Towards Automated Activity Recognition in an Intensive Care Unit

In this work we develop a new system for automated activity recognition in an Intensive Care Unit. Our system is capable of recognizing a set of 7 high-level activities using only 3D perception. We take a three step approach to classifying actions from depth images. First, temporal action sequences are created by segmenting and tracking people throughout several hours of video footage. Second, for each sequence we derive features based on position, orientation, and multi-person interactions. Lastly, a comparative evaluation is provided for Support Vector Machine and Decision Forest classification techniques. We achieve 75% overall accuracy on data collected at a Pediatric ICU and show that this approach could realistically be used for increased safety and efficiency.