Hierarchical Visual Topological Mapping for Mobile Robots

Abstract

Building a visual representation of an environment is challenging since different factors such as variable lighting conditions, different viewpoints, mobility of robots, dynamic and featureless appearances can all affect the representation. This thesis proposes a novel method for appearance-based visual topological mapping using low-resolution omnidirectional images. Successive images captured as the mobile robot traverses its environment are compared (using a pixel by pixel strategy) to previous images to select some as nodes of a graph that constitutes the topological map. We propose a sampling method that automatically adapts to the evolution of the environment as it is perceived by the robot navigating it. We propose two strategies to control the sampling rate (and compare them to a fixed rate). The edges of the map initially correspond to the succession of images as they are captured. These edges are complemented by loop closure detections as the map is being built. Loop closures are detected using a pixel by pixel comparison in a strategy that adapts to the local appearance of the environment. The method first uses a global search from the existing map for loop closures. This search affects the time complexity of the loop closing process as it becomes worse as the size of the map increases. For efficient searching, a hierarchical structure is built on top of the map by recursively grouping nodes with similar visual properties into an average node representing their group. The process creates a forest of trees of such averaged nodes. This hierarchy is built in parallel with the actual map, and searches performed for loop closures use the hierarchy to speed up the process. The use of the hierarchy for searching improves the time complexity and scalability for localising new nodes in the existing map. The work is therefore a solution to the SLAM process that uses the appearance of the environment and produces a visual topological map. To evaluate the proposed methods, we have collected multiple datasets captured under changing weather conditions along various paths. Each image in the dataset is stamped with its GPS location as ground truth for measuring the nodes catchment area and visualisation only. To evaluate the map’s quality, we propose two novel metrics called catchment area coverage and relative length of the map. We compare our method to state of the art methods (F AB-MAP 2.0, BINM ap and HTM ap) using code provided by the authors of these works ran on our datasets. Our method performs better under the mentioned condition against the other methods.

Date of Award	2022
Original language	English
Awarding Institution	Aberystwyth University
Supervisor	Myra Wilson (Supervisor) & Fred Labrosse (Supervisor)