Selective object stabilization for home video consumers

This paper describes a unified approach for video stabilization. The essential goal is to stabilize image sequences that consist of moving foreground objects, which appear frequently in today's home videos captured by hand-held consumer cameras. Our proposed techniques mainly rely on the analysis of motion content. Three major components are: initialization, segmentation and stabilization. In motion initialization, we propose a novel algorithm to efficiently search for the best possible frame in a sequence to start segmentation. Our segmentation algorithm is based on expectation-maximization (EM) framework which provides the mechanism for simultaneous estimation of motion models and their layers of support. Based on the framework of Kalman filter and EM motion estimation, our proposed algorithm has the flexibility of allowing selective stabilization with respect to background or/and foreground objects, subject to the preferences of customers.