Place recognition based on a spinning radar

Background

Place recognition, detecting when a sensor has revisited a location, is a central task in robotics localization. Spinning radar has recently become compact and accurate to be used for localization. The sensor is resilient to dust and operates well in all weather, and has great potential to enable highly robust robotic perception systems. Some work has been published targeting spinning radar for place recognition [1-2].

Motivation and scope

It is generally hard to understand the uncertainty of predictions from state-of-the-art methods for place recognition. Deep convolution networks are learned to be overconfident of their predictions. Hence, there is a need to introduce uncertainty of predictions. This can possibly be done using Bayesian Neural Networks.

Specific tasks

The task is to reimplement [2] and extend the method to include prediction uncertainty. Optionally, instead of predicting uncertainty, this method for place recognition can be integrated with MCL (Monte Carlo Localization) for localizing on prebuilt maps on radar data.

Benefits

You will work on machine learning that achieves state-of-the-art results for place recognition, a core task in robotics. Practical experience on machine learning is a much desired competence within the industry. 

Necessary skills

Good programming skills, Machine learning, Python

References

[1] https://arxiv.org/pdf/2109.13494.pdf

[2] https://arxiv.org/abs/2110.02744

Contact

Daniel Adolfsson, daniel.adolfsson@oru.se