Vision Aided Inertial Navigation

Abstract

Camera measurements can be used to replace or supplement GPS measurements for aiding inertial navigation systems. This is especially useful in cases where GPS is unavailable or unreliable, whether it is because the signals are blocked by the environment or intentionally corrupted. In this thesis a system is developed for aiding inertial navigation by a simple vision system consisting of a single camera, and computer simulations are done to test its performance. The system is represented as a state space model which is used in an extended Kalman filter to estimate the navigation states.

The system model is created by relating the acceleration and angular velocity to the rate of change of the orientation, velocity and position of a vehicle and to the relative position of a stationary landmark, which is represented by a pinhole camera projection. Noise values used for simulating inertial sensor noise are calculated based on specifications obtained from a user manual. The state space model is based on considering the inertial measurements as control inputs, directly affecting the states, whle the camera and GPS measurements are outputs of the system. An extended Kalman filter is then implemented, which uses measured inputs and outputs to estimate the state based on a linearized model.

Simulations are done which show that the system works as intended and is able to restrain the drift in position from the inertial navigation. Monte Carlo simulations show that the best results are achieved when the vehicle swings from side to side as it approaches the landmark, however the filter has problems estimating the standard deviations correctly. Suggested further work includes using more landmarks and replacing the extended Kalman filter with an unscented filter.