Improving tracking behavior with virtual sensors in the loop

Any tracking algorithm can be cast as an estimation problem, i.e., x(t|s(t)). It estimates the target position x(t), from clutter sensor measurements s(t). The estimation process, by nature, is an open loop; that is once the target position is estimated, there is no means of re-acquiring the track, if the estimation results in track-loss. In this paper, we propose virtual sensors to deal with the problem of track-loss. Virtual sensors allow the system to recognize mismatch between the actual and the estimated target positions. They achieve this without knowing where the target is or even an estimate of it. This mismatch can be quantified and fed back to the tracking system to correct the estimation, so as to rescue (in the events of significant mismatch) the tracking process from collapsing. The proposed system is investigated in an experiment where a target is moving in a two-dimensional scene that is populated by “cheap” sensors. Two similar landmark based tracking (LMT) algorithms that differ only in the use of virtual sensors have been tested. Their results are compared and contrasted for different types of target motion