Hi,
It'll be more helpful to answer your second question first. To start, see here for the definition of a multimodal distribution:
https://en.wikipedia.org/wiki/Multimodal_distribution
When applied to modeling background pixels, a multi-modal model can capture the different states a background pixel can take over time. For example, in an image with a lot of sky, they sky is usually thought of as the background. The pixels representing the sky will change from blue to black as you go from day to night. Blue and black are the two modes of the sky pixel values; both represent background modes.
A unimodal model only captures one mode, so in the sky example, only blue pixels would be considered background. When it gets dark, the unimodal model is not able to capture the fact that black pixels are also background.
For your first question, one situation to increase it's value is if your video has a lot of background motion. In this case a high value will ensure most of the background changes are captured by one of the modes of the background model. If it is set to a low value, then the background will be modeled by fewer modes, and any motion will be detected as foreground. This can lead to an increase in false positives.
This paper has more details about the algorithm:
'MinimumBackgroundRatio' corresponds to the "T" in equation 5 within the paper.
Hope that helps, Birju
