An Adaptive Shape Subspace Model for Level Set-Based Object Tracking

Shape priors have been widely used for level set-based tracking to solve some difficult problems, such as noisy data, partial occlusions and weak contrast at the boundaries. In this paper, we propose a two-layer hierarchical level set-based tracking framework in which color and shape information are fused sequentially. In the first layer, the initial contour is evolved only with the color feature, then the Mahalanobis distance-based discriminant criterion is adopted to determine whether the shape model is needed. If the shape model is needed, in the second layer the contour is evolved with the shape constraint continuously. For the second layer, a weighted shape distance term (WSDT) is introduced into the pixel-wise contour evolution equation to fuse the global shape information and the local color information. Principal Component Analysis (PCA) subspace of shape samples is trained off-line and updated using an on-line algorithm. The experimental results on several real video sequences demonstrate the robustness and the effectiveness of our method.