Unsupervised Human Motion Capture
PhD thesis
Click here to apply on line for this PhD offer
Thesis objectives and approach
The vast majority of human motion capture methods use visual observations to estimate the parameters of a detailed kinematic body model. This is a difficult problem for several reasons. The kinematic parameters are not directly observable and hence one has to solve an inverse kinematic problem — see for instance the recent work of D. Knossow et al 2008. The latter has a valid solution only if it properly initialised. Even if a good initialization is available, all the body parts are not visible and hence some of the parameters cannot be estimated at all.
In this work we propose to consider a radically different approach that does not make use of a prior kinematic model. Instead, this model will be "discovered" on line within an approach that combines segmentation [Cuzzolin et al. 2007] with registration [Mateus et al. 2007].
In this PhD proposal we suggest to use 3-D observations such as a surface mesh (or a set of voxels). These observations are easily obtained with a multi-camera system (such as the one available in our laboratory) and with a 3-D reconstruction method based on silhouettes. The shape of this mesh evolves over time as the articulated body moves. Hence, there will be several problems to be solved:
1. Segment the data into body parts,
2. Register the data over time,
3. Find the appropriate rigid motions associated with each body part, and
4. Associate an adhoc kinematic model with the observed articulated motions.
Bibliography
We suggest to address these issues within the framework of unsupervised learning. For example, one can apply embedding techniques (local-linear embedding, Laplacian eigenmaps, etc.) to represent the shape in its natural dimension. Therefore, segmentation and registration could be applied to the embedded shape. The registration problem will be addressed in the framework of the expectation-maximization method.
[Knossow et al. 2008] Knossow, D., Ronfard, R. and Horaud, R. P., "Human Motion Tracking with a Kinematic Parameterization of Extremal Contours", International Journal of Computer Vision, 2008.
[Cuzzolin et al. 2007] Cuzzolin, F., Mateus, D., Boyer, E. and Horaud, R. P. "Robust Spectral 3D-bodypart Segmentation along Time", Second Workshop on Human Motion, Understanding, Modeling, Capture and Animation. LNCS vol. 4814, pp 196-211, 2007.
[Mateus et al. 2007] Mateus, D., Cuzzolin, F. Horaud, R. P. and Boyer, E. "Articulated Shape Matching by Robust Alignment of Embedded Representations", IEEE Workshop on 3D Representation for Recognition (3DRR 2007), Rio de Janeiro, Brazil. October 2007.
Click here to apply on line for this PhD offer
Eligibility: The PhD candidate should have a strong academic record as well as expertise in computer vision, computer graphics, probability theory, and machine learning.
Start date: 1 October 2008
Contact person: Radu Patrice HORAUD
Deadline: 15 May 2008
