Maximum-Likelihood Methods in Target Tracking & Fundamental Results on Trackability

If a GLR (generalized likelihood ratio) test cannot make a good decision, then there is no good decision to be made. If the test is as to whether or not a VLO target is present in heavy clutter, the GLR should be the maximum-likelihood probabilistic data association (MLPDA) tracker. The MLPDA is very effective, but has several operational shortcomings that its close cousin, the maximum-likelihood probabilistic multi-hypothesis tracker (MLPMHT) avoids. We will discuss and compare both algorithms, plus show some fortuitous new MLPMHT developments. Perhaps most interesting, we are now able to set the MLPMHT threshold accurately and confidently, as would be a requirement for real-time operation. And since one cannot do better than ML, we are now able to make fundamental statements about which targets can be tracked and which cannot: these statements are essentially a bound, as opposed to algorithm-specific performance experience.