Feature detection for optic flow on carpet/patterns/rough surfaces

edit

Hello All,

I am running the optic flow algorithm(LK) using feature points from good features to track(Shi tomasi). However this method for feature detection doesnt seem to work very well when there are repetitve patterns(tiles), rough surfaces(such as roads), carpets. Flow from LK goes haywire in such cases. Can feature detection/flow calculation be improved in some way in atleast rough surfaces using some other method or by using some pre-processing before finding features/flow? As I need to calculate flow in real time, I cannot really afford dense methods, as it might hamper performance. Could you please suggest some ways to go about tackling this problem?