Bodo Rosenhahn's research

Pose Estimation Project

My main research project deals with the 2D-3D pose estimation problem. A modified version of this page can also be found at the official 2D-3D pose estimation project page from the cognitive systems group at the University of Kiel.
Pose estimation itself means to estimate the relative position and orientation of a 3D object with respect to a reference camera system.

The main focus of my work concentrates on the geometric modeling and application of the pose problem. Therefore I use as mathematical language the conformal geometric algebra, which is a universal Clifford algebra. It provides a homogeneous model for stereographically projected points on a hypersphere and therefore couples kinematics with projective geometry. This representation is useful for many computer vision or robot vision tasks like object manipulation or selflocalization of mobile robots.

For the pose scenario, 3D object models are treated two-fold, feature-based and free-form based:

The reason for this approach is that starting from simple features (e.g. point features) an extended scenario is derived, which deals with higher order features such as lines, planes, circles, spheres, kinematic chains or cycloidal curves. Cycloidal curves are defined as circles rolling on circles and are treated as 3D entities within the pose scenario. This scenario extends to general free-form contours by interpreting contours generated with 3D Fourier descriptors as n-times nested cycloidal curves.
The following movies visualize some properties of the algorithms I developed (click on the images):

The algorithm for pose estimation of kinematic chains is robust and can be used for different tasks, e.g. for visual remote controlling of a manipulator.

My free-form pose estimation algorithms are able to detect outliers during pose estimation and to deal with multiple and partial hidden object contours as well.

The next sequence shows pose results of two nested contours. The other sequence further shows the possibility to model slight object deformations by fusing kinematic chains within the approach for free-form object modeling.