Backward Motion for Estimation Enhancement in Sparse Visual Odometry

Visual Odometry (VO) aims to estimate camera position and orientation based on the images acquired by the camera. This area is rapidly gaining traction due to the dissemination of utonomous robots (cars, drones etc). It is well understood that different camera translations have different impacts on camera pose estimation. Movement along the optical axis (forward or backward) produces lower parallax than movement perpendicular to the optical axis (sideway translations). In this work, projection characteristics of sparse features for a camera moving away from the scene (backward movement) are discussed and a monocular backward looking camera VO system is suggested to take advantage of those characteristics. Implementation results tested with the KITTI dataset, a forward motion application processed in reversed order to emulate backward movement, show that the proposed system presents superior precision then forward looking published solutions and is suitable for real time operation.