Detecting and Tracking Multiple Stealthy Targets: Comparison of PHD Filter and Track-Before-Detect Approaches
Trevor Wood · 2011
The traditional approach to target tracking involves extracting target detections from pixel amplitude data which are then passed to a tracker. This is known as the detection step. For low SNR targets the information loss in the detection step may be significant. Track-Before-Detect is a Bayesian tracking method founded on the idea that for low SNR targets it is better to work with the full set of pixel amplitudes. An alternative approach is the probability hypothesis density (PHD) filter, an approximate multitarget filter derived using finite set statistics (FISST). The PHD filter, like traditional methods, relies on a detection step but has the advantage of handling the multitarget aspects of the tracking problem by rigorously incorporating explicit statistical models for phenomena such as target appearance/disappearance, missed detections and false alarms. Recently, an extension to the PHD filter was developed which permits the inclusion of target amplitude information through the detection step. In this paper, both of these methodologies are presented and it is shown that the TrackBefore-Detect method is closely related to a special case of the PHD filter with target amplitude information. Indicative results are presented for a test using simulated data. It is demonstrated that in the case where there are either 0 or 1 targets present at all times, the performance of Track-Before-Detect and PHD filter methods are similar. The PHD filter with amplitude information is also deployed on a scenario with more than one target, showing good results. It is not possible to deploy Track-Before-Detect methods on such a scenario in a straightforward and theoretically rigorous manner. Given the close correspondence between the two methods, the PHD filter with amplitude information provides a way to obtain the benefits of the Track-Before-Detect methodology rigorously in multitarget scenarios.