Omnidirectional Camera Motion Estimation

Abstract

We present an automatic technique for computing relative ca mer motion and simultaneous omnidirectional image matching. Our technique works for small as well as larg e motions, tolerates multiple moving objects and very large occlusions in the scene. We combine three prin ci les and obtain a practical algorithm which improves the state of the art. First, we show that the correct motion is found much sooner if the tentative matches are sampled after ordering them by the similarity of their descriptors. Secondly, we show that the correct camera motion can be better found by soft voting for t he direction of the motion than by selecting the motion that is supported by the largest set of matches. Fi nally, we show that it is useful to filter out the epipolar geometries which are not generated by points recon structed in front of cameras. We demonstrate the performance of the technique in an experiment with 189 image pairs acquired in a city and in a park. All camera motions were recovered with the error of the motion di rection smaller than 8 ◦, which is 4 % of the 183◦ field of view, w.r.t. the ground truth.