Dynamic Shape Prior Based Object Segmentation in Video Stream
Ping Sheng · Journal of Sichuan University · 2009
Accurate object subtraction is a fundamental step of machine vision applications.However,segmentations with intensity information alone are prone to fail for objects with diffuse edges,in clutter,or under occlusion.To address this problem,a novel segmentation method was proposed for deformable object in monocular videos.Firstly,shape prior term was treated as a distance of deformation under Markov random field framework by explicitly estimating integral motion of the whole contour.Then a potential function was formulated by incorporating the shape term with spatial-temporal image information,and the Graph Cut algorithm was applied to got its global optimization;Simultaneously the dynamic shape model was presented to characterize the trends of nonlinear shape variations by building the autoregressive model for the latent variables,which were obtained by projecting the detected shapes onto the orthogonal basis learned from training sets by principal component analysis.Consequently,both the orthogonal basis and the autoregressive model parameters were updated online to capture model evolutions.Experimental results demonstrated the potentials and robustness of the proposed method in intermediate computer vision applications with respect to noise,clutter,and partial occlusions.