Towards Automatic Visual Obstacle Avoidance

Abstract

This report describes ongoing research on a working system which drives a vehicle through cluttered environments under computer control, guided by images perceived through an onboard television camera. The emphasis is on reliable and fast low level visual techniques which determine the existence and location of objects in the world, but do not identify them. Included are an interest operator for choosing distinctive regions in images, a correlator for finding matching regions in similar images, a camera solver which determines camera
displacement and distance to objects from stereo information [Gennery, D.B., this Proceedings] and an automatic geometric distortion corrector for camera nonlinearities. Many of these use pictures reduced in linear dimension by powers of 2 by summation of pixels. Other operators are a high pass filter, a point noise remover, a contrast normalizer, a vertical roll corrector, a picture comparator and an operator for reducing pictures by other than powers of two.