Human activity recognition using symbolic sequences

Human activity recognition research is an active area in an early stage of development. We present two approaches to activity recognition based on symbolic representations of multivariate time series of joint locations in articulated skeletons.One approach uses pairwise alignment and nearest-neighbour classification, and the other uses spectrum kernels and SVMs as classifiers. We tested both approaches on three datasets derived from RGBD cameras (e.g., Microsoft Kinect) as well as ordinary video, and compared our results with those of other researchers.