Monocular Visual Odometry with Cyclic Estimation

Monocular Visual Odometry (MVO) estimates the camera position and orientation, based on images generated by a single camera. In this paper a new sparse MVO system for camera equipped vehicles is proposed. Three view cyclic Perspective-n-Point with adaptive threshold is used for camera pose estimation, perspective image transformations are used to improve tracking, and a multi-attribute cost function selects ground features for scale recovery. Results using the KITTI dataset show that the proposed system achieves 1.29% average translation error and average rotation precision of 0.0029 degrees per meter.