Motion Research Progress Report: 2010-04-19

I had to take a break from the motion research project because I had to do some work on developing EDDE 221. I just finished module 4, which is pretty big for a module. It's one of the biggest modules I've ever developed at UPOU. Anyway, there's still a ton of work to do there.

So. Back to Weka.

I think what I need to start doing is divide this Progress Report into the subsystem reports. Here are the subsystems so far. ADU stands for Acceleration Data Unit.

•  Wearable Hardware System
• I ordered an XBee USB Explorer and a FTDI basic breakout. Tomorrow I will see whether that helps me with my hardware woes
• Hardware-Software Interface
• I haven't taken a look at what Mary did, and I'm not sure if I will at this point, at least not for IAT 888. She used MaxMSP to do all the serial reading. I know I can use the Processing code that I already have.
• Acceleration Data Cleaner
• Haven't even touched this yet. I'm Greg mentioned a kind of filter that he says is always used in motion cleaning. Damn I can't remember it.
• Motion Segmentation System
• Also haven't started working on this, but I'm looking at what Mary did and I'm supposing I should first try the general approached she did (sliding windows  of various sizes and different weights) to be used for the LEC
• Motion Feature Extractor
• Well, I have candidates for these features! Now I just need to actually derive them based on acceleration data.  :{
• Laban Effort Classifier
• I have a working classifier! Right now I'm using a toy domain to test the classifier ("SHould I play or stay in based on various weather conditions?").  I have a working GUI demonstration.

I put a lot of work learning about Weka, and I'm hoping that this will pay of in the long run in that I am not married to using a neural network to do the classification, and Weka will provide me a more opportunities to try other machine learning methods
Now I know that I will get the best performance out of my super rudimentary system if I build the MFE first and use the motion features for the classification. But I need to have something to demonstrate on Thursday, so I will see what will happen if I feed direct accelerometer values into the neural network. At least what I can then demonstrate is that I can recognize ... perhaps shapes drawn in the air?  That's a good start, and it will help with Space classification, at least.

So it looks like the next step is for me to work on the motion segmentation system. Conveniently enough, Mary's approach builds precisely on what Philippe told me to do, which is to use sliding windows.

Labels: metacreation, motion research