Proceedings of the International Conference on Advanced Robotics (ICAR) , August 2007.
Abstract
We present an integrated system to localize a
mobile robot in rough outdoor terrain using visual odometry. Our previous work [1] presented a visual odometry solution
that estimates frame-to-frame motion from stereo cameras and integrated this incremental motion with a low cost GPS.
We extend that work through the use of bundle adjustment over multiple frames. Bundle adjustment helps to reduce the error significantly, thereby making our system more robust and accurate while still operating in real-time. Our new system can keep the robot well localized over several hundreds of meters to within 1% error. We present experimental results for our system over a 300 meters run in a challenging environment and compare it with ground truth Real Time Kinematic (RTK) GPS.
Learning Applied to Ground Robots
LAGR is a DARPA project to develop autonomous offroad navigation, using techniques based on realtime vision and learning.
