Abstract
An extended differential stereo vision algorithm designed to perceive an opaque object as well as superposition of transparent surfaces is described. It is based on a principle what we call a probabilistic accumulation of the binocular disparities. In the accumulation system, the outputs of a differential disparity detector are histogrammed, then its occurency peaks are detected to determine individual heights of the surfaces. To prove this principle, we derive first a relation between true surface heights and a response to them of the differential disparity detector. It is shown to be an average height weighted by their transparency gradients. Then by analyzing its probability density function, we show the response values tend to distribute in a multimodal manner such that their peaks correspond to the individual heights. The condition is pointed out to be related intuitively to a spatial inhomogeneity of the transparency variations. Several results of computer simulation are shown to examine a performance of the algorithm.
