Mr Vitalijs Butenko has been working on the movement recognition for the last 6 months and gave his final presentation today in EPFL. We could do a live demo with the recognition of 5 classes triggering sounds.

Here is a table taken from Mr Butenko’s research showing the error rate on the recognition of the 7 movements recorded by Nicholas Petit here in January.

  W=50 W=75 W=100 computation time
k-means 42% 38% 36% < 0.5s
DTW 27% 20% 12% > 1s
SVM 0.8% 0.8% 0.5% < 1s
Error rates for 3 window sizes @ 256 samples/second

From this table, we see that support vector machines looks very promising with less the 1% error ! And the best is that there is a free library with ruby wrappers available : libsvm.

As soon as the rubyk editor starts to work, we will jump into this new classifier.

Many thanks to Mr Butenko for his great work.

pdf document

2.7 Mb

butenko_report

Movement recognition report from Vitalijs Butenko for project home.