Optimal Image-Aided Inertial Navigation

The utilization of cameras in integrated navigation systems is among the most recent scientific research and high-tech industry development. The research is motivated by the requirement of calibrating off-the-shelf cameras and the fusion of imaging and inertial sensors in poor GNSS environments. The three major contributions of this dissertation are The development of a structureless camera auto-calibration and system calibration algorithm for a GNSS, IMU and stereo camera system. The auto-calibration bundle adjustment utilizes the scale restraint equation, which is free of object coordinates. The number of parameters to be estimated is significantly reduced in comparison with the ones in a self-calibrating bundle adjustment based on the collinearity equations. Therefore, the proposed method is computationally more efficient. The development of a loosely-coupled visual odometry aided inertial navigation algorithm. The fusion of the two sensors is usually performed using a Kalman filter. The pose changes are pairwise time-correlated, i.e. the measurement noise vector at the current epoch is only correlated with the one from the previous epoch. Time-correlated errors are usually modelled by a shaping filter. The shaping filter developed in this dissertation uses Cholesky factors as coefficients derived from the variance and covariance matrices of the measurement noise vectors. Test results with showed that the proposed algorithm performs better than the existing ones and provides more realistic covariance estimates. The development of a tightly-coupled stereo multi-frame aided inertial navigation algorithm for reducing position and orientation drifts. Usually, the image aiding based on the visual odometry uses the tracked features only from a pair of the consecutive image frames. The proposed method integrates the features tracked from multiple overlapped image frames for reducing the position and orientation drifts. The measurement equation is derived from SLAM measurement equation system where the landmark positions in SLAM are algebraically by time-differencing. However, the derived measurements are time-correlated. Through a sequential de-correlation, the Kalman filter measurement update can be performed sequentially and optimally. The main advantages of the proposed algorithm are the reduction of computational requirements when compared to SLAM and a seamless integration into an existing GNSS aided-IMU system