A novel translation estimation for essential matrix based stereo visual odometry

Visual Odometry (VO) plays an important role in autonomous navigation systems for vehicle localization. For traditional stereo visual odometry (SVO), we can estimate the rotation and translation of camera motion either simultaneously or separately where 3D information reconstructed from the stereo image is used as the input of the translation estimation. The accuracy of pose estimation is dependent on the uncertainty of 3D features as well as their portion used. This paper presents a novel translation estimation for essential matrix-based SVO to avoid the effectiveness of 3D feature uncertainty from stereo disparity. The rotation is extracted accurately from essential matrix of each pair of consecutive image frames on the left side; with a pre-estimated rotation matrix, the translation is rapidly and accurately estimated by solving a proposed linear closed-form only using 2D features as input with one-point RANSAC. The experimental results on the autonomous driving testing dataset (KITTI) indicate that the proposed approach enhances 20 % accuracy compared to traditional approaches in the same experimental scenario.