Cleaning Wi-Fi CSI data: filtering, calibration and baselines

If your first CSI capture looked like static, you are not doing anything wrong. Raw CSI is noisy by nature, because it reacts to every tiny change in the radio channel plus a stack of hardware quirks. Turning it into something a detector or a model can use is a fixed sequence of steps, and once you know the order it is routine. Here is that order.

01Why cleaning matters

Raw CSI is a blend of several things, and only one of them is useful. There is the real channel, which carries the movement you want. On top of it sit empty subcarriers that carry nothing, sharp outliers from radio glitches, phase that is scrambled by clock errors, and slow drift as the channel wanders over minutes. Feed that blend straight into a threshold or a neural network and the noise buries the signal.

Cleaning is about stripping away everything that is not information and stabilizing what remains, so that a genuine event, a person entering or a chest rising, stands clearly above a quiet background. The steps below are the ones almost every practical CSI project uses, in roughly this order. You do not always need all of them, but this is the full menu. For the raw capture that produces the input, see how to capture Wi-Fi CSI on an ESP32.

02Drop the null and pilot subcarriers

The first step is the cheapest and the most important. The channel contains subcarriers that carry no payload: the guard bands at the edges and the DC bin at the center are empty, and the pilot subcarriers carry fixed references rather than channel information. Their CSI values are meaningless for sensing, so keep the data subcarriers and discard the rest.

On an ESP32 20 MHz link, that usually means trimming a handful of bins from each end of the reported vector and ignoring the near-zero center. Exactly which indices to drop depends on the format and bandwidth, so inspect a quiet capture first and remove the subcarriers that sit flat near zero. The structure behind this is explained in CSI subcarriers and OFDM explained.

03A Hampel filter for outliers

Radios produce occasional sharp spikes, single samples that are wildly wrong. A plain moving average smears those spikes into their neighbors instead of removing them, so the standard tool here is a Hampel filter, a median-based outlier remover.

For each sample it looks at a small window around it, takes the median and the median absolute deviation of that window, and if the sample lies more than a few deviations from the median it is replaced by the median. Because it works on medians it is robust: one bad value cannot drag the estimate, and genuine motion, which is smooth and sustained, passes through untouched while isolated glitches vanish.

04Smoothing the result

After outliers are gone, the signal still has high-frequency jitter that no real body motion can explain. Human movement is slow in radio terms: walking, gestures and breathing all live below a few hertz. So a gentle low-pass filter or a short moving average over time removes the jitter above that band while keeping the movements you care about.

Keep the window modest. Too much smoothing and you blur the very events you are hunting; too little and the jitter survives. A window spanning a fraction of a second is a sensible starting point, which you then tune against real captures. For breathing specifically you can band-pass around the 0.1 to 0.5 Hz range to isolate the rhythm.

05Normalize, then learn a baseline

Two calibration steps make the signal comparable over time. First, amplitude normalization puts subcarriers and captures on a common scale, for example by dividing each subcarrier by its own mean or standard deviation, so that a naturally strong subcarrier does not dominate a naturally weak one.

Second, and most important for presence work, learn an empty-room baseline. Record CSI for a short period while the space is unoccupied, average it per subcarrier, and store that as the reference. From then on you compare live CSI against the baseline: subtract it, or divide by it, so that only departures from the empty room, meaning activity, remain. Without a baseline you are guessing what "quiet" looks like; with one, the quiet room is defined and anything above it is a candidate event. The honest limits of this approach are covered in our complete guide to Wi-Fi CSI.

06Phase sanitization, if you use phase

If your project only uses amplitude, you can skip this step. If you use phase, it is mandatory, because raw phase is dominated by carrier frequency offset, sampling frequency offset and a random per-packet offset. Sanitization unwraps the phase and then removes the linear component, the slope and offset the clock errors add, typically by fitting a straight line across the subcarriers and subtracting it. What survives is close to the true channel phase and safe to use for fine motion, distance or breathing. The reasoning is laid out in amplitude vs phase in Wi-Fi CSI.

07The pipeline in code

Here is the amplitude path as a short NumPy script: drop the edge subcarriers, remove outliers with a Hampel filter per subcarrier, learn a baseline from the first unoccupied seconds, and subtract it. It is deliberately small so you can read every line.

import numpy as np

def hampel(x, k=5, n_sigma=3.0):
    # Replace spikes in a 1D series using a sliding median window.
    y = x.copy()
    scale = 1.4826                       # MAD -> std for Gaussian noise
    for i in range(len(x)):
        lo, hi = max(0, i - k), min(len(x), i + k + 1)
        med = np.median(x[lo:hi])
        mad = scale * np.median(np.abs(x[lo:hi] - med))
        if mad > 0 and abs(x[i] - med) > n_sigma * mad:
            y[i] = med                     # outlier -> local median
    return y

# amp: CSI amplitude, shape [packets, subcarriers]
amp = amp[:, 6:-5]                        # 1) drop null/guard subcarriers
amp = np.apply_along_axis(hampel, 0, amp)  # 2) outliers, per subcarrier over time
amp = amp / amp.std(axis=0, keepdims=True)  # 3) normalize each subcarrier

baseline = amp[:200].mean(axis=0)       # 4) learn the empty room (first packets)
clean = amp - baseline                   # 5) keep only departures from empty

motion = np.abs(clean).mean(axis=1)      # one motion value per packet
print("motion range:", motion.min(), motion.max())

Add smoothing with a short moving average over the motion series, and for phase insert a sanitization step before step 3. For public captures to practice on, our roundup of Wi-Fi CSI datasets and tools lists sources you can load straight into a script like this one.

08FAQ

SGP Lab — written by the team that designs and builds SGP hardware in Europe. We test every example on our own boards before publishing.