Error concealment has attracted extensive interest from both research and application communities, for it is an economical approach at the decoder side to ensuring an acceptable and stable video quality in case of packet erasure or loss. The common strategy is to employ the available spatial or temporal information as much as possible. However, for heavily corrupted sequences, satisfactory techniques are still pretty scarce. In the present work we put forward a new framework based on the unscented Kalman filter accompanied by a modified bilinear Motion Field Interpolation (MFI) method. Experiments on typical sequences with high macroblock loss rates unfold its average superiority over the boundary matching algorithm and the conventional MFI algorithm by up to 4.32 dB.