Tracking multpile distributed regression motions with the EM-algorithm
by
M.B. Malyutov
Northeastern University, Boston, MA, USA
Coauthors: A.N. Nikiforov (SUNY at Stony Brook, USA), M. Lu (Northeastern University, Boston, MA, USA)

We deal with tracking the motion of several indistinguishable targets described by parametric regression models in noise. The input data consists of a sequence of frames which are the results of observations of moving objects. The observations are subject to noise and some random targets loss and false targets appearance due to clutter. The latter is modeled as a stationary random field satisfying certain regularity conditions. On the basis of all frames, we estimate all trajectories using the EM-algorithm and its robust versions. Our algorithms are simplified versions of so-called PMHT algorithms

We compare by simulation standard and a robust version of the EM-algorithm based on least median of squares for estimating multiple targets in noise and clutter. Simulation shows an essentially more stability of a robust EM version in strong clutter.

The traditional methods of multiple target tracking (MTT) are based on preliminary data association (PDA) (data measurements and the objects are associated at each frame according to some criterion) (Bar Shalom and Li, Multitarget Multisensor tracking, YBS, 1995)). For large samples, these methods are exponentially (in the number of frames observed) consuming computationally requiring supercomputers. Additionally, they are inconsistent (even theoretically) if the distances between targets are comparable with the standard errors of measurements. As a result, whatever amount of data and computation is available, the algorithm resolution cannot be made better than certain nonvanishing limits.

Date received: October 16, 2000