Modeling high-level descriptions of real-life physical activities using latent topic modeling of multimodal sensor signals.

We propose a new methodology to model high-level descriptions of physical activities using multimodal sensor signals (ambulatory electrocardiogram (ECG) and accelerometer signals) obtained by a wearable wireless sensor network. We introduce a two-step strategy where the first step estimates likelihood scores over the low-level descriptions of physical activities such as walking or sitting directly from sensor signals and the second step infers the high-level description based on the estimated low-level description scores. Assuming that a high-level description of a certain physical activity may consist of multiple low-level physical activities and a low-level physical activity can be observed in multiple high-level descriptions of physical activities, we introduce the statistical concept of latent topics in physical activities to model the high-level status with low-level descriptions. With an unsupervised approach using a database from unconstrained free-living settings, we show promising results in modeling high-level descriptions of physical activities.