Human Activity Recognition Using Smartphone Sensor Data

Abstract

The hot topic in recent times is recognition of human activities through a smartphone, smart home, remote monitoring and assisted healthcare. These fall under ambient intelligent services. This also includes recognition of simple activities like sitting, running and walking, and more research is being held for semi-complex activities such as moving upstairs and downstairs, running and jogging. Activity recognition is the problem of predicting the current action of a person by using the motion sensors worn on the body. This problem is approached by using supervised classification model where a model is trained from a known set of data, and a query is then resolved to a known activity label by using the learned model. The exigent issue here is whether how to feed this classification model with a set of features, where the input provided is a raw sensor data. In this study, three classification techniques are considered and their accuracy in predicting the correct activity. In addition to the systematic comparison of the results, a comprehensive evaluation of data collection and some preprocessing steps are provided such as filtering and feature generation. The results determine that feeding a support vector machine with an ensemble selection of most relevant features by using principal component analysis yields best results.