REFINING 3D RECONSTRUCTIONS - A THEORETICAL AND EXPERIMENTAL-STUDY OFTHE EFFECT OF CROSS-CORRELATIONS

In robot navigation a model of the environment needs to be reconstruct ed for various applications, including path planning, obstacle avoidan ce, and determining where the robot is located. Traditionally, the mod el was acquired using two images (two-frame structure from motion) but the acquired model were unreliable and inaccurate. Recently, research has shifted to using several frames (multiframe structure from motion ) instead of just two frames. However, almost none of the reported mul tiframe algorithms have produced accurate and stable reconstructions f or general robot motion. The main reason seems to be that the primary source of error in the reconstruction-the error in the underlying moti on-has been mostly ignored. Intuitively, if a reconstruction of the sc ene is made up of points, this motion error affects each reconstructed point in a systematic way. For example, if the translation of the rob ot is erroneous in a certain direction, all the reconstructed points w ould be shifted along the same direction. The contributions of this pa per include mathematically isolating the effect of the motion error (a s correlations in the structure error) and showing theoretically that these correlations can improve existing multiframe structure from moti on techniques. Finally it is shown that new experimental results and p reviously reported work confirm the theoretical predictions. (C) 1994 Academic Press, Inc.