This paper presents a method for modeling and then tracking the 2D planar size, location, orientation, and number of individuals of an animal aggregation using Autonomous Underwater Vehicles (AUVs). It is assumed that the AUVs are equipped with sensors that can measure the position states of a subset of individuals from within the aggregation being tracked. A new aggregation model based on provably stable Markov Process Matrices is shown as a viable model for representing aggregations. Then, a multi-stage state estimation architecture based on Particle Filters is presented that can estimate the time-varying model parameters in real-time using sensor measurements obtained by AUVs. To validate the approach, a historical data set is used consisting of >100 shark trajectories from a leopard shark aggregation observed in the La Jolla, CA coast area. The method is generalizable to any stable group movement model constructed using a Markov Matrix. Simulation results show that, when at least 40% of sharks are tagged, the estimated number of sharks in the aggregation has an error of 6%. This error increased to 27% when the system was tested with real data.