Tech Report CS-90-09

Applications of Pyramid Structures to Multiscale Optical Flow

Ted Camus

August 16, 1990


An important problem in machine vision is the description of motion of a 3-D world as projected on a 2-D retina (optical flow). Most current optical flow algorithms are not suited for practical implementations such as tracking because they require massively parallel supercomputers, highly specialized hardware, or up to an hour with a scientific workstation. This paper solves this problem in three ways. First, we provide a fast serial implementation for a recent optical flow algorithm that can run up to ten frames per second for practical image sizes on a conventional workstation. Second, an efficient method for abstracting a high-level description of the motion in an image is described, along with an application for tracking. Finally, we provide a heuristic that speeds up the original optical flow algorithm by a factor of three using a pyramid architecture.

(complete text in pdf)