RGB-D based daily activity recognition for service robots using clustering with Gaussian Mixtures and FastDTW

Abstract

Recognizing human daily activities will enhance the capability of a service robot that interacts with humans. Such a task is challenging due to the complex and dynamic nature of human gestures in interaction, which brings about much difficulty in the reliability and real-time performance of activity recognition. In this paper a reliable and fast RGB-D based daily activity recognition method for service robots is proposed. The proposed method uses clustering with Gaussian Mixtures to model spatial features of motion patterns instead of using raw joint position information, which improved the reliability of activity recognition. Meanwhile, in order to deal with the spatial and temporal variance caused by individual nature of the activities, FastDTW algorithm is utilized for recognizing the time series of skeleton trajectories, which greatly improves the computational speed of activity recognition. Experimental results indicate that the method can fully satisfy accuracy rate requirement and computational speed requirement of dynamic activity recognition for human-robot interaction applications.