Gesture Recognition Based On MEMS-IMU

There are two main types of three-dimensional gesture recognition devices: machine vision and motion sensors.Motion sensors have unique advantages over machine vision,such as easy-to-use applicability and device portability.With the rapid development of wearable devices such as smartphones,smartwatches and healthmonitoring bracelets,MEMS-IMU sensors are getting stronger and cheaper.Therefore,this paper will study the gesture recognition based on MEMS-IMU.The methods of gesture recognition using MEMS-IMU as gesture data sensors mainly include motion trajectory reconstruction,template matching,hidden Markov model,gesture data statistics and artificial neural network.In this paper,artificial neural network and deep learning are applied to gesture recognition.The recurrent neural network(RNN)using long short term memory module(LSTM)can effectively solve the problem of inconsistent gesture length,and it can also integrate acceleration,gyroscope and attitude information for gesture recognition.This paper defines and collects eight one-hand gestures and eight two-hand gestures,all of them are dynamic gestures.Use the sensor on the right hand to collect one-handed gestures,such as left,right,circle,etc.The two-handed gesture requires the cooperation of both hands,the left and right hands are each wearing a sensor to collect gesture.Eight kinds of traffic gestures are used in this paper.After the gesture collection is completed,the coordinate system transformation needs to be completed,and then some column preprocessing processes are performed to make the gesture data meet the input requirements of the LSTM-RNN model.Appropriate preprocessing can improve the performance of the gesture recognition system.The transformed gesture data is subjected to sample choose,normalization,filtering,and motion segmentation.The pre-processed gesture data was input into LSTM-RNN for training,and different hidden nodes were tried.The optimal recognition rate of one-hand gestures is 97.22%,and the optimal recognition rate of two-hands gestures is 96.98%.Finally,it is verified by experiments that the integrity of gesture data is very important for the recognition of complex gestures.The recognition rate of two-hand gesture model using acceleration and gyroscope data is very low.After adding quaternion information,the recognition rate is greatly improved.The experiment proves that the gesture recognition method used in this paper has excellent performance and has great advantages in complex gesture recognition.