Pose estimation
A system to estimate the pose of a human is implemented. Therefore, the human is filmed by a camera and segmented in the image. Then, the visible hull is compared to a database of precomputed visible hulls of humans is various poses. As the human can turn arbitrarily, the visible hull for all poses are pre-computed in 3D volumetric data sets.
Abstract
Pre-processing
- Visible Hull A voxelization of a real-world object is computed based on photographs from different point of views on a circle or sphere around the object. This voxelization algorithm is very similar to the cone beam reconstruction algorithm known from the medical imaging field. One visible hull is computed for each pose.
- Triangulation Once the visible hulls has been computed a triangular mesh is generated using the marching cubes algorithm.
Run-time system
- Image acquisition A camera films a moving person in front of a blue screen.
- Matching The real-time system segments and matches of the pose of a person with the pre-processed data sets. The image with the closet match is the desired pose.
Requirements
- Good C/C++ programming skills
- DirectX9 and HLSL programming expertise
- Interest in cameras, image segmentation and numerical methods
Supervisors: T. Schiwietz, J. Schneider, J. Georgii
