the poor generalization ability of the action recognition model trained by the sensor data in the scene of a single field.
[0079] Specifically, refer to Figure 3 First, the source domain sensor data H1 (including data D1) and the target domain data H2 (including data D2) are converted into a common feature space (i.e., the new feature space) through transfer learning (e.g., σ(D1), δ(D2)), and then the action recognizer cls is trained based on the converted source domain sensor data H1. a′ (equivalent to the action recognition model) is used to predict the corresponding action recognition results based on the converted target feature data (test data).
[0080] For example, in the new feature space, the distance between the source domain and the target domain can be minimized as much as possible, which can be expressed as
[0081] Figure 3 The left box shows the source domain sensor data, target domain data, and action recognizer cls before transfer learning. a, the right box refers to the step of training the initial action recognition model (equivalent to the action recognition model before training) after transfer learning, and finally the target action recognition model is obtained, which can be used to predict the action recognition result of the target feature data in step S30. Among them, on the basis of steps A10 to A30, the target feature data also needs to be updated, and the target feature data after conversion to the common feature space is used for action recognition. This is because the action recognition model is trained by the source domain sensor data (i.e., the training feature data) after the distance with the target feature data in the common feature space is minimized. The use of the updated target feature data can make the action recognition result output by the action recognition model more accurate.
[0082] In particular, in a feasible embodiment, when the action recognition of the object to be recognized is performed for the first time, if there is no sensor data and description data in the target domain for training the action recognition model, the pre-stored source domain sensor and label data can be called to combine with the currently acquired target feature data for transfer learning, and a target action recognition model suitable for the object to be recognized can be trained to recognize and predict the action of the object to be recognized. Therefore, the technical solution of the embodiment of the present application can be used across scenarios and is also suitable for the scenario of the first use by new users, which effectively improves the universal adaptability and applicability of the action recognition system.
[0083] Further, in a feasible embodiment, the common feature space is a reproducing kernel Hilbert space;
[0084] The step of converting the source domain sensor data and the target feature data into a common feature space to obtain training sensor data corresponding to the source domain sensor data comprises:
[0085] Step A21, mapping the source domain sensor data and the target feature data to the reproducing kernel Hilbert space;
[0086] Step A22, with the goal of minimizing the distance between the source domain sensor data and the target feature data in the reproducing kernel Hilbert space, updates the source domain sensor data and the target feature data, and determines the training sensor data corresponding to the source domain sensor data and the updated target feature data.
[0087] The embodiment of the present application provides a method for transfer learning based on transfer component analysis, specifically, transfer component analysis (TCA). The TCA method aims to find a high-dimensional space so that the data distribution of the source domain and the target domain in the space is closer. This helps to solve the problem of decreased model generalization ability caused by differences in data distribution between different domains. And by minimizing the difference between the source domain and the target domain, TCA can significantly improve the effect of cross-domain learning, so that the action recognition model trained in the source domain can better adapt to the data distribution of the target domain, thereby improving the accuracy of the action recognition model in predicting the target feature data in the target domain.
[0088] Specifically, the kernel method is first used to map the source domain sensor data and the target feature data to a high-dimensional reproducing kernel Hilbert space (RKHS), and a low-dimensional embedding is found in this space to minimize the distribution difference between the source domain sensor data and the target feature data in the embedding space. The optimization process needs to consider optimizing the distance metric, which is mainly achieved by optimizing the maximum mean difference (MMD) distance. Among them, MMD is an effective method to measure the difference between two data distributions. In the TCA process, the difference between the source domain and the target domain can be reduced by minimizing the MMD distance, and finally the training sensor data and the updated target feature data are obtained.
[0089] It should be noted that the training sensor data and the updated target feature data are respectively used to train the initial action recognition model and are input into the target action recognition model to predict the current action recognition result of the object to be recognized.
[0090] In another feasible embodiment, the step of converting the source domain sensor data and the target feature data into a common feature space to obtain training sensor data corresponding to the source domain sensor data includes:
[0091] Step A23, initializing a transformation matrix, transforming the source domain sensor data and the target feature data into a common feature space through the transformation matrix, to obtain updated source domain sensor data and updated target domain data;
[0092] Step A24, iteratively optimizing the transformation matrix with the goal of minimizing the marginal distribution and conditional distribution between the updated source domain sensor data and the updated target domain data, and updating the updated source domain sensor data and the updated target domain data based on the optimized transformation matrix;
[0093] Step A25, when the preset convergence condition is met, stop the iterative optimization, set the current updated source domain sensor data as the training sensor data, and set the current updated target domain data as the updated target feature data.
[0094] The embodiment of the present application also provides a method for transfer learning based on joint distribution adaptation (JDA). This is a distribution adaptation method that considers marginal and conditional probability distributions. The main idea of JDA is to adapt the source domain sensor data to the target domain data by projecting the source domain sensor data and the target domain data into a subspace.
[0095] Exemplarily, the step of transforming the source domain sensor data and the target feature data based on joint distribution adaptation includes: first, initializing a transformation matrix A for transforming the source domain sensor data and the target feature data to a new space to obtain updated source domain sensor data xs and updated target domain data xt, with the purpose of reducing the difference in joint probability distribution between the source domain and the target domain; then, performing marginal distribution adaptation, by optimizing the marginal distribution difference between the updated source domain sensor data xs and the updated target domain data xt, with the goal of minimizing the marginal distribution and conditional distribution, adjusting the transformation matrix A so that the transformed source domain sensor data ATxs and the target domain data ATxt are as close as possible in marginal distribution; then, performing conditional distribution adaptation, first, using the classifier trained with the updated source domain sensor data to predict the target domain data to obtain a pseudo label y^t, and then, by optimizing the conditional distribution difference between the source domain sensor data xs and the target domain data xt with the pseudo label, further adjusting the transformation matrix A; repeating the aforementioned transformation matrix adjustment and optimization steps until the convergence condition is met or the preset number of iterations is reached; and setting the latest updated source domain sensor data and updated target domain data as training sensor data and target feature data, respectively.
[0096] In the above steps, the aforementioned marginal distribution adaptation and conditional distribution adaptation are usually achieved by minimizing the maximum mean difference (MMD) distance. In addition, a regularization term can be introduced to prevent overfitting and ensure that the transformed data maintains a certain variance. The resulting transformed source domain sensor data will be used to train the action recognition model in order to accurately predict the updated target feature data.
[0097] In another feasible embodiment, before the step of inputting the target feature data into a preset target action recognition model, the method further includes:
[0098] Step B10, obtaining source domain sensor data in a source domain and label data corresponding to the source domain sensor data;
[0099] Step B20, determining the optimal subspace dimension based on the source domain sensor data and the target feature data;
[0100] Step B30, constructing a geodesic streamline kernel function between the source domain sensor data and the target feature data;
[0101] Step B40, converting the source domain sensor data to the target domain where the target feature data is located according to the subspace dimension and the geodesic streamline kernel function to obtain training sensor data;
[0102] Step B50, training a preset initial action recognition model based on the training sensor data and the label data to obtain a target action recognition model.
[0103] In the embodiment of the present application, a method for transfer learning based on the idea of geodesic flow kernel (GFK) is provided. The idea of GFK is to first determine the optimal subspace dimension and construct the geodesic, then calculate the geodesic flow kernel function, and finally convert the source domain sensor data to the target domain for model training. Unlike TCA and JDA, the idea of the GFK method is to directly convert the source domain data to the target domain, and then train the recognizer on the converted data and make predictions.
[0104] Exemplarily, firstly, the optimal subspace dimension is selected to determine the optimal dimension d of the embedded domain subspace so as to maintain the maximum consistency between domains while reducing the dimensionality; principal component analysis (PCA) is performed on the data of the source domain S and the target domain T (source domain sensor data and target feature data) respectively to obtain PCA subspaces P_S and P_T; then the source domain sensor data and the target feature data are merged and PCA analysis is performed to obtain the merged subspace P_{S+T}; then the subspace inconsistency measure (SDM) is used to calculate the angles α_d and β_d between the source domain, the target domain and the merged domain, and then the total measure D(d) = 0.5[sinα_d+sinβ_d] is obtained; a greedy algorithm can be used to determine the optimal dimension d, that is, taking as large a d as possible to ensure the number of subspaces, but avoiding orthogonality of the two domain subspaces. In addition, it is also necessary to construct a geodesic streamline kernel function connecting the source domain and the target domain on the Grassmann manifold; specifically, the bases P_S and P_T of the source domain and target domain subspaces are regarded as two points on the manifold, and mapped to two extremes 0 and 1 through the geodesic streamline kernel function Φ; and for any point t in the interval [0,1], calculate the value of Φ(t), that is, Φ(t) = P_SU_1Γ(t)-R_S U_2Σ(t), where R_S is the orthogonal complement of P_S, and U_1, U_2, and ΓΣ are obtained through singular value decomposition (SVD).
[0105] In the process of calculating the geodesic streaming kernel, all points on the geodesic are used to calculate the streaming kernel by integration to capture the smooth changes between the source domain sensor data and the target domain where the target feature data is located. For example, for any point t on the geodesic, the projection Φ(t)^T x of the feature vector x in the subspace Φ(t) at that point is calculated. After accommodation, all the projected points on the geodesic are integrated into a kernel function through integration or other kernel techniques for subsequent classification tasks.
[0106] The calculated geodesic flow kernel is used to convert the source domain sensor data to the target domain, and then the initial action recognition model is constructed based on the training sensor data after conversion to the target domain. The target action recognition model is trained on the training sensor data in the target domain using traditional machine learning methods (such as support vector machine SVM) combined with the geodesic flow kernel. Since the geodesic flow kernel captures the smooth changes from the source domain to the target domain, it can train a target action model with high prediction accuracy. The trained target action recognition model can show good performance in predicting target feature data.
[0107] In summary, in cross-user application scenarios, the transfer learning method mentioned in the aforementioned embodiments can project the source domain sensor data (pre-existing in the system) and the target feature data (acquired new user data) into a shared subspace. Then, the action recognition model will be trained on the basis of the projected source domain sensor data, and the action recognition result of the current identified object will be predicted based on the input projected target feature data. Therefore, the action recognition method based on multi-source data of the embodiment of the present application can adapt to a variety of scenarios, has strong general adaptability and generalization ability, and combines various types of sensor data to accurately identify the actions of the object to be identified, thereby improving the user experience in scenarios such as health monitoring, motion analysis, and human-computer interaction.
[0108] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the motion recognition method based on multi-source data of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.
[0109] The present application also provides a multi-source data action recognition device, referring to Figure 4 , the motion recognition device based on multi-source data includes:
[0110] A data acquisition module 10, for extracting raw sensor data from an electromyographic sensor, a force sensor, a gyroscope, and an acceleration sensor, respectively, wherein the electromyographic sensor, the force sensor, the gyroscope, and the acceleration sensor are all worn on the object to be identified;
[0111] A feature selection module 20, configured to select target feature data from each of the raw sensor data based on a preset sliding window;
[0112] The action recognition module 30 is used to input the target feature data into a preset target action recognition model, and output a current action recognition result corresponding to the object to be recognized through the target action recognition model.
[0113] In one embodiment, the raw sensor signal includes electromyographic signal data, force signal data, gyroscope data and acceleration data; the data acquisition module 10 is further used for:
[0114] The raw sensor data corresponding to the electromyographic sensor is rectified and integrated to obtain pre-processed electromyographic signal data;
[0115] Smoothing the raw sensor data corresponding to the force sensor, the gyroscope, and the acceleration sensor respectively through a preset filter to obtain preprocessed force signal data, gyroscope data, and acceleration data;
[0116] Based on the preprocessed electromyographic signal data, force signal data, gyroscope data and acceleration data, the step of selecting target feature data from each of the raw sensor data based on a preset sliding window is performed.
[0117] In one embodiment, the action recognition module 30 is further used to: obtain source domain sensor data of a source domain and label data corresponding to the source domain sensor data;
[0118] The source domain sensor data and the target feature data are converted into a common feature space to obtain training sensor data and updated target feature data corresponding to the source domain sensor data, wherein the updated target feature data is used to predict the action recognition result of the object to be recognized;
[0119] Based on the training sensor data and the label data, a preset initial action recognition model is trained to obtain a target action recognition model.
[0120] In one embodiment, the common feature space is a reproducing kernel Hilbert space; the action recognition module 30 is further used for:
[0121] Mapping the source domain sensor data and the target feature data to a reproducing kernel Hilbert space;
[0122] With the goal of minimizing the distance between the source domain sensor data and the target feature data in the reproducing kernel Hilbert space, the source domain sensor data and the target feature data are updated, and the training sensor data corresponding to the source domain sensor data and the updated target feature data are determined.
[0123] In one embodiment, the action recognition module 30 is further used for:
[0124] Initializing a transformation matrix, transforming the source domain sensor data and the target feature data into a common feature space through the transformation matrix to obtain updated source domain sensor data and updated target domain data;
[0125] Iteratively optimizing the transformation matrix with the goal of minimizing the marginal distribution and conditional distribution between the updated source domain sensor data and the updated target domain data, and updating the updated source domain sensor data and the updated target domain data based on the optimized transformation matrix;
[0126] When the preset convergence condition is met, the iterative optimization is stopped, the current updated source domain sensor data is set as the training sensor data, and the current updated target domain data is set as the updated target feature data.
[0127] In one embodiment, the action recognition module 30 is further used for:
[0128] Acquire source domain sensor data of a source domain and label data corresponding to the source domain sensor data;
[0129] Determining an optimal subspace dimension based on the source domain sensor data and the target feature data;
[0130] Constructing a geodesic streamline kernel function between the source domain sensor data and the target feature data;
[0131] Converting the source domain sensor data to the target domain where the target feature data is located according to the subspace dimension and the geodesic streamline kernel function to obtain training sensor data;
[0132] Based on the training sensor data and the label data, a preset initial action recognition model is trained to obtain a target action recognition model.
[0133] The motion recognition device based on multi-source data provided by the present application adopts the motion recognition method based on multi-source data in the above-mentioned embodiment, which can solve the technical problems that the current motion recognition scheme has limitations and the motion recognition accuracy is low. Compared with the prior art, the beneficial effects of the motion recognition device based on multi-source data provided by the present application are the same as the beneficial effects of the motion recognition method based on multi-source data provided by the above-mentioned embodiment, and the other technical features of the motion recognition device based on multi-source data are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.
[0134] The present application provides an electronic device, which includes: 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 so that the at least one processor can execute the action recognition method based on multi-source data in the above-mentioned embodiment 1.
[0135] Reference below Figure 5 , which shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present application. The electronic devices in the embodiments of the present application may include but are not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 5 The electronic device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.
[0136] like Figure 5 As shown, the electronic device may include a processing device 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the electronic device are also stored. The processing device 1001, ROM1002, and RAM1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the electronic device to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows an electronic device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or have alternatively.
[0137] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.
[0138] The electronic device provided by the present application adopts the motion recognition method based on multi-source data in the above-mentioned embodiment, which can solve the technical problems that the current motion recognition scheme has limitations and the low accuracy of motion recognition. Compared with the prior art, the beneficial effects of the electronic device provided by the present application are the same as the beneficial effects of the motion recognition method based on multi-source data provided in the above-mentioned embodiment, and the other technical features in the electronic device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.
[0139] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.
[0140] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
[0141] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, and the computer-readable program instructions are used to execute the action recognition method based on multi-source data in the above-mentioned embodiment.
[0142] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.
[0143] The computer-readable storage medium may be included in the electronic device, or may exist independently without being installed in the electronic device.
[0144] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by an electronic device, the electronic device: extracts raw sensor data from an electromyographic sensor, a force sensor, a gyroscope and an acceleration sensor respectively, wherein the electromyographic sensor, the force sensor, the gyroscope and the acceleration sensor are all worn on the object to be identified; based on a preset sliding window, selects target feature data from each of the raw sensor data; inputs the target feature data into a preset target action recognition model, and outputs the action recognition result corresponding to the object to be identified through the target action recognition model.
[0145] Computer program code for performing the operations of the present application may be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0146] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.
[0147] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.
[0148] The readable storage medium provided by the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned action recognition method based on multi-source data, and can solve the technical problems that the current action recognition scheme has limitations and the action recognition accuracy is low. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present application are the same as the beneficial effects of the action recognition method based on multi-source data provided by the above-mentioned embodiment, and will not be repeated here.
[0149] The present application also provides a computer program product, including a computer program, which implements the steps of the above-mentioned action recognition method based on multi-source data when executed by a processor.
[0150] The computer program product provided by the present application can solve the technical problems that the current motion recognition scheme has limitations and the motion recognition accuracy is low. Compared with the prior art, the beneficial effects of the computer program product provided by the present application are the same as the beneficial effects of the motion recognition method based on multi-source data provided by the above embodiment, which will not be repeated here.
[0151] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.
Claims
1. A method for action recognition based on multi-source data, characterized in that: The action recognition method based on multi-source data includes: Extracting raw sensor data from an electromyographic sensor, a force sensor, a gyroscope, and an acceleration sensor, respectively, wherein the electromyographic sensor, the force sensor, the gyroscope, and the acceleration sensor are all worn on an object to be identified; Based on a preset sliding window, selecting target feature data from each of the raw sensor data; The target feature data is input into a preset target action recognition model, and the action recognition result corresponding to the object to be recognized is output through the target action recognition model.
2. The action recognition method based on multi-source data according to claim 1, characterized in that: The raw sensor signals include electromyographic signal data, force signal data, gyroscope data and acceleration data; After the step of respectively extracting raw sensor data from the electromyographic sensor, the force sensor, the gyroscope, and the acceleration sensor, the method further comprises: The raw sensor data corresponding to the electromyographic sensor is rectified and integrated to obtain pre-processed electromyographic signal data; Smoothing the raw sensor data corresponding to the force sensor, the gyroscope, and the acceleration sensor respectively through a preset filter to obtain preprocessed force signal data, gyroscope data, and acceleration data; Based on the preprocessed electromyographic signal data, force signal data, gyroscope data and acceleration data, the step of selecting target feature data from each of the raw sensor data based on a preset sliding window is performed.
3. The action recognition method based on multi-source data according to claim 1, characterized in that: Before the step of inputting the target feature data into a preset target action recognition model, the method further includes: Acquire source domain sensor data of a source domain and label data corresponding to the source domain sensor data; The source domain sensor data and the target feature data are converted into a common feature space to obtain training sensor data and updated target feature data corresponding to the source domain sensor data, wherein the updated target feature data is used to predict the action recognition result of the object to be recognized; Based on the training sensor data and the label data, a preset initial action recognition model is trained to obtain a target action recognition model.
4. The method for motion recognition based on multi-source data as claimed in claim 3, characterized in that: The common feature space is a reproducing kernel Hilbert space; The step of converting the source domain sensor data and the target feature data into a common feature space to obtain training sensor data corresponding to the source domain sensor data comprises: Mapping the source domain sensor data and the target feature data to a reproducing kernel Hilbert space; With the goal of minimizing the distance between the source domain sensor data and the target feature data in the reproducing kernel Hilbert space, the source domain sensor data and the target feature data are updated, and the training sensor data corresponding to the source domain sensor data and the updated target feature data are determined.
5. The method for motion recognition based on multi-source data as claimed in claim 3, characterized in that: The step of converting the source domain sensor data and the target feature data into a common feature space to obtain training sensor data corresponding to the source domain sensor data comprises: Initializing a transformation matrix, transforming the source domain sensor data and the target feature data into a common feature space through the transformation matrix to obtain updated source domain sensor data and updated target domain data; Iteratively optimizing the transformation matrix with the goal of minimizing the marginal distribution and conditional distribution between the updated source domain sensor data and the updated target domain data, and updating the updated source domain sensor data and the updated target domain data based on the optimized transformation matrix; When the preset convergence condition is met, the iterative optimization is stopped, the current updated source domain sensor data is set as the training sensor data, and the current updated target domain data is set as the updated target feature data.
6. The method for motion recognition based on multi-source data as claimed in claim 1, characterized in that: Before the step of inputting the target feature data into a preset target action recognition model, the method further includes: Acquire source domain sensor data of a source domain and label data corresponding to the source domain sensor data; Determining an optimal subspace dimension based on the source domain sensor data and the target feature data; Constructing a geodesic streamline kernel function between the source domain sensor data and the target feature data; Converting the source domain sensor data to the target domain where the target feature data is located according to the subspace dimension and the geodesic streamline kernel function to obtain training sensor data; Based on the training sensor data and the label data, a preset initial action recognition model is trained to obtain a target action recognition model.
7. A motion recognition device based on multi-source data, characterized in that: The motion recognition device based on multi-source data comprises: A data acquisition module, used to extract raw sensor data from an electromyographic sensor, a force sensor, a gyroscope, and an acceleration sensor, respectively, wherein the electromyographic sensor, the force sensor, the gyroscope, and the acceleration sensor are all worn on the object to be identified; A feature selection module, used to select target feature data from each of the raw sensor data based on a preset sliding window; The action recognition module is used to input the target feature data into a preset target action recognition model, and output the current action recognition result corresponding to the object to be recognized through the target action recognition model.
8. An electronic device, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the action recognition method based on multi-source data as described in any one of claims 1 to 6.
9. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the action recognition method based on multi-source data as described in any one of claims 1 to 6 are implemented.
10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the action recognition method based on multi-source data as claimed in any one of claims 1 to 6 are implemented.
Citation Information
Cited By
Gesture and force synchronous recognition method and device for realizing nerve driving through feature enhancement based on VMD denoising processing
CN121879586A
