My knowledge about the matter is almost certainly out of date, but when I was working on image recognition over a decade ago the process I went through was like this: translate image to grayscale, run an edge detection algorithm, and then look for certain features. Lighting and dark colors had a big effect on the detection quality, and the whole thing wasn't very robust. I can see how skin color could be an issue if there isn't enough contrast in the image to identify features.