Non-iterative visual odometry using a monocular camera -

Abstract

This thesis presents a visual odometry system for ground vehicles using a single downward-facing camera and two tilt sensors. Conventional visual odometry algorithms use the probabilistic and iterative RANSAC to remove outliers and calculate the motion. They suffer from different problems including dynamic obstacles, changes in lighting conditions, inaccuracy in depth estimation and the high computational cost. The proposed method calculates the motion from a monocular camera without using any probabilistic (non-deterministic) or iterative routines. It makes use of the constant distance between the camera and the ground to impose the depth and improve the accuracy. Moreover, it makes use of the concept of a downward looking camera and the known depth to implement the inliers-detection method as a substitute for RANSAC to remove outliers. This improves the speed of the algorithm and decreases the computational cost. The algorithm is validated for real data sets and shows competitive accuracy and robustness with a loop closure error reaching as low as 1.25percent for a run of 461 meters.