Physics-Aware Deep Learning for RF-Based Gait Analysis and Sign Language Recognition

Atlanta, Georgia, United States, Virtual:

. In recent years, advances in machine learning, parallelization and the speed of graphics processing units (GPUs), combined with the availability of open, easily accessible implementations, have brought deep neural networks (DNNs) to the forefront of research in many fields. Likewise, deep learning has offered significant performance gains in the classification of radar micro-Doppler signatures, paving the way for new civilian applications of RF technologies that require a greater ability to recognize a larger number of classes that are similar in nature. However, RF data has a fundamentally different phenomenology from other application areas of deep learning, such as computer vision or natural language processing. The signal received from radar is a complex time-series, whose amplitude and phase are related to the physics of electromagnetic scattering and kinematics of target motion. Radar data oftentimes undergoes a number of filtering and analysis stages before being presented to a DNN as two- or three-dimensional data. Thus, DNN approaches, popular in other fields, do not necessarily translate to the RF domain and their benefits must be separately reassessed. A related challenge, which further necessitates the exploitation of a physics-based, knowledge-aided signal processing approach to DNN design and training, is the small amount of measured data that is available in RF applications. This talk will discuss the unique aspects of application of deep learning to radar micro-Doppler classification, focusing on the problem of DNN training under low sample support. Generative Adversarial Networks (GANs) have achieved great success in synthetic data generation; however, we show how in the case of radar micro-Doppler signatures, misleading signatures with poor kinematic fidelity can also be generated and degrade classification performance. Methods for improving the quality of synthetic signatures generated through adversarial learning are discussed. Results and benefits for two case studies are considered: biomedical gait analysis and recognition of American Sign Language. Linguistic and kinematic insights into the application of deep learning as well as ways in which machine learning and radar signal processing can illuminate linguistics study via RF sensing are presented. Speaker(s): Dr. Gurbuz, Atlanta, Georgia, United States, Virtual: