To solve the inaccurate problem of global motion estimation due to local motion in video stabilization, a real-time stabilization method was proposed. The method proposes an approach of feature points iterative filtering algorithm based on minimal spanning tree and applies spanning tree similarity of successive frames to measure feature matching and discard wrong matched points and points in moving foreground. Then, an adaptive weighted method was applied to correct the transformation matrix to solve the wobble problem due to few feature points caused by foreground occlusion. Finally, an modified dual Kalman filters based on motion queue was proposed to correct measurement noise covariance adaptively and adjust smoothess of filter dynamically. It can process videos with intended camera motion and random jitters effectively. Experimental results show that the proposed method can perform well even in situation of local foreground motion and intended camera motion. Under circumstance of Intel Core i5 3.30 GHz CPU, the proposed method can stabilize videos of 640×360 resolution at a rate of 40FPS. And it has real-time advantage since next frame information is not needed during computation.