Heavy-tailed noise modeling versus probabilistic data association for robust filtering

A popular approach to robust filtering is to replace the standard Gaussian distribution for measurement noise with a heavier-tailed distribution. A recent example of this approach is presented and compared via simulations with the benchmark probabilistic data association filter. While both filters often perform similarly and clearly outperform the standard Kalman filter when outliers are present, the heavy-tailed noise modeling approach provides better consistency performance, is less prone to diverge, and is less sensitive to poor initialization in the majority of examples considered.