Graph-Based Relative Path Estimation Using Landmarks for Long Distance Ground Vehicle Following

Abstract

This dissertation presents a graph-based sensor fusion framework for the localization of automated ground vehicles for leader-follower path duplication. A specific focus is given to scenarios where the following distance between vehicles is high. Localization accuracy is critical for any automated path following task. While high localization accuracy can be achieved using GPS corrections or a priori maps, these resources are not available in certain scenarios such as remote areas with limited infrastructure and a priori information. In this dissertation, a novel method is proposed for solving this localization problem that does not depend on built infrastructure or a priori information. A graph-based framework is used to estimate the path taken by the lead vehicle using relative measurements of differential GPS between vehicles, odometry from onboard sensor measurements, and exchanged landmark observations. The graph-based estimation framework developed in this dissertation allows for ad-hoc, nonlinear measurements for estimating a near-optimal path solution. Each sensor input provides complementary benefits: differential GPS allows for centimeter-level accuracy, vehicle odometry allows for spanning GPS outages and gaps in landmarks, and landmark observations bound path errors with respect to following distance. The findings of an observability analysis are presented to show failure conditions for certain vehicle and landmark configurations. A simulation study is developed and used to demonstrate the success of the proposed method and characterize the estimator’s performance in a number of controlled scenarios. The findings from the simulation study are validated with experimental data from a pair of Class 8 tractor-trailers, each equipped with a GPS receiver, a multi-channel lidar, wheel encoders, an inertial measurement unit, and a Dedicated Short Range Communications (DSRC) radio. An overview is presented for each method used to generate the measurements used in the graph-based estimator. This includes an overview and error characterization of Time Differenced Carrier Phase (TDCP) and Dynamic-Base RTK (DRTK) used for precise odometry and inter-vehicle relative positions, respectively. An odometry model is provided for determining vehicle motion from wheel speed and yaw rate sensors. Additionally, two unique approaches are presented for detecting road sign and pole-like objects from 3D point cloud data for use as landmark observations. Results show that the presented method improves performance when compared to existing methods in terms of both accuracy and availability. Compared to many existing map-matching approaches, the presented approach requires a relatively small number of landmarks (~11 per km) to achieve the accuracy target regardless of GPS availability. Under nominal conditions, the presented method is successful in meeting the estimation accuracy required for path control with average lateral path position errors of 0.94 cm and path orientation errors of 0.14 degrees. It is also shown that the path errors remain bounded with respect to following distance when a sufficient number of landmarks are present, allowing for large gaps between vehicles.