A data processing pipeline is described for gaze analytics featuring velocity-based signal filtering leading to inferential statistics of fixation transitions. The approach is demonstrated on data collected from a virtual environment study conducted in Unity 3D testing the visibility of a wayfinding aid. Analysis of human spatial cognition when navigating within physical or virtual environments can be bolstered by collecting eye movements with the help of an eye tracker. In physical reality, a head-mounted tracker can be worn during navigation with recorded gaze data often mapped to the screen coordinates of a forward-facing camera. In virtual reality, gaze data can be recorded with a so-called remote, or table-mounted eye tracker placed in front of the display screen on which the virtual environment is presented. In both cases gaze data will eventually need to be processed in order to infer insights about human visual attention to elements in the environment. Application of inferential statistics to collected gaze data often relies on characterization of the raw data into fixations, usually derived from some form of filtering, e.g., dispersion-based or velocity-based. Unfortunately, most commercial software packages provide only a limited choice of fixation detection algorithms (i.e., filters), often hiding implementation details or filter parameters from the user. Some systems (e.g., Ogama) still rely on the dispersion-based “fixation pickers” [7] which have been shown to be less than reliable, particularly when evaluating data captured on di↵erent platforms, at di↵erent sampling rates [10]. Velocity-based filters, or “saccade pickers”, while perhaps more dicult to tune, o↵er a more reliable alternative. Beyond the lack of sucient control over filtering parameters, commercial packages often do not include flexible means for statistical analysis. This is hardly surprising, however, since eye tracking vendors can hardly be expected to anticipate all possible experimental designs for which their devices are used. In this paper we describe a gaze analytics pipeline through which raw gaze data is processed. The pipeline consists of the following steps: 1. denoising and filtering raw gaze data gi = (xi, yi, ti), and classifying raw gaze into fixations fi = (xi, yi, ti, di), where (xi, yi) coordinates indicate the position of the gaze
(a) No wayfinding aid (b) Wayfinding aid at bottom-right point or centroid of the fixation, with ti indicating the timestamp of the gaze point or fixation and di the fixation’s duration, 2. collating fixation-related information for its subsequent statistical comparison, 3. interpreting and visualizing statistical tests conducted on processed data. Visualization of the data at each stage of the pipeline is particularly helpful in finetuning parameters, such as threshold levels for velocity-based filtering.
We demonstrate the utility of the analytics pipeline on data collected from a
wayfinding study in a virtual environment (see Fig. 1) [
The eye tracker outputs a stream of gaze points (xi, yi, ti). Typically, this data is noisy
and requires smoothing (see Fig. 2). Treating xi or yi independently, smoothing or
differentiating (to order s) is achieved by convolving 2p+1 inputs with filter hti,s and 2q+1
(previous) outputs x˙i or y˙i with filter gti,s at midpoint i [
Following Andersson et al. [
Fine-tuning of the velocity threshold in degrees per second depends on viewing distance and screen resolution (e.g., in dots per inch). In the exemplar fixations of Fig. 3, an 1824⇥1026 display (80 00 diagonal) was viewed at 7000. 3
Quantitative analysis of filtered fixation data generally depends on application of
inferential statistics, e.g., comparison of means via analysis of variance (ANOVA). Typical
eye movement metrics include number of saccades, saccade length and duration,
saccadic amplitude, convex hull area, spatial density, number of fixations, fixation
durations, and a fixation/saccade ratio [
Assigning a character label to each of the 4 ⇥ 3 grid cells leads to a {a, b, c, d} ⌦ {a, b, c} labeling scheme (with cell aa at bottom-left in Fig. 3). Accumulating single fixation transitions between cells and normalizing to the source leads to a first-order Markov model of gaze transitions represented by transition matrices visualized in Fig. 4. The matrix representing the viewing condition with the wayfinding aid (Fig. 4(b)) shows a higher probability of transitions to the bottom-right cell where the aid was present (compare columns da in the matrices). A critical question is whether these transitions, on average, di↵ered significantly under the given experimental conditions.
Considering the set of AOIs as S = {1, . . . , s}, transition matrices a↵ord computation of Shannon’s entropy Ht = Pi2S ⇡ i P j pi j log pi j where pi j denotes the probability of transitioning from the ith to the jth A2SOI, which in turn allows statistical comparison of matrices. In this particular instance, a Welch two sample t-test shows lack of significance of fixation transition entropies between the two wayfinding aid conditions (t = 0.85, p = 0.41, n.s.). Lack of significance is likely due to the relatively low number 0.30 0.09 0.10 0.08 0.03 0.0 0.01 0.01 0.0 0.0 0.01 0.0
Heatmap
Heatmap l)s684 iit(rxepeondao -cy342 (a)
Data from session without wayfinding aid (b)
Data from session with wayfinding aid 3. Representative (individual) scanpaths with no wayfinding aid (1543 fixations) from filtering corresponds of smoothed data. The to the raw and smooth scanpath capdata in Fig. 2. of participants. Presence of the wayfinding aid resulted in a higher mean transition entropy (M = 0.65, SD = 0.12) than when it was absent (M = 0.62, SD = 0.04), suggesting that users tended to make (slightly) more transitions with the wayfinding aid present than without it, e.g., without the aid, viewers tended to transition to the screen center (grid cells bb and cb, c.f. Figs. 4(a)–4(b) and Fig. 4(c)). Power analysis suggests that the e↵ect of the wayfinding aid on gaze transitions may reach significance ( p < 0.05) with about 400 participants (per between-subjects group). 4
A data pipeline was described for processing raw gaze data through filtering and velocitybased fixation classification followed by collation of fixations into transitions for entropybased statistical comparison.
With gaze data recorded in individual XML files, the entire processing pipeline, complete with data visualizations, is readily implemented in Python and R, the free software environment for statistical computing.