Cross-Frequency Classification of Indoor Activities with DNN Transfer Learning

Remote, non-contact recognition of human motion and activities is central to health monitoring in assisted living facilities, but current systems face the problems of training compatibility, minimal training data sets and a lack of interoperability between radar sensors at different frequencies. This paper represents a first work to consider the efficacy of deep neural networks (DNNs) and transfer learning to bridge the gap in phenomenology that results when multiple types of radars simultaneously observe human activity. Six different human activities are recorded indoors simultaneously with 5.8 GHz and 25 GHz radars. Firstly, the bottleneck feature performance of the DNNs show that a baseline of 76% is achieved. On models trained only with 25 GHz data when 5.8 GHz data is used for testing 81% accuracy is achieved. in absence of a large dataset for radar at a certain frequency, we demonstrate information from a different frequency radar is better suited for generating the classification models than optical images and by using time-velocity diagrams (TVD), a degree of interoperability can be achieved.