We present in this paper a general-purpose approach for articulated object recognition. We split the recognition process in two distinct phases. In the former we use standard model-based techniques in order to recognize and localize in the input image the rigid components the articulated object is composed of. In the second phase the spatial configurations formed by the recognized components are analyzed and compared with the valid configurations of the object we are searching. The comparison is based on a constraint satisfaction method which can deal with both missing components and false positives. The proposed method is based on a redundant set of constraints which represent the valid spatial configurations of the object's components. Such constraints are not embedded in the system nor are domain-specific but they are learned during a suitable training phase.

We show how this approach can be used in different scenarios with different kinds of articulated objects and we present a case study concerning a robotic application.