Deep transfer learning-based gaze tracking for behavioral activity recognition

Computational Ethology studies focused on human beings is usually referred as Human Activity Recognition (HAR). Specifically, this paper belongs to a line of work the identification broad cognitive activities that users carry out with computers. The keystone kind systems noninvasive detection subject’s gaze fixations in selected display areas. Noninvasiveness ensured by using conventional laptop cameras without additional illumination or tracking devices. ethograms, composed sequences fixations, are basis identify user activities. To determine fixation areas highest accuracy, explores use transfer learning approach applied several well-known deep network (DLN) architectures whose input eye area extracted from face image,and output computer screen. Two different datasets created and used validation experiments. We report encouraging results may allow general system.