RF-Sensing: A New Way to Observe Surroundings

Radio frequency sensing (RF-sensing) is an emerging field assisting vision technology for object detection, tracking, and various such use cases. RF-Sensing technology uses radio signals and their reflections in order to capture the surrounding details and then applies artificial intelligence algorithms to sense the objects. RF-sensing has advantage over the existing vision technology when it comes to low light conditions, privacy concerns, and far-range scenarios. In this paper, we propose different RF-sensing methods for person identification (PI), human activity identification (HAI), and surrounding-location identification (SLI). Here, we make use of data generated from millimeter-Wave sensors and pre-process it using signal processing techniques before feeding it to deep learning (DL) based classifiers. For PI and SLI, we propose a DL network consisting of convolutional neural network (CNN) and long short term memory (LSTM) blocks. Here, first the input is passed through the CNN layers and then followed by the LSTM. For HAI, we propose a hierarchical classifier consisting of two stages. First, similar activities are clubbed together and the coarse level classifier is trained to distinguish among the non-similar activities. This is followed by fine level classifiers, which are trained to distinguish between those activities which are deemed to be similar while training the coarse level classifier. While testing, we employ the coarse level classifier unconditionally. The fine level classifier is employed only when the coarse level classifier’s output class corresponds to that of similar activity. We use the output of the hierarchical classifier to determine the activity. Further, we show that the proposed algorithms for the PI and SLI are able to predict accurately among 5 persons and among 5 different surrounding-locations respectively and the proposed hierarchical classifier for HAI, which is used to classify among 11 different activities, shows an improvement of
$\approx 4 \%$
compared to other DL based architectures in the literature.