Multiple Manoeuvering Target Tracking: A Memory-constrained Minimum Risk Approach.

Abstract

One of the main challenges of modern tracking systems, as a matter of fact a robotics problem, lies in the ability to remain efficient versus highly maneuvering targets. The classical approach, from the tracker point of view, consists in modelling the target as a hybrid dynamical system, i.e., to consider the current behavior of the target to belong to a finite family of possible models, each model being characterized by a different evolution equation. In the pseudo-bayesian approach, each observation performed by a sensor leads to the updating of the current probabilities of each possible model and of the global state estimate as an adequate combination of the estimates delivered by the different model-based filters. The underlying hypothesis is to consider that the past global estimate is a suitable reference. We choose here a memory-constrained bayesian approach, which means that we try to store a maximum-depth tree corresponding to all possible sequences of models from the current “initial time” (the tree root) to the current time (the tree leaves). The tree depth is constrained by the processor memory and by the fact that many targets are tracked at the same time, which leads to a concurrential multi-users trade-off. The decision to prune a tree is optimized through two different approaches.