Deep mm-Wave Gait Analysis: Challenges and Opportunities

Deep learning algorithms have shown exceptional performance in processing of high-dimensional complex datasets especially in computer vision and natural language processing tasks. This has motivated several recent works leveraging the power of deep neural networks to unlock the potential of rich but unintuitive signals generated by mm-wave sensors. And yet, some of these studies are either limited to very controlled settings that do not easily generalize to real-world applications, or suffer from small sample sizes that inevitably result in biased conclusions with limitations on scalability. In this talk we first investigate common challenges in applying deep learning methods on data from unconventional sensors (eg non-natural images or speech) and pinpoint main barriers in data-driven analysis of mm-wave signals. We then demonstrate how one could overcome such challenges in training and inference of deep neural networks in the context of a real-world mm-wave gait dataset. Finally we discuss open problems and potential research avenues towards generalizable deep mm-wave gait analysis.