Investigating spatio-temporal Markov models for 3D tracking

Master thesis

During this master thesis, we will consider the problem of tracking an articulated object in multiple camera environments.

In such a setting, at a given instant of time, one may build a good kinematic description of the moving object by fusing information from several 2D images [1,2] (see images below). However, one recurrent problem of 3D tracking in the time consistency of this skeleton. A well-known approach to incorporate motion a priori is to use Sequential Monte Carlo methods (also know as Particle filter). However, the resulting trajectories suffer form jittering. During this master thesis, we propose to keep the Hidden Markov modelization, and to investigate how spatio-temporal Markov fieds can be used in that setting.

In that context, several challenging topics will be adressed:
  Explore the use of medial axis to estimate a skeleton for a set of 2D images
  Lear the correlation graph of joints for some motions (walking / kicking)
  Use this a priori information in conjonction with the observed skeletons to estimate a proper skeleton trajectory.

[1] J.-S. Franco and E. Boyer "Fusion of Multi-View Silhouette Cues Using a Space Occupancy Grid" International Conference on Computer Vision (ICCV 05)

[2] C. Menier, E. Boyer, B. Raffin "3D Skeleton-Based Body Pose Recovery", 3DPVT’06

---

---

Contact

This work will be carried out within the PERCEPTION research group. Experiments will be done using our multiple-camera laboratory (the GrImage platform).

  Elise Arnaud, Elise.Arnaud@inrialpes.fr, +33 (0)4 76 61 55 59

Start date: 1 March 2009

Contact person: Elise ARNAUD

Deadline: 1 December 2008