Multiresolution Active Contour Models in Textured-Stereo Images

This paper presents a method for generating sparse range data from textured surfaces which have structured light projected onto them. The work is motivated by the need to measure 3-D road defects rapidly and reliably. Traditional approaches to computing range from stereoscopic images have replied on either smooth or finely textured surfaces when using structured light. Conventional techniques that take advantage of the inherent texture in the images are not applicable. This is because corresponding stereoscopic road surface views are dissimilar due to the geometry of the cameras and the surface texture. The method described places initial edge points in a low resolution version of the intensity image. These points are used to initialise open active contour models or snakes which are propagated via a pyramid to a higher resolution. At this higher resolution, internal and external constraints are applied to the snake; the internal constraint being a smoothness functional and the external one being based on a maximum likelihood estimate of the edge strength across each light stripe. Computation is spatially localised at each stage and thus this algorithm could easily be parallelised.