Multiple motion estimation and segmentation in transparency

In this paper a mechanism for computing multiple motion models along with the corresponding regions of support for a sequence of images is presented. The mechanism is an indirect parametric motion estimation technique. Firstly, a robust technique based on the fundamental constraint equation of multiple optical flow is used to estimate a dense multiple vector field of the scene. Secondly, from the recovered low-level data, motion models and their corresponding regions of support are estimated through a variant of the expectation-maximisation (EM) algorithm. The proposed algorithm is shown to provide good motion estimates and regions of support.