Deformable object tracking with statistical models

We propose to track an object of interest in video sequences based on a statistical model. The object appearance is modeled with kernel elements that are induced by a normalized non-parametric density in the local regions. The choice of these kernel elements is based on the stable appearance or distinctive features such as discriminative characteristics from background. This allows imposing weight factors to signify certain features in the object searching process, which is extended from the template match by relaxing the matching pixels' correspondence so as to handle more effectively the local appearance changes caused by the object deformation or illumination changes. The object extraction process is less computational because fewer matching pixels are actually needed. Experiments show that this approach can well handle the local appearance change for a deforming object. As an alternative method, a Bayesian framework is applied to derive the posterior probabilities for the tracked object. The object likelihood and the background likelihood for a given pixel are calculated by the non-parametric density estimators, and are incorporated with a spatial probability model to optimize the statistical location of the current object.