MSc thesis project proposal

[2020-21] Good data for free with GANs: too good to be true?

Radar is investigated for the classification of human activities and movements, as knowing what activities someone performs, when, where, and how often, can be a powerful indicator of his/her health status, both in terms of physical mobility and cognitive wellbeing.
As for any classification problem, recent techniques developed by the deep learning community are also of great interest to the radar researchers to process the radar data as images (using convolutional networks) or as time series of samples (using some form of recurrent networks).
However, a fundamental issue is that radar data are intrinsically different from optical images, video, audio, or language data. Radar data encode in their pixels and/or samples kinematic information about the targets, their position, their velocity, their bearing. Hence, simply “recycling” networks and algorithms designed for different types of data may not be the winning strategy. This project and its companion project investigate two interesting outstanding questions out of the many!

Project 1-> One of the limitations of using radar data and deep learning for classification is that there are not enough data to train deep and complex neural networks for classification. For example, if we think of human gestures, it is hard to collect experimentally the tens of thousands of samples of different gestures from different people that would be required.
Here is where techniques for data augmentation on radar data, for example transfer learning approaches and Generative Adversarial Networks (GANs) can be useful. GANs can be trained to generate synthetic data starting from a small number of experimental data, but are these new data any good? What is the best format of the radar data for this augmentation? And What is the best structure and training approach to have these GANs working well? These are some of the research questions to tackle in this project, where we also seek to collaborate with Dr M. Hoogendoorn at VU University Amsterdam, where they investigate GANs for wearables’ data.

Requirements

Radar basics, MATLAB, Python (very desirable), basis of pattern recognition & AI

Contact

dr. Francesco Fioranelli

Microwave Sensing, Signals and Systems Group

Department of Microelectronics

Last modified: 2021-01-19