Exploiting the residual redundancy in motion estimation vectors to improve the quality of compressed video transmitted over noisy channels

Many of the existing and emerging low-bit-rate video coding techniques employ motion compensation to exploit the high correlation between successive frames in real-world image sequences-the translation being represented by motion estimation vectors (MEVs). When transmitted over a noisy channel, errors in the MEVs can severely degrade performance, especially when the MEVs are entropy coded using a variable-length code prior to transmission. Techniques commonly used to mitigate the effect of these errors often employ some form of spatio-temporal error masking which relies on the relatively high degree of correlation between MEVs in neighboring macroblocks. This paper presents an alternate approach which exploits this correlation to reduce the probability of errors, rather than try and mask them when they do occur. The new approach is to perform maximum a posteriori probability (MAP) detection using a new method for joint source-channel MAP decoding applicable to data encoded using a variable length code followed by an FEC code. A first-order Markov model is used to model the inter-frame correlation between MEVs. Results presented show that the proposed approach may result in significant improvement in performance at low-to-mid SNR