Top head detection from overhead camera

I have a camera set up开发者_开发问答 on a ceiling , capturing people standing in line. I need to count the amount of people. Currently I do that with background subtraction and Hough transform. The detector works well for contrast heads, but if hair color is close to the cloth color, the performance is poor. Any advices would be greatly appreciated. Example image here http://old.nabble.com/file/p32380283/1.png

Hough transform is generally good for geometric shapes that are partly obscured - wouldn't be my first choice for heads.

Assuming the camera is fixed I would start with anything that doesn't match a stored background image.

Then I would just look for edges and find a center of gravity and area of each blob. If the blob is large enough to be a head and it's c of g moves at about the right rate for a person (ie it isn't sub-pixel stable for minutes at a time or moves metres/second) then I would count that as a person.