Learning-Based Posture Detection Using Purely Passive Magneto-Inductive Tags


Henry Schulten, Florian Wernli, and Armin Wittneben


IEEE Global Communications Conference (Globecom), Madrid, Spain , pp. 7, 2021-12-07.

[BibTeX, LaTeX, and HTML Reference]


We propose a novel low-cost and low-complexity approach that uses the magnetic near-field detuning caused by passive coils in order to detect and classify movements. This idea can enable ubiquitous health monitoring applications on or within the body. In this work, we employ our idea with the specific goal of detecting human body postures with only a few measuring nodes. More precisely, we place resonantly loaded passive coils (tags) on a person’s limbs and investigate how the sole presence of these coils affects the input impedances of secondary coils (anchors) which are located on the torso. Due to the strong magnetic near-field coupling of all coils, each different posture changes the inter-coil links and hence leads to a unique set of impedance measurements. These impedance patterns can conversely be learned to use them for classification purposes. In order to verify this idea, we simulate a joint-based human body model and use numerical approximations for the magnetic-near field coupling to generate realistic noisy data sets of impedance measurements. Based on these data sets, we train support vector machines, which we subsequently use to classify the postures on a separate testing set. We investigate how robust this procedure is against increasing noise levels and variations of the postures. In this process, we identify appropriate operating points that lead to an average classification accuracy of 94 percent.


Near Field, Posture Detection, SVM, WBAN

This publication is currently not available for download.