OPTICAL FLOW THROUGH RELAXATION IN THE VELOCITY SPACE

A new multiconstraint approach to the estimation of optical flow is presented, which is based on constraint relaxation in the velocity space. An architecture for a parallel implementation of the algorithm is suggested, in which relaxation is accomplished by a mesh grid of loosely coupled simple computational units, one for each image point where optical flow has to be computed. Optical flow smoothing is performed by a vector median filter, whose nonlinear nature is helpful in preserving motion boundaries. Results on both synthetic and real-world sequences are presented, which demonstrate that the algorithm behaves well in the presence of noise and occlusions.