WiiGee Gesture Recognition

[Umeer]

Hello,

I have been using the wiigee gesture recogniton and have been impressed with the work. I have been looking at the code and been looking to expand it. I have used it and it seems fairy accurate in terms of detecting gestures that have been recorded. The issue I am having with it is that some very random movements also get detected as gestures even though they don't even remotley resemble the gesture wiigee says it has recognised. I was wondering if this has always been the case, and if so, is there a way around it, or is it just my experience with wiigee.

Regards,

Umeer

[BePo]

Hi Umeer,

what you're experiencing is called 'false positives'. This is a common issue of such recognition algorithms. There are some approaches around to solve this issue, you may can have a look at them. Finally, wiigee's big PRO is that it can recognize trained gestures and do not ignore unknown movements. ;-)

Best Regards,
Benjamin

[robomotic]

Dear Benjamin,
I just tried the WiiGee and it's fantastic and I noticed that the gestures are not recognized if executed with a different orientation.
Do you plan to add an orthonormalization phase so that the gesture is recognized no matter what the orientation is?
I have found the method on a paper.
I can post if if you are interested.
cheers

[BePo]

Hi Robomotic,

thanks for your offer. I have thought about stuff like that, but thought it could reduce wiigees recognition rates. Probably this can be realised within an additional Filter. Would be great if you let me know the paper title that I can grab it from the web. However, I can not guarantee that I will integrate it in the next time as there is only a very small amount of time left for wiigee extension.

Thanks in advance.
Benjamin

[robomotic]

Hi,
sorry i didn't check the forum.
The paper is available here:
http://www.springerlink.com/content/v7yyaeyt37ttaqr7/
if you cannot have access I can send to your email.

[BePo]

Thank you. I've access to Springer and will read the paper soon.

Best Regards,
Benjamin

[robomotic]

I'm more than happy to help you with the implementation.
The paper is not well written though, because the author used the transpose notation T but with the same notation he also counts the number of samples of the 1 matrix (why he couldn't call them N?).
I understood that the first operation to do is the average of the x,y,z acceleration over N samples, then after finding r2 with the projection operation one can find the r1 and r3 vector that composes the R matrix for the matrix rotation.