Personalized Real Time Human Activity Recognition

Abstract

Human activity recognition is possible using small top K compressed personalized training dataset. However, unavailability of personalized dataset weakens the personalization of human activity. To address the shortage of personalized dataset an integrated training and testing algorithm was implemented. The algorithm extract and label top K time domain feature used to train the implemented hybrid Naive Bayes Nearest Neighbor algorithm to detect human activity in real time. The hybrid algorithm was implemented using HTML5, JavaScript and Cordova multi-platform for Android. Three subjects were recruited and requested to put in the Samsung Smartphone their front pocket as they train and test the model. Each subject was requested to perform human activities during training and testing. During testing each subject was requested to perform each human activity three times, therefore 21 comma separated values files were collected and analyzed using confusion matrix based on precision, recall and accuracy. Overall the results reveal a balanced precision, recall, f-measure and accuracy of 79%, 75%, 75% and 70% respectively for all subjects. Also the results indicate that it possible to train and test the data intensive algorithm with small training dataset.