Improving Long-Term Localisation with Exponential Decay and Neural Feature Filtering

  • Alexandros Rotsidis

Student thesis: Doctoral ThesisDoctor of Engineering (EngD)


The work presented in this thesis is a joint work between Dc-activ, an industrial partner and the University of Bath. The industrial end goal is to develop an augmented reality platform for retail shops where virtual intelligent avatars can interact with customers. The virtual avatars must be able to act as shopping assistants and training staff, which entails these avatars must be intelligent. The platform must be capable of dealing with the dynamic nature of retail shops in order to provide a pleasant augmented reality for its users. In this work, I first present a number of industrial application prototypes that showcase the feasibility and potential of deploying intelligent avatars in a retail environment. Lastly, I present the two research contributions that improve the efficacy of such a platform.

Determining the camera pose, i.e. localising, is an important stage for a complete augmented reality experience. Previous research has put effort into localising a camera in static environ- ments. The real world though is a dynamic place with moving objects, such as cars, people and lighting changes caused by seasons or different times of the day. Estimating a camera pose in a dynamic environment, such as the world around us, is called long-term localisation. Long-term localisation is a problem more relevant and appropriate to the dynamic nature of our world. Recently the problem of long-term localisation has been one of the new interest areas to receive a lot of attention.

In this thesis, I first present several intelligent virtual avatar prototypes that prove the feasibility of deploying augmented reality intelligent avatars in a retail shop.

Then, I first address the task of long-term localisation in a more specific indoor environment, retail shops. To tackle this problem, I introduce a method based on exponential decay that improves long-term localisation in a retail shop. The proposed method is additionally tested on outdoor environments as well where it also shows to improve long-term localisation.

The last task that this thesis addresses is improving the localisation speed while aiming to the performance intact. I propose a learning-based method that reduces the amount of data needed to estimate a camera pose. This is achieved by using a neural network trained to classify and predict static points that are more useful for localising. The method can be easily integrated into existing pipelines.
Date of Award28 Jun 2023
Original languageEnglish
Awarding Institution
  • University of Bath
SupervisorWenbin Li (Supervisor), Christof Lutteroth (Supervisor) & Christian Richardt (Supervisor)

Cite this