We present an approach for the computation of single-object velocity statistics in a noisy fluorescence image series. The algorithm is applied to molecular imaging data from an in vitro actin-myosin motility assay. We compare the relative efficiency of wavelet and curvelet transform denoising in terms of noise reduction and object restoration. It is shown that while both algorithms reduce background noise efficiently, curvelet denoising restores the curved edges of actin filaments more reliably. Noncrossing spatiotemporal actin trajectories are unambiguously identified using a novel segmentation scheme that locally combines the information of 2-D and 3-D segmentation. Finally, the optical flow vector field for the image sequence is computed via the 3-D structure tensor and mapped to the segmented trajectories. Using single-trajectory statistics, the global velocity distribution extracted from an image sequence is decomposed into the contributions of individual trajectories. The technique is further used to analyze the distribution of the and components of the velocity vectors separately, and it is shown that directed actin motion is found in myosin extracts from single skeletal muscle fibers. The presented approach may prove helpful to identify actin filament subpopulations and to analyze actin-myosin interaction kinetics under biochemical regulation.