Non-rigid tracking method based on layered elastic motion analysis

In this paper, we present a spatial-temporal layered elastic motion tracking method to estimate long-term stable structures and short-term local motions for non-rigid targets. First, the sequence is segmented into several pieces by sequential shape clustering based a piece-wise elastic motion tracking model, the correspondence among frames in the same segment and the mean shape of all clusters are calculated by piece-wise elastic motion tracking. Then, we use a Bayesian network based global search method to find the correspondence of mean shapes of adjoining clusters and extract the common shape of the entire sequence. The proposed method, which does not require prior shape models, is adaptive to and descriptive for general objects. The experiments on non-rigid targets validate both the long-term stability and the detailed accuracy of our proposed method.