Abstract: A new model based global motion estimation method is proposed by fitting optical flow field for aerial video imagery. Prior to motion model fitting, optical flows are first estimated based on a multi-resolution mechanism. However, due to noise-sensitive defect of standard linear regression technique, the uncertainties and independent motions of optical flow estimation would ruin the fitting result of global motion model. Therefore, instead of fitting the whole optical flow field, an array of non-overlapping flow regions is used to derive an initial set of motion models by fitting each region separately. Regions with fitting errors greater than a prescribed threshold are rejected since they do not provide a good description of motion within the region. Finally, the remaining optical flows are fitted again, which results in a more accurate motion model. Experimental results of both aerial image registration and independent motion detection show the effectiveness of the algorithm proposed.

Key words: global motion estimation, motion model fitting, optical flow, image registration, motion detection