Where is the bug ?

Help me catch the big hidden bug that spoils all the results from the current movement recognition setup for home.

When properly setup, PCA object & Svm object are powerful tools. Svm recognizes these ugly digits:

If you do not believe me, you can read about HeroSvm.

Doing some google walking, I fell on a paper on Transformation Invariance in Pattern Recognition and thought this might be the main weakness of our setup: the movement is only recognized during a very short window and once the movement is past, noise is recognized. If I had a way to make the recognition sticky (using a probability threshold for example), I could easily filter the “noisy recognitions” but I could not find a way to get a probability out of SVM. I tried using the “votes” (how many times a class “wins” when competing two by two with the other classes) but it is not really convincing (too few classes for the moment).

It might be time to ask the specialists for some help… Or try Hidden Markov Models.