Normalization and cross-correlation - what am I doing wrong?

Hi NG,

I have a problem regarding normalization when performing
cross-correlation:

I have a relatively short signal/waveform ("A") which I'm
cross-correlating with a longer timeseries ("B") to look for this
waveform. Since there are large fluctuations in B, I have decided to
apply some sort of normalization scheme to both A and B. So, for A I
divided all the samples with the squareroot of the average power of the
signal, i.e. sqrt(sum(A(i)^2)/N). As for B, I do the same thing FOR
EACH WINDOW OF "B" THAT "A" PASSES IN THE CORRELATION. That is, before
the windows are multiplied and summed in the correlation, the segment
of B in question undergoes the same treatment as described above.

The result looks very strange, and is definately not correct. Do you
see any flaws in my method?


Thanks!

Best,

M.L.

Normalized correlation is:

sum(A(i)B(i))/sqrt(sum(A(i)^2) sum(B(i)^2))

Where all summations are over the same number of elements N (regardless 
of the actual length of each signal, which may have elements that do not 
participate in the correlation).

M.L. wrote:
> Hi NG,
> 
> I have a problem regarding normalization when performing
> cross-correlation:
> 
> I have a relatively short signal/waveform ("A") which I'm
> cross-correlating with a longer timeseries ("B") to look for this
> waveform. Since there are large fluctuations in B, I have decided to
> apply some sort of normalization scheme to both A and B. So, for A I
> divided all the samples with the squareroot of the average power of the
> signal, i.e. sqrt(sum(A(i)^2)/N). As for B, I do the same thing FOR
> EACH WINDOW OF "B" THAT "A" PASSES IN THE CORRELATION. That is, before
> the windows are multiplied and summed in the correlation, the segment
> of B in question undergoes the same treatment as described above.
> 
> The result looks very strange, and is definately not correct. Do you
> see any flaws in my method?
> 
> 
> Thanks!
> 
> Best,
> 
> M.L.
> 

the description of your problem is not easy to understand.

why normalize with power?

I would do it like this

a_new=a/max(a);
b_new=b/max(b);

cross-correlate a_new with b_new and look for peak(s) in cross-correlation.

can you upload screenshots to some webpage so we can see why it looks 
"strange"?

which language are you programming in?

Hi,

Thanks - both of you. I changed my code (MatLab by the way) to use the
formula stated in twains post and now it seems to work much better. I
think my mistake was that I divided the sqrt(power) with the number of
samples in the window. Apart from that, I had the same calculation..

Thanks! I appreciate it..

Best,

M.L.