CBR-ACE: Counting Human Exercise using Wi-Fi Beamforming Reports

Sorachi Kato, Tomoki Murakami, Takuya Fujihashi, Takashi Watanabe, Shunsuke Saruwatari

Abstract

Repetitive exercise monitoring is a key application for indoor human sensing. While Wi-Fi-based sensing has shown promise, it typically relies on Channel State Information (CSI), which is difficult to extract from modern commercial off-the-shelf (COTS) devices. This paper presents CBR-ACE, a novel system that leverages Compressed Beamforming Reports (CBR)—a standard feature in 802.11ac/ax protocols—to count repetitions of exercises like push-ups and arm swings. By using a Raspberry Pi-based sniffer and a robust processing pipeline centered on Kernel Density Estimation (KDE), CBR-ACE bridges the gap between lab prototypes and real-world deployment. Experimental results demonstrate high accuracy across various devices, achieving a Mean Absolute Error (MAE) of 1.73 repetitions on average.

Introduction

Wireless sensing utilizes ambient Wi-Fi signals to detect human movement without requiring wearable sensors or cameras. Traditionally, researchers have relied on "Full CSI," which provides fine-grained data on how signal subcarriers are affected by the environment. However, accessing CSI usually requires specialized, legacy hardware (e.g., Intel 5300 NICs) and custom firmware, creating a significant "deployment gap" between laboratory research and practical application.

Alternative metrics like Received Signal Strength Indicator (RSSI) are widely available but often too coarse for counting subtle repetitive movements. CBR-ACE identifies Compressed Beamforming Reports (CBR) as a viable middle ground. CBRs are standard in modern Wi-Fi and contain phase information used for beamsteering, which fluctuates predictably with human motion.

Comparison of sensing methods showing the advantages of CBR in terms of accessibility and granularity.

System Design: CBR-ACE

The CBR-ACE architecture consists of two primary components: WiPiCap for data acquisition and ACE for repetition estimation.

WiPiCap: Data Acquisition

WiPiCap is a sniffer tool implemented on a Raspberry Pi 3 B+ using the Nexmon firmware patch. It passively intercepts CBR frames transmitted between an Access Point (AP) and a client device. These reports contain angles ($\phi$ and $\psi$) that describe the channel state. Although CBR is sampled at a lower frequency (10–20 Hz) compared to full CSI, it remains sufficient for capturing the 0.5–1.5 Hz frequency range typical of human exercise.

ACE: Activity Count Estimator

The ACE pipeline processes the raw CBR data through several stages:

  1. Filtering: An Exponential Moving Average (EMA) removes high-frequency hardware noise, followed by a bandpass filter to isolate exercise-related signals.
  2. Frequency Analysis: A Fast Fourier Transform (FFT) is applied to each subcarrier to identify the dominant repetition frequency.
  3. Aggregation via KDE: Instead of simple averaging, which is susceptible to outliers from multipath interference, ACE uses Kernel Density Estimation (KDE). This statistical method treats each subcarrier's estimate as a "vote" to find the most probable global repetition count.
The CBR-ACE system architecture, illustrating the flow from Wi-Fi packet capture to final activity count estimation.

Experimental Evaluation

The system was evaluated in a real-world office environment using a variety of COTS devices, including a MacBook Pro, an iPad, and a Samsung Galaxy Note 10. The dataset comprised 890 trials of activities such as arm swings, stepping in place, and torso twists.

Key Results

  • High Accuracy: CBR-ACE achieved an average MAE of 1.73. Performance was best on the MacBook Pro, with an MAE of 0.97.
  • Robustness of KDE: The researchers found that traditional subcarrier selection methods based on variance were ineffective for CBR data. The KDE-based aggregation proved essential, significantly outperforming simple averaging (which had an error of 8.6).
  • Device Agnosticism: The system successfully counted repetitions across devices from different manufacturers (Apple and Samsung), confirming the universality of the CBR-based approach.

Conclusion and Future Work

CBR-ACE demonstrates that standard Wi-Fi beamforming reports are a powerful and accessible source for human activity sensing. While the system currently assumes a relatively constant exercise rhythm, future work will explore deep learning techniques to handle varying speeds and to classify specific activity types. This research paves the way for ubiquitous, device-free health monitoring using existing home Wi-Fi infrastructure.

Summary updated

BibTeX
@article{kato2022cbr,
  author = {Sorachi Kato and Tomoki Murakami and Takuya Fujihashi and Takashi Watanabe and Shunsuke Saruwatari},
  title = {{CBR-ACE: Counting Human Exercise using Wi-Fi Beamforming Reports}},
  journal = {J. Inf. Process.},
  year = {2022},
  doi = {10.2197/ipsjjip.30.66},
}
Figure