Advancing SLAM with multi-features, multi-sensor fusion and deep learning algorithms for indoor environments

Abstract

The increasing integration of service robots into indoor environments, such as offices and public facilities, has led to the development of low-intelligence robots that operate in controlled settings, such as food delivery, cleaning, and freight handling. However, the increasing demand for higher accuracy and robustness is driving advancements in Simultaneous Localisation and Mapping (SLAM) systems.

This thesis introduces a series of innovations in SLAM tailored for indoor environments. The proposed methodologies address key challenges, including multi-feature integration, sensor fusion, dynamic environment adaptation, and semantic mapping.

Thesis is embargoed until 30th June 2027

First, a novel enhanced SLAM framework integrates multiple geometric features, such as points and planes, leveraging their inherent relationships to improve trajectory accuracy and mapping quality in structured indoor settings.

Based on the first finding, a sensor fusion SLAM framework is developed. It integrates a monocular camera and Light Detecting and Ranging (LiDAR) sensor to address scale ambiguity in monocular SLAM. The approach extracts complementary features and extends plane-based methods with line features, enhancing robustness in challenging indoor environments. An extended study further refines LiDAR-based plane detection using an image segmentation network, improving feature quality and accuracy for effective indoor and outdoor operation.

Following the idea of using a deep learning algorithm, the final research incorporates an objection detecting network for semantic mapping and dynamic filtering. A novel dynamic filtering method is developed, and a semantically labelled octomap is generated.

The proposed methodologies are extensively evaluated using open-source datasets and sequences from Ulster University. The results demonstrate significant advancements over reference algorithms. In terms of robustness, the methods address tracking loss in the challenge dataset and achieve the most stable performance across various dynamic sequences. For scale recovery, the approach eliminates reliance on dense depth data sources without compromising tracking accuracy. Regarding tracking accuracy, the proposed methods achieve the lowest errors across most sequences.

Thesis is embargoed until 30th June 2027

Date of Award	Jun 2025
Original language	English
Supervisor	Jose Santos (Supervisor), Haiying Wang (Supervisor) & Huiru (Jane) Zheng (Supervisor)