Research And Implementation Of Human Activity Recognition System Based On Inertial Sensors

Human activity recognition based on inertial sensors is a hot research direction in machine learning and pattern recognition.It has important applications in the field of medical rehabilitation,somatosensory games and intelligent home.In recent years,great breakthroughs have been made in this direction.However,there are still some key technical problems that need to be solved urgently in some important hardware design and signal process parts,including how to design a wireless signal acquisition platform with high stability,light weight,small size,low cost and easy to wear;how to design a reasonable feature extraction method for practical application in order to achieve better recognition accuracy;how to explore a more effective feature selection method to reduce the computational complexity and improve the recognition accuracy.Based on above,this thesis designs a wearable wireless signal acquisition platform and focuses on the section of feature extraction and feature selection in human activity recognition.In feature extraction,two methods of Empirical Mode Decomposition(EMD)and Ensemble Empirical Mode Decomposition(EEMD)are focused on.With their excellent signal analysis capability,an improved adaptive EEMD feature extraction method is proposed.By filtering different Intrinsic Mode Functions(IMF)for different behaviors and extracting novel features such as windowed mean difference,it's expected to obtain more accurate and effective feature information at different resolutions.To verify the performance of the proposed method,two classifiers of K-Nearest Neighbor(KNN)and Support Vector Machine(SVM)are trained by typical time domain features,frequency domain features and features extracted by this method,respectively.Besides,the Leave-One-Out(LOO)method is used to test.The experiment results indicate that compared with the original method,this method not only increases total average recognition accuracy of three easily confused behaviors including walking,upstairs and downstairs by 29.22% and 15.79%,but also increases the accuracy of 7 human behaviors by 17.20% and 10.19%,reaching a high total average accuracy of 95.11% and 93.14% when using KNN and SVM for classification,respectively.In feature selection,due to the poor performance that traditional filter feature selection methods select features by using only one correlation measure,an improved feature selection method based on the max-Relevance and min-Redundancy(mRmR)is proposed in this thesis.In the framework of mRmR,this method firstly eliminates some redundant features and irrelevant features to obtain the preliminary feature subsets by combining multiple correlation measures and considering the possible positive influence on feature selection when analyzing the correlation between the features to be selected and the selected features under consideration of the classification category.Then,this method encodes the preliminary features by binary code and searches for optimal or suboptimal feature subsets by using genetic algorithm.The results indicate that compared with other methods,the proposed method has the highest total average recognition accuracy of 7 behaviors in the experiment.The total average recognition accuracy of 7 behaviors classified by SVM and KNN reaches 97.02% and 95.73% and compared with the traditional mRmR method,it is also increased by 13.72% and 9.92%,respectively.