Skip to Main Content
Causal estimation of multiple feature points trajectories by using a switching state space model is proposed. The state vector of the model consists of the position of each feature point, the velocity of each rigid object, and some indicator variables for each feature point. Ther are two types of indicator variables: an object indicator representing the association between the feature point and rigid object, and an aperture indicator representing the attribute of the point, e.g. aperture or not. By estimating the state vector using a Rao-Blackwellized particle filter, smooth trajectories of feature points, velocity of objects, object indicators, and aperture indicators are obtained simultaneously. Performance on a real image sequence is presented by comparing to a Kalman filter being given true indicators.