适用于便携式设备的深度神经网络眼动跟踪

Eye-movement Tracking Based on Deep Neural Network for Portable Devices

针对目前眼动跟踪方法难以适用于智能手机、平板电脑等便携式设备的问题,提出一种基于大规模数据集的眼动跟踪方法。首先,通过众包法构建大规模数据集;然后,使用该数据集训练一个深度神经网络,用于端对端的预测。最后,训练一个更小更快的网络进行优化,使所提方法在移动设备上的运行具有一定的实时性。实验结果表明,与其他类似方法相比,所提方法具有更好的跟踪鲁棒性以及数据泛化能力。在移动设备中的运行速度可达10~15帧/s。在未校正的情况下,该方法在手机和平板电脑中的预测误差分别是1.71 cm和2.53 cm。校正后,误差分别降至1.34 cm和2.12 cm。

Aiming at the problem that the current eye-movement tracking methods can not be applied to intelligent mobile phones,tablet computers and other portable devices,an eye-movement tracking method based on large-scale data sets is proposed.Firstly,a large-scale data set is constructed by crowd-sourcing method.Then a deep neural network is trained with the data set for end-to-end prediction.Finally,a smaller and faster network is trained to optimize,which makes the proposed method run in real-time on mobile devices.Experimental results show that the proposed method has better tracking robustness and data generalization ability than other similar methods.The speed of running in mobile devices can reach 10~15 frames per second.The prediction errors of this method are 1.71 cm and 2.53 cm respectively in mobile phone and tablet computer without correction.After calibration,the errors are reduced to 1.34 cm and 2.12 cm respectively.