光流資訊在電腦視覺與圖形識別領域中，是連續影像的一種重要的對應資 訊。有別於多數的傳統光流演算法僅使用於前後兩張影像，在本篇論文中，我們 提出了一個結合多張影像的新式光流估測法，此法具有下列特徵：(1)以微分為 基礎，不受區塊比對法的子像素限制；(2)以特徵點為基礎，可以在影像各點獨 立運作；(3)時間軸的資訊加入使得歧義對應的減少，因此也適用於以光流場為 基礎的應用。在實驗結果中，我們驗證了這個方法在整體光流場而言，具有較少 的歧義對應及較低的平均估測誤差。若是在好的特徵點上，估測值會更為準確。 為了進一步驗證我們所提出的光流估測法的實用性及有效性，在本論文中， 我們將此光流估測法應用在兩個實際問題上。第一，在智慧型交通運輸系統中， 藉由影像中車輛特徵點的光流估測值，投影至道路平面，可以用來估測車輛的真 實速度。實驗結果顯示在特徵點準確的光流估測之下，可以得到與車輛真實速度 接近的估測速度。第二，我們提出了一套基於運動模式的真假臉辨認系統。藉由 觀察真人臉與照片臉的光流場分布差異，我們提出了基於線性區分分析與直方統 計圖比對兩種不同的辨識方法來區分照片臉的偽裝。實驗結果顯示，利用我們所 提出的光流估測法，可以在真臉運動與假臉運動之間產生顯著的差異，並可產生 準確的辦識率。最後，我們對於提出的多圖像光流估測法作出結論並提出未來可 以改進的方向。 Optical flow reveals important correspondence information in the fields of computer vision and pattern recognition. Different from the traditional methods which only use two successive frames, we propose a novel optical flow method by integrating multiple frames. This method has the following characteristics: (1) It is a gradient-based method so that it will not be constrained by the subpixel matching problem, (2) It is a feature-based method so that it can estimate independently of each image point, (3) The reduction of ambiguous matching because of the temporal information included, so that it can also be adopted in the applications based on the dense optical flow field. In the experimental results, we have verified that the proposed method will produce less ambiguous matching and estimation error. Moreover, the estimation results will be more accurate at good feature points. To further verify the practicability and effectiveness of the proposed optical flow method, we apply this method to two practical problems in this dissertation. First, in an intelligent transportation system, the real vehicle speed can be estimated by optical flow at feature points through an image-road mapping. Experimental results show that if optical flow can be successfully and accurately estimated, the speed estimationresults will be close to the real speed. Second, we propose a system to distinguish true faces and face photos based on their motion models. By observing the difference of both models, an LDA-based method and a histogram-based method are proposed to detect the falsification by using face photo. Experimental results demonstrate that if the multi-frame optical flow method is adopted, the motion difference between true face and face photo is obvious so that the satisfactory verification rate can be obtained. Finally, concluding remarks of the proposed method are given and the improvement methods for future works are listed.
