Accurate object subtraction is a fundamental step of video surveillance and synthesis vision applications. However, segmentation with intensity information alone is prone to fail for objects with diffuse edges, in clutter, or under occlusion. In this work, a novel segmentation method is presented for object with deformation in monocular videos. First, we formulate shape prior term as a distance of deformation by explicitly estimating integral motion of the whole contour, which can then be incorporated with spatial-temporal image information under Markov random field energies framework to be minimized by the Graph Cut algorithm. Then we proposed the concept of dynamic shape to characterize the highly variable nonlinear shape priors. The deformation is represented in an orthogonal basis learned by principal component analysis from training sets and the latent variables are expressed as autoregressive model. Simultaneously, both the orthogonal basis decomposition and the autoregressive model parameters are updated on-line to capture model evolutions. Finally, Promising experimental results demonstrate the potentials of the proposed segmentation framework with respect to noise, clutter, and partial occlusions.
