Person object tracking using optical flow and possibly k-means clustering?

The camera is fixed, above a doorway (Wish to implement this real-time). Basically I am trying to build a people counting program. My approach -so far- is as follows:

1.) Identify 'blobs', where movement is detected. I'm using foreground detection, this is proving problematic so far as the walls's corners are identified as the 'foreground' for some reason.

2.) I want to store the bounding boxes and use them as ROIs for the SURF algorithm. QUESTION 1: How to use the SURF algorithm on an image but only search within the specified ROIs.

3.) I then want to take all of the points found and cluster them using k-means in order to best identify 'people' below. I would rather associate a person with a set of feature points than with a blob using blob area or whatever becase often 2 people bump into each other, form one blob, and then mess everything up. QUESTION 2: K-means clustering requires a predefined value for K. Any Ideas?

4. Tracking with multiple KLT tracers won't be a problem after step 3 (hopefully).

Much Appreciated!