Action recognition in bed using BAMs for assisted living and elderly care

There is a large interest on performing elderly care monitoring using Computer Vision. It has the potential to provide a better scene understanding than current sensing approaches at an affordable price, but there are still considerable practical challenges that have limited its deployment. The BAM descriptor is a privacy-conscious, calibration-free representation of a single-person bed obtained from a depth camera, and thus is very practical for uninterrupted monitoring. It has been used to recognize static and time-invariant phenomena such as sleeping position and agitation with great success. In this work, we explore BAM-based feature representations for higher level scene understanding. To this end, we created a database of 17 actions typical for elderly care which we use to evaluate our approach demonstrating promising results. We hope that this level of high scene understanding would allow the prediction of accidents in elderly care before they happen, instead of triggering an alarm after they happen.