基於二階層式支持向量機之即時注視區域分析

Abstract

眼動追蹤過去經常被使用在學術研究方面，近年來由於技術的進步眼動追蹤也被應用在醫療以及交通方面，抑或是應用於駕駛或課堂學生專注度的分析等。然而，過去相關的研究技術許多會利用侵入性的紅外線設備照射眼睛，或是利用較為昂貴的眼動儀輔助，雖然可提高注視點分析辨識率及準確度，但卻忽略了對人體可能的潛在傷害或是無法為大眾輕易取得的缺點。 本研究提出一個使用筆記型電腦內建之低解析度的網路攝影機即時偵測眼睛與注視點分析方法，實現以低成本且可輕易取得之設備達到正確偵測眼睛與注視點分析的目的。本研究主要方法分成兩大部分，首先利用Adaboost的人臉及人眼偵測獲得眼睛影像，接著加入光線濾波，利用眼睛區域平均灰階值過濾過強的光線，並且記錄使用者的眼睛特徵資訊（包含眼睛開合高度、上眼瞼斜率以及瞳孔位置）；其次記錄使用者於不同注視區塊的眼睛資訊，透過本論文提出之二階層式支持向量機（2-Layer Support Vector Machine），建構使用者相對於當下環境的注視點模型，藉由比對測試資料及模型資訊以達到注視區塊的決策。 注視區塊決策準確度在注視輔助點固定的狀況下平均可達84%，比使用單一層支持向量機之準確度高出9.4%，而在注視輔助點是隨機出現的情況下平均約為80%。Eye-gaze estimation has been used in attention analysis and human behavior research. Eye tracker has been an expensive tool used in these researches. In recent resent researches, intrusive and non-intrusive methods for eye-gaze estimation have been proposed to replace eye tracker. In this research, we proposed a non-intrusive real-time gaze estimation system using webcam as input device. A 2-layer support vector machine (SVM) is proposed to determine the human eye gaze region. We had 9 gaze regions within the monitor, which was divided into 3×3 grids. Two types of eye features were used in the 2-layer SVM: 1st layer SVM applied shape type features (eye height and eyelid slope) to determine which row the gaze located, and 2nd layer applied location type features (pupil location) to finally determine the gaze region. Experiments with 7 subjects showed an 84% average accuracy on fixed-point gaze estimation, and 80% average accuracy on random-point gaze estimation.