DIRSAC: a directed sample and consensus algorithm for localization with quasi-degenerate data

Abstract

Autonomous robotic platforms are gaining interest in the scientific and military communities as well as in many industrial and commercial applications. For tractable generic autonomous operation, an autonomous system must maintain an accurate estimate of position and orientation (pose) within a local environment, commonly referred to as ego-motion estimation. Most state of the art robotic systems have some range sensing capability which can be used to perform obstacle detection and/or avoidance, do model inspection and generation, or perform system navigation. Over the last decade the robotics and automation community has started to take advantage of this underutilized sensing modality to not only perform mapping, but also localization by matching scan data during motion. Because real world sensing is noisy and models are imperfect, this typically works best if wrapped in some type of robust estimator capable of rejecting outliers such as RANSAC. While RANSAC has been shown to be a powerful and ubiquitous tool, it often suffers in practice from overwhelmingly redundant and non-constraining data. However, we can improve the robustness and convergence of the common RANSAC algorithm by replacing the purely random search with a more directed version especially when the data is highly degenerate. In this thesis, we derive, implement and evaluate an improved algorithm for point selection that analyzes the effect of each point on reducing the sensor's pose covariance. This is done by developing a simple Jacobian relating the sensor's measurements to the sensor's pose. Based on this Jacobian, we can translate uncertainties in the sensor's measurements to uncertainties of the pose and actively remove redundancies in the data. We identify redundant points by computing the Mutual Information between points and analyzing the relative information content overlap with previously selected points. We do this while keeping the flavor of the random sampling from RANSAC which maintains robustness to outlier contamination. We demonstrate increased performance with more reliable convergence by comparing our modified approach to that of common RANSAC in several quasi-degenerate cases. We also compare our approach to a similar approach which operates on quasi-degenerate data in a slightly different context of a 3D homography. We have also developed a metric by which we can score the level of degeneracy of the collected data from the perspective of constraining the pose solution.