裝置在行動載具上的相機拍攝連續影像例如，汽車、機車、自行車、電動輪椅、..等，由於道路品質不佳或行動載具避震效果不好，連續影像會有不自主晃動及震動的現象，而使得連續影像不適合人眼觀看，也不適合後續的3D電腦視覺分析。因此在本論文的研究中，我們分析影像像素的移動向量 (motion vector)，藉以尋找補償影像不自主晃動及震動的問題。 影像中的移動向量有主要三個來源：i.景觀中自主運動 (autonomous motion) 的物體，ii.行動載具帶動相機左右及上下自我轉動 (ego-motion)，iii.道路品質不佳或行動載具避震效果不好所造成的相機不自主晃動及震動 (involuntary motion)。本研究目的只是要消除影像中的第三種不自主晃動及震動，並盡量保留第一、二種本來就存在的運動資訊。 本研究進行步驟有比對區塊及估計區塊移動向量、篩除不可靠的移動向量、篩除自主運動物體的移動向量、估計影像的等距轉換 (isometric transformation) 參數與補償影像並保留自我移動向量。 首先利用PMVFAST (predictive motion vector field adaptive search technique) 快速比對區塊並取得區塊移動向量。接著，利用邊強度與SAD (sum of absolute difference) 值篩除不可靠區塊移動向量以減去不準確移動向量的影響。第三步驟，我們使用篩選後的移動向量，利用最小平方誤差估計法計算出初始影像轉換參數，接著利用初始影像轉換參數轉換後的移動向量與原始移動向量之間的差異值來篩除自主運動物體的移動向量。第四步驟，將所有保留下來的區塊移動向量再使用相同的最小平方誤差估計法，估計目前影像的等距轉換參數。第五步驟，分析影像移動向量並估算最大不自主震動量，以其為門檻值來辨別自我轉動。然後再依據此門檻值來決定是否修正影像補償向量，以保留自我轉動。最後依據補償向量補償影像，也就是補償不自主震動，使得前後時刻影像內的背景能夠保持穩定。 我們在不同的載具上測試我們的系統，穩定後影像的PSNR (peak signal to noise ratio) 值平均提升約20.85%，而具有適應航向時的系統PSNR值則高於無適應航向約53.4%。我們也比較了固定門檻值搭配最低個數百分比與固定個數百分比之間的做法，而利用前者確實可以使系統得到較好的穩定效果，若以數據表示則約為0.97%。 Because of poor quality of roads or bad shock absorber of moving platform, there will be shaking or vibration in the image sequences captured by camcorder installed in moving platforms such as cars, motorcycle, bicycle, and power wheelchair. Then these image sequences are not suitable for watching with eyes or analyzing with computer. So in this thesis, we estimate the motion vector of image pixels to overcome the involuntary shaking and vibration problems of these image sequences. There are three kinds of motion vector in image sequences. The first kind is from object with autonomous motion in the scene. The second kind is camera ego-motion made by moving platform. And the third kind is involuntary motion made by poor quality of roads or back shock absorber of moving platform. Our purpose is only compensating third kind of motion and keeping first and second kinds of motion in the stabilized image sequences as many as possible. The stabilization system consists five step processes. First we use PMVFAST (predictive motion vector field adaptive search technique) to get motion vectors of every image block. Then we filter out unreliable block motion vectors by edge responses and SAD (sum of absolute difference) value. Then we estimate the first isometric transformation parameter by least-squares. Then we use the first isometric transformation parameter to estimate new block motion vector of and then compare new block motion vector with original ones to filter out the autonomous motion of moving object. Then we can estimate the isometric transformation parameter by least-squares again with remained block motion vectors. Then we analyze image motion vector to estimate the max involuntary shaking value as a threshold to determine compensation motion vector. Then we use the compensation motion vector to compensate involuntary motion and preserve camera ego-motion. So finally we keep the background stabilized in the output image sequences. In our experiment, we have test five different moving platforms. The PSNR (peak signal to noise ratio) value of output image sequences is 20.85% larger than the original image sequences in average. The frame rate is about 30 frames per second.
