Accurate detection of pupil position in successive frames of eye videos is finding applications in many areas including assistive systems and E-learning. Processing the big datasets of eye videos in such systems requires robust and fast eye-tracking algorithms that can predict the position of eye pupil in consecutive video frames. As a major technique, particle filters provide adequate speed but have a low detection rate. To solve this problem, the present paper suggests the use of genetic algorithms in the sampling step of the particle filter technique. As a result, in each frame, the variety of particles required for predicting the pupil position in the next video frame is maintained and their uniformity is reduced. Finally, the speed and detection rate of the proposed method, as well as the basic particle filter method in predicting the pupil position in video frames are calculated and compared for various populations. The experimental results indicate that, in comparison with the basic particle filter algorithm, the proposed algorithm detects the pupil more accurately and in a shorter time. Also, by achieving an average detection rate of 79.89% in estimation of the pupil center with an error of five pixels on a variety of eye videos with different situations of occlusion and illumination, not only the robustness of the proposed method is assessed but also its superiority to the state-of-the-art methods is evinced.

中文翻译：

---

基于鲁棒且精确的基于粒子滤波的瞳孔大数据集瞳孔检测方法

在眼动视频的连续帧中准确检测瞳孔位置正在许多领域中得到应用，包括辅助系统和电子学习。在这样的系统中处理眼睛视频的大数据集需要强大且快速的眼睛跟踪算法，该算法可以预测连续视频帧中瞳孔的位置。作为一项主要技术，粒子过滤器可提供足够的速度，但检测率较低。为了解决这个问题，本文建议在粒子滤波技术的采样步骤中使用遗传算法。结果，在每一帧中，维持了预测下一视频帧中的瞳孔位置所需的粒子的多样性，并且降低了它们的均匀性。最后，提出的方法的速度和检测率，并针对各种人群计算并比较了用于预测视频帧中瞳孔位置的基本粒子滤波方法。实验结果表明，与基本的粒子滤波算法相比，该算法能够在更短的时间内更准确地检测出瞳孔。此外，通过在瞳孔中心的估计中实现平均检测率达79.89％，在具有不同遮挡和照明情况的各种眼睛视频上误差为五个像素，不仅可以评估该方法的鲁棒性，还可以评估该方法的鲁棒性证明了相对于最新方法的优越性。提出的算法可以在更短的时间内更准确地检测出瞳孔。此外，通过在瞳孔中心的估计中实现平均检测率达79.89％，在具有不同遮挡和照明情况的各种眼睛视频上误差为五个像素，不仅可以评估该方法的鲁棒性，还可以评估该方法的鲁棒性证明了相对于最新方法的优越性。提出的算法可以在更短的时间内更准确地检测出瞳孔。此外，通过在瞳孔中心的估计中实现平均检测率达79.89％，在具有不同遮挡和照明情况的各种眼睛视频上误差为五个像素，不仅可以评估该方法的鲁棒性，还可以评估该方法的鲁棒性证明了相对于最新方法的优越性。