You might check with Dr. Elisa Gironzetti of Texas A&M.

[email protected]

She is accomplished in eye-tracking.

Hey Rik,

since you are sampling at 60Hz, your signal already has a low-pass filter on it, so I would not go for much more smoothing (e.g. like a moving average). If you have trouble with mains noise, apply a soft low-pass filter around 45Hz (e.g. a low-order Butterworth filter).

Depending on the literature, fixed velocity values for saccade detection are around 20 to 50 deg/sec - although these values might not directly apply to low-sampling data (since peak velocity is lowered).

The work by Nyström et al has some comparison values in it. If I were you, I'd go for dispersion-based fixation detection (use a moving-RANGE operator to compute dispersion and apply a threshold to find fixation breaks). This is a quite easy method to implement and is not so dependent on the sampling rate. Salvucci and Goldberg have a paper on that. In the book by Holmqvist et al a superficial description is given on the method, which might be a good starting point (Chapter 5.5.1).

Greetings, David

Article An adaptive algorithm for fixation, saccade, and glissade de...

Article Eye Tracking: A Comprehensive Guide To Methods And Measures

Conference Paper Identifying fixations and saccades in eye-tracking protocols

To offer a slightly 'lower brow' perspective to David's, if it were me I'd just try the simplest thing first, and see if that works.

In this case, the simplest approach would be to perform a moving-average filter (e.g., http://www.mathworks.com/matlabcentral/fileexchange/41859-moving-average-function ), and looking for instances were N contiguous points fall farther than X pixels from a given location. The moving average will require you to 'hold' a few frames in a buffer (i.e., 'look ahead'), so it will add a small lag to your program (e.g., 3/60 seconds, with a window of +/- 3). But that should be fine for your application.

I wouldn't bother trying to do saccade identification online using a 60 Hz eye-tracker. It's not impossible, but it is not robust, and doesn't sound necessary for what you're trying to do.

Unless you can find a paper using the exact same setup, there would be relatively little point relying on previously reported values, since they will be dependent on a large number of uncontrolled variables. If you wanted to be really rigorous, you could always get trusted observers to fixate various spots, and construct PDFs to quantify the likelihood of fixating point XY given observations xy1, xy2,..,xyN. This would be overkill in my opinion though, and should only be done if the simple approach proves deficient.