Estimating egomotion in a 3D cluttered scene

In our optical snow research, we showed how to estimate motion parallax in cluttered 3D scenes.  Such estimates are useful for solving for egomotion, namely the instantaneous 3D camera motion.   Previous techniques for computing have assumed that one can reliably estimate pointwise image velocities.  Unfortunately, this assumption is dubious in cluttered scenes, since occlusion boundaries and spurious features such T-junctions are dense and classical optical flow methods do not perform well in these conditions.  We developed an alternative method estimating local motion parallax directly from image intensities (see optical snow), which avoids these prior pointwise estimates, and shown how to combine these local estimates to estimate the observer's egomotion.

Selected Publications
bush          bowties       vxvy