Saccade detection using a particle filter

Detection of normal and slow saccades using implicit piecewise polynomial approximation

The quantitative analysis of saccades in eye movement data unveils information associated with intention, cognition, and health status. Abnormally slow saccades are indicative of neurological disorders and often imply a specific pathological disturbance. However, conventional saccade detection algorithms are not designed to detect slow saccades, and are correspondingly unreliable when saccades are unusually slow. In this article, we propose an algorithm that is effective for the detection of both normal and slow saccades. The proposed algorithm is partly based on modeling saccadic waveforms as piecewise-quadratic signals. The algorithm first decreases noise in acquired eye-tracking data using optimization to minimize a prescribed objective function, then uses velocity thresholding to detect saccades. Using both simulated saccades and real saccades generated by healthy subjects and patients, we evaluate the performance of the proposed algorithm and 10 other detection algorithms. We show the proposed algorithm is more accurate in detecting both normal and slow saccades than other algorithms.

MAD saccade: statistically robust saccade threshold estimation via the median absolute deviation

Saccade detection is a critical step in the analysis of gaze data. A common method for saccade detection is to use a simple threshold for velocity or acceleration values, which can be estimated from the data using the mean and standard deviation. However, this method has the downside of being influenced by the very signal it is trying to detect, the outlying velocities or accelerations that occur during saccades. We propose instead to use the median absolute deviation (MAD), a robust estimator of dispersion that is not influenced by outliers. We modify an algorithm proposed by Nyström and colleagues, and quantify saccade detection performance in both simulated and human data. Our modified algorithm shows a significant and marked improvement in saccade detection - showing both more true positives and less false negatives - especially under higher noise levels. We conclude that robust estimators can be widely adopted in other common, automatic gaze classification algorithms due to their ease of implementation.

Human-level saccade detection performance using deep neural networks

Saccades are ballistic eye movements that rapidly shift gaze from one location of visual space to another. Detecting saccades in eye movement recordings is important not only for studying the neural mechanisms underlying sensory, motor, and cognitive processes, but also as a clinical and diagnostic tool. However, automatically detecting saccades can be difficult, particularly when such saccades are generated in coordination with other tracking eye movements, like smooth pursuits, or when the saccade amplitude is close to eye tracker noise levels, like with microsaccades. In such cases, labeling by human experts is required, but this is a tedious task prone to variability and error. We developed a convolutional neural network to automatically detect saccades at human-level accuracy and with minimal training examples. Our algorithm surpasses state of the art according to common performance metrics and could facilitate studies of neurophysiological processes underlying saccade generation and visual processing. NEW & NOTEWORTHY Detecting saccades in eye movement recordings can be a difficult task, but it is a necessary first step in many applications. We present a convolutional neural network that can automatically identify saccades with human-level accuracy and with minimal training examples. We show that our algorithm performs better than other available algorithms, by comparing performance on a wide range of data sets. We offer an open-source implementation of the algorithm as well as a web service.

Unsupervised detection of microsaccades in a high-noise regime

Micromovements of the eye during visual fixations provide clues about how our visual system acquires information. The analysis of fixational eye movements can thus serve as a noninvasive means to detect age-related or pathological changes in visual processing, which can in turn reflect associated cognitive or neurological disorders. However, the utility of such diagnostic approaches relies on the quality and usability of detection methods applied for the eye movement analysis. Here, we propose a novel method for (micro)saccade detection that is resistant to high-frequency recording noise, a frequent problem in video-based eye tracking in either aged subjects or subjects suffering from a vision-related pathology. The method is fast, it does not require manual noise removal, and it can work with position, velocity, or acceleration features, or a combination thereof. The detection accuracy of the proposed method is assessed on a new dataset of manually labeled recordings acquired from 14 subjects of advanced age (69-81 years old), performing an ocular fixation task. It is demonstrated that the detection accuracy of the new method compares favorably to that of two frequently used reference methods and that it is comparable to the best of the two algorithms when tested on an existing low-noise eye-tracking dataset.

Bayesian microsaccade detection

Microsaccades are high-velocity fixational eye movements, with special roles in perception and cognition. The default microsaccade detection method is to determine when the smoothed eye velocity exceeds a threshold. We have developed a new method, Bayesian microsaccade detection (BMD), which performs inference based on a simple statistical model of eye positions. In this model, a hidden state variable changes between drift and microsaccade states at random times. The eye position is a biased random walk with different velocity distributions for each state. BMD generates samples from the posterior probability distribution over the eye state time series given the eye position time series. Applied to simulated data, BMD recovers the “true” microsaccades with fewer errors than alternative algorithms, especially at high noise. Applied to EyeLink eye tracker data, BMD detects almost all the microsaccades detected by the default method, but also apparent microsaccades embedded in high noise—although these can also be interpreted as false positives. Next we apply the algorithms to data collected with a Dual Purkinje Image eye tracker, whose higher precision justifies defining the inferred microsaccades as ground truth. When we add artificial measurement noise, the inferences of all algorithms degrade; however, at noise levels comparable to EyeLink data, BMD recovers the “true” microsaccades with 54% fewer errors than the default algorithm. Though unsuitable for online detection, BMD has other advantages: It returns probabilities rather than binary judgments, and it can be straightforwardly adapted as the generative model is refined. We make our algorithm available as a software package.

Topology for gaze analyses - Raw data segmentation

Recent years have witnessed a remarkable growth in the way mathematics, informatics, and computer science can process data. In disciplines such as machine learning, pattern recognition, computer vision, computational neurology, molecular biology, information retrieval, etc., many new methods have been developed to cope with the ever increasing amount and complexity of the data. These new methods offer interesting possibilities for processing, classifying and interpreting eye-tracking data. The present paper exemplifies the application of topological arguments to improve the evaluation of eye-tracking data. The task of classifying raw eye-tracking data into saccades and fixations, with a single, simple as well as intuitive argument, described as coherence of spacetime, is discussed, and the hierarchical ordering of the fixations into dwells is shown. The method, namely identification by topological characteristics (ITop), is parameter-free and needs no pre-processing and post-processing of the raw data. The general and robust topological argument is easy to expand into complex settings of higher visual tasks, making it possible to identify visual strategies.

Switching between two targets with non-constant velocity profiles reveals shared internal model of target motion

Several experiments have shown that smooth pursuit and saccades interact while tracking an object moving across the visual scene. It was proposed two decades ago that the amplitude of saccades triggered during smooth pursuit ('catch-up saccades') were corrected by a delayed sensory signal to account for the ongoing target displacement during catch-up saccades. However, recent studies used targets with non-constant velocity profiles and suggested that the correction of catch-up saccade amplitude must be done through an internal model of target motion. It is widely accepted that an internal model of target motion is also used by the central nervous system (CNS) to cancel inherent delays between visual input and smooth pursuit motor output, ensuring accurate tracking of moving targets. Our study proposes a new paradigm in which the target switches unexpectedly from one target with a non-constant periodic velocity profile to another with a non-constant aperiodic velocity profile. Our results confirm the hypothesis that the CNS uses an internal model of target motion to correct catch-up saccade amplitude. In addition, we reconcile the sensory delayed and the internal model of target motion hypotheses and show that a common internal model of target motion is shared within the CNS to control smooth pursuit and to correct catch-up saccade amplitude.