A new and general approach to signal denoising and eye movement classification based on segmented linear regression

Abstract

We introduce a conceptually novel method for eye-movement signal analysis. The method is general in that it does not place severe restrictions on sampling frequency, measurement noise or subject behavior. Event identification is based on segmentation that simultaneously denoises the signal and determines event boundaries. The full gaze position time-series is segmented into an approximately optimal piecewise linear function in O(n) time. Gaze feature parameters for classification into fixations, saccades, smooth pursuits and post-saccadic oscillations are derived from human labeling in a data-driven manner. The range of oculomotor events identified and the powerful denoising performance make the method useable for both low-noise controlled laboratory settings and high-noise complex field experiments. This is desirable for harmonizing the gaze behavior (in the wild) and oculomotor event identification (in the laboratory) approaches to eye movement behavior. Denoising and classification performance are assessed using multiple datasets. Full open source implementation is included.

Human-level saccade detection performance using deep neural networks

“…Performance was then tested on the trials with labels from two coders, which makes our result comparable with previously reported results 32 . Note that we were not able to reproduce the interrater measures reported by Andersson et al 29 in line with the results of Pekkanen and Lappi 32 .…”

“…U'n'Eye: new state-of-the-art eye movement classifier. In order to compare our algorithm to state-of-the-art methods for eye movement classification, we evaluated its performance on a benchmark dataset 31 , which has previously been used for the comparison of 12 eye movement classifiers 29,32 . The dataset comprises 500 Hz eye tracking recordings from humans watching videos, images, or moving dots, and it contains human labels for fixations, smooth pursuits, saccades, PSO (Fig.…”

“…Of the few algorithm that are capable of detecting saccades as well as PSO 31,32,42 , U'n'Eye achieves highest performance. Note that Zemblys et al 42 reported higher absolute Cohen's kappa values for saccade and PSO detection, but obtained these on cleaner data than the benchmark dataset from Andersson et al 29 .…”

“…However, this algorithm, like others, may still mislabel some microsaccades due to high eye tracker noise (as is typical with video-based eye trackers) as well as small catch-up saccades occurring during smooth pursuit. Other existing algorithms 31,32 have the added advantage of providing additional labels for fixations and post-saccadic oscillations (PSO).…”

“…Despite their success, several shortcomings still render the use of existing algorithms either less reliable than desired or, at the very least, cumbersome. While the performance of many published algorithms is promising 29,32 , it does not reach the level of trained human experts. Also, none of the existing algorithms shows convincing performance for all eye movement-related events that may need to be analyzed (e.g.…”