In this PhD defense talk a novel localization and mapping solution following a bottom-up approach is presented. The methodology starts with an Attitude and Heading Reference System(AHRS), continues with a 3D odometry dead reckoning solution and builds up to a full graph optimization scheme which uses visual odometry and takes into account rover traction performance, bringing scalability to modern SLAM solutions.
A design procedure is presented in order to incorporate inertial sensors into the AHRS. The procedure follows three steps: error characterization, model derivation and filter design. A complete kinematics model of the rover locomotion system is developed in order to improve the wheel odometry solution. Consequently, the parametric model predicts delta poses by solving a system of equations with weighed least squares. In addition, an odometry error model is learned using Gaussian processes (GPs) in order to predict non-systematic errors induced by poor traction of the rover with the terrain. The odometry error model complements the parametric solution by adding an estimation of the error. The information gained by the non-parametric GP model serves to adapt the localization and mapping solution to the current navigation demands (domain adaptation). The adaptivity strategy is designed to adjust the visual odometry computational load (active perception) and to influence the optimization back-end by including highly informative keyframes in the graph (adaptive information gain). Following this strategy the solution is adapted to the navigation demands, providing an adaptive SLAM system driven by the navigation performance and conditions of the interaction with the terrain. The proposed methodology is experimentally verified on a representative planetary rover under realistic field test scenarios. This thesis introduces a modern SLAM system which adapts the estimated pose and map to the predicted error.
The system maintains accuracy with fewer nodes, taking the best of both wheel and visual methods in a consistent graph-based smoothing approach.
