A crowdsourced system for robust eye tracking

Abstract

Eye tracking is widely used in modern intelligent applications, such as HCI, somatosensory game and fatigue driving. Traditional eye tracking system based on Haar-like features or external hardware, which is loss of accuracy and complicated. It is obviously that human gaze point is related to head pose. However, the label of head pose in most dataset is ambiguous. So in this paper, we propose a crowdsourced system which can collect large-scale dataset for eye tracking. For better performance, we leverage head guidance point and random dot instead of fixed dot as the concern when capture frames from camera. And different illumination, poses and persons also considered for robust performance. And we propose a two-phase CNN training strategy for combining head pose and eye angles. The proposed CNN architecture can reduce the overfitting when we train eye tracking models with head pose directly. The experimental results show that our proposed method can perform well in eye tracking.

Introduction

With the development of computer vision, eye tracking is an indispensable technique [3], [33]. Eye tracking can be used in HCI, somatosensory or fatigue driving. HCI is an interface for communication between humans and computers, so eye tracking technique can be easily used for better experiences [31], [32], [33]. Eye tracking aims at considering the direction of gaze, so it is widely used in fatigue driving, which can provide alert when drivers are fatigued because of their changing direction of gaze. Besides, in modern researches, biology and psychology experiments show that human only focus on few objects when they observe a scene or images [5], [7], [9]. The gaze shifting paths (GSPs) can reflect the sequences of saliency regions within an image [2]. As shown in Fig. 1, human tend to focus on the most saliency regions within an image. Eyetracker II is a hardware which can capture human gaze shifting paths when human observe in front of the computer. It has a very high accuracy, but the cumbersome hardware is a limit.

Traditional method for eye tracking always based on Haar-like features. Haar-like feature is widely used in face recognition. Haar-like feature aims at the difference between different regions and consider the difference as the feature. It is low-level feature, so it is loss of accuracy. For example, it is always affected by different illuminations such as dark light. A better method is to use external hardware such as infrared equipment. It can leverage fusion feature of eyes to pursue a better performance. In this way, eye tracking can be achieved in real-time and some applications based on this can be possible. But it is obviously that relying on hardware devices is complicated and expensive.

As far as we know, there is no related dataset in this domain. So in this paper, we propose our dataset collection framework based on crowdsourced system. And a two-phase training strategy is proposed for better performance. In the first phase training, we train head pose and gaze angle respectively, which does not need very precise labels. It undoubtedly reduces the difficulty of collecting data and increases the robustness. In our second phase training, we combine the models of the first phase training to continue finetune. In this phase, the label is gaze point which obtained in our data collecting system. Our two-phase training strategy can reduce overfitting compared with training head pose directly.

The main contributions of our work can be summarized as follows:

• Most related works demonstrate that training eye tracking model with head pose directly will lead to overfitting. The label of head pose is not easy to obtain. So in order to obtain more accurate result, researches always need precise equipment, which is undoubtedly complicated and expensive. We propose a two-phase training strategy, which does not require very precise labels during the first phase of training.

• In this paper, we propose a succinct crowdsourced system for collecting dataset for eye tracking. In our implement, we consider most of the environment so our dataset is robust.

• As far as we know, there is less related dataset for eye tracking. We believe that our work can promote the development of eye tracking.