Motion Estimation with Affine Flow (Roerelseestimering med Affin Floedesmodell)

摘要

When an algorithm in a system searches for targets in an image from some sort of211u001esensor (e.g., video camera or IR-camera), the algorithm must be fast enough to 211u001ematch the image rate and not prone to false target alarms. The first step towards 211u001esuch an algorithm is to reduce background information without affecting the 211u001epossible target signatures. For this purpose motion estimation is needed. If the 211u001ebackground movement can be defined and is applied when two consecutive frames are 211u001esubtracted, the result will be the objects that are moving differently in respect 211u001eto the background (possible targets). The report describes a motion estimation 211u001ealgorithm. The algorithm uses a hierarchical approach to eliminate false matches 211u001eby using 'large scale' structures, and to achieve a computationally efficient 211u001eestimation. The estimation model is an affine flow model and is implemented with 211u001esix parameters. The implemented algorithm uses linear interpolation to calculate 211u001ethe compensated images. Future work is to improve the algorithm to be able to 211u001edistinguish more than one motionfield to get a more robust estimation.