Systematic Analysis of the PMBM, PHD, JPDA and GNN Multi-Target Tracking Filters

Multiple-target tracking has increasingly gained attention over the last 60 years, with the data association task being one of the most challenging aspects in sensor data fusion due to its computational burden. Hence, a plethora of algorithms has been proposed to solve this data association problem. However, most approaches are solely evaluated in comparison to algorithms of the same class. Therefore, this paper tries to give an overview and intends to evaluate systematically the four main classes of multiple-target tracking filters, namely the non-Bayesian data association filters, the Bayesian data association filters, the intensity filters and the multi-Bernoulli filters. These four classes are exemplified by respective filters, namely the Global Nearest Neighbor filter, the Joint Probabilistic Data Association filter, the Probability Hypothesis Density filter and the Poisson multi-Bernoulli mixture filter. These four filters are evaluated on two challenging simulated scenarios comprising situations with false measurements as well as birth and death of targets. The performance is assessed using the well-established GOSPA metric. It is shown that the Poisson multi-Bernoulli mixture filter outperforms the other three filters regarding the GOSPA metric in these scenarios, yielding a smaller mean error and deviation in its position estimates. This accuracy comes at the cost of a higher runtime performance, as the other three filter types require less computational time to produce their state estimates.