Dear Colleague,

Thank you for introducing this kind of research and for asking an interesting question. Although we do not have a direct answer for you right now, this situation reminds me of the clustering of spatiotemporal data in gait research.

Our (upper- and lower-limb functional anatomy) research group dealt with evaluating gait frequencies (in primates and in other mammals as well). We recently compared such frequencies in man during health, and during the progression and the rehabilitation of neuropathic diseases. Some papers about it are, or will be on our ResearchGate site.

So in spite of all differences between upper and lower limb (interesting enough), your classifying problem may be redefined as a clustering problem. I add one seminal literature example in attachment (Kienast et al., 1999) about lower limb data.    Asking (bio-)statisticians may be helpful, in order to redefine your question. They will possibly suggest to you the right formulas and algorithms and / or software for this.

Wishing you success in continuing your project,

with kind regards,

Koos Jaap van Zwieten

Dear  Stavros,

in the figures you attached, I see a three dimensions space with acceleration (m/sec^2 ?), Rotation (angle values), and EMG which stands for electromyography, I suppose.

To my knowledge EMGs create very noisy data, depending on position, experimenter's ability , and on the condition of skin and muscle. So it can be very difficult to use these data.

Apart this, generally speaking, when you use sensors on people, some data may be less reliable than others and they can be more noisy and less usable, of course. 

This is a common situation. I tried to figure out a solution by proposing a general method to deal with this problem.

You could  reduce noise, using a feature selection approach and then count the frequencies of the intraclass features, to improve the validity of data for the  specific movement, exploiting repetitions; then also count  inter-class frequencies of the same features, reducing the importance of the less selective ones, according to a formula (see attachment). 

This could answer to your vision of across validation.

For  more specific details,  you may eventually look at my paper in attachment.

The results that I personally  obtained on a public dataset (created by U.C.Berkeley)  were pretty impressive even on very noisy data, reducing of four time the error rate of Berkeley researcher on their own dataset .

The paper is quite preliminary so if you need, I am available by e-mail, for further explanations.

It could be necessary to pass from value space to a feature space, but maybe it is not stricly required: the method could be adapted on the given space, I think.

About real time it depends on how fast it is the required reaction, and why.

I applied my method in realtime, on continuous acquisitions of several minutes,  but this requires a segmentation operation. And segmentation may add “noise”.

I hope this can be of some help.

My best,

Stefano

Conference Paper Movement Recognition with Intelligent Multisensor Analysis, ...

Dear Stefano,

Many thanks for your response.  The 3 dimensions (signals) shown in the figure are just 3 of the 16 total; about 8 of them are EMG signals and, yes, they are noisy but not overly so.  My main concern comes from the requirement of your approach for segmentation (I'm assuming of individual movements).  This can be done during the training phase.  But during the classification phase, our application requires continuous classification of movements, even when they are not "completed".  Would your algorithm work with only segments of the movement having been performed?

Dear Stavros, thank you for your time.

I proposed the method  during my Phd Thesis, I was concerned on the generality and soundness  of the method more than ever. But I personally did apply it in realtime applications with success (in internal trials). For all the time my general aim was to create something usable in realtime. Not just a "scientific concept”.

But realtime is a fuzzy term, for it is “function” of the request. E.g. for a package shipping from NY to Washington, an hour for the delivery it is "real-time", is superfast. For an airbag sensor,  to react in less than 30 milliseconds means realtime. So, which  times of reactions are required?

For my realtime tests my rationale was speech recognition:  the method suggested to me by my former advisor Roberto Bisiani that worked for some years on realtime speech recognition matters, when he was a researcher and then professor at Carnegie Mellon University.

In speech recognition you divide a temporal acquisition in tiny  (circa 10 millisec) temporal slices and then you extract features  for (e.g.) HMM analysis. I do not use HMM, only my own method.

We are dipendent from Nyquist sample theorem: I used slices at regular rate of  ¾ sec (it was just a test, slide should be of different dimensions in my opinion).

I think  the methodology can be adapted to the time series for quasi-immediate reactions.

We could work together on the question, and eventually collaborate, at your ease, if you want.

I am reachable at [email protected] for any question.

My Best,

Stefano