Open Computer Vision Library

Symptom:
When updating the background model (icvUpdateGaussianBGModel()), in partial
window condition, when one pixel has a value of 0, that pixel becomes 0 in
the background model and so remains forever. For the tests I used 1-channel
frames.

Cause:
The icvMatchTest() function matches the zero pixel to the first
uninitialized gaussian component whose mean value is 0 (mean is initially
set to 0 by the cvCreateGaussianBGModel() function). Afterwards
icvUpdatePartialWindow() sets to 1 the match_sum of that gaussian component
and the variance to 0 (icvUpdateFullWindow() does not so), and then
icvGetSortKey() calculates the sort key to inf (division by zero), making
that gaussian component dominant forever.

Proposed solution:
To prevent such a division by zero, the gaussian components that have never
been matched don't have to be matched in icvMatchTest(). Just put the
following test around the contents of the for-cycle inside icvMatchTest():

if( g_point->g_values[k].match_sum > 0 ) {
...
}