Particle filtering has become a standard tool for nonparametric estimation in computer vision tracking applications. It is an instance of stochastic search. Each particle represents a possible state of the system. Higher concentration of particles at any given region of the search space implies higher probabilities. One of its major drawbacks is the exponential growth in the number of particles for increasing dimensions in the search space. We present a graph based filtering framework for hierarchical model tracking that is capable of substantially alleviate this issue. The method relies on dividing the search space in subspaces that can be estimated separately. Low correlated subspaces may be estimated with parallel, or serial, filters and have their probability distributions combined by a special aggregator filter. We describe a new algorithm to extract parameter groups, which define the subspaces, from the system model. We validate our method with different graph structures withing a simple hand tracking experiment with both synthetic and real data.