Event Detection in Eye-Tracking Data for Use in Applications with Dynamic Stimuli

Abstract: This doctoral thesis has signal processing of eye-tracking data as its main theme. An eye-tracker is a tool used for estimation of the point where one is looking. Automatic algorithms for classification of different types of eye movements, so called events, form the basis for relating the eye-tracking data to cognitive processes during, e.g., reading a text or watching a movie. The problems with the algorithms available today are that there are few algorithms that can handle detection of events during dynamic stimuli and that there is no standardized procedure for how to evaluate the algorithms. This thesis comprises an introduction and four papers describing methods for detection of the most common types of eye movements in eye-tracking data and strategies for evaluation of such methods. The most common types of eye movements are fixations, saccades, and smooth pursuit movements. In addition to these eye movements, the event post-saccadic oscillations, (PSO), is considered. The eye-tracking data in this thesis are recorded using both high- and low-speed eye-trackers. The first paper presents a method for detection of saccades and PSO. The saccades are detected using the acceleration signal and three specialized criteria based on directional information. In order to detect PSO, the interval after each saccade is modeled and the parameters of the model are used to determine whether PSO are present or not. The algorithm was evaluated by comparing the detection results to manual annotations and to the detection results of the most recent PSO detection algorithm. The results show that the algorithm is in good agreement with annotations, and has better performance than the compared algorithm. In the second paper, a method for separation of fixations and smooth pursuit movements is proposed. In the intervals between the detected saccades/PSO, the algorithm uses different spatial scales of the position signal in order to separate between the two types of eye movements. The algorithm is evaluated by computing five different performance measures, showing both general and detailed aspects of the discrimination performance. The performance of the algorithm is compared to the performance of a velocity and dispersion based algorithm, (I-VDT), to the performance of an algorithm based on principle component analysis, (I-PCA), and to manual annotations by two experts. The results show that the proposed algorithm performs considerably better than the compared algorithms. In the third paper, a method based on eye-tracking signals from both eyes is proposed for improved separation of fixations and smooth pursuit movements. The method utilizes directional clustering of the eye-tracking signals in combination with binary filters taking both temporal and spatial aspects of the eye-tracking signal into account. The performance of the method is evaluated using a novel evaluation strategy based on automatically detected moving objects in the video stimuli. The results show that the use of binocular information for separation of fixations and smooth pursuit movements is advantageous in static stimuli, without impairing the algorithm's ability to detect smooth pursuit movements in video and moving dot stimuli. The three first papers in this thesis are based on eye-tracking signals recorded using a stationary eye-tracker, while the fourth paper uses eye-tracking signals recorded using a mobile eye-tracker. In mobile eye-tracking, the user is allowed to move the head and the body, which affects the recorded data. In the fourth paper, a method for compensation of head movements using an inertial measurement unit, (IMU), combined with an event detector for lower sampling rate data is proposed. The event detection is performed by combining information from the eye-tracking signals with information about objects extracted from the scene video of the mobile eye-tracker. The results show that by introducing head movement compensation and information about detected objects in the scene video in the event detector, improved classification can be achieved. In summary, this thesis proposes an entire methodological framework for robust event detection which performs better than previous methods when analyzing eye-tracking signals recorded during dynamic stimuli, and also provides a methodology for performance evaluation of event detection algorithms.