Olivier Toupet
NASA Jet Propulsion Laboratory, USA
Title: Fusing wheel and visual odometry with inertial measurements: An Extended Kalman Filter for Unmanned Ground Vehicles in GPS-Denied Enviroments
Biography
Biography: Olivier Toupet
Abstract
Localization is a central component of a robotic system as many other technologies providing autonomy rely upon it, such as
planning, control, mapping, etc. Smooth and accurate real-time pose estimation is thus a key enabler and many algorithms
exist to provide that capability. One of the most successful and widespread approach is the Extended Kalman Filter (EKF),
which can fuse asynchronous, time-delayed measurements from heterogeneous sensors. It is commonly used in Inertial
Navigation Systems (INS) for combining GPS and IMU measurements. In order to estimate the position and orientation
of an unmanned ground vehicle as it travels through an unknown, unstructured, GPS-denied environment, we leveraged
the EKF framework to fuse the information provided by the vehicle’s onboard gyroscopes, accelerometers, wheel encoders
and cameras. Our implementation is customized for ground vehicles, where the non-holonomic kinematics makes the lateral
motion diffi cult to observe. We use angular rate measurements provided by the IMU’s gyroscopes together with the forward
linear velocity measurement provided by the wheel encoders to predict the vehicle’s motion at a high rate. We then use Visual
Odometry (VO) as a relative pose measurement together with accelerometer measurements provided by the IMU to update
and correct those predictions over time. Our pose estimation system was demonstrated on a real unmanned ground vehicle, a
modifi ed military Humvee, in both short and long-drive scenarios and achieved a performance of less than one percent error
over distance travelled with accumulated yaw drift being the major source of localization error