Predictive video coding is based on motion estimation. In such systems, the temporal correlation is exploited at the encoder, whereas at the decoder, the correlation between the previously decoded frames and the current frame is never exploited. In this chapter, we propose a method for motion estimation at the decoder. Based on the prediction residue and on the already decoded frames, the decoder is able to partially reconstruct the motion field, which therefore can be skipped in the encoded stream. The proposed approach is based on Least Square Estimation (LSE) prediction, and is suitable for low bit-rate video coding, in which the transmission of the motion field has a significant impact on the overall bit-rate. The same technique could also be useful in case of high definition video coding, where a detailed and accurate motion field is required. Preliminary results seem to be very promising.