基於影像處理及藉由深度資訊輔助的快速眼球追蹤算法

Abstract

本論文將呈現一種快速的眼球追蹤演算法，該演算法將接收影像的深度資訊以及灰階紅外線資訊，再透過傳統的影像處理演算法，以快速(每秒可處理超過90張圖片)又不失精確的方式，從這張含有一張人臉的圖片中，定位出眼球在以相機為原點之現實座標系中的位置。除了呈現結果之外，本論文也會詳加說明這種演算法的限制與優缺點。 大體而言，本方法利用了深度資訊來快速定位出臉部的位置，進而縮小尋找眼球的範圍，以同時達到高速執行以及降低錯誤率的目的。在定位出了臉部位置之後，出於速度考量，本方法將繼續使用運算較不複雜的影像處理演算法，在臉部的圖片中定位出眼球在圖片中的位置，最後以針孔相機模型[1]，將圖片上的像素座標轉換為現實座標。

In this thesis, we will purpose a fast eye-tracking method which takes depth image and gray-scale IR image as input, and put them into a traditional-image-processing-algorithm- based system. Given an IR image which containing exactly one face and the corresponding depth image, the purposed method can locate the real-world coordinate which the camera is at origin with high speed (>90 frames per second) with acceptable error. Basically, this method takes advantage of the depth information to quickly locate face, which can shrink the range we detect the eyeballs and then decrease the error rate and accelerate the operation. After finding the face region, in need of high execution speed, we apply some less complicated computer vision algorithms to locate the eyeball’s position and finally transform the pixel coordinate on the image to the real-world coordinate though pinhole camera model[1].