When the same motion looks different: Wireless channel ambiguities
Sensors are made to capture a quantity of interest. Cameras capture images; thermometers capture temperatures. These devices are meant to be faithful, and their readings directly reveal what we want to know. In Wi-Fi sensing, we do not design the sensor but instead repurpose a communication byproduct: wireless channel measurements.
The concept of Wi-Fi sensing is simple: channel measurements are affected by motion, and we can analyze them to infer things. Unfortunately, unlike camera images, wireless channels are a poor motion primitive. Minor changes in position, body orientation, or device placement can make the same motion look very different. Conversely, different motions may produce very similar channel patterns.
How the channel is formed
To understand that ambiguity, we need to look at how a channel is formed. As Wi-Fi signals move through a room, they reflect and scatter from walls, furniture, and people. These copies overlap at the receiver. Because each copy has traveled a different distance, each arrives with a different phase shift. The channel is therefore frequency-dependent and path-based, often written as
where is the strength of path and is its delay. Wi-Fi sensing therefore does not work with a clear motion primitive. It works with the sum of all propagation paths that motion happens to perturb.
From this expression, the sensitivity to context follows directly. Move a person by a few centimeters and several path lengths already change by a meaningful fraction of the wavelength. Rotate the body and the affected paths change as well. Move the transmitter or receiver and the full path geometry changes again. Therefore, even when the motion stays the same, its channel signature may not. Specifically, what we measure depends not only on the motion, but also on where it happens, how the body is oriented, and where the devices are placed.
How current work deals with it
Most current work treats this as a generalization problem: once the person changes position or orientation, or the devices move, the sensing system is effectively operating under different conditions than the ones it was trained for [1]. One response is to build a representation that is less tied to one specific geometry. Widar3.0 [2] is a representative example. It derives a body-coordinate velocity profile instead of working on raw CSI directly, with the goal of making gestures look more stable across users, positions, and environments. A second response is to generate or translate data across configurations. WiAG [3] does this for position and orientation by generating virtual samples for unseen configurations from data collected in one known configuration, while CrossSense [4] synthesizes training data for a new site from measurements collected elsewhere. A third response is to model the effect of geometry more explicitly. Work based on Fresnel-zone and CSI-ratio models for example explain why position, posture, orientation, and placement enter the channel so strongly in the first place [5].
Why this remains hard
Those three lines of work all help, but they also point to the same underlying fact: the instability is already present in the measurement. The channel changes with the person, with the room, and with the devices all at once. Thus, reproducibility is hard to get. And without reproducibility, inference becomes hard as well. That is why ambiguity sits so close to the center of wireless channel-based sensing.
References
[1] C. Chen, G. Zhou, and Y. Lin, “Cross-Domain WiFi Sensing with Channel State Information: A Survey,” ACM Computing Surveys, vol. 55, no. 11, 2023.
[2] Y. Zhang, Y. Zheng, K. Qian, G. Zhang, Y. Liu, C. Wu, and Z. Yang, “Widar3.0: Zero-Effort Cross-Domain Gesture Recognition with Wi-Fi,” IEEE Transactions on Pattern Analysis and Machine Intelligence.
[3] A. Virmani and M. Shahzad, “Position and Orientation Agnostic Gesture Recognition Using WiFi,” in Proceedings of the 15th Annual International Conference on Mobile Systems, Applications, and Services (MobiSys ’17), New York, NY, USA: Association for Computing Machinery, 2017, pp. 252-264.
[4] J. Zhang, Z. Tang, M. Li, D. Fang, P. Nurmi, and Z. Wang, “CrossSense: Towards Cross-Site and Large-Scale WiFi Sensing,” in Proceedings of MobiCom, 2018.
[5] D. Wu, Y. Zeng, F. Zhang, and D. Zhang, “WiFi CSI-based device-free sensing: from Fresnel zone model to CSI-ratio model,” CCF Transactions on Pervasive Computing and Interaction, vol. 4, no. 1, pp. 88-102, 2022.
An article by Fabian Portner