This paper presents a technique that reliably tracks large numbers of particles undergoing dense antiparallel motion and frequent appearance and disappearance. Such techniques are essential to many applications of fluorescence cellular and molecular imaging for automated quantitative analysis of dynamic cellular functions. The basic tracking algorithm of this technique integrates motion models at particle, local and global levels. It establishes correspondence between particles based on state similarity and resolves correspondence conflicts using optimal graph assignment. A statistical and robust approach for algorithm parameter setting is developed through establishing the equivalence of the algorithm to a Kalman-filtering based tracker under assumptions that are biologically supported. Online track initiation and propagation depend critically on computing the global vector field of particle flow using a new optimal-flow minimum-cost graph algorithm. Vector field denoising and interpolation are performed using anisotropic filtering after clustering. The technique has been experimentally verified and successfully applied to the tracking of Fluorescent Speckle Microscopy images of live cells.