Research On Digital Image Stabilization Technology Based On Vector Clustering And Kalman Filtering

As electronic imaging devices are widely used in military and civil fields, the quality requirements of video are increasingly high. Especially in some special shooting situations, such as hand-held camera systems, monitoring system in earthquake and military UAV shooting system, the captured video is usually accompanied by a large jitter, which is not conducive to the manual observation and the extraction of useful information. Digital image stabilization technology aimed at removing this interference noise is a video stabilization technology, which use methods related to digital image processing to eliminate the camera jitter, and eventually achieve the stability of video sequences. This paper mainly studies the core part of digital image stabilization process, and improves the motion estimation algorithm and the motion separation algorithm respectively.The current development of digital image stabilization technology has been researched at the beginning of this paper. By comparing the strengths and weaknesses of different algorithms, the similarity matrix is selected as the image transformation model eventually; besides, feature points are extracted by the combined algorithm based on Harris corner detection and KLT tracking, and the Kalman filter with good real-time performance and accuracy is used to separate the motion parameters.Aiming to deal with the interference of moving object for motion estimation, an improved algorithm resistant to foreground is proposed in this paper. Firstly, motion vectors are defined based on the matching feature points of adjacent images, which will be divided into different categories through the improved k_means clustering algorithm, and then complete the classification of all feature points in the image. According to the proposed elimination criterion of moving object in continuous video images, the blind separation of background is achieved effectively, which makes the local motion vectors of reserved feature points consistent. Finally, the reserved motion vectors of feature points are used to compute the global motion parameters, which ensure that the obtained camera motion is more accurate and close to reality.Furthermore, a hybrid scheme based on improved Kalman prediction filtering is proposed in this paper. This method combines the smooth performance of low-pass filter and the prediction characteristic of Kalman filter respectively, which can more effectively achieve the separation of movement, and filter out the camera jitter. Moreover, the improved Kalman filter can adaptively update statistical properties of noise according to the actual motion of camera, and further enhance the robustness of this algorithm. In addition, Kalman prediction algorithm used in this paper can ensure the filtering effect while reducing computation time, thus the real-time performance of this algorithm is greatly improved.Experimental results show that the motion estimation method proposed in this paper has good resistance performance to foreground, and the improved motion separation method can obtain more smooth and effective active scanning motion. The quality of compensated video is obviously improved after applying the separated motion parameters to the part of motion compensation.