Why More Sensors Can Make Sensing Worse
Motivation
WiFi can see you. Walk past a router and the signal changes, so a radio can tell that someone is there without a camera, in the dark, and through a wall. We have nine small radios in a room at the lab, and between them run 72 invisible tripwires, every one of which twitches when a person crosses it.
Problem statement
You would expect 72 tripwires to beat 3. They do not. Using all 72, the system guesses right 64% of the time, while using only the best 3 it gets 79.5%, because only a handful of tripwires actually cross the person and the rest just add noise on top. We cheated to get that number though, since we chose the best 3 already knowing where the person was standing. A real system has no idea where anyone is, as that is the whole thing it is trying to work out. Choosing the right tripwires needs the answer, and the answer needs the right tripwires.
The project
You will break that circle. The system has to decide where to pay attention by itself, sampling broadly, guessing, concentrating, and changing its mind when the person walks somewhere else. The radios, the firmware, and 312 hours of recordings already exist, so you start on the interesting part instead of soldering, and you will see it work in a live view of the room showing which tripwires the system is listening to. Both ends of the scoreboard are known, since 64% is what you have to beat and 79.5% is the best anyone could do. No wireless background is needed, the maths is a few lines of Python, and the hardware is provided. The specific goal of this project is to find out how much of that 79.5% a system reaches while picking its own tripwires, and how fast it reacts when the person moves.
Objectives
- Understand how WiFi senses people, and why one link only covers part of a room (the Fresnel zone).
- Reproduce the result on recorded data, comparing all 72 links against the best 3.
- Implement a strategy that picks links at runtime, without knowing where the person is.
- Measure what it reaches against the 64% floor and the 79.5% ceiling.
- Show how fast it adapts when the person moves, and where it fails.
Previous theses in this line
References
[1] Sigg, S. (2025). Feeling the Room: A Practical Study on WiFi-based Occupancy Detection. Bachelor’s thesis, University of St.Gallen. https://sensing-group.com/files/theses/ba-simon-sigg.pdf
[2] Miao, F., et al. (2025). Wi-Fi sensing techniques for human activity recognition. ACM Computing Surveys, 57(5), 1-30.
[3] Hernandez, S. M., & Bulut, E. (2022). WiFi sensing on the edge. IEEE Communications Surveys & Tutorials, 25(1), 46-76.
[4] Wang, H., Zhang, D., Ma, J., et al. (2016). Human respiration detection with commodity WiFi devices: do user location and body orientation matter? Proc. ACM UbiComp 2016, 25-36.
[5] Krause, A., Singh, A., & Guestrin, C. (2008). Near-optimal sensor placements in Gaussian processes: theory, efficient algorithms and empirical studies. Journal of Machine Learning Research, 9, 235-284.
[6] Auer, P., Cesa-Bianchi, N., & Fischer, P. (2002). Finite-time analysis of the multiarmed bandit problem. Machine Learning, 47, 235-256.
Interested? Email the advisors: Bruno Rodrigues, Karim Khamaisi.