Mosaicking Techniques for Deep Submergence Vehicle Video Imagery - Applications to Ridge2000 Science

Severe attenuation of visible light and limited power capabilities of many submersible vehicles require acquisition of imagery from short ranges, rarely exceeding 8-10 meters. Although modern video- and photo-equipment makes high-resolution video surveying possible, the field of view of each image remains relatively narrow. To compensate for the deficiencies in light and field of view researchers have been developing techniques allowing for combining images into larger composite images i.e., mosaicking. A properly constructed, accurate mosaic has a number of well-known advantages in comparison with the original sequence of images, the most notable being improved situational awareness. We have developed software strategies for PC-based computers that permit conversion of video imagery acquired from any underwater vehicle, operated within both absolute (e.g. LBL or USBL) or relative (e.g. Doppler Velocity Log-DVL) navigation networks, to quickly produce a set of geo-referenced photomosaics which can then be directly incorporated into a Geographic Information System (GIS) data base. The timescale of processing is rapid enough to permit analysis of the resulting mosaics between submersible dives thus enhancing the efficiency of deep-sea research. Commercial imaging processing packages usually handle cases where there is no or little parallax - an unlikely situation for undersea world where terrain has pronounced 3D content and imagery is acquired from moving platforms. The approach we have taken is optimized for situations in which there is significant relief and thus parallax in the imagery (e.g. seafloor fault scarps or constructional volcanic escarpments and flow fronts). The basis of all mosaicking techniques is a pair-wise image registration method that finds a transformation relating pixels of two consecutive image frames. We utilize a rigid affine model with four degrees of freedom for image registration that allows for camera translation in all directions and camera rotation about its optical axis. The coefficients of the transformation can be determined robustly using the well-established and powerful featureless Fourier domain-based technique (FFDT), which is an extension of the FFT-based correlation approach. While calculation of cross-correlation allows the recovery of only two parameters of the transformation (translation in 2D), FFDT uses the Phase shift theorem of the Fourier Transform as well as a log-polar transform of the Fourier magnitude spectrum to recover all four transformation coefficients required for the rigid affine model. Examples of results of our video mosaicking data processing for the East Pacific Rise ISS will be presented