Using Statistical Shape Priors in Geodesic Active Contours for Robust Object Detection

A novel statistical shape prior model based on level set representations is proposed in this paper for robust object detection by geodesic active contours. This prior model is able to accommodate multiple shape states of objects. The level set representations (signed distance map) of the shapes are considered to form distinct clusters in a low dimensional feature subspace and a Gaussian mixture model (GMM) is employed to fit the feature distribution in the subspace. A Bayesian classifier is used to assign the currently detected object to the most similar shape cluster. A shape prior is then constructed from the statistical properties of that cluster and is used to drive the geodesic active contour curve towards it in the subsequent evolution. Experiments demonstrate the effectiveness of our shape prior model