Motion trajectories

A simple algorithm for selecting and linking interesting flow vectors across a sequence of frames for computing motion trajectories is presented. Tokens are tracked that have both interesting pixel gray values in the spatial domain and in the optical flow field in the temporal domain. This effectively removes some redundant trajectories. Due to errors introduced during the computation of optical flow, and the linking of such flow vectors across a sequence of frames, the resultant trajectories are not always smooth. A Kalman filtering based approach is discussed for smoothing the trajectories. Isolating the trajectories into sets belonging to individual objects is an important first step that should be taken before any type of shape or motion interpretation can be done. Therefore, a simple algorithm for segmenting motion trajectories is also discussed. When motion trajectories belonging to a single translating object are extended, they intersect at a single point called the focus of expansion (FOE). If the motions of objects are assumed to be independent, each FOE represents one object. Therefore, FOE can be used to segment trajectories belonging to individual objects. A simple but highly robust algorithm for partitioning motion trajectories is presented that does not require the exact location of FOE, but uses some useful properties of FOE. The authors have applied their method for computing and segmenting motion trajectories to a number of real sequences, and have obtained very encouraging results