Have you considered training your system with various light situation (i.e. sunset, sunrise, daylight, cloudy, darksky, night). It should help.

Have you considered training your system with various light situation (i.e. sunset, sunrise, daylight, cloudy, darksky, night). It should help.

You can use illumination normalization techniques such as histogram equalization,adaptive histogram equalization,Log transformations, homomorphic filter,contrast stretching,Retinex filter,...etc

Try considering HSV or similar color models, Value depends on the lightning conditions, Saturation - is less dependant on the lightning but Hue never changes (unless You get totally underexposure or overexposure)

You have three possibilities, they are:

1) Use HSV representation and use the Hue

2) Train the machine learning algortihm with different illumination scenario

3) Normalize/Equalize the image