數位影像穩定技術及其應用

Abstract

本論文為提出數位影像穩定技術研究成果及其應用，本文中提出如何在手持式、車用或固定式監視攝影機，因人為不可預期的抖動、車輛行駛巔岥、方向盤轉動的效應或風力、外力等的影響而造成影像的抖動，以數位影像處理的方式去除不必要的抖動而保留必要的移動。數位影像穩定技術主要可分為兩部份：(1)如何從影像序列中有效率地估算出準確、可靠的全域移動向量；(2)如何從所取得的移動向量在邊界的限制下補償出一平滑的影像運動軌跡。 本文在移動向量估測上提出倒三角方法以尋找區域移動向量、最佳化代表點選擇以降低尋找移動向量的計算量、產生精煉型移動向量以應用在乏特徵的狀況仍可估算全域移動向量、另提出天際線的檢測方法與以背景為基礎的對等演算法以求得較可靠的全域移動向量。 本文在移動向量補償上提出移動軌跡的繪製與平滑指標的計算以驗證所提出移動向量補償方法改善數位影像穩定的定量分析，同時在補償方法中提出在回路中加上一內部回授積分器以改善鏡頭在定速移動時造成補償效果不良的問題。最後則以模糊推論的機制，以兩種不同的方法透過模糊推論，選擇較佳的補償方式，以適應數位影像防振在各種情況的應用。 經實驗的結果，本文所提出的方法可適用在不同狀況的影像序列如乏特徵、重複圖樣、大移動物件及大區域低對比的影像狀況，而能估算出準確的全域移動向量。在移動向量補償上則解決在定速移動下所造成補償效果降低的問題，並以模糊推論整合方式有效地改善補償向量的補償效果，從平滑指標與移動軌跡圖上均可驗證本文所提出的方法有效地改善數位影像穩定的效能。

In this dissertation, a digital image stabilization (DIS) technique and its applications are proposed as a way to remove the unwanted shaking phenomena in the image sequences captured by hand-held, in-car or fixed-type surveillance camcorders, without being affected by moving objects in the image sequence or by the intentional panning motion of the camera. DIS contains two major parts: (1) How to estimate an efficient, precise, and reliable global motion vector. (2) How to use the existing GMV to compensate for a smooth motion trajectory within the window shifting allowance boundary. For motion estimation (ME), an inverse triangular method is proposed to look for the local motion vector (LMV). An optimization of the representative points is proposed to reduce computation complexity, and a refined motion vector is proposed to apply to any ill-conditions of the GMV estimation. Skyline detection for in-car applications and background based peer to peer evaluation are proposed to enforce the reliability of the GMV estimation as well. For motion compensation (MC), a plot of the motion trajectory and a smoothness index evaluation are proposed to quantitatively verify the analysis that shows the improvement of the MC. An inner feedback-loop integrator has been applied to the MC to improve image stabilization during the camera’s panning motion. Finally, Fuzzy inference digital image stabilization (FIDIS) is proposed to adaptively determine better motion compensation methods through the use of two different MCs. Experimental results show that the proposed methods of this dissertation adapt to different conditions of image sequencing, such as a lack of features, repeated patterns, large moving objects and large low-contrast areas in the image and that they can also estimate the GMV precisely as well. The degradation of image stabilization during the panning motion is solved by adding an inner feedback-loop integrator. The proposed FIDIS also shows effective improvements in different conditions of image sequence through the evaluations of the smoothness index and the motion trajectory.