The effect of sampling rate and lowpass filters on saccades – A modeling approach

Abstract

The study of eye movements has become popular in many fields of science. However, using the preprocessed output of an eye tracker without scrutiny can lead to lowquality or even erroneous data. For example, the sampling rate of the eye tracker influences saccadic peak velocity, while inadequate filters fail to suppress noise or introduce artifacts. Despite previously published guiding values, most filter choices still seem motivated by a trial-and-error approach, and a thorough analysis of filter effects is missing. Therefore, we developed a simple and easy-to-use saccade model that incorporates measured amplitude-velocity main sequences and produces saccades with a similar frequency content to real saccades. We also derived a velocity divergence measure to rate deviations between velocity profiles. In total, we simulated 155 saccades ranging from 0.5°to 60°a nd subjected them to different sampling rates, noise compositions, and various filter settings. The final goal was to compile a list with the best filter settings for each of these conditions. Replicating previous findings, we observed reduced peak velocities at lower sampling rates. However, this effect was highly non-linear over amplitudes and increasingly stronger for smaller saccades. Interpolating the data to a higher sampling rate significantly reduced this effect. We hope that our model and the velocity divergence measure will be used to provide a quickly accessible ground truth without the need for recording and manually labeling saccades. The comprehensive list of filters allows one to choose the correct filter for analyzing saccade data without resorting to trial-and-error methods.

MAD saccade: statistically robust saccade threshold estimation

“…One future direction would be to test thresholding algorithms that use MAD at different sampling frequencies. We expect larger improvements when sampling at lower frequencies, because (1) this data tends to be noisier, (2) peak velocities cannot be reliably recovered (Mack et al, 2017), and (3) outliers exert more influence with less data. However, this remains to be tested.…”

“…A critical and common step in algorithmic saccade detection is thus the choice of the threshold. However, variability in saccadic profiles, the presence of other gaze events such as fixations or smooth pursuits (Larsson et al, 2013), measurement noise (Dai, Selesnick, Rizzo, Rucker, & Hudson, 2016;Holmqvist, Nyström, & Mulvey, 2012), or sampling frequency (Mack, Belfanti, & Schwarz, 2017), all make it difficult to reliably detect saccades algorithmically. Thus, saccadic detection can be improved if the threshold is estimated robustly from the data itself, and may even adapt to changing conditions (A. T. Duchowski, 2003;Engbert & Kliegl, 2003;Nyström & Holmqvist, 2010).…”

Model-Based Separation, Detection, and Classification of Eye Movements

“…To evaluate the estimation performance of our approach, we also estimated the eye position and velocity profiles by the traditional approach of lowpass filtering and numerical differentiation (denoted by 'Filt'). We set the cutoff frequency and filter order following [8], where these values are chosen such that the peakvelocity is optimally recovered.…”

“…S2. with peak-velocity optimized parameters from [8]. Note that choosing larger (or lower) cutoff frequencies of the lowpass filter results in overestimation (or underestimation) of the peak-velocity.…”

“…Traditionally, kinematic signals (eye position, velocity and acceleration) have been obtained by lowpass filtering and numerical differentiation of the measured eye position [6]. Unwary filtering, however, can substantially alter characteristics of eye movement recordings, in particular saccadic peak velocities [7], [8]. Nonetheless, filtered position and velocity profiles are commonly used for detecting saccades and extracting relevant parameters.…”