PODPose: Integrating Proper Orthogonal Decomposition and EITPose

TL;DR

The paper addresses robustness and practicality in EITPose by introducing two POD-based strategies: a sensitivity-volume–informed electrode-placement method that projects POD bases onto the mesh to select informative electrode locations, and a data-projection method that reconstructs full measurements from partial data via truncated POD coefficients. The electrode-placement approach yields minimal improvement over evenly spaced electrodes, while the data-projection method achieves accurate data reconstruction with missing channels, demonstrating a practical path for real-time, wearables-grade EIT systems. Together, these contributions enable more reliable, scalable EIT-based gesture sensing and open opportunities to deploy higher-resolution models on devices with fewer electrodes. The work has potential impact for consumer wearables and future EIT applications where electrode contact is variable or scarce.

Abstract

This work examines two ways of using proper orthogonal decomposition (POD) to enhance the prior work of EITPose, a device which uses electrical impedance tomography (EIT) to detect posture by way of a band of electrodes on the forearm. First, an electrode placement algorithm is described, which employs the sensitivity volume method and a POD basis to choose the combination of electrode locations that spans the POD basis most effectively. Next, a data placement algorithm is introduced, which uses a POD basis to account for electrodes that are providing poor data. Analysis is conducted on these two algorithms using the same techniques as the original EITPose paper, and it is shown that the electrode placement has little effect, but the data projection algorithm is very accurate when synthesizing data. The data projection algorithm represents a novel technique for adapting EIT devices live to poor electrodes, and can be applied to future implementations of the sensing technique.

Figures (4)

• Figure 1: The mesh projections of the first three POD bases. Each image has the same color bar range; note the first basis is smaller in magnitude. Each circle is a cross-section of the forearm, with the left of the circle lining up roughly with the ulna. Interestingly, the edges of the projections have a large magnitude, indicating that much of the difference in data comes from electrode contact.
• Figure 2: An image of the best electrode configuration according to the electrode placement method used in this paper. Each of the dark lines on the sides is an electrode.
• Figure 3: The MPJPE for each possible different hand poses for evenly spaced electrodes (a) and algorithm-determined electrode placement (b), showing the error for each pose. Blue is trained on 20 minute session and tested on 10 minute session, while orange is the opposite. The two graphs each look quite similar, meaning the two electrode placements have little difference in error.
• Figure 4: A histogram of each joint's error when tested on projected data. The distribution is relatively uniform, but the data projection algorithm is able to communicate some joints better than others.