SLAM with adaptive noise tuning for the marine environment

This paper presents an alternative formulation for the Factorised Solution to the Simultaneous Localization and Mapping (FastSLAM) algorithm using an Adaptive Extended Kalman Filter based approach. The FastSLAM algorithm jointly estimates the pose of the robot and the location of landmarks on the feature based map by factorising the SLAM posterior into a localisation component which is implemented using a particle filter, and a mapping component implemented via independent Kalman Filters. To facilitate non-divergent state estimates, the process model noise statistics of the robot must be modelled correctly, and maintain validity in all the conditions encountered by the vehicle. This is infeasible in cases where the vehicle's model may degrade over time, or in applications involving complex, noisy, highly nonlinear environments such as marine environments. This paper proposes an algorithm to recursively estimate the state and the motion model noise covariance simultaneously, using a moving window of previous estimates. The algorithm proposed is then tested on an Autonomous Surface Craft (ASC) in a marine environment, and the results obtained are compared to current state of the art algorithms. Results from the experiments show promising performance for the proposed SLAM framework, especially in highly noisy environments with nonlinear process models.