A good way to analyse head tracker measured data

Good morning,

I am sitting in front of a pile of head tracker data that we collected in a study. I would now like to analyze whether there were typical movement patterns that our participants made, since all test persons had the same task.

The only thing that is important to me is the rotation of the head to the left and right. The three-dimensional movement is not critical. I have the movements axis-separated. So I could make an analysis only about the Z-axis.

My question: is there a good tool in Matlab to analyze this data and identify patterns? Basically the movements are curves.

Thanks a lot

[Adam Danz]

Typically the positional data are converted to angular velocity using the sampling rate. Then you can plot out the angular velocity for epochs / subjects / conditions or whatever you're analyzing. If you expect a strong pattern you should see it in the plot. You may need to eliminate noisy trials and then do some curve fitting to get the underlying pattern and it's error. There may be several patterns that emerge in which case you could do some time series clustering and then fit each cluster. It may be helpful to segment the analysis into temporal epochs since the variability is likely to increase in time (assuming all participants start at a fixed position).

Perhaps instantaneous orientation is the more relevant variable in which case you could analyze those data instead of angular velocity (the two can easily be converted assuming a fixed sampling rate).

[Christian Sky]

Thank you,

I will give it a try. At the moment I have something like that.

The red dottet line is my "zero" line and all the other are the actual movement of the head. At the moment there is to much noise in all the data. I need a way to clear this up.

Second, I need to find something, that analyses the curves, if there is a spezific pattern in it.

[Adam Danz]

The pattern of increasing error with time and increasing variability with time are expected. You may want to choose a cuttoff point due to the decreasing number of samples after ~1000ms. A time series average would certainly smooth things out but you'd lose single trial data (see below). There's lots of options but those decisions should be guided by whatever you plan to do with the data and how they were generated.

Here's similar data I've been working with for years. The y-axis is time (ms) and the x-axis is angular velocity which could easily be converted to angular orientation. Each line is a single trial sampled at 200 Hz and the black dotted line is the average across all samples for each temporal bin (5ms). The abrupt ending at ~1950ms is because that's where I cut off the data which typically extended up to ~3000ms but each trial had a different duration so the number of trials that contain data decreases after ~2000ms. Just FYI, this is not head rotation - it's the angular velocity of primates steering down a curved path in VR which can be interpreted as full-body rotation.

By just eyeballing your data, my guess for it's time series average is also below (the dashed black line overlaid on your data is not the result of a curve fit - just my guess on what the result would look like). It also probably makes theoretical sense from a control theory perspective since the expected pattern is likely an initial head rotation, an over-shoot, and then a correction which itself has an overshoot (example).