Color Separation for Machine Vision

Accurate color discrimination is vital for autonomous driving systems to detect traffic signals, road markings, and emergency vehicle lights. Traditional color separation metrics like ΔE or ΔC are insufficient because they overlook sensor noise and changes from image signal processing, focusing only on human visual perception. Euclidean distance metrics also fail under linear transformations like those introduced by color correction matrices (CCMs) and white balancing, leading to misleading assessments of color separability—especially when transparent color filter arrays (CFAs) like RCCB and RCCG amplify cross-channel noise. This presentation examines various CFAs and illumination sources, comparing pre- and post-CCM color separation. Findings show Mahalanobis distance more reliably measures color separation by considering noise distribution characteristics. The impact of non-linear transformations such as conversion to CIELAB and gamma correction on this metric are also presented.