How to Speed Up EigenBackground Calculation?

Question

I'm trying an optimization of this algorithm for Adaptive Background Detection with Moving Camera Video:

http://www.springerlink.com/content/4j2m3885r3133202/fulltext.pdf

This algorithm speed up the upgrade of the background model (SVD is too heavy for a realtime application). There are 3 attribute that I can manage for speed up the algorithm. The attributes are:

1. N the number of frame for the first SVD calculation
2. M the number of eigenvector for describing the background model (M<=N)
3. P the number of frame to wait for doing a batch update

Do you know how I can find out a relationship between the processing time and these 3 attributes?

Do you know how find out the optimal values?

The heavy part of this algorithm is the QR decomposition to orthonormalize the incoming frame during the batch update. This function takes lots of time.

Do you know faster algorithm for calculate orthonormalization?

Here the function QR of m_cT (matrix U'=[U|E] in the paper) that I'm using:

  for (int i = m_iNFrame; i < m_iNFrame+m_iNUpdate; i++)
    {
    m_cT.col(i).copyTo(m_cQ.col(i));
    for(int j = 0; j < i; j++)
    {
      m_cR.at<double>(j,i) = m_cQ.col(j).dot(m_cT.col(i));
      m_cQ.col(i) = m_cQ.col(i) -  m_cR.at<double>(j,i)*m_cQ.col(j);
    }
    m_cR.at<double>(i,i)=norm(m_cQ.col(i));
    m_cQ.col(i)=m_cQ.col(i).mul(1.0/m_cR.at<double>(i,i));
  }

I noticed during the test that high frequency detail in the frame create lots of false positive during the motion of the camera.

Do you think that the application of a smooth filter before the SVD calculation can get a better detection BG/FG?

Thank you for your help.