Particle Filter methods are one of the dominant tracking paradigms due to its ability to handle non-gaussian processes, multimodality and temporal consistency. Traditionally, the exponential growth on the number of particles required (and therefore in the computational cost) with respect to the increase of the state space dimensionality means one of the major drawbacks for these methods. The problem of part based tracking, central nowadays, is hardly tractable within this framework. Several efforts have been made in order to solve this problem, as the appearance of hierarchical models or the extension of graph theory by means of the Nonparametric Belief Propagation. Our approach relies instead on the use of Auxiliary Particle Filters, models the relations between parts dynamically (without training) and introduces a compatibility factor to efficiently reduce the growth of the computational cost. We did run the experiments presented without using a priori information.