Runz, Martin; (2022) Object-Aware Tracking and Mapping. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Runz__thesis.pdf - Other Download (68MB) | Preview

Abstract

Reasoning about geometric properties of digital cameras and optical physics enabled researchers to build methods that localise cameras in 3D space from a video stream, while – often simultaneously – constructing a model of the environment. Related techniques have evolved substantially since the 1980s, leading to increasingly accurate estimations. Traditionally, however, the quality of results is strongly affected by the presence of moving objects, incomplete data, or difficult surfaces – i.e. surfaces that are not Lambertian or lack texture. One insight of this work is that these problems can be addressed by going beyond geometrical and optical constraints, in favour of object level and semantic constraints. Incorporating specific types of prior knowledge in the inference process, such as motion or shape priors, leads to approaches with distinct advantages and disadvantages. After introducing relevant concepts in Chapter 1 and Chapter 2, methods for building object-centric maps in dynamic environments using motion priors are investigated in Chapter 5. Chapter 6 addresses the same problem as Chapter 5, but presents an approach which relies on semantic priors rather than motion cues. To fully exploit semantic information, Chapter 7 discusses the conditioning of shape representations on prior knowledge and the practical application to monocular, object-aware reconstruction systems.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item