Motion estimation for video compression using Kalman filtering with adaptive adjustment

In this work, a new block-based motion estimation method using the adaptive Kalman filtering (KF) based on one-dimensional (1-D) and two-dimensional (2-D) autoregressive (AR) models is proposed to improve the motion estimates. Conventional block-matching algorithms (BMAs) are utilized to obtain the measured motion vectors (MVs). Autoregressive models are employed to characterize the motion correlation between neighbouring blocks and predict motion information. The measured and predicted motions are utilized to generate an optimal estimate of motion vector. To improve the performance of the adaptive Kalman filtering, in our method, a new error function is defined and the statistics of MV measurements are utilized to dynamically adjust the state parameters during each iteration. The experimental results indicate that the proposed approach performs better than the existing methods in comparison, in terms of the peak-signal-to-noise-ratio (PNSR) of the motion compensated images. The developed adaptive KF method is independent of motion estimation, thus it can be applied to any block-matching motion estimation scheme for performance improvement. Additionally, after the Kalman filtering, the fractional-pixel accuracy of motion vector is achieved with no additional bits for MV.