基於單應性轉換與支持向量回歸之注視預測研究

Abstract

視覺是人體接收外界資訊最重要的感官之一，扮演著獲取外界資訊的重要角色，近年來，科技技術的進步，透過視覺能達到人機互動，在教育、廣告、醫療、娛樂或遠距教學等領域也已經有不少應用。 當前，眼動追蹤設備大多需要頭戴式裝置或是基於瞳孔中心角膜屈光 ( Pupil Center Corneal Reflection, PCCR ) 技術，此種裝置在使用規範上仍有不少限制，如系統不允許使用者頭部移動、環境中不可有其他干擾光源等，為了解決上述的問題，本論文提出基於機器學習方法預測使用者注視區域，並允許使用者在特定程度下自由移動頭部，且僅需使用筆記型電腦內建的攝影機擷取影像。 在系統流程上，首先以單應性轉換矩陣 ( Homography Transformation Matrix, HT )計算使用者注視位置，當使用者頭部偏移時，會導致轉換矩陣預測注視位置偏移，為了實現頭部的自由移動，採用機器學習方法補償頭部位移量，最終預測注視位置為單應性矩陣計算值及支持向量回歸 ( Support Vector Regression, SVR ) 補償量之疊加。 透過實驗結果，證明此演算法對於瞳孔中心偵測準確率達94%，在注視區域預測準確率達90%，深入探討不同頭部與鏡頭距離以及頭部移動幅度，對於該系統注視預測的影響，表明在特定頭部移動程度下，使用者輕微晃動或距離鏡頭30至40公分，將達到最佳的注視預測效果。

Vision plays an important role in obtaining information from the outside world. In recent years, gaze tracker has been a tool used in many fields, such as education, advertising, medical treatment, entertainment or remote teaching. Most gaze tracking devices require head-mounted devices or are based on PCCR method. In some traditional methods, users’ heads are even asked to remain still. In order to solve the above problems, our paper presents a gaze estimation algorithm based on homography transformation matrix and support vector regression to provide accurate gaze estimation under free head motion. Above all, we utilize least-squares to train a transformation from pupil features to screen coordinates. However, the transformation matrix cannot obtain accurate gaze point when the user’s head moves. Secondly, we calculate gaze estimation error caused by head movement using trained Support Vector Regression to compensate for gaze point. We give discussion about the influence of different head-to-webcam distances and the performance analysis under head motion in the end. Experimental results show an 94% accuracy on pupil center detection, and 90% average accuracy on gaze estimation.