A bayesian approach to markerless motion capture and activity recognition of multiple people

This work presents a general framework for tracking simultaneously the body posturas of multiple people from non-intrusive visual sensors. The method is based on a training-then-tracking philosophy, with the main addition of being able to handle more than just one person. We train the body postura from labelled motion capture datasets. The training process is based on popular non-linear dimensionality reduction techniques. Then, a mixed, discrete and continuous state particle filter is used to simultaneously detect the postura and the kind of motion performed by each of the human bodies. The resulting hypotheses, automatically selected from the particle distribution, are then refined using non-linear optimization methods with statistical priors. The whole framework is tested using a simple but standard method based on comparing articulated cylindrical models with SfS volumes, taken from several cameras. Our accuracy in public available datasets is near to the state-of-the-art works that do not take into account multiple people in the problem.