實境場景中基於外觀的遠端眼動追蹤全面架構： 系統實作與資料集整合

Abstract

眼動追蹤技術是一項非常重要的工具，廣泛應用於科學研究及商業應用。其不僅能提供關於個體注意力、動機以及生理/心理狀態等訊息，更能在人機互動介面中提供更自然的方式，貼近個體需求。想像一下，若能透過眼睛便能操控電腦，取代傳統繁複的滑鼠操作，這種直覺且快速的使用者輸入方式，不僅將能為世界帶來顯著的改變，更能為更多族群提供友善與便利。 然而，市面上現有的眼動追蹤技術，多需額外的儀器或複雜的安裝及步驟設定，才能在個人電腦上使用。為了簡化這繁瑣的過程，此領域已有相關研究應運而生，希望能利用手機、平板及筆記型電腦內建的網路相機來捕捉使用者的臉部外觀，並直接預測其視線目標，進而實現即時控制滑鼠的便利性。 有鑑於眼動追蹤技術具有巨大的未來潛力及應用發展層面，我們架設了一套利用內建相機的即時眼動追蹤系統，提供最直接、最簡單的使用方式。另外，我們收集了能幫助我們設計之演算法訓練的資料集，並提出了一個加入注意力機制的輕量化模型，以達到在移動裝置上更具應用潛力的效能。在實驗中，我們的架構在三個著名的公開資料集中實現了最先進的性能，並在實驗各個指標上展現了我們提出的方法的實用性與可應用性。我們期望我們提出的即時眼動追蹤系統，能成為未來人機互動領域的重要里程碑。

Eye tracking is an important tool in various fields, including scientific research and commercial applications. However, existing eye tracking devices in the market often need extra equipment and complicated setup procedures to be used on personal computers. In an effort to make this process simpler, recent research in this field aims to use the built-in cameras of smartphones, tablets, and laptops to capture users'' face appearance. The goal is to directly predict where they are looking, making it easy to control the computer in real-time.

Recognizing the huge future potential and developmental possibilities of eye tracking technology, we have developed a practical real-time eye tracking system using built-in cameras, providing the most direct and simple user experience. Additionally, we collected our own dataset to help train algorithms designed by ourselves. We also propose a lightweight model that includes an attention mechanism to work better on portable devices. Our proposed method demonstrates practicality and applicability across various indicators in the experiments. We hope that our proposed eye tracking system will become an important step forward in how people interact with computers in the future.