Research On Electronic Image Stabilization Based On Cluster Analysis And Kalman Filtering

The demand for a stable and trustworthy video is quickly increasing because of artificial intelligence techniques with real-time image recognition such as intelligent surveillance,military reconnaissance system and private camera which have been gradually applied.Video sequences captured from camera are often subject to annoying random jitter due to camera shaking caused by vibrations or unstable external environment.The unstable sequences acquired by camera have a great impact on further processing.Electronic image stabilization technology is to remove the interference between adjacent frames in order to obtain a good visual stability of the video.Firstly,the development of video stabilization technology in recent years has been researched.By comparing the advantages and disadvantages of the commonly used methods and their application fields,the affine transformation is chosen as the camera motion model.Then the clustering notion is applied to the algorithm of eliminating incorrect-matched points,and the real-time Kalman filter has been exploited to remove jitter components,which retains the actual camera movement at the same time.The following improvements in motion estimation and motion filtering have been made in this paper.For motion estimation,SURF combined with KLT algorithm is used to speed up the feature matching between adjacent frames at the beginning,then motion vectors have been calculated by matching points,which are exploited to establish two-dimensional feature space.In order to speed up the convergence of the clustering,the improved K-means algorithm has been adopted to cluster feature points,which contains the group with the largest number of motion vectors,has the highest probability of having global motion.Finally,the sample consistency algorithm and the cascade affine transformation model are used to refine the motion parameters of adjacent frames,which eliminate most of the local motion caused by the incorrect-matched points,and the adaptability of motion estimation in different camera scenes is enhanced.Aiming at the motion filtering,an adaptive filter based on constant ratio of parameters has been designed in the analysis of traditional Kalman filter.Firstly,the speed of low frequency scanning motion between frames can be regarded as a controllable variable,so that the gain of improved Kalman filter has been proved to be only related to the ratio of system noise and observation noise without concern for the value of the parameter.In order to adapt to different camera motion changes,two corresponding feedback factors are used to describe the real time filter performance.The ratio of parameters have been updated by consistency and stability of filtering result,which achieves a self-adjusting parameters of the motion filtering algorithm to accurately describe the true camera movement and adapt to its changes.