Interpreting Images of Small-bore Urban Sewers

We present an overview of a project that aimed to develop and implement computer vision techniques capable of extracting quantitative, three-dimensional descriptions of the shape of the surveyed pipe from sequences of images captured by an uncalibrated camera moving along a small-bore urban sewer. The resulting techniques are capable of use both on-line, during actual sewer surveys, and offline, to process video records of previous surveys. Novel techniques were developed for the recovery of optic flow fields from images of steep surfaces under high velocity camera motions and moving illumination. Extrinsic camera parameters were recovered from large-scale image features.