Research On Human Lower Limb Movement Recognition System Based On The Fusion Of SEMG And IMU

Human motion recognition technology has become a research hotspot in related fields,which is widely used in exoskeleton control,medical rehabilitation and many other fields.Human electrophysiological signals and motion signals contain rich information.Thanks to the rapid development of hardware technology,these information can be easily collected,processed and analyzed.Human action recognition methods based on these information have been widely studied by researchers,but most of the current recognition methods ignore the association information between human muscles and joints,which limits the further improvement of recognition accuracy.On the other hand,with the increasing demand of human-computer interaction,the action recognition system which can be applied to portable mobile platform has broad prospects.All around these issues,this paper takes sEMG signal and IMU signal as data sources,and introduces graph convolution neural network into the research of action recognition method by constructing the topological graphs of sEMG signal and IMU signal.Based on the action recognition method,this paper builds the action recognition software system for mobile platform,and completes the recognition of 9 lower limb actions,including standing up,sitting down,walking,running,climbing stairs,descending stairs,turning left,turning right and jumping.The main contents of this paper are as follows(1)The acquisition of human lower limb movement signal and the software system of action recognition.In this paper,combined with the relevant research theory,the collection points of sEMG signal and IMU signal of human lower limbs are determined,and the recognition software platform is built by raspberry pie,EMG signal conditioning unit,analog-to-digital conversion module and inertial measurement unit.The recognition software is written in C/C++language and based on Linux multithreading.Its functions include sEMG signal and IMU signal acquisition,data synchronization,active segment extraction and action recognition.(2)Data preprocessing.Aiming at the noise contained in SEMG signal,this paper uses Butterworth filter,notch filter and spectrum interpolation to denoise it.The noise content of IMU signal after hardware filtering is less,so in order to retain its effective information to the greatest extent,software filtering and noise reduction is not done.In this paper,combined with sliding window and dual threshold,an active segment detection method is implemented to complete the data active segment detection.(3)The construction of topological graph and characteristic matrix.In this paper,the muscle topological network graph of sEMG signal is constructed by calculating the feature correlation among the channels of sEMG signal,and the topological skeleton graph of IMU signal is constructed by the human skeleton map applied to posture recognition.The constructed topology graphs contain the connection information between muscle and joint acquisition points.Through feature engineering,the characteristic matrix of each graph network of two kinds of signals is constructed.(4)The construction of recognition algorithm and the verification experiments.Based on the third generation GCN theory and topological graphs,this paper constructs a GCN network model which can be applied to sEMG and IMU signals.Combined with the difference of the two signals,this paper adopts parallel GCN branch architecture,and selects multi-layer fully connected network as the fusion strategy to complete the recognition of 9 kinds of daily movements of lower limbs.Compared with the recognition results of many traditional recognition algorithms,the parallel GCN fusion network architecture takes the connection information between human muscles and joints into account,and has higher recognition accuracy.Through the test of real-time recognition verification experiment,the action recognition system based on the recognition architecture achieves the expected recognition results.