That formulation would be for a non-causal filter, i.e. process the data off-line after you have collected it. For a causal implementation, i.e. using only data you have collected up to the current time, you would be loading all of the incoming data into a buffer (a sliding window) of length M. If M is odd then you would be indexing from the centre of that window (for m =-K...+K). The m=0 point in that loop would be delayed from the current time by  K samples, if M = 2*K+1. So just shift everything so you don't get negative indexing. In doing this you have just applied a delay of K samples to make your processing causal.     

Rakesh, this is just a notation and this thing must be implemented by starting with index (0,0) if you are working in C otherwise with index (1,1) if you are working with Matlab...

It is because when you are comparing an image with a template, you start from the center point. Negative coordinate means, (left in x, up in y) positive means (right in x, down in y).

So you may want to start from not 0, 0.

Instead

tx = templatewidth / 2;

ty = templateheight / 2;

for(int x = tx; x < width - tx; x++)

  for(int y = ty; y < height - ty; y++)

    .. so on..