Pose Estimation of an Autonomous Car by Visual Feature Correspondence and Tracking

This paper presents a pose estimation method based on 1-Point RANSAC EKF (Extended Kalman Filter) framework. The method fuses the depth data from a LIDAR and the visual data from a monocular camera to estimate the pose of an Unmanned Ground Vehicle (UGV) in a GPS denied environment. Its estimation framework continually updates the vehicle’s 6D pose state and temporary estimates of the extracted visual features’ 3D positions. In contrast to the conventional EKF-SLAM (Simultaneous Localization And Mapping) frameworks, the proposed method discards feature estimates from the extended state vector once they are no longer observed for several steps. As a result, the extended state vector always maintains a reasonable size that is suitable for online calculation. The fusion of laser and visual data is performed both in the feature initialization part of the EKFSLAM process and in the motion prediction stage. A RANSAC pose calculation procedure is devised to produce pose estimate for the motion model. The proposed method has been successfully tested on the Ford campus’s LIDAR-Vision dataset. The results are compared with the ground truth and the estimation error is ∼1.9% of the path length.