Abstract
In this work, we propose a novel method that utilises video-based depth and optical flow information of human body movement for human activity recognition. The recognition method utilises independent components (ICs) of depth silhouettes and optical flow-based motion features from a series of depth images and discrete hidden Markov models (HMMs) for recognition. The IC features are extracted from a collection of depth silhouettes containing various human activities, over which linear discriminant analysis (LDA) is performed for better classification. Furthermore, to improve the recognition performance, optical flow-based motion features extracted from the consecutive depth silhouettes are utilised in an augmented form. In addition, discrete HMMs are employed to recognise and model the time-sequential features of human activities. Our results show that the depth silhouette feature-based approach provides better human activity recognition than most generally used binary silhouette-based approaches, and that the augmented depth IC and optical flow features provide additional improvements.
This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MEST) (No. 2011-0001031).