Abstract: Rotary jitter compensation is a difficulty in video image stabilization. The feature point matching and inaccurate point filtering were studied. An affine model with 6 parameters was established for moving images, and an over-determined equation to estimate motion parameters was derived. The least squares iterative algorithm was used to remove error feature points judged by sum of absolute difference (SAD) matching algorithm. A pyramid-style Lucas-Kanade (LK) algorithm based on optical flow was adopted for feature point matching of rotary video. All algorithms were programmed. After detecting feature points by using the gradient matrix of the feature window (KLT) method, SAD and LK algorithms were used to match feature points respectively, and the rotation matrix parameter errors were obtained and compared. The reasons of matching error were analyzed. Kalman filter was used to smooth rotation parameters. All algorithms were implemented on an autonomous robot. The experiment results show that LK can get less matching errors than SAD for rotation jitter, and Kalman filter makes the maximum 8.37° rotation jitter less than 3.68°, thus it can compensate the rotary video effectively.

Key words: information processing, feature point, optical flow, jitter compensation, robot