Online Autonomous Extrinsic Calibration of an Inertial Measurement Unit using Gaussian Radial Basis Function Neural Networks

Abstract

This thesis presents a fully online and autonomous method to extrinsically calibrate an inertial measurement unit (IMU) to the body frame of a vehicle. Extrinsic sensor calibration is an important step in obtaining valid information in the vehicle frame, without which an autonomous vehicle cannot function properly. Traditionally, a manual calibration routine must be performed by a set of trained experts to a high degree of precision before the vehicle can be safely operated. This procedure costs time and money and limits the design of the sensor suite. An online and autonomous calibration method would eliminate this constraint, saving time, and allowing for the dynamic reconfiguration of the sensor suite. The autonomous IMU-to-Vehicle calibration procedure presented in this thesis is conducted in a two-stage process. First, a gaussian radial basis function neural network is used to emulate the output of a virtual IMU in the vehicle frame. Then, a maximum likelihood search algorithm estimates the extrinsic calibration parameters by performing an IMU-to-IMU calibration between the IMU on the body of the vehicle, and the emulated IMU. The IMU emulation method obtains high-fidelity acceleration estimates on both simulated and experimental data sets. The maximum likelihood search method obtains sensor position estimates within 2 mm of the true sensor location in each direction and within 0.2 degrees of the true sensor orientation for a battery of tests in simulation. In experimental tests, this method estimated the true lateral and longitudinal sensor positions to within 3 cm, and the true sensor orientation to within 0.25 $\circ$ in each direction.