Extracting objects by clustering of full pixel trajectories

We propose a novel method for the segmentation of objects and the extraction of motion features for moving objects in video data. The method adopts an algorithm called two-dimensional continuous dynamic programming (2DCDP) for extracting pixel-wise trajectories. A clustering algorithm is applied to a set of pixel trajectories to determine objects each of which corresponds to a trajectory cluster. We conduct experiments to compare our method with conventional methods such as KLT tracker and SIFT. The experiment shows that our method is more powerful than the conventional methods.