Eye-Tracking Algorithm for Low Webcam Image Resolution Without Calibration

Abstract

Applications of eye-tracking devices aim to understand human activities and behaviors, improve human interactions with robots, and develop assistive technology in helping people with some communication disabilities. This paper proposes an algorithm to detect the pupil center and user’s gaze direction in real-time, using a low-resolution webcam and a conventional computer with no need for calibration. Given the constraints, the gaze space was reduced to five states: left, right, center, up, and eyes closed. A pre-existing landmarks detector was used to identify the user’s eyes. We employ image processing techniques to find the center of the pupil and we use the coordinates of the points found associated with mathematical calculations to classify the gaze direction. By using this method, the algorithm achieved 81.9% overall accuracy results even under variable and non-uniform environmental conditions. We also performed quantitative experiments with noise, blur, illumination, and rotation variation. Smart Eye Communicator, the proposed algorithm, can be used as eye-tracking mechanism to help people with communication difficulties to express their desires.

Improved Smart Eye Tracker Communicator Using Low Resolution Webcam

“…An evolution of the Smart Eye Communicator (SEC) [9] software, called SEC-II, was proposed. SEC-II is an algorithm that proposes the detection of the gaze direction through a facial aligner, mathematical calculations, and facial landmarks predictors, which locate the contours of the eyes and eyelids, as shown in Figure 1.…”

“…As in the previous version of the [9] system, the classification of the gaze direction is based on the vertical and horizontal coordinates of the center of the pupil from the landmarks, after trying to identify the "Right", "Left" and " Up", and none are identified, the system moves to the "dominant eye" identification attempt step. A method has been developed to solve the "divergent pupil centers" problem, that is, when the pupil center is not captured correctly, in one eye, and the two centers are misaligned, as shown in Figure 5.…”

“…To evaluate the performance of the SEC-II, the same dataset of the initial version [9] was used. Bearing in mind that the base is composed of 2,908 frames, encompassing females and males, with different lighting conditions in different locations, with or without glasses, and in different positions in relation to a laptop webcam.…”

“…This work proposes the improvement of a calibration-free digital image processing strategy to identify the gaze direction [9], based on five states: right, left, up, center, and eyes closed. The system has been improved to apply facial alignment strategies, before the steps of capturing the center of the pupil, as well as a new technique to search and use the coordinates of the "dominant eye", which influences the classification of the gaze direction.…”

“…The system has been improved to apply facial alignment strategies, before the steps of capturing the center of the pupil, as well as a new technique to search and use the coordinates of the "dominant eye", which influences the classification of the gaze direction. A database of 2,908 images was used, the same one created for the SEC [9], as it covers images with people in different positions, skin tones, and lighting. Accuracy of 86 % and 0.89 of the F1 score metric was achieved.…”