Merging-based forward-backward smoothing on Gaussian mixtures

Paper in proceeding, 2014

Conventional forward-backward smoothing (FBS) for Gaussian mixture (GM) problems are based on pruning methods which yield a degenerate hypothesis tree and often lead to underestimated uncertainties. To overcome these shortcomings, we propose an algorithm that is based on merging components in the GM during filtering and smoothing. Compared to FBS based on the N-scan pruning, the proposed algorithm offers better performance in terms of track loss, root mean squared error (RMSE) and normalized estimation error squared (NEES) without increasing the computational complexity.