Master thesis: Aerial View Goal Localization with Reinforcement Learning

RISE RESEARCH INSTITUTES OF SWEDEN
November 20, 2022
Contact:N/A
Offerd Salary:Negotiation
Location:N/A
Working address:N/A
Contract Type:Other
Working Time:Negotigation
Working type:N/A
Job Ref.:N/A
Master thesis: Aerial View Goal Localization with Reinforcement Learning

Master thesis at RISE Research Institutes of Sweden (spring 2023): Aerial View Goal Localization with Reinforcement Learning – Towards Emulating Search- and-Rescue

Climate-induced disasters are and will continue to be on the rise, and thus search-and-rescue (SAR) operations, where the task is to localize and assist one or several people who are missing, become increasingly relevant. In many cases the rough location may be known and a UAV can be deployed to explore a confined area to precisely localize the people. Due to time and battery constraints it is often critical that localization is performed as efficiently as possible.

In a previous mater thesis (currently under review at an AI-for-climate workshop; see preprint https: // arxiv.org/abs/2209.03694), this type of problem was abstracted in a framework that emulates a SAR-like setup without requiring access to actual UAVs. In this framework, an agent operates on top of an aerial image (proxy for a search area) and must localize a goal that is described through visual cues. To tackle this task, a reinforcement learning (RL)-based model, AiRLoc, was proposed. Experiments showed that AiRLoc outperforms various baselines and humans, and that it generalizes across datasets, e.g. to disaster-hit areas without seeing a single disaster scenario in training.

A lot of work however remains to make AiRLoc a viable approach for real-world use cases, and this master thesis will explore some of these next steps. Examples of relevant directions to study:

Automatic stopping criterion. AiRLoc should itself understand when it has reach its goal location (currently, an episode simply terminates as soon as the agent reaches a goal, independently of whether AiRLoc has understood that it has reached it or not).

More flexible visual goal specifications. The goals are currently specified as top-view patches within the same search area that is to be explored by AiRLoc, and the patches are also from the same point in time. This is highly unrealistic in real-world scenarios – one would expect goals to be specified at different circumstances and even from different perspectives (e.g. ground-level images instead of top-view ones).

Larger search areas and/or and/or continuous movements and/or movement across multiple elevations. The current setup considers mainly 5 x 5 or 7 x 7 grid-like, discrete setups (note however that humans found it difficult to solve it even for 5 x 5), and the agent ‘hovers' at a constant elevation. It would be interesting to explore how to train models that operate well on larger grids (or, ideally, get rid of the grid setup altogether and consider continuous actions), and/or that can inspect the search area at various elevations.

Who are you? The work requires skilled students within computer vision and machine learning. Ideally a thesis would result in a research submission to some conference / workshop at the intersection of AI and climate change. We expect you to have the following skills and background:

• Experience of implementing image analysis / machine learning models. • Taken courses in machine learning and/or image analysis and/or statistics. Strong grades is meriting. • Programming skills, preferably with some experience of relevant frameworks such as PyTorch or Tensorflow. Coding will likely be done exclusively in Python, as most code bases that are used in research are written in that language. PyTorch will be used extensively, as the code base from the previous thesis will be used as starting point. • Experience in writing scientific papers / text is meriting.

Applying in a pair with someone else is encouraged – in such cases, all the above mentioned prerequisites may not be required per person, as long as the respective skill- sets are complementary and together cover most of the prerequisites.

Welcome with your application! Last day of application is 7th of November, 2022. If you have any questions, please contact Aleksis Pirinen, PhD in computer vision, machine learning researcher at RISE, aleksis.pirinen@ri.se. Office space may be provided in the Lund office (Ideon).

About RISE RISE is Sweden's research institute and innovation partner. Through our international collaboration programmes with industry, academia and the public sector, we ensure the competitiveness of the Swedish business community on an international level and contribute to a sustainable society. Our 2,800 employees engage in and support all types of innovation processes. RISE is an independent, State-owned research institute, which offers unique expertise and over 100 testbeds and demonstration environments for future-proof technologies, products and services.

RISE Center for Applied AI Research connects AI research within RISE Research Institutes of Sweden. We are around 60 researchers working on machine learning related tasks within different fields including natural language processing, computer vision and network analysis.

The Deep Learning Research Group is connected to RISE Center for Applied AI Research working on modern AI and machine learning. We have solid expertise in the field of deep learning, computer vision, federated learning, uncertainty quantification, and privacy-preserving machine learning.

Om jobbet Ort

Lund

Anställningsform

Visstidsanställning 3-6 månader

Job type

Student - examensarbete/praktik

Kontaktperson

Aleksis Pirinen +46102284004

Referensnummer

2022/577

Sista ansökningsdag

2022-11-07

Skicka in din ansökan

From this employer

Recent blogs

Recent news