Hybrid Algorithms for Multitarget Tracking using MHT and GM-CPHD

The Gaussian mixture cardinalized probability hypothesis density (GM-CPHD) is a new original algorithm for multitarget tracking adapted to false alarms, nondetection and closely spaced objects. It models the target set as a random finite set (RFS) and estimates the target state as the first-order moment of a joint probability distribution. In the classical version no track assignment is implemented; this is a limit to scene understanding in a multitarget context. A technique for choosing the peak to track association is therefore proposed. With this implementation the main strength of the GM-CPHD is shown: it drastically improves the performances concerning the estimation of the number of targets and gives acceptable performances concerning the state of each individual target even if targets are close together, but it cannot rival an interacting multiple model estimator with multiple hypothesis tracking (IMM-MHT) in regards to velocity estimation, which is also the case with other multitarget tracking algorithms not equiped with IMM. However, MHT performance decreases due to poor estimation of the number of targets when targets are close together. It is worth noting that combining a probability hypothesis density (PHD) filter with a multiple-model approach should improve the velocity estimation but is unnecessary because we have developed a hybrid algorithm, combining the precision of the estimation of the number of targets given by the GM-CPHD, used in a labeled version, with the precision of the estimation of each individual state given by the MHT. These noteworthy performances can be observed for individual targets as well as for convoys. This hybrid algorithm is extended by using an IMM-MHT with road constraints.