Study Of Electronic Stabilization Technique For Video Sequences

Electronic Image Stabilization(EIS) is a stability technique for video image sequences based on computer, which purpose is to separate motion vectors with digital image processing technology, identify the image offset difference, and finally get the stable image sequence after image compensation. EIS has been widely used in the field of civil, industrial and military investigation for its many advantages, such as high accuracy, low cost, small volume, etc.EIS system is mainly composed of three parts, Image preprocessing, motion estimation, and motion compensation. Among them, the motion estimation, which includes motion detection and motion determining, plays a key role in the EIS system for its accuracy, robustness and real-time, it affects the performance of the whole system greatly.In this thesis, we firstly analyzed the common EIS algorithms and their evaluation criteria. On this basis, two algorithms were proposed according to the different requirements on accuracy and real-time, namely the high-precision EIS algorithm bases on SIFT feature matching and the real-time one bases on Kalman filter. Both algorithms were used to stable the actual captured image sequences and good results were achieved.For the image sequences with stringent requirements of accuracy, we firstly got the matching points with the SIFT feature extraction method. Secondly, the RANSAC-based estimation algorithm was used to select the precise points, estimate affine transformation parameters. Finally the images were compensated. The EIS algorithm bases on Kalman filter was used for the image sequences which need to be real-time stabled. We firstly extracted the feature points, and secondly simulated the trajectory of camera with the Kalman filter, which purpose is to detect the conscious and unconscious movement of camera, and improve the algorithm speed at the same time. Thirdly, we established the Gaussian pyramid to realize coarse to precise estimation. According to the idea of optical flow, the inter-frame translation, rotation, scaling vectors were estimated with the error approximation algorithm. Finally, the affine transformation parameters were improved with the method of groups feature points and weight parameters calculation. In this thesis, the two algorithms above were deeply analyzed and studied, and the stabilization effects were objective evaluated with the actual captured videos, experimental results show that the PSNR between the frames after stabled image sequence are significantly improved, so the two algorithms in this thesis are effective.Finally, a video stabilization system was established with VC++ language and different videos were stabled with the system. The experiment shows that this system can stabilize a variety of videos and basically meet the requirements of stabilization for common kinds of videos.