Activity recognition using accelerometer data of wrist-worn smart phone

The need to get the exact information about the activities and movements of the people in everyday life has been increasing day by day. Various studies have been made on activity recognition in the fields such as medical, military, security and etc. for this purpose. In this thesis, activity recognition has been conducted by using the acceleration values obtained from the accelerometer sensor at the smartphones. By using a mobile phone application developed for this purpose, two data sets were created for two different phone positions (in the pocket of the pants and at the wrist) and nine different activities (sitting, standing, running, lying, eating, going up stairs, going down stairs, driving car, walking). The obtained raw data were processed and transformed into a data set with twenty-two features (mean, standard deviation, correlation, etc.) and then activities were classified by using four different classification methods (k-nearest neighbors, Näive Bayes, Random Forest, and Support Vector Machines). The accuracy percentages, that were obtained by applying the methods, are examined with various parameters and the findings were written under three headings. The most striking one among these findings is that the position of the mobile phone significantly affects the accuracy percentages obtained. It is observed that the accuracy percentage obtained from the data set created by carrying the smartphone at the wristband is significantly greater than the percentage of accuracy obtained from the data set created by carrying in the pocket of the pants. It shows that the motion intensity in the wrist region provides extra information for activity recognition.