Detailed Activity Recognition with Smartphones

Sensors embedded in smart handheld can be ex-tremely useful in providing information on people's activities and behaviors. Human exact activity recognition through pos- ture identification is increasingly used for medical, surveillance. Existing works mostly uses one or more specific devices (with embedded sensor) for activity recognition and most of the time the detected activities are coarse grained like sit or walk rather than detailed like sit on chair or brisk walk. Consequently, in this paper we propose a detailed activity recognition system that uses smartphone accelerometer (available in almost every smartphone) thus does not need any special device to be carried by the user. The system applies inexpensive (in terms of resources consumed) feature extraction and learning mechanism to detect detailed activities, for instance, stow walk and brisk walk. We introduce a new feature based on jerk to detect both detailed static activities (sit on chair) and detailed dynamic activities (brisk walk). Implementation of the framework with real devices indicates 95% accuracy with state-of-the-art machine learning techniques while using a minimal set of features.