藉由眼動資訊、九軸感測訊號及生理訊號估測專注度;Attention Estimation using Eye Movement, Nine-axial Sensing and Physiological Signals

[[abstract]]本研究提出以眼動資訊、九軸感測訊號及生理訊號進行專注度估測，透過三種不同難度的持續性操作測驗誘發專注力，實驗進行時，眼球動作、九軸訊號及心電圖資訊會被同步地記錄下來。 於眼動追蹤部分，本研究透過智慧眼鏡之眼球攝影機獲得眼影像，以影像處理方式計算虹膜中心座標，影像處理過程中透過大津演算法取得自適應的二值化閥值，得到二值化影像後再以高斯模糊及形態學方式消除雜訊，接著以霍夫轉換及感興趣區域的條件限制擬合出最正確虹膜中心。最後以紅外線打光之影像比較其視角誤差為3.02度；於九軸部分，使用內建於智慧眼鏡之九軸模組獲得九軸原始數據；於生理資訊部分，邀請具有生醫訊號處理背景的同學協助心電圖訊號的蒐集；透過上述三種訊號蒐集並擷取特徵，擷取出12項眼動特徵、37項九軸特徵及51項生理特徵。本研究以循序浮動正向選取方式對三種訊號特徵進行特徵挑選，再以支持向量機進行專注與放鬆的辨識，並以K-fold方法驗證其可靠度。其三者資訊混合之辨識率雖只有78.3%，但將情境分開討論，兩兩資訊的組合卻能得到不錯的效果，在CPT1及CPT2情境下，眼動資訊加上九軸資訊辨識率可達89.9%及93.4%，而在CPT3情境下，九軸資訊加上生理資訊之辨識率亦可達93.8%。本研究之最大創新點與貢獻度在於將三種不同訊號做同時進行訊號蒐集，並對不同的信號組合進行分析與比較。 This study proposed attention estimation using eye movement, nine-axial sensing and physiological signal. Through three different difficulty of the Continuous Performance Test to induce attention, in the experiment, eye movement, nine-axial signal and ECG information will be recorded synchronously. In the eye movement detection, this study obtains the eye images from the eye camera of the smart glasses, and the iris center coordinates are calculated by image processing. In the image processing process, we obtain the adaptive binarization threshold through the Otsu algorithm. When getting binarized images, we use Gaussian blur and Morphology to eliminate noise. Next, through the Hough Transform and the Region Of Interest to fit out of the most correct iris center, and finally, compared with images with infrared light, the angle deviation is 3.02 degrees. In the nine-axial, the raw data of nine-axial can be recorded through the built-in nine-axial sensor module for smart glasses. In the physiological information section, we invite a student who has biomedical signal processing background to assist the collection of ECG signals, We collect and retrieve features from the three signals described above, and produced 12 eye movement characteristics, 37 nine- axial characteristics and 51 physiological characteristics. We use the sequential floating forward selection method for all signals to do feature selection, and then we use support vector machine to identify attention and relax. Finally, we use K-fold cross validation to verify its reliability. Although the accuracy combined three signals only 78.3%, combining two signals can get good effect when the situation is discussed separately. In the CPT1 and CPT2 situation, using eye movement and nine-axial sensor information, the accuracy up to 89.9% and 93.4%. In the CPT3 situation, we use nine-axial sensor and physiological information, the accuracy is 93.8%. The biggest innovation and contribution of this research is to collect three different signals at the same time, and analyzes and compares the signal combination pattern