Although there are many feature matching and tracking methods so far

the side effect of moving foreground object

which will cause global motion estimation error

is still an open problem. In order to avoid features

located on the foreground objects

participating in motion estimation

feature effectiveness evaluation is employed to improve feature reliability for the features extracted by the traditional KLT method. Effective features are utilized to estimate global motion and obtain accurate motion parameter

based on which video frames are compensated. However

motion compensation will cause undefined area. There are some approaches to reconstruct the undefined area; nevertheless

they have not considered the effect of fast moving objects in the foreground of the video

which will decrease the video quality after stabilization and content completion. In our proposed algorithm

optical flow between defined areas of current frame and neighbor frame is first calculated

and then it is used as a guide to erode unknown areas. Finally

mosaicking on the base of reference frame is used to obtain a complete video stabilization sequence. Experiment results show that the proposed method is robust to moving foreground objects and is able to realize video frames stabilization with complete content.