Abstract:
Activity Recognition (AR) is an active area of research that has direct applications on life quality and health of human beings. Related studies aim to classify di erent daily activities of people with high accuracy rates using various types of sensors. Becoming an essential part in our daily lives, smartphones are now suitable tools that enable people to make use of AR technologies without being obliged to use or wear some extra device. However, due to power and computational constraints of these devices, it becomes a challenging task to attain accurate results by using power and CPU-intensive classi ers. In this study, we present an e cient selection of features to attain high accuracies in recognizing ve daily activities with a lightweight classi- er, K Nearest Neighbors (KNN). Since previous studies in this area show that it is possible to obtain high recognition performance with the KNN classi cation algorithm, we focused on the problem of feature selection to see how far this performance can be enhanced by employing the most appropriate feature sets for the KNN algorithm. We use some well-known features together with some more speci c features and in order to keep the system energy-e cient, all features are extracted from the readings of a single accelerometer on a smartphone that is carried in the trousers' pocket with di erent orientations. In this study, we also evaluated the e ect of di erent window lengths and window functions that are used for segmenting the data prior to feature extraction. The results show that by having an e cient selection of features it is possible to obtain promising accuracy rates with a simple classi cation algorithm like KNN which facilitates online and real-time activity recognition on smartphones.
