Detecção acurada da pupila mediante oclusões por cílios com suporte temporal estendido

Abstract

Context: An eye-tracker is a device that estimates the eye position, and using a calibration process, is also able to estimate the point of gaze. Several techniques have been proposed for tracking de eyes, including special contact lenses and electrodes coupled around the eyes. Among the techniques, the use of video stands out for its ease of use and comfort. In a video-based eye-tracker, biological features of the eye are typically exploited, such as the pupil border and reflexes on the surface of the cornea. Commercial devices are still expensive (worth over R$ 3000.00), which has motivated several low-cost initiatives. However, such proposals have limited speed and accuracy, usually result of the use of off-the-shelf cameras and the high processing requirements. Typically, the processing is within a region of interest (ROI) defined around the expected pupil position. To improve such region, recent works exploit the pupil movement in what is called extended temporal support (ETS). Eyelashes and eyelids can also affect the accuracy as they may occlude part of the pupil. As the pupil center is generally used in the gaze mapping, a robust occlusion handling method is highly desirable. Objective: This work aims to study a pupil detection approach with the presence of eyelid occlusions exploiting an ETS to improve the pupil detection accuracy. Method: Study and implementation of a new approach to pupil detection with ETS. The proposed ETS exploits an exponential eye movement model. Comparison of the accuracy of the proposed method with annotated pupils from Świrski et al. (2012). The metric used in the comparison of accuracy is the Hausdorff distance between ellipses. Results: The accuracy of the pupil detection was improved. There was a reduction in processing time with the reduction of the processed region. The performance of the prediction of the eye movements using an exponential model showed feasible only at high frame rates. Conclusions: The accuracy was improved in the presence of eyelash occlusions and the computational effort per frame was reduced. However, the reduction in the ROI must follow an accurate prediction of the pupil position. During saccades, estimating eye movements using an exponential model showed feasible only at high frame rates. There was a reduction in the processing time caused by the reduction in the processed area, also providing a reduction in the negative impact of eyelash occlusions.