A simple framework for spatio-temporal video segmentation and delayering using dense motion fields

To represent object movement, the conventional approach is to embed the motion information into parametric and/or statistical motion models. However, the inherent complexity of this description is ill suited for application requiring rapid and automatic, albeit significant, responses. Surveillance applications belong to this class and are emerging as the most active field of research in computer vision. For this purpose, we observe that the differential invariants obtained by dense optic flow are capable of accurately describing complex object motion without requiring setting up and initializing models. In this letter, we demonstrate the novel use of such motion descriptors for spatio-temporal object segmentation and delayering of sequences. Our results show the ability of this approach to describe simply and accurately differently moving objects and to be incorporated in segmentation processes that deliver a hierarchical description of object, producing evident improvements to the segmented objects while being computationally efficient