Human action capturing and classification
Feng Zhou · 2006
Recognizing humans and activities in video is a key enabling technology for ultra-intelligent environments.These environments, which we call sentient spaces, are envisioned to be fully capable of sensing the motives and desires of occupants, and proactively reacting to them in ways to dramatically improve work productivity and quality of life.In this report, a vision-based framework is proposed for learning and inferring occupant activities at different levels.These levels range from short temporal interval movements, to intermediate level events and long temporal term activities.Our research is focused on using a combined tracking-classification framework for the unsupervised classification of human action.An initial comparative study was done to evaluate several existing foreground segmentation methods that employ background modeling.Our own probabilistic foreground-background segmentation method is proposed to extract human-centric reference frames.Based on the human-centric reference frames, a principled analysis of the correspondence problem leads to a novel probabilistic action representation called the correspondence-ambiguous feature histogram array (CAFHA) that is robust to variations across similar actions.CAFHA is shown to be effective in unsupervised action classification and quasi real-time action inference.A novel feature selection method is proposed to select the optimal features to improve the CAFHA representation, such that the best discrimination between different action clusters may be found via unsupervised spectral clustering.Finally, a number of potential future directions are proposed that are targeted at further improvements to our framework and creating new research methods required to recognize human activity at longer temporal scales.