Research And Implementation Of Gesture Recognition Algorithm For Wearable Devices

With the advancement of semiconductor technology,computer hardware capabilities have been greatly improved,and many new research technologies such as somatosensory games and virtual reality have been developed rapidly.Traditional human-computer interaction methods are increasingly unable to meet the requirements of these technologies,and gestures,as an easy-to-learn interactive method,can give users a higher sense of immersion.At the same time,the wearable gesture recognition device can overcome the influence of the external environment such as light occlusion and achieve higher recognition accuracy.Traditional gesture interaction research work is mostly focused on the research of gesture recognition algorithms to improve recognition accuracy,ignoring the differences in user wearing styles and biological signals between different users,and not paying attention to the robustness of models migrating to new devices.As a result,the accuracy of wearable devices in actual use does not match expectations,which also limits the practical application of wearable gesture recognition devices.At the same time,traditional gesture recognition is mostly realized by machine learning methods,which require a large amount of annotation data to achieve higher algorithm accuracy,which requires a lot of time to collect data,and the time cost is too high.In view of the above problems,this article focuses on the research of gesture recognition algorithms for wearable devices,and verifies the robustness of the algorithm through the gesture acquisition device designed in this article.The main research contents of this article include:1.Design a wearable gesture signal acquisition device to study the robustness of the model migration to the new device.The equipment includes a data acquisition module,a control module,and a power supply module.The data acquisition module uses 8 groups of dry electrode EMG sensors to collect EMG signals,and inertial sensors collect inertial signals.The equipment control core chooses STM32F407 to meet the equipment interface requirements.It is powered by a lithium battery to meet the power supply requirements of the entire device.The final device realizes the collection and transmission of gesture signals.2.Research on a highly robust gesture recognition algorithm based on deep transfer learning,using EMG and inertial signals to identify dynamic and static gestures.First,after detecting the active segment of the gesture signal based on the sliding window threshold method,the preset gesture threshold is used to determine whether the electromyographic electrode has a large range of electrode offset,and the electrode return is realized through the data rearrangement algorithm.Then the dynamic time warping algorithm is used to realize the classification of dynamic and static gestures,and complete the recognition of dynamic gestures.If it is a dynamic gesture,the recognition result is output;otherwise,a static gesture is recognized based on deep transfer learning.Design and train the CNN model on the Ninapro DB5 data set.After curing some network layers of the CNN model,train and update the uncured network layer weights of the CNN model by collecting a small amount of new data to achieve high robust recognition of new data.The highly robust gesture recognition algorithm implemented in this paper has an average recognition accuracy of 95.7% for 6 dynamic gestures,an accuracy of 100% for classification of dynamic and static gestures,and an average recognition accuracy of 88.69% for 12 static gestures.In view of the small-scale data offset,the average recognition accuracy of the algorithm is increased by 15%;for the influence of different forces,the average recognition accuracy of the algorithm is increased by 4%.Verify the robustness of the gesture recognition algorithm based on deep transfer learning under the influence of electrode displacement and different forces.3.Research the multi-stream fusion gesture recognition algorithm based on GAF conversion to preserve the temporal correlation between data.Both the 8-channel EMG signal and the 9-channel inertial signal are processed as one-dimensional time series to realize the two-dimensionalization of each channel of the gesture signal based on GAF conversion.After multi-stream fusion and feature splicing of the two-dimensional multi-channel gesture signal,the simultaneous discrimination of dynamic gestures and static gestures is completed.In the end,the multi-stream fusion gesture recognition algorithm based on GAF conversion proposed in this paper has an effective recognition rate of over 86.67% for 18 gestures(including static and dynamic gestures),and the recognition accuracy rate after electrode offset only drops by 2.2%.The accuracy of gesture recognition under the force is only reduced by 2.76%,which verifies the high robustness of the multi-stream fusion algorithm based on GAF conversion.