GSoC Accelerometer-based Gestures Update 4

[Paul-Valentin Borza]

I've been busy the previous week as I had to finish writing my bachelor's
thesis on the gestures.
Daniel Willmann - my mentor - sent the Neo FreeRunner to me on Friday. It
still hasn't arrived, but it's now in Hungary and will arrive this week.
That's definitely good news :)
Like I've said to Mickey and Daniel while we were in the proposal phase of
GSoC, I have to implement some tweaks in HMMs to achieve better results.
State duration (a.k.a. duration modeling) will be used to prune unpromising
models in the recognizer.

Like I've said, each model (gesture) has a number of states; each can have
its own number of states and these states differ from one model to the
other.
Since I'm using a left-to-right model topology, the decoded sequence must be
something like this: 1 2 3
where 1 is the initial state, and 3 is the final state.
If the last state is not the final state (the model doesn't reach the final
state) then it's clearly not a match (trivial).

Also here's what happens with models that aren't a match. Note that the
length of the state sequence is the length of the observation sequence (HMM
assumption).
12333333333333333333333333333333333333333333344
It's clear that the state duration should be proportional with the length
and a model that generates a state sequence like that should be pruned.
Instead, I'm looking for something like
1111111122222222223333333333444444444

This is just a simple optimization that can improve the accuracy. After I
prune those models, I look at the posterior probabilities and decide on the
best match.

Paul
-- 
http://www.borza.ro
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.openmoko.org/pipermail/openmoko-devel/attachments/20080623/2d4f345c/attachment.htm