|RESEARCH - Offline Global Alignment|
The online mosaic generated during the data acquisition phase of an exploration mission will provide only a first, rough approximation of the final mosaic image. This is mainly due to the fact that time constraints during do not allow us to apply sophisticated data processing techniques. Nevertheless, the online mosaic will provide scientists with a suitable base to evaluate the quality of the acquired data immediately after the mission. Mosaics will be then refined applying global alignment techniques. In contrast to the rough motion estimations obtained during online mosaicing, global alignment calculates a globally consistent mosaic image considering all the images of a given sequence at the same time, in addition to the navigation data.
Global alignment subsumes several steps. Initially the motion between two consecutive frames is reconstructed given rough initial estimates from the georeference data as well as from the online processing from previous phase. These estimates are then refined by taking visual image data into account. The global arrangement of all sequence images is derived from these pairwise motions by transferring the local motion into a global coordinate frame. Finally, based on this, image pairs that are temporally non-consecutive, however, spatially overlapping can be identified. The final global alignment is done by optimizing all pairwise motion of spatially adjacent images while taking the global consistency of the mosaic into account.
In addition to the registration algorithms needed to generate a globally consistent mosaic image, additional capabilities require development and investigation during this phase. Since underwater images often suffer from various distortions (e.g., geometric distortions due to light refraction and lens distortion, or radiometric distortions due to vignetting or non-uniform illumination of the scene), appropriate image preprocessing steps will be carried out to compensate for the distortions before starting the global alignment.
With regard to biological and geological interpretations of the data, also color yields an important source of information. However, underwater images acquiring natural light suffer from light attenuation due to the water column and real colors of underwater objects are not preserved in visual images. We will address this problem by including color correction techniques that allow recovering the original color in the absence of absorption and turbidity. For this the high-resolution photographic camera mounted on the robot in addition to the video cameras will be used in conjunction with multi-spectral correction algorithms.
Finally, after correcting and globally aligning all the images the resulting mosaic image has to be rendered adequately to allow an easy exploration by scientists. On the one hand this subsumes the investigation of blending techniques that allow to fuse all sequence images into one single frame, e.g., omitting any visible seams between different images and preserving individual color information. On the other hand, rendering a mosaic includes the automatic choice of an appropriate coordinate frame for representation of the mosaic that minimizes perspective distortions.
The software that has to be developed within the context of this work package will heavily rely on work that was done in the VICOROB group before. In particular, local as well as global registration algorithms already exist that will be extended with regard to the requirements of the new field of application. Also, for mosaic rendering and visualization the "MosaicViewer" of the group will be applied and further adapted for scientific use.