A new adaptive Kalman filtering method for block-based motion estimation

Abstract

This paper presents a new adaptive Kalman filtering method to improve the performance of block-based motion estimation. In our work, measured motion vectors are obtained by a conventional block-matching algorithm (BMA). A first order autoregressive model is employed to fit the motion correlation between neighboring blocks and then to achieve the predicted motion information. To further improve the performance, a new approach is proposed for adaptively adjust the state parameters of the Kalman filter during each filtering iteration according to the statistics of the motion vector (MV) measurements. The experimental results indicate the effectiveness of the proposed method in terms of peak-signal-to-noise-ratio (PSNR) of the motion compensated images with smoother motion vector fields. Furthermore, since the Kalman filtering is independent of motion estimation, it can be applied to any block matching motion estimation algorithm for performance improvement with little modification. Another benefit from the Kalman filtering is the fraction pixel accuracy of motion vectors with no additional bits to transmit for MV.