Trajectory retrieval from Monte Carlo data association samples for tracking in fluorescence microscopy images

The dynamic behavior of sub-cellular structures is of great interest to bio-medical research. We propose to use probabilistic multi-target tracking to analyze the dynamics of stress granules (SGs) to alleviate the effort of manual analysis. Inherent to multi-target tracking approaches is the combinatorial problem to associate observations to underlying targets. Rao-Blackwellized Monte Carlo Data Association circumvents this problem by sampling in the space of associations. As each sample provides its own hypothesis of SG trajectories, we employed a graph partitioning algorithm to extract one single set of trajectories. This is shown to outperform the sample with maximum probability on both synthetic data and fluorescence microscopy images.