在駕駛人注意力偵測、臉部辨識、及擴增實境的應用中，從連續影像中估計出臉部姿勢是很重要。在本論文的研究中，我們藉由卡曼濾波器，從連續影像獲得臉部的旋轉及位移參數。此外，在臉部辨識系統中，我們可以將非正面臉影像利用已獲得的臉部旋轉資料轉換成正面臉影像以利辨識。 臉部姿勢估計主要有三個步驟，臉部偵測、臉部特徵點偵測、及臉部姿勢估計。首先，利用臉部相似圓型的特性在二值化影像中偵測臉部。在第二步驟中，我們從臉部影像中偵測眼睛及嘴巴區塊，分別在這些區塊中偵測角點，接著根據眼角和嘴角的特性，定位出四個眼角及二個嘴角。在第三步驟中，我們以三維人臉模型對應使用者的臉部，取得使用者的三維特徵點座標；接著，對每張影像偵測臉部特徵點位置，利用影像平面座標與空間座標的對應關係，粗略估計臉部姿勢。最後再以卡曼濾波器追蹤臉部轉動，期望可以準確地估計臉部姿勢。 在臉部轉正的應用上，我們以獲得的臉部姿勢及一個通俗的人臉模型將非正面臉影像轉換成正面臉影像，以改進臉部過於偏轉不利於辨識，藉此期望增加人臉辨識率的問題。估計臉部姿勢的實驗分析中，我們發現參數設定及三維特徵點座標對於估計旋轉角度的正確率有很重要的影響。在實驗中，我們用已知轉角的臉部影像來估計臉部旋轉的角度；大概轉角誤差都在正負五度內。 Acquiring the location and orientation of faces from an image sequence is important to driver attention detection, face recognition, and augmented reality applications. In this thesis, we propose a Kalman filter-based method to acquire the face location and orientation from contiguous images. Moreover, we need to transform the non-front-views face to a front-views face based on the acquired orientation data for face recognition application. There are three steps for face position estimation. The steps include face detection, facial feature points detection, and face position estimation. First, a face is detected based on a circle model. Second, we detect the region of eyes and mouth from face image and then detect the corners from those regions. Finally, the corresponding facial feature points in 3D and 2D coordinates are used to estimate the 3D face pose. To get the accurate 3D coordinates of facial feature points, a 3D face model is constructed. To more accurately detect the 3D face pose, we employ a Kalman filter to detect and track the 3D face pose. In face recognition, front-view faces are commonly used. A side-view face is generally hard to recognize. In this study, we try to transform the non-front-view faces into the front-view ones for recognition. We hope the transformation may improve the recognition rate. In experiments, we found the accuracy of the proposed pose estimation method is heavily dependent on the parameters and coordinates of the 3D face model. With a set of more accurate parameters and coordinates of the 3D face model. We can get error rate of plus or minus of five degrees or less with estimated rotation angle and known rotation angle for face pose estimation.
