GSoC Accelerometer-based Gestures Update 2

Paul-Valentin Borza paulvalentin at borza.ro
Sun Jun 8 15:16:51 CEST 2008


As promised, the classifier is ready. Does anyone have a Wii Remote near by?

Check out the files from SVN and compile with Eclipse CDT.
svn checkout https://svn.projects.openmoko.org/svnroot/gestures

In gesphone folder, there's a shell script.
Run it with ./gesphone.sh --load call/call.ges
Follow the instructions on the screen. You should be able to see some
accelerometer values when a sudden motion is determined.
The classifier actually does this: separated normal moves from sudden
moves (gestures).

If you think it's too sensitive, you can always re-estimate it for
your own taste with the gesapp
In gesapp folder, there's also a shell script.
Run it with ./gesapp.sh --new-class call/noise.gauss (for noise) and
./gesapp.sh --new-class call/motion.gauss (for motion)
To train ./gesapp.sh --train-class call/noise.gauss and ./gesapp.sh
--train-class call/motion.gauss (for motion)

There's still a problem with the training: when more training sessions
occur, the values in the covariance matrix will converge to 0.0 (as
expected) and then when the determinant is computed it will end up to
be 0.0 and when the probablity density function divides with this
determinant which is 0.0, it will give nan. I'll have to enforce a
minimum value for the covariance matrix.

Thanks,
Paul
http://www.borza.ro



More information about the openmoko-devel mailing list