A Minimax solution for stereo based motion estimation

In this paper we propose a new way of improving motion estimation based on L-infinity nonlinear minimization. The proposed approach which is a frame to frame motion estimation algorithm for consecutive image stereo pairs is based on three stages. Firstly image processing for detection and match of the remarkable features which are then reconstructed to 3D. Secondarily from the 3D point of cloud an initial estimate of the relative motion during the time step which separates the two frames is performed then used to filter the 3D point with a statistical outlier rejection scheme. Thirdly the minimization of this estimation within the filtered point cloud is conducted to reach a more accurate estimation of the motion. Minimax proposed method turns out to be even more efficient than commonly used L2 nonlinear minimizing approach especially for certain configurations of non-homogeneous distribution of the features. Cases which are quite frequent in visual odometry. Details of the algorithm and results will be discussed in the paper.