Stochastic temporal models of human activities

Human activities are characterised by the spatio-temporal structure of their motion pattern. Such structures are probabilistic and often rather ambiguous. Modelling such spatio-temporal structures as static templates can be very sensitive to noise and cannot capture variations in observation measurements caused by different subjects performing the same act. In this paper we introduce the concept of modelling temporal structures by statistical dynamic systems using first-order Markov process descriptions. Prior knowledge is learned from training sequences and recognition is performed through continuous propagation of density distributions. Taking current observations into account to temporarily augment the learned prior leads to more accurate recognition with less computational costs.