Lidar-Visual-Inertial Odometry via Modifications and Improvements to Super Odometry

Abstract

The main focus of this thesis involves improvements and extensions to Super Odometry, a preexisting method for lidar-inertial odometry. This was done in the context of the DARPA RACER program as a member of Carnegie Mellon's DEAD Fast team, aiming to provide reliable state estimation for an autonomous off-road ground vehicle. Super Odometry was modified to take simultaneous input from multiple Velodyne VLP-32C lidars, and a more complex "dewarping" method was added to fully account for motion of the vehicle during a given lidar scan. Significant effort was put into streamlining and optimizing the implementation of Super Odometry to maintain real-time performance and provide a stable baseline for future work on the RACER program. In addition, a secondary pose graph was added to fuse the odometry solution with absolute position data from GPS. This provided a full 6-DoF pose estimate in UTM coordinates to the rest of the vehicle's autonomy stack.
The second thrust of this work focused on merging TartanVO, a learning-based visual odometry method, with Super Odometry. By design, Super Odometry can easily accept another source of pose estimates in addition to lidar odometry. These poses are added as relative pose factors in its factor graph. Significant implementation efforts were put into producing a working lidar-visual-inertial odometry method under the Super Odometry framework.