It should be emphasized that visual tracking is different from motion estimation. The latter concerns only two successive frames; while the former concerns the motion estimation of a whole video sequence. While a naive tracker estimates motion recursively frame by frame, tracking is more than motion estimation as the above naive tracker is prone to drift away. To overcome drifting, the tracking algorithm does the following: (i) updates the appearance model strategically and/or (ii) performs temporal smoothing/fusion.
--- from paper : BoostMotion: Boosting a discriminative similarity function for motion estimation
