Localization and Posture Recognition via Magneto-Inductive and Relay-Aided Sensor Networks
Open access
Author
Date
2022Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
Body-centric sensor networks play a crucial role for future eHealth systems and are envisioned to constantly monitor vitals, to provide local in-body treatments and to warn the user about dangerous conditions. The magneto-inductive physical layer, which is based on magnetic near-field coupling, has been shown to be promising for the communication of such sensor networks, primarily due to the low interaction of the human tissue with low-frequency magnetic fields. It also comprises the possibility of purely passive relaying, which is a key enabler for communication applications. However, many body-centric applications do not only require a stable connectivity and high capacity between the sensors, but also a precise knowledge of the sensors’ locations and orientations. In this thesis we hence strive to extend the benefits of magneto-inductive body-centric networks to on-body and in-body localization. To this end, we first develop a theoretical framework, which is based on circuit theory and draws heavily on the Cramér-Rao lower bound. The scaling behavior of magneto- inductive localization with multiple anchors (observing infrastructure coil sensors) and agents (coil sensors that are to be localized) is analyzed and the impact of practical parameters such as the coil dimensions, the transmit power and the operating frequency is quantified.
It becomes evident that the distance dependency of the magnetic near field does not only lead to quickly decreasing channel gains, which is expected, but rather that it limits the position root-mean-square error by making it highly directional. It is this directional asymmetry, which often dominates the overall position root-mean-square error. This thesis proposes two different means to mitigate this asymmetry, improve the position root-mean-square error by orders of magnitude, and generally enhance magneto-inductive localization. The first approach uses purely passive relays (resonantly loaded coils), which provide additional signals paths from the agents to the anchors and may be exploited for the localization. The second approach introduces cooperation between all agents, which allows to share inter-agent channel state information. Both approaches are investigated by simulation and their individual drawbacks are extenuated by simple functionality extensions, such as a load-switching of the passive relays or a proper initialization for the highly-dimensional cooperative localization. Another contribution of this work is the derivation of an analytic closed-form expression for the maximum likelihood position estimator for a single pair of three-axis coils.
Based on the theoretical framework, a novel concept for a magneto-inductive human posture recognition system is proposed. This system relies on anchors that are centralized on the human torso and on purely passive coils that are placed on the extremities. Note that the approach is inherently low-power (purely passive relays) and low-complexity (low-frequency impedance measurements). The magnetic coupling between all coils, which depends on the body posture, leads to a detuning of the anchors’ input impedances. The relationship between anchor input impedances and postures can consequently be learned via supervised classifiers to enable posture recognition capabilities. For this concept neither the location of the anchor coils nor of the purely passive coils need to be known in advance.
With the goal in mind to design and implement an experimental system, we study and optimize this concept via simulation for different body types, coil designs, coil placements, and also for different types of noise. For realistic noise levels, these simulations yield a classification accuracy of more than 90 %, even when only considering single-frequency impedance measurements. Based on these results an experimental system is implemented with low-cost materials. An extensive measurement campaign in a real-world office environment is conducted. It confirms the simulations and an excellent classification accuracy is achieved. However, this classification accuracy degrades substantially when unaccounted posture variations and coil displacements disturb our testing data. To increase the robustness in this practical setting, it is proposed to measure the anchor impedance at multiple frequencies. With respect to computational complexity, different types of classifiers are compared. Moreover, we consider different feature spaces for the machine learning algorithms, which have practical implications for the hardware complexity. We find that the using the magnitudes of the impedances as features (instead of the complex impedances) already leads to a satisfactory performance while enjoying the benefit of reduced complexity. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000574826Publication status
publishedExternal links
Search print copy at ETH Library
Publisher
ETH ZurichSubject
Magnetic Induction; Wireless Sensor Networks (WSN); Wireless Communications; Wireless Positioning; Localization; RFID; posture recognition; Medical sensors; On-Body Sensors; Human centric wireless (HCW); Machine Learning; ClassificationOrganisational unit
03608 - Wittneben, Armin (emeritus) / Wittneben, Armin (emeritus)
More
Show all metadata
ETH Bibliography
yes
Altmetrics