Aerial Visual-Inertial Odometry Performance Evaluation

Daniel J. Carson


With the cheapening of optical camera technology and an increasing interest in reducing GPS dependency, computer vision-based navigation algorithms have grown in popularity. This experiment implements five visual odometry algorithms into a common framework, evaluating their accuracy in a variety of situations on an unprecedented level. A variety of techniques are compared and contrasted including FAST-features versus gridded pixel-features, inertial-aided and inertial-independent rotation estimation, the effectiveness of image histogram equalization, the benefits of image rectification, and two-frame versus multi-frame visual odometry.