Hierarchical recognition of articulated objects from single perspective views

This paper presents an approach to the recognition of articulated 3D objects in monocular video images. A hierarchical object representation models objects as a composition of rigid components which are explicitly connected by specific kinematic constraints, e.g., rotational and/or translational joints. The recognition task follows this tree-like structure by first estimating the 3D pose of the static component (root) and afterwards determining the relative 3D pose of the remaining components recursively. This method limits the search space for the actual correspondences between image and model features and copes with the problem of self-occlusion. Experiments in the context of autonomous, mobile robots show the practicability of this approach