Efficient Computation of Intensity Profiles for Real-Time Vision

Abstract

For the EMS-vision system realized on distributed general- purpose processors with a set of video cameras on an active gaze control platform, an efficient method for exploiting area-based image information has been developed (as opposed to edge features preferred in real- time vision systems up to now). It relies on the same oriented intensity gradient operators as have been used for edge localization in the past (K(C)RONOS). However, the goal achieved now is fast derivation of one-dimensional intensity profiles with piecewise linear shading models. First, regions of large intensity changes (so-called ‘non-homogeneous’ regions) are separated from ‘homogeneous’ ones containing at most moderate intensity changes (to be specified by a threshold parameter). The average intensity values and ternary mask responses in these areas yield information for a coarse linear (first order) intensity model. Then, in the homogeneous regions, the one-dimensional equivalent of a pyramid (a triangle-) representation is derived for the residues between the actual intensity values and the coarse linear model. Depending on the size of the homogeneous region and the number of intensity peaks, a certain triangle level for further processing is selected. Again, a (different) ternary mask operator is used for intensity gradient computation and for finding the zero-crossings of the gradient. This information is sufficient for determining the fine structure of regions with linear shading models. Examples are given for road and vehicle detection and recognition.