A Practical Activity Recognition Approach Based on the Generic Activity Framework

Abstract

In spite of the obvious importance of activity recognition technology for human centric applications, state-of-the-art activity recognition technology is not practical enough for real world deployments because of the insufficient accuracy and lack of support for programmability. The authors introduce a generic activity framework to address these issues. The generic activity framework is a refined hierarchical composition structure of the traditional activity theory. New activity recognition algorithms that can cooperate with the proposed activity framework and model are proposed. To be practical, activity recognition technology should also be programmable. The hierarchical aspects of our generic activity framework help to improve activity recognition programmability. The generic activity framework decouples the observation subsystem (i.e., the sensor set) from the rest of the activity model. The authors demonstrate the value of this decoupling by experimentally comparing the level of user effort needed in making sensor changes and the ramifications of such changes on model updates. They compare the level of effort required by the authors’ model to the requirements of previously reported approaches.