In this theme, we will advance the frontiers of SLAM and 3D-reconstruction through the creation of efficient and robust methods for obtaining high-quality 3D virtual models of complex physical objects and environments from heterogeneous sensing modalities. Few existing SLAM solutions can perform real-time mapping and at the same time inform replanning algorithms. Existing techniques running on board robotic platforms are either restricted to small areas, produce low-resolution maps, or perform poorly in complex, changing environments. Most of the existing onboard mapping solutions inform local planning and navigation rather than inspection and survey.
Our research will focus on four key areas:
Computationally efficient mapping algorithms with real-time feedback on the quality of the estimation. In survey and monitoring applications, where robotic deployment is expensive, real-time feedback on the quality of the estimation is highly relevant in the mapping process and allows for local replanning of the mission to maximize data collection. Efficient SLAM and 3D-reconstruction techniques exist, but none is able to compute the uncertainty of the estimation in real-time, especially when object detection and semantic segmentation are used for high-level map abstractions. We will develop novel methods to enable robust 3D-reconstruction of dynamic objects and to improve mapping where there are structural and appearance changes in the scene.
Long-term change modelling and spatio-temporal representation. How to model change? How to embed a series of changes/events into the map? How to detect trends? How to support model query? How to best represent such data? We will research novel ways to represent how the environment can change over time and explore methods on to integrate such changes into a digital twin of an asset.
Change detection, relevant change selection. The challenge here is to distinguish between relevant and irrelevant changes, which often requires expert knowledge and is impossible to hand-code. Existing, predominantly pixel-based change detection techniques are highly sensitive to image registration errors and noise. To overcome these limitations, we will develop learning-based and probabilistic techniques to identify changes. We will also investigate how accurate modelling of uncertainty in spatiotemporal mapping and map registration processes can assist with change detection.
Dynamic scene representation. Dynamic scene graphs are successfully used in creating digital, multi-level spatiotemporal representations of indoor (e.g. office) environments, to facilitate communication, planning and decision-making. We will extend this type of representation to more complex environments (e.g. underwater structures, bridges), where we will use it to represent not only dynamic objects but also structural and appearance changes.