Feature Extraction And Activity Recognition Based On Phone Acceleration Data

Nowdays,with the rapid development of mobile Internet,smartphones have become a very important part in the people's lives,and people carry them almost anywhere and anytime.The smartphones today are mostly equipped with a set of sensors and the human activity recognition technology based on smartphone sensors has aroused the interest of many researchers.The human activity recognition technology based on smartphone sensors has a very broad application prospects and it can be predicted to be a hot research topic in the future.In this thesis,we propose a recognition scheme of human activity according to smartphone acceleration data.In the scheme,by using the smartphone,we collect the linear acceleration data concerning with human activities and then apply the common sliding window technique to produce activity instance from the original sensor data and extract features form activity instance.For the step of feature extraction,we propose a new method of feature extraction based on interval weight.At last,the unsupervised learning algor ithm,such as MCODE clustering algorithm and hierarchical clustering algor ithm,is utilized to carry out activity recognition research.In the experiment,we also introduce two other feature extraction methods which are targeted on the time domain and time-frequency mixing domain,respectively.The experimental results show that,from the linear acceleration data concerning with six kinds of daily human activity,adopting the feature extraction method based on interval weight and the hierarchical clustering algor ithm are fairly well recognize "walking","jogging","ascending stairs" and "descending stairs",for which the overall recognition rate reached 99.18%.However,the method is unable to distinguish between "sitting" and "standing" which are completely recognized as a same activity in the results.Among the recognition to the six kinds of daily human activity,adopting the feature extraction method based on interval weight combined with the MCODE clustering algorithm gain the recognition efficiency similar to the hierarchical clustering algor ithm,but the input parameter selection for MCODE clustering algor ithm is relatively trouble.The human activity recognition results corresponding to each feature extraction method also show that the feature extraction method based on interval weight is superior to the other two introduced methods of feature extraction.