| With the vigorous development of intelligent technology,it has been widely applied in industries,healthcare,sports,entertainment,and other fields.The requirement for the convenience of human-computer interaction is higher and the types of interaction are more diverse,so human motion recognition had become a hot research topic in this field.This project conducts research on human motion recognition technology based on multi-dimensional MEMS sensors,and the specific research work is as follows:Firstly,we analyzed the structure and motion mode of human body,summarized the motion mode of human joints,and constructed a simple human joint skeleton model.Compared the advantages and disadvantages of several common human motion capture devices.The MEMS acquisition core chip integrated by acceleration,gyroscope and geomagnetic sensor was selected,and the multi node layout was used to collect the motion of 60 subjects,and the human behavior data set was established.Next,we conducted comparative experimental analysis on two general attitude calculation algorithms,and it was found that Euler angle method can intuitively represent changes in attitude angle compared to quaternion method;Next,the quartile distance method is used to remove outliers from the sensor data,and an ⅡR Chebyshev Ⅰ low-pass filter is constructed to smooth and normalize the sensor data;We extracted time-domain features such as mean and median,as well as frequency-domain features such as amplitude mean and frequency variance.Principal component analysis was used to reduce the dimensionality of the features for subsequent support vector machine action classification.Then,support vector machine(SVM)classifier was used to classify and recognize 18 kinds of human daily actions,and two parameters of fitness maximum penalty factor c and radial basis function kernel parameter g in the SVM model were determined.Two optimization algorithms based on particle swarm optimization(PSO)and genetic algorithm(GA)were compared.By using these two algorithms to find the optimal parameters of the SVM model,the final average recognition result obtained by PSO is better than that obtained by GA,and it was found that the average recognition rate of SVM classifier for simple actions is 94%,and the average recognition rate for complex actions is 88%,indicating that the recognition rate of SVM classifier for complex actions is not good.The reason is the error caused by attitude calculation and the incomplete data caused by dimensional reduction.Finally,in order to improve the recognition rate of complex actions and avoid errors caused by pose calculation,the LSTM network was introduced into the convolutional neural network to solve the problem of insufficient data time dependence of the convolutional neural network.A hybrid LSTM convolutional neural network was proposed.Through experimental comparison,it was found that the hybrid LSTM convolutional neural network model outperformed the SVM classifier in terms of average recognition rate for complex actions on both public and self test datasets.Moreover,when comparing three different dimensions of data on the self test dataset,it was found that the convolutional network had higher recognition accuracy for data with higher dimensions.Overall,it was found that,The hybrid LSTM convolutional neural network model can better complete the task of sensor human motion recognition. |
