Human body pose estimation or tracking using computer vision is a difficult task owing to the high dimensionality of the pose space. Learning based aproachs, especially methods using embeding spaces such as LLE [RS00] or GPLVM [Law03] can cope with this di culty but are restricted to cyclic motions [UFF06]. Other methods proceed in comparing a test image to a learning base. Among them, PSH [SVD03] is usefull to speed up the selection of a subset of nearest neighbours from large learning bases. However, even when pose regression is used to extrapolate new poses from the learned ones, a su cient coverage of the pose space is diffi cult to reach with learning-based approaches [TSDD06]. Other ways consist in using deterministic or stochastic algorithms. The former kind of methods often provide suboptimal solutions because they get stuck on local minima owing to monocular vision ambiguities [PF03]. Stochastic approachs are used to explore the posterior probability function but once again, the high dimensionality of the pose space, especially in the case of simulation-based methods using sampling, requires a huge number of samples to explore the main mode. An interesting solution consists in using a loose-limbed body model [SBR+04] where the likelihood of each limb is evaluated independently. In this manner the dimension of the explored space is reduced to the number of dof of a limb. In uence between limbs is estimated by belief propagation [KFL01] to provide consistent body poses. This last solution is adopted in this thesis in association with particle ltering to provide a discrete space where the beliefs are computed [BCMC06]. This method is prefered to a parametric modelling of beliefs using Gibbs sampler, a method derived from the PAMPAS algorithm [Isa03] involving heavy computational load. However, in addition to this solution, robust human body tracking, even in two dimensions [NB07b], requires to use several images cues. Thus, hypotheses likelihood is evaluated from gradient and color based cues combined with a background subtraction [NB06] and a motion detection. A main diffi culty in monocular 3D tracking is the depth estimation making the fusioned images cues mentioned before unable to constrain suffi ciently the pose. However, owing to articulation constraints, the real pose space covered by human motion is much smaller than the theoretical one. Associating the fusioned images cues with articulation constraints implemented on the belief propagation step result in suitable algorithm performances even on unconstrained environments (light, clothes...) [NB07a]. A more effi cient occlusion handling is provided adding a learning-based hypotheses compatibility term. With the used body model [SBR+04], the learning base consists in limbs exemplars instead of full body poses permitting a wider coverage of the pose space with the same amount of exemplars. Belief propagation provides consistent body poses and the selection of similar limbs from the learning base can be speeded-up by PSH [SVD03].
