Daily Routine Recognition with Visual Interactive Labeling by Fusing Acceleration and Audio Signals
Thomas Kuebert, Henning Puder, Heinz Koeppl · 2019
This work focuses on a semi-supervised learning scheme for daily routine recognition. To annotate and process hours of sensor data, an efficient processing workflow is established for a two-step classification: first propagating the labels and second recognizing the daily routine. The first contribution is the extension of visual interactive labeling (VIL) to activity recognition challenges and time series data exploration. With the help of a coarse time diary and a graphical user interface, the data is interactively labeled, and a first classifier propagates these annotations. Afterwards, a second classifier is trained and evaluated on the cross-validated VIL label set. As an extension of previous work, the second contribution is, that not only acceleration sensors are utilized to analyze the routine, but also audio features enrich the environmental description power. The third contribution is the comparison to a topic model on a public data set, where our approach outperforms on average recall of 84.3% and precision of 88.7%. Moreover, the data is recorded by a hearing aid, which ultimately aims to personalize the device configuration to the daily routine of the wearer.