On robustness and localization accuracy of optical flow computation from color imagery

Accurate and efficient optical flow estimation is a major step in many computational vision problems, including tracking and 2-D/3-D mapping applications. Processing of grayscale images has been the dominant approach, with only a few studies investigating selected aspects in the use of color imagery. In a physics-based analysis to study the impact of the spectral-dependent medium attenuation on the color channels, we have shown merit in the use of color cues in the computation of optical flow for underwater imagery- the primary motivation of the investigation. Comparisons among various color representations and traditional intensity component on the optical flow computation are given, suggesting that the HSV representation could be the most suitable. For both underwater and terrestrial imagery, even where data in the 3 color channels are highly correlated, one expects multiple constraints from color channels to give increased robustness due to the independent channel noises. Results of experiments are given to demonstrate improved localization and accuracy.