人臉辨識中，如何正確取得人臉位置是一個重要的研究，因此我 們提出結合追蹤與偵測的方法來找到需要辨識的區域，進一步，辨識 的部分又可以分為人的身分辨識以及人的行為辨識，後者我們又可以 分析眼睛注視的方向，可見光的瞳孔偵測上應用上相當廣泛，像是行 車是否有專心注視前方或是分析行人注視廣告看板中的位置等等，先 行的方法已經被提出了許多，有些是分析梯度、有些是分析曲率等 等，但是這些狀況很容易受到不是眼睛區域的影響，甚至有可能找到 頭髮的區域，因此，我們提出了一個的方法來改進結果，這個方法包 含了膚色偵測，主要是避免找到瞳孔位置是落在膚色的區域中，而且 這個方法結合了哈爾特徵級聯分類器以及臉部對齊的資訊，透過兩者 資訊來找到接近眼睛位置，再根據人的雙眼聯動的條件，對於位移以 及角度上進一步的修正，這部分實驗在BioID 臉部影像資料庫以及自 己建立AIPR LAB 人臉影片資料庫。更進一步，我們比較了我們以及 其他的方法，在結果上有著顯著的正確率以及穩定性。;Face-locating in videos is an important issue in face recognition. Therefore, we propose a solution which combines tracking and detecting methods in order to find and recognize the face regions. In face recognition, there are many researches about face perception such as emotion, gaze estimation, identity analysis,etc. Our research we focus on the gaze estimation in the visible light due to the visible cameras are universal devices everywhere. This technology can be applied on many different situations. For example, to alert the driver while he/she is not driving carefully; to show Ads where the passengers are looking at on the lcd ad board. Those are widely used applications nowadays. Some researches have been proposed for that such as calculate by gradients or by curvature, etc. However, sometimes it is possible to detect non-eye regions by those methods. The non-eye regions might be hair or eyebrow. To avoid those problems we propose a algorithm to correct the location of eyes. The algorithm includes skin detections, haar cascade classifier, face alignment and a correction based on conjunctive eye movements. We test our result on BioID image database and our video samples. Finally, we compare our proposed algorithm to other methods, and ours has better performance with high accuracy and robustness.
