On robustness and localization accuracy of optical flow computation for underwater color images

Color-image optical flow processing may offer little merit in terrestrial imagery, mainly due to high correlation among the color channels. However, spectral-dependent environmental factors reduce the degree of correlation in underwater imagery, enriching visual motion cues. Additionally, use of multiple motion constraints can increase estimation robustness and noise immunity, which is significant for overcoming higher underwater image noise from various sources. Despite high variability in the conditions of various bodies of water, a simplified image model allows us to draw general conclusions on the computation of visual motion from color channels, based on average common medium characteristics. In particular, the model offers insight into: (1) information encoded in various color channels; (2) advantages in the use of a certain color representation over others; (3) consistency between conclusions from the theoretical study and from experiments with data sets recoded in various types of ocean waters and locations. The study concludes that optical flow computation based on the HSV representation typically provides more improved localization and motion estimation precision relative to other color presentations. Results of various experiments with underwater data are given to assess the accuracy.