Poster: Attention-based Spatio-Temporal Model for HAR Using Multivariate Time Series.

Recognizing human activities in a sequence is a challenging area of research in ubiquitous computing. Modeling spatiotemporal dynamics from multivariate time series plays a key role in recognizing human activities, unfortunately, the existing approaches could not extract spatial and temporal dynamics simultaneously. This poster presents ABST, a deep learning framework that extracts spatiotemporal dynamics for discriminating human activities. We demonstrate how to utilize two GRUs along two dimensions of inputs to automatically learn features. In addition, considering that each time step pays different levels of attention to the final prediction, on the topmost of framework, we apply an attention-gated recurrent layer. Experimental results show a promising recognition performance, and outperforms the state-of-the-art methods.