Robust and Efficient Smoothing for Underwater Navigation

Publication Type:
Issue Date:
Full metadata record
Computationally efficient and robust data fusion algorithms are valuable in navigation (or localisation) applications using low-cost sensors. This thesis considers the problem of enhancing the robustness and efficiency for 6-DoF (Degree of Freedom) underwater IMU (inertial measurement unit)-vision based navigation (or localisation). The emphasis is placed on algorithms that are robust and operate efficiently using low-cost inertial-visual sensors in underwater environment where it is vulnerable to outlier measurements. Such capability is desirable for autonomous underwater navigation. One major factor that degrades the navigation accuracy are outlier measurements. In particular, inertial-visual underwater navigation is susceptible to wrong observation measurements. As a result, online and constant time robust state estimation techniques are valuable to provide the smoothed and enhanced vehicle trajectory. Existing solutions have mainly focused on increasing the robustness of the (extended) Kalman filter ((E)KF). They often require tuning the motion and prediction model noise covariance matrices that are fairly involved. The contributions of this thesis arise from proposing a robust Biswas-Mahalanobis Fixed lag smoother (BMFLS) by utilising EKF, and a robust sliding window filter (RSWF) using the nonlinear least-squares (NLS) optimisation approach. The robust-BMFLS solution performs outlier rejection using the Chi-square test through reclassification and iterative smoothing. The limitation is it requires more iteration in time intervals with high ratio of outliers. While, in the NLS optimisation approach works by assigning a weight to each observation, that are iteratively computed from the robot pose prediction error and observation error, outliers are detected and rejected by classification expectation-maximisation. However, solving the NLS optimisation using full-batch estimation is an offline process. By introducing the RSWF, a constant time and online solution is presented. This is an incremental and online robust solution which is computationally efficient to robust full-batch estimation. The impact of different optimisation window sizes and update periods are studied on the navigation performance. This is useful to determine the optimum window size and update period using RSWF for localisation.
Please use this identifier to cite or link to this item: