Nonlinearly Optimized Dual Stereo Visual Odometry Fusion

Abstract

Visual odometry (VO) is an important problem studied in robotics and computer vision in which the relative camera motion is computed through visual information. In this work, we propose to reduce the error accumulation of a dual stereo VO system (4 cameras) computing 6 degrees of freedom poses by fusing two independent stereo odometry with a nonlinear optimization. Our approach computes two stereo odometries employing the LIBVISO2 algorithm and later merge them by using image correspondences between the stereo pairs and minimizing the reprojection error with graph-based bundle adjustment. Experiments carried out on the KITTI odometry datasets show that our method computes more accurate estimates (measured as the Relative Positioning Error) in comparison to the traditional stereo odometry (stereo bundle adjustment). In addition, the proposed method has a similar or better odometry accuracy compared to ORB-SLAM2 and UCOSLAM algorithms.