A self-tuning system for real-time optical flow detection

Abstract

The work presented in this paper is aimed towards the real-time analysis of image sequences acquired from a moving vehicle. The authors' goal is to extract the optical flow field from the image sequence. The information obtained is used to detect obstacles in front of a moving vehicle and to compute the time to impact for approaching objects. These requirements mean that special care must be taken with the response time of the algorithm. In order to obtain a fast response, a special purpose massively parallel architecture is used. Furthermore, the algorithm uses an heuristic approach rather than an analytical one. In the approach presented, the detection of optical flow is driven by a high level process, by tuning the algorithm's parameters. >