An adaptive learning framework for event-based remote eye tracking

초록

Event cameras are next-generation asynchronous image sensors that detect only changes in light intensity. Because event cameras can capture fast-moving objects without motion blur, they have gained attention as a suitable technology for tracking the human eye, the fastest-moving part of the body. While there has been significant progress in near-range eye tracking, event-based remote eye tracking is still in its early stages due to the challenge of limited spatial information, resulting in fewer studies in this area. In this paper, we propose a novel framework for remote eye tracking using event cameras, incorporating the application of artificial intelligence (AI). Our framework addresses the challenge of limited remote eye tracking datasets by transforming frame video into event streams and generating eye annotations. We also select optimal eye keypoints suitable for event-based tracking and predict their displacement using an event-based feature tracking network. The method detects initial keypoints from a single frame, generates a reference patch, and combines it with event patches for event feature tracking. To validate our model, we conducted comprehensive evaluations on a self-collected dataset, which includes various face angles and lighting conditions, including low light environments. Our proposed method achieves a feature age of 0.550 and an expected feature age of 0.549, demonstrating promising results for event-based remote eye tracking.

키워드