Re: Ant's homebrew motion detector

[Ant Skelton <ant@xxxxxxx>]

James Derrick wrote:
>
> Your images are interesting (not least for the test images!) when the
> temporal difference is compared with the detected motion. Temporal
> noise (changing lighting?) is filtered out nicely, but the detected
> motion seems to be rather insensitive.

The detected motion frames on that page are frames which have both
temoral and background motion, whereas they should probably be frames
with EITHER temporal OR background motion, which would show much more
sensitivity in the given example, but reduced immunity to lighting
changes.

Having said that, the algorithm is parametric (there are four things you
can fiddle about with to adjust sensitivity etc), and I just picked some
numbers out of thin air. I'm investigating how they affect the
usefulness of the algorithm at the moment.

The technique I used to generate the web page images is extremely
cumbersome and time consuming, so I'm knocking up a KDE2 app to make the
whole process easier.

> I take it that the background
> 'statistical model' is something like a weighted mean of past temporal
> diffs?

Yes, pretty much. The model comprises a weighted mean and a weighted
standard deviation. It's updated every frame from the adaptation region,
which is the region of the frame for which no temporal motion has been
detected in the last n frames. Background motion is excluded from the
adaptation process to prevent propagation of errors into the model.

At the moment I initialise the background model with 'suitable values',
but it will be possible to train it with a 'ground truth' image
sequence, i.e. one in which no motion occurs.

I hope to have some more pages up shortly, and maybe a *severely* alpha
version of the app  and library.

This is all taking much longer than I anticipated - I'm meant to be
finishing my DIY curtain automation! Oh well. ;)

cheers

ant
ant@xxxxxxx

-------------------------- eGroups Sponsor -------------------------~-~>
eGroups eLerts
It's Easy. It's Fun. Best of All, it's Free!
http://click.egroups.com/1/9698/3/_/2065/_/974317747/
---------------------------------------------------------------------_->