Error concealment via particle filter by Gaussian mixture modeling of motion vectors for H.264/AVC

We use particle filtering for correcting the erroneous motion vectors (MVs), which are derived from the boundary matching algorithm (BMA) in packet video communications. Assuming a two-state Markov channel model for transmission, the error of the extracted MVs by BMA is shown to be modeled by the Gaussian mixture distribution. Formulating the problem in the state space, we deploy particle filtering for denoising the erroneous MVs. The main challenge of using particle filters is high computational complexity that is directly related to the number of particles. The proposed particle filtering scheme is efficient even if the number of particles is decreased. Experimental results are provided to show the efficiency of this filtering approach compared to a recent scheme based on Kalman filtering. The experiments show meaningful increase in the quality of the recovered video sequences in terms of PSNR up to 3 dB compared with the other error concealment techniques. Also, the computational complexity of the proposed scheme is discussed.