A brief look through the literature didn't reveal any algorithms that have been directly applied to the online prediction of saccade landing sites while a saccade is in progress.

While there would surely be considerable noise in any such estimate, the direction vector, velocity and acceleration profiles should contain sufficient information to make a reasonable guess. Is anyone aware of any papers which have attempted such an approach?

Also, how good an estimate of landing site could be derived, and how early prior to landing could this estimate be made. What methods would be suggested for making such an estimate?

  • $\begingroup$ Your question is pretty vague. At what point during the saccade would you like to calculate the landing position? The main problem is distance because at some point all saccades over a certain distance reach maximum velocity and while the saccade is in that state the landing point is inestimable. $\endgroup$
    – John
    May 21, 2013 at 14:03
  • $\begingroup$ @JohnChristie Thanks for your comments. This question is primarily aimed to inquire whether previous work on this has been done. Also, I ask the question "how early prior to landing could this estimate be made." From this, the implication is that I would be interested in knowing landing site as early as possible (if possible at all). If it is known that peak velocity has been reached, and the current distance travelled is known would this not constrain the likely endpoint at least to some reasonable degree? $\endgroup$
    – skleene
    May 21, 2013 at 14:42
  • $\begingroup$ I don't know anything about this topic, but recently during a presentation at work I heard mentioning that they can predict landing sites to some degree while the saccade is in progress. They were using this technique during an experiment. Definitely an interesting question! $\endgroup$
    – Steven Jeuris
    May 21, 2013 at 15:13
  • $\begingroup$ There is a lot of work on predicting fixation patterns during reading, scene perception/exploration, and naturalistic tasks (making a sandwich, navigating around obstacles, etc.). As far as I know, these predictions are typically made based on the structure of the text/scene/task, not during each saccade. Is there a reason you want to make the prediction specifically during a saccade? $\endgroup$
    – Dan M.
    May 21, 2013 at 16:51
  • 1
    $\begingroup$ @skleene Interesting. I've seen eye-tracking demos with gaze contingent manipulations like a gray box at gaze position, which creates the entertainingly frustrating experience of being able to see the scene in one's peripheral vision but blocking out whatever the participant tries to fixate. But that may not be what you're looking for. $\endgroup$
    – Dan M.
    May 22, 2013 at 11:57

2 Answers 2


Interesting question! Trying to interpret your question, I think what you are after is a model that can predict the saccade endpoint based on eye tracking data?

The following answer is more of a basic neuroscientific approach:

The neuronal responses (spike rates) in the superior colliculus (SC) can be accurately used to predict the total saccade as a vector. Basically, when you model the response properties of all the SC neurons involved in a saccade, one can calculate the vector sum of all these different neurons. With this vector sum the saccade can be totally described in terms of its direction and amplitude. With the vector sum, hence, the end point can be predicted.

If this subject is of interest to you, the early work of Van Gisbergen and John van Opstal (based in Nijmegen, the Netherlands) are a good source.

Reference: Van Gisbergen et al., 1987


This is an old question, but the following paper might be useful. The authors describe an algorithm of real-time prediction of saccadic landing point.

The algorithm they use is based on a compressed exponential function. Basically, you collect the position samples from an eye-tracker, check if the speed is more than some threshold (e.g., 20 deg/s), then mark this position as a saccade start. Then you continue to collect the samples and as soon as you get enough (at least three) you try to predict the position via non-linear curve fitting. If the fit is good, you use it. If not, you just use the latest available position.

Han, P., Saunders, D. R., Woods, R. L., & Luo, G. (2013). Trajectory prediction of saccadic eye movements using a compressed exponential model. Journal of vision, 13(8), 27.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.