Uncalibrated computer vision techniques in the underwater environment

Computer vision research has focused mainly on calibrated vision systems, i.e., for which painstaking procedures have been followed in order to estimate the values of the systems internal parameters. Unfortunately, the calibration process is a very slow and difficult procedure, often of inadequate accuracy and requiring systems to be taken off-line for large amounts of time, thus leading to increased costs. Moreover there is no guarantee that the estimated parameters will not change, for instance due to mechanical or environmental reasons, or if the camera uses a zoom mechanism. Uncalibrated computer vision dispenses with the whole calibration procedure, extracting all the information it needs from the images themselves. Hence, it allows for greater versatility since no assumptions about the camera setup are made, and greater robustness since one possible source of error, i.e., the calibration procedure, is eliminated. In particular the field of underwater applications of computer vision where thermal and mechanical strains on the camera are the norm, along with a usually unknown and highly unstructured environment, badly illuminated scenes and noisy images, provides an environment where the use of uncalibrated computer vision techniques could prove very beneficial. This paper demonstrates early results of the application of uncalibrated computer vision techniques in the underwater environment. The robustness and versatility of the techniques are demonstrated through synthetic data and real underwater sequences