Abstract

Attempts to derive 3-D self-motion information from 2-D image sequences have been hampered by the difficult problem of reliably measuring the 2-D image motion. The small size of the motion sensors relative to the length of many of the contours in the scene results in measurements of the normal component of the velocity vectors, not the correct image velocity (the aperture problem). This is problematical for many self-motion estimation algorithms, especially those that rely on vector differences to decompose the translational and rotational components of the flowfield. Current techniques for solving the aperture problem are not readily applicable to the type of image motion generated during observer motion. We know however that the normal component of the image motion is constrained to lie within ±90° of the true direction. Motion sensitive cells, such as those in the area MT of the primate brain, often have direction bandwidths of this size.

© 1991 Optical Society of America