A bit of clarification on Lewis' skin detection. (If anyone cares.) Since data came from the camera in YUV format, we didnt need to do any color-space conversions. YUV is similar to HSL insofar as one component cooresponds to the luminosity of the pixel (Y,L respectively) and the other two components describe the essential color. (Note: all skin is essentially red.) Since variation in skin tone due to different races or glare/shadows only really affects the luminosity, this component is ignored. The remaining UV pair is then checked against a look-up table of known skin chromosities. It turns out that skin occupies a fairly small region in the chromosity plane, so this technique is somewhat effective.
Two caveats: first, changing lighting conditions shift where skin falls in the chromosity plane, so the UV look-up table needs to be retrained for new environments or automatically shifted to account for lighting. Second, we found that because we ignore the Y component, certain shades of red are labeled as skin even though they dont resemble skin to a human eye. As such, for more robust skin detection the Y component cannot be completely ignored.
Michael Dixon Media and Machines Lab
Computer Science and Engineering
Washington University in St.Louis
Slashdot Mirror
Two caveats: first, changing lighting conditions shift where skin falls in the chromosity plane, so the UV look-up table needs to be retrained for new environments or automatically shifted to account for lighting. Second, we found that because we ignore the Y component, certain shades of red are labeled as skin even though they dont resemble skin to a human eye. As such, for more robust skin detection the Y component cannot be completely ignored.
Michael Dixon
Media and Machines Lab
Computer Science and Engineering
Washington University in St.Louis