IBLoc-UAV: Inferring On-Body Location in UAV-to-Ground Channels

Drones frequently communicate with devices on the ground, often carried in different ways by users. A user might hold their device near their chest while flying a drone, walk with their phone in their pocket, or walk facing away from the drone while it is tracking them. Despite the proliferation of drones and possibility of such scenarios, the prediction of user orientation and the user equipment (UE) location on or near the user's body in Drone-to-Ground (D2G) channels has not received adequate attention in literature. In scenarios where visibility is no longer available to the drone - due to adversarial attacks or harsh weather - the wireless signal can be used to detect the user's presence, their orientation, and even the location of the device on or near their body. This is the objective of this work. We study how the baseband I/Q samples, converted into spectrogram images that span a relatively short period of time, can be used to predict the on-body device location. We leverage Convolutional Neural Networks (CNNs) to classify three different use cases of holding a device operating at two different carrier frequencies (2.5 GHz and 900 MHz). Specifically, we investigate three on-body locations: near chest while facing the drone, in-pocket while facing the drone, and near chest but facing away from the drone. We show that, using only spectrogram images as input, we can predict these use cases with an average overall accuracy of 87% and 85% at 2.5 GHz and 900 MHz, respectively. We also investigate the classification performance on a dataset that belong to a hovering location that was not seen by the model, and show that the CNN model was able to correctly classify all images belonging to user orientation and 91% of the images that belong to near-chest vs in-pocket for the same orientation. Finally, we study the application of transfer learning and CNNs on classifying the on-body location at a different carrier frequency from the one on which they were trained, and show that while one use case can be correctly predicted, more complex models and hyper-parameter tuning is needed to achieve this goal. This work could be useful for building real-time deep learning models that help drones to make intelligent decisions and adapt to changes in user postures and on-body locations in air-to-ground (A2G) channels.