Vision based metric-topological localization for UGV
Date of Issue2017
School of Electrical and Electronic Engineering
Robotics Research Centre
This thesis studies vision based localization methods for unmanned ground vehicle (UGV) to achieve accurate and robust positioning in GPS challenging environments. Efforts are made from the perspective of topological and metric localization. Due to the incremental nature, visual odometry belongs to metric localization category. For a monocular visual odometry system, drift and scale ambiguity are the main issues that restrict it from extensive applications in autonomous navigation. In this thesis, a metric localization approach based on the fusion of visual odometry and road constraints is proposed. The drift and scale ambiguity of monocular visual odomtry are both considered as measurement uncertainties and incorporated into a presented Gaussian-Gaussian Cloud model. The geometric shapes of road networks are considered as constraints to assist with position estimation. Shape matching method is utilized to evaluate the alignment between historical trajectory from visual odometry and road shape from digital map. As a typical topological localization approach, place recognition is playing important roles in mobile vehicle navigation. Most of the current place recognition methods are designed for the application in a particular environment (e.g. indoor or urban environment). In this thesis, a place recognition method which is applicable to various environments is presented. A modified vocabulary tree with the ability of merging multiple kinds of features is designed to customize different combination of features for different environments. The downsides of pure place recognition and road-constrained metric localization are obvious. Place recognition approach suffers from its discontinuous output, while road constrained metric localization suffers from the on-road assumption as well as the tough initialization. To play their respective advantages, a metric-topological localization approach based on the integration of place recognition, visual odometry and road constraints is proposed. Topological and metric modules run in a parallel way and a mutual check scheme is utilized to ensure the consistency of the positioning results. When information sources from other sensors are available, a proper sensor fusion technique is required. To this end, a fusion approach is proposed to localize vehicles by integrating a visual odometry, a low-cost GPS, and a two-dimensional digital road map in this thesis. The concept of artificial potential field that is widely used for obstacle avoidance is leveraged to represent measurements and constraints, respectively. Position measurements from visual odometry and low-cost GPS are modelled with a potential well function, while road constraints from digital map are modelled with a potential trench function without additional map matching. By searching for the minimum of the combined potential field, the position can be estimated. All the approaches developed in this thesis are extensively and successfully verified on real world datasets.
DRNTU::Engineering::Electrical and electronic engineering