An Adaptive Gaussian Particle Filter based Simultaneous Localization and Mapping with dynamic process model noise bias compensation

Abstract

Simultaneous Localization and Mapping (SLAM) is a fundamental component of all autonomous robotics systems, which probabilisticaly fuses information from an exteroceptive sensor and a proprioceptive sensor to simultaneously estimate the robot's trajectory and the map. Inputs from the pro-prioceptive sensor are fed into the estimation algorithm via a process model corresponding with the vehicle kinematics, while a measurement model is used to process inputs from the exteroceptive sensor. Most SLAM algorithms assume known, fixed model estimate bias. This assumption does not hold true for systems with wrongly modeled estimate bias, or those affected by component fatigue due to applications requiring long term autonomy. This paper will display the adverse effects of mismodeled process model bias using a simulation. An adaptive algorithm employing Adaptive Gaussian Particle Filter based process model bias compensation will be deployed in tandem with a particle filter based FastSLAM algorithm. The algorithm will be compared favourably with existing state of the art SLAM algorithms in controlled simulations. Experimental data from a marine environment will be used to validate the efficacy of the algorithm.