Recognition of human gestures and behaviour based on motion trajectories

Abstract

Human activities are characterised by the spatio-temporal structure of their motion patterns. Such structures can be represented as temporal trajectories in a high-dimensional feature space of closely correlated measurements of visual observations. Models of such temporal structures need to account for the probabilistic and uncertain nature of motion patterns, their non-linear temporal scaling and ambiguities in temporal segmentation. In this paper, we address such problems by introducing a statistical dynamic framework to model and recognise human activities based on learning prior and continuous propagation of density models of behaviour patterns. Prior is learned from example sequences using hidden Markov models and density models are augmented by current visual observations.