Batch Continuous-Time Trajectory Estimation

Abstract

As the influence of autonomous mobile robots grows stronger on the lives of humans, so too does the importance of robust and accurate localization (and control). Although motion-estimation techniques utilizing passive cameras have been a core topic in robotic research for decades, we note that this technology is unable to produce reliable results in low-light conditions (which account for roughly half the day). For this reason, sensors that use active illumination, such as lidar, are an attractive alternative. However, techniques borrowed from the fields of photogrammetry and computer vision have long steered the robotics community towards a simultaneous localization and mapping (SLAM) formulation with a discrete-time trajectory model; this is not well suited for scanning-type sensors, such as lidar. In this thesis, we assert that a continuous-time model of the trajectory is a more natural and principled representation for robotic-state estimation. Practical robotic localization problems often involve finding the smooth trajectory of a mobile robot. Furthermore, we find that the continuous-time framework lends its abilities quite naturally to high-rate, unsynchronized, and scanning-type sensors. To this end, we propose novel continuous-time trajectory representations (both parametric, using weighted basis functions, and nonparametric, using Gaussian-processes) for robotic state estimation and demonstrate their use in a batch, continuous-time trajectory estimation framework. We also present a novel outlier rejection scheme that uses a constant-velocity model to account for motion distortion. The core algorithms are validated using data from a two-axis scanning lidar mounted on a robot, collected over a 1.1 kilometer traversal.