Filtrage conditionnel pour la trajectographie dans des séquences d'images - Application au suivi de points Conditional filters for image sequence tracking - Application to point tracker

In this paper, we propose a new conditional formulation of classical filtering methods dedicated to image sequence based tracking. These conditional filters allow to consider a state model and a measure model which both depend on the image sequence data. On this basis, we derive two filters for the point tracking problem, which authorize to cope with trajectories exhibiting abrupt changes and occlusions. They combine a dynamic relying on the optical flow constraint and measures provided by a matching technique. The first tracker is linear, well-suited to image sequences exhibiting global dominant motion. This filter is deduced through the use of a new estimator called the conditional linear minimum variance estimator. The second one is a nonlinear tracker, implemented from a particle filter. This latter allows to track points whose motion may only be locally described.