Localization and mapping are fundamental tasks in robotics. Many localization and mapping algorithms are based on Bayesian filtering technique. However, optimal Bayesian filter is computationally intractable due to the requirement to maintain the full conditional probability density function (PDF). Generally, this requires an infinite number of parameters. Approximation methods such as the sum of Gaussian and particle filter were commonly used to address the above problem. Two limitations of these methods are that they require positive definiteness of the Gaussian PDF, and difficulty in ensuring the numerical symmetry of the covariance matrix. We propose Minimal Iterated Bayesian Estimator (MIBE) – a novel, minimal representation estimator that effectively approximate the PDF with positive semi-definite property. Two application areas of MIBE will be presented.
First, MIBE is applied to improve the triangulation of 3D scene points in visual SLAM. Other challenges of monocular visual SLAM are in tracking the image feature positions accurately and determining the translational scale. Most state-of-the-art methods track sparse features, assumed to have homogeneous, isotropic Gaussian error. They also rely on known motion model of the camera and trained ground segmentation to improve tracking accuracy. These make their methods not applicable in estimating general camera motion. We propose new Bayesian Dense Flow algorithm along with Mahalanobis eight-point algorithm to robustly estimate the camera pose undergoing general 6 degrees of freedom motion.
Second, MIBE is applied to highly non-linear radio-based localization problems. Most existing radio-based time-difference-of-arrival (TDOA) and frequency-difference-of-arrival (FDOA) localization methods use batch processing, where all input measurements are jointly optimised. We propose a new efficient measurement-wise recursive TDOA and FDOA localization algorithm incorporating MIBE. Our algorithm also recursively identifies and discards outlier measurements.
Simulations and experiments on real data for visual SLAM and radio-based localization demonstrate the improved performance.
Yonhon Ng was born in Penang, Malaysia. He completed his Bachelor of Engineering (with First Class Honours and University Medal) in mechatronic systems from the Australian National University. He is currently a PhD student in the Research School of Engineering, Australian National University.
His current research interests include Bayesian estimator, visual odometry, simultaneous localization and mapping (SLAM), optical flow and radio-based localization.