Multiwindow least squares approach to the computation of optical flow with discontinuities

The use of optical flow fields in image sequence analysis allows us to perform motion-based segmentation as well as 3-D reconstruction. Many techniques for optical flow estimation are based on some global or piecewise global smoothness assumption. Other techniques compute the flow field based only on local information. A local algorithm explicitly addressing the problem of evaluating a reliable optical flow field at motion boundaries is presented. Velocity vectors are computed as solutions of a multiwindow least-squares problem; the field is then regularized by a vector median filter. The algorithm is noniterative and nonparametric. Results on both synthetic and real-world sequences are shown; a performance comparison with two well-known techniques demonstrates the effectiveness of the algorithm in terms of noise rejection, motion boundary preservation, and speed.