Find in Library

Supervised activity recognition algorithms require labeled data to train classification models. Labeling an activity can be performed trough observation, in controlled conditions, or thru self-labeling. The two first approaches are intrusive, which makes the task tedious for the person performing the activity, as well as for the one tagging the activity. This paper proposes a technique for activity labeling using subtle gestures that are simple to execute, and that can be sensed and recognized using smartwatches. The signals obtained by the inertial sensor in a smartwatch are used to train classification algorithms in order to identify the gesture. We obtained data from 15 participants who executed 6 proposed gestures in 3 different positions. 208 characteristics were computed from the accelerometer and gyroscope signals and were used to train two classification algorithms to detect the six proposed gestures. The results obtained achieve a precision of 81% for the 6 subtle gestures, and 91% when using only the first 3 gestures.