A particle system formulation for tracking deformable structures

This work presents image processing tools that can be helpful in the modeling of deformable structures encountered in oceanography. This paper tackles the problem of modeling and tracking deformable structures in remotely sensed image sequences. To take into account deformation (small and large), the authors focus on level-set methods, as they permit the accurate modelling of deformable structures. They can also easily take into account topological changes. The classical level-set formulation is modified by incorporating a particle system aimed at controlling the level-sets. Different kinds of internal energies can be superimposed on the particle system to accurately represent different physical behaviours of the level-set. According to image processing methodology, one tries to derive an external force field coming directly from the image data, mostly from local extrema of the gradient norm. This data is used to provide an external force field acting upon the particle system, and the level set is attracted toward the boundary of a structure. The external energy incorporates a specific "collar" energy used to minimize the number of undesirable features coming from the computation of the gradient's local extrema. A suitable initialization is performed using the skeleton of the distance map. Experimentation shows that such an initialization is stable w.r.t. simple affine motion. One of the advantages of this particle system formulation is that no partial differential equation is solved, and the minimization just requires the use of a classical gradient descent.