Abstract

Most current optical flow algorithms are not suited for practical implementations, such as tracking, because they either require massively parallel supercomputers, specialized hardware, or up to several hours of computing time on a scientific workstation. A recent optical flow algorithm¹ has been shown to be robust and noise-resistant, but until now it could only run in anything close to real-time on a massively parallel computer (such as a Connection Machine). We describe a modified algorithm that uses a pyramid architecture that is fast enough to power a real-time tracking system and requires only a conventional workstation. Three main techniques are employed. The first is a simple but powerful abstraction mechanism, based on a pyramid architecture, that can quickly extract the information from an optical flow field, which is necessary for real-time tracking. The second technique uses this abstraction mechanism to control focused fine-scale optical flow processing, and it results in a detailed optical flow field at a fraction of the normal cost. Third, by a controlled image-sampling rate, real objects can be tracked over a wide range of velocities.

© 1990 Optical Society of America