Learning to Look at Humans -- What Are the Parts of a Moving Body?
AMDO '08 Proceedings of the 5th international conference on Articulated Motion and Deformable Objects
Acquisition of articulated human body models using multiple cameras
AMDO'06 Proceedings of the 4th international conference on Articulated Motion and Deformable Objects
| Hi-index | 0.00 |
The use of models for visual tracking provides robustness towards image noise, clutter and occlusion and offers the ability to recover three-dimensional pose information of objects even from monocular images and image sequences. Model-based approaches hence have a tremendous value for computer vision algorithms that are aimed at analyzing videos containing articulated targets. Past research that has focused on the problem of model-based tracking of articulated targets has neglected to address the problems of model-acquisition and initialization. However, for model-based approaches to ever become practical and autonomous, these important issues need to be addressed. Towards this goal, this thesis presents a comprehensive set of methodologies that acquire models from video. The presented work is able to estimate the structure, shape, and appearance of articulated models from monocular image sequences. In addition, the initialization problem is solved by estimating pose information for at least one frame of a sequence, allowing subsequent model-based tracking. The presented work is based on basic assumptions and hence not restricted towards specific types of targets. It has in particular the ability to process human as well as non-human targets and makes no assumptions with respect to kinematic tree structure or complexity. The overall problem addressed in this thesis is divided into a number of sub-problems. A target and link-segmentation algorithm is responsible for decomposing an observed target into rigidly moving components and for obtaining initial motion estimates. Based on the obtained information, articulated structure estimation is performed, determining the joint constraints between observed target components. Finally, the 3D shape and pose is estimated by utilizing a novel model-based tracking framework for obtaining pose information across a set of uncalibrated keyframes and by performing shape from silhouette estimation together with 3D joint localization. This work hence presents a set of systematic solutions to the problems of model-acquisition and initialization that bridge the gap between state of the art model-based tracking approaches and practical applications.