This paper presents a Bayesian framework for a new model-based learning method, which is able to track nonrigid objects in the presence of occlusions, based on a dynamic shape description in terms of a set of corners. Tracking is done by estimating the new position of the target in a multimodal voting space. However, occlusion events and clutter may affect the model learning, leading to a distraction in the estimation of the new position of the target as well as incorrect updating of the shape model. This method takes advantage of automatic decisions regarding how to learn the model in different environments, by estimating the possible presence of distracters and regulating corner updating on the basis of these estimations. Moreover, by introducing the corner feature vector classification, the method is able to continue learning the model dynamically, even in such situations. Experimental results show a successful tracking along with a more precise estimation of shape and motion during occlusion events.