A data-driven algorithm for offline pupil signal preprocessing and eyeblink detection in low-speed eye-tracking protocols

Abstract

Event detection is the conversion of raw eyetracking data into events-such as fixations, saccades, glissades, blinks, and so forth-that are relevant for researchers. In eye-tracking studies, event detection algorithms can have a serious impact on higher level analyses, although most studies do not accurately report their settings. We developed a data-driven eyeblink detection algorithm (Identification-Artifact Correction [I-AC]) for 50-Hz eye-tracking protocols. I-AC works by first correcting blink-related artifacts within pupil diameter values and then estimating blink onset and offset. Artifact correction is achieved with data-driven thresholds, and more reliable pupil data are output. Blink parameters are defined according to previous studies on blink-related visual suppression. Blink detection performance was tested with experimental data by visually checking the actual correspondence between I-AC output and participants' eye images, recorded by the eyetracker simultaneously with gaze data. Results showed a 97% correct detection percentage.

Quantifying saccades while walking: Validity of a novel velocity-based algorithm for mobile eye tracking

“…Similar to previous work which assessed blink number during eye-tracking [14], frame by frame visual inspection of the eye movement videos from the experimental trials with ten different individuals served as the ground truth for evaluating the detection performance of the algorithm ( Table 2).…”

Oculometric variations during mind wandering

“…We designed a two-step artifact removal and blink identification procedure inspired by Pedrotti et al (2011): the first step aims to define statistics of the blink-free signal which is used to detect and remove non-blink related artifact and the second step uses a fine grain blink detection procedure to compute blink duration and interpolate signals during blinks. The automated procedure we developed to achieve this is summarized in Figure 2 and detailed below.…”

Automatic Stress Classification With Pupil Diameter Analysis

“…Because eyetracking systems deal with this problem (loss of information) in a variety of ways (Gitelman, 2002), it is impossible to create a universal procedure to recover missing information. Several algorithms for eye-blink detection have been proposed, by both researchers directly interested in the eye-blink phenomenon and researchers faced with eye-blink artifacts (Pedrotti, Lei, Dzaack, & Rötting, 2011). Once blink onset and offset have been identified, missing/corrupted pupil data are usually estimated using linear (or cubic) interpolation, or even more sophisticated techniques such as moving average or support vector regression (see Nakayama, Yamamoto, & Kobayashi, 2012).…”

“…Moreover, other blink markers in the eye-tracking protocol-such as the momentary loss of gaze position during blink-were combined to foster correct blink detection percentage (for a detailed description of the algorithm, see Pedrotti et al, 2011). Blink onset was defined as the third sample (60 ms) preceding the first zero observation: At this point, the lid starts its descent until the pupil is covered (in 79% of blinks; see Pedrotti et al, 2011). Blink offset was defined as the first valid sample after a blink: At this point, the pupil is visible to the eye tracker camera.…”