Improved pel-recursive motion estimation algorithms

The pel-recursive approach to motion estimation is studied. Two algorithms that model the spatial gradient as an estimate plus noise are derived. The first algorithm deals with the usual two-frame case, except that damping terms are introduced in the gradient covariance matrix. Simulations show that this algorithm is stable enough that no restriction on the maximum allowable updating step size is required. The second algorithm is an extension of the first algorithm to the general multiframe case. The algorithms are tested and compared using synthetic and real video sequences, and it is shown that using more than two frames simultaneously can track motion more accurately in noisy environments. >