Deep Learning-Based Visual Localisation and Navigation for the Voyager Mobile Robotic Platform

Autonomous ground vehicles (AGVs) are becoming increasingly popular in industrial and research environments and are employed to perform repetitive or hazardous tasks without direct human intervention. One major challenge in autonomous navigation lies in reliably estimating a robot’s position and orientation, known as localisation. This challenge is particularly notable in indoor environments where satellite-based positioning systems, such as Global Positioning System (GPS), are unavailable. This thesis presents a deep learning-based visual localisation method and navigation stack implementation for the Voyager mobile robotic platform, which aims to solve a variation of the kidnapped robot problem. A convolutional neural network (CNN), specifically a fine-tuned ResNet-50 model, is trained for visual place recognition using images captured from the robot’s environment. The model extracts discriminative feature embeddings which are stored in a FAISS index alongside corresponding position and orientation data obtained during a LiDAR-based (Light Detection and Ranging) mapping phase. During localisation, the robot captures a live image, from which an embedding is extracted using the fine-tuned ResNet-50 model and queried with stored index to estimate the most likely pose of the mobile robot. This pose serves as an initial pose estimate for the popular Real-Time Appearance-Based Mapping (RTAB-Map) Simultaneous Localisation and Mapping (SLAM) system, which then refines the localisation estimate using LiDAR measurements and provides continuous, accurate pose estimates for navigation. The localisation system is integrated with Robot Operating System 2 (ROS2) and its native navigation stack, Navigation 2 (Nav2), enabling the robot to plan and execute collision-free paths within the known environment. Development and verification was performed both in the Gazebo simulation environment and on the physical Voyager platform. Experimental results show that the proposed system successfully provides accurate initial pose estimates, enabling transition to RTAB-Map-based localisation and subsequent autonomous navigation, while avoiding unseen obstacles. The findings confirm that deep learning-based visual place recognition can effectively comple-ment traditional SLAM methods, providing global localisation robustness in GPS-denied indoor environments.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Daneels, Alexander Luke

Advisor dc:contributor.advisor

• Engelbrecht, J. A. A.

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/135711

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/135711