Research On Gesture Recognition Method Based On Feature Fusion Of Multi-mode Information

Surface EMG is one of the most valuable bioelectrical signals.Often used to measure the activity state of muscles,it can provide useful information for many research in rehabilitation medicine and intelligent prosthetics.However,the EMG signal is a very weak electrical signal,which makes the gesture recognition effect unsatisfactory.When used in combination with other types of signals,these shortcomings can be compensated.The acceleration signal can accurately reflect the direction,speed and displacement of the movement.The gyroscope is a sensor that can detect the angular velocity of the object when it is moving.These are ideal input signals.Therefore,based on the surface electromyography signal(s EMG),combined with the acceleration signal(ACC)and the gyroscope signal(GYRO)to characterize the motion characteristics of gestures,and carry out gesture action recognition research.In this paper,the non-invasive wearable acquisition device MYO armband is used to collect surface EMG signals,acceleration signals and gyroscope signals at the position of 7 subjects near the elbow joint of the forearm.In the process of signal acquisition,a large number of non-action segment signals and noise signals are included,which reduces the accuracy of gesture recognition.Therefore,it is necessary to filter the signal to eliminate noise interference.On the basis of completing the signal filtering,the signal data of the action segment is extracted.In order to realize the recognition of gesture actions,it is necessary to perform feature extraction on the three signals corresponding to each valid activity time period.However,a feature is usually only sensitive to the change of some features of the signal,but not to the change of other features,and the recognition rate of a classifier trained based on a single feature is not high.In order to solve the above problems,the paper proposes a feature-level fusion method,that is,a weighted global canonical correlation analysis fusion algorithm,which directly extracts the original features through a custom feature set,and then performs global canonical correlation analysis on the feature collection to obtain a generalized canonical projection vector.The final feature space is created by weighting the projection vector,which is used as input for classification recognition in the classifier.Considering that the extreme learning machine(ELM)can train faster than the traditional neural network while ensuring the gesture recognition rate,the extreme learning machine is used to build the gesture action classification model,but some parameters of the extreme learning machine are randomly generated.,resulting in a low accuracy of the prediction model.Therefore,the genetic algorithm is used to optimize the ELM network.The experimental results show that the recognition rate of 14 gestures using the extreme learning machine model is 90.82%.Under the same conditions,the recognition rate of the GA-ELM classification model is 94.89%,which is 4.07% higher than that of the ELM model.This shows that the classification accuracy of the optimized network is higher,which further verifies the feasibility and efficiency of the GA-ELM gesture classification model.