A Non-Contact Human-Computer Interaction Method Based on Gaze

Gaze interaction has a broad application prospect in the field of human-computer interaction. Traditional gaze interactive sensing devices are expensive and inconvenient to use. The gaze point estimation model based on deep learning is complex and poor in generalization, and difficult to deploy on intelligent terminals. This paper proposed a gaze point estimation network based on deep learning and its personalized calibration method which can adapt to different devices and users. The gaze point estimation network used eye images, eye corners and the camera relative to screen position as inputs to extract feature embeddings. Then the feature embeddings were fed to the trained personalized SVR model to obtain the final gaze point estimation. Tests were conducted on both the self-built gaze dataset and the open source dataset to verify the effectiveness of the proposed model in terms of accuracy and latency. On this basis, a simple human-computer interaction application was implemented by analyzing the fixation point and identifying the user's control behavior.