A Practical Approach to Motion Estimation for Omnidirectional Vision

Compared with conventional perspective camera, data noise for omnidirectional camera is more serious due to its non-linear projection. For linear estimation of motion parameters with noise corrupted data, it is necessary to establish sufficient number of correspondences and use a appropriate robust technique. In this paper, our interests concentrate on two parts. First, we propose a new method to establish a large number of reliable point correspondences between omnidirectional images. Classical affine invariant region detecting and matching algorithms are initially employed, then more distinctive point correspondences inside or near the matched regions are extended. Our second work focuses on the RANSAC based linear estimation technique. Based on the simulation experiment we give a strategy to gradually select more and more inliers from a noise-corrupted data set, by which the initial guess of motion parameters improved in precision and stability. Experiment results with true data show that our method works well in real applications and makes the performance of linear estimates of motion parameters improved greatly.