The joint analysis of movement and deformation is crucial in many computer vision applications. This thesis proposes a stochastic non-linear filter to track a free curve in time. The proposed approach is implemented through a particle filter including colorimetric measurements characterizing respectively the target and the background. The involved dynamics is formulated as a stochastic differential equation. This allows a continuous representation of the curve trajectory, and thus the possibility to deduce the deformation between images. The curve is defined by an implicit level set, on which the stochastic dynamics is expressed. This takes the form of a stochastic differential equation with a Brownian motion of small dimension. We combined in these evolution models a local motion information extracted from the images and a model of the uncertainty of the dynamics. The associated filter proposed for curve tracking thus belongs to the family of conditional particle filters. Its capabilities are tested on different sequences containing highly deformable objects.
