Improved gradient-based methods for motion estimation in image sequences

This paper shows that two conventional gradient-based motion estimation techniques can be significantly improved by using unit gradient vectors instead of image intensities. Since unit gradient vectors are insensitive to changing image intensities, the proposed motion estimation techniques are far more robust to irregular lighting conditions than the conventional techniques. In addition, the proposed techniques demonstrate much better motion estimation performance where image gradients are close to one-directional because the proposed methods utilize the measure of the varieties of local gradient directions as a weighting factor in the computation of motion estimates.