An Optical Flow Algorithm Preserving Motion Boundaries and its Parallel Implementation

One of the main techniques for the analysis of image sequences is the optical flow computation. The optical flow is a vector field representing the apparent motion, on the image, of the projected points of the objects in the scene. Many algorithms have been developed to compute the optical flow. In this paper one of these algorithms is reviewed, the performance enhanced and a possible parallel implementation presented. The main advantage of the resulting algorithm is its good behavior in presence of noise or motion boundaries; this is obtained by means of a local technique which doesn't require global optimization and by applying a median filter to restore the vector field. Furthermore a pyramidal approach is used to increase the range of interframe displacements computable by the algorithm. To enhance the time performance a parallel implementation on a network of transputers has been studied; a theoretical analysis of the efficiency of the implemented architecture is discussed and its agreement with the experimental results is reported.