EAAI Journal 2026 Journal Article
Event-based low-power spiking gaze estimation
- Zhipeng Sui
- Weihua He
- Fei Liang
- Yongxiang Feng
- Xiaobao Wei
- Qiushuang Lian
- Ziyang Zhang
- Guoqi Li
Event camera has emerged as a powerful alternative to frame-based camera in gaze estimation, which generally has stringent requirements on power consumption in potential Augmented Reality and Virtual Reality application scenarios. However, existing event-based eye tracking relies on either hybrid modality or complex illuminating equipment, leading to high power consumption. Here, we propose a fully event-based algorithm pipeline for gaze estimation to reduce power consumption by minimizing sensor modality and algorithm computational complexity. The pipeline features with five modules, including wake-up, hibernation, eye segmentation, eye-movement tracking, and gaze mapping. In designing these modules, we take advantage of the sparse and dynamic nature of event data to achieve both low computation and error. In particular, the wake-up module determines the eye state through the input event data, and directs the pipeline to one of the three modules of hibernation, eye segmentation, or eye-movement tracking, considering both computational complexity and accuracy. A lightweight spiking neural network instead of deep neural network is adopted for eye segmentation to reduce the computational complexity by an order of magnitude. Furthermore, morphological operations involving sparse event data are used for eye segmentation, requiring extremely low computation enabled by minimal update. We conduct experiments on available event-based gaze dataset proposed by Angelopoulos, and compared to their implementation, our approach shows better accuracy (approximately 50% reduction in average angle error), and lower power consumption (about 68% decrease, at 100-milliwatt level). We believe that our method would facilitate eye tracking applications in power-sensitive scenarios.