Can barometric pressure and motion signals synchronize wearable recordings offline?

Clocks drift apart when devices cannot talk

Combining data from several wearable devices requires knowing which sample on one device corresponds to which moment on another. Each device timestamps its measurements with its own real-time clock, and those clocks are built on crystal oscillators. The paper notes typical accuracy of 10–100 ppm, which means two clocks that matched at the start can deviate by several seconds after hours of runtime. Repeated synchronization is therefore needed before recordings can be analyzed jointly.

The usual fixes carry costs that matter for long, standalone collections. GPS can provide a precise time reference, but it is unreliable without a clear view of the sky and draws milliwatts of power. Wireless sensor-network protocols such as TPSN, FTSP and RBS are optimized for accuracy rather than very low power. BMAR takes a different route: it treats the recorded signals themselves as the synchronization evidence, so devices never need to communicate while recording.

Air pressure provides the coarse shared reference

BMAR's first stage uses barometric pressure, which the paper describes as inexpensive, low-power and commonly included in wearables. Devices worn by the same person share the surrounding air pressure, so their pressure traces rise and fall together even when the person is not moving. BMAR compares two pressure traces across all possible alignments and picks the lag that best matches them, using either a Huber-loss difference or a minimum standard deviation between the traces.

This stage does two jobs. It verifies that two recordings were actually acquired at the same time, rejecting traces whose mean sample-by-sample pressure difference exceeds a threshold set to the sensor's 100 Pa absolute tolerance. It also produces a coarse pre-alignment. That pre-alignment is deliberately treated as a constant lag, because its error is generally larger than the clock skew between the two devices. The coarse step exists to avoid locking onto a local rather than global alignment before the finer stage runs.

Motion refines the lag and tracks clock skew

The second stage uses accelerometer magnitudes to sharpen the alignment. BMAR splits the trace into windows, computes a lag for each window by maximizing cross-correlation, and fits a line through those window lags with ordinary least squares, removing outliers via the Bonferroni test. The fitted line is a time-dependent lag, which is how the method accounts for clock skew that accumulates over a recording rather than staying constant.

The number of windows is not fixed. BMAR starts at four windows and doubles the count until the window falls below one second or the count exceeds 512, then selects the approximation with the best adjusted r-squared. The paper's reasoning is that local misalignments within some windows will not look linear across the whole trace, so fit quality signals alignment quality. Refinement is bounded to plus or minus five seconds around the pre-alignment; a larger pre-alignment error cannot be fully corrected, and widening the range raises the risk of misalignment.

Active sessions land within tens of milliseconds

The evaluation used custom coin-sized devices measuring 23 mm by 22 mm, worn at the forehead, sternum, wrist or above the ankle. Recordings covered sleep, office work, cycling and skiing, with three to five devices at once and runtimes up to about ten hours. Ground truth came from a wired apparatus that sent a button-press trigger to every device at the start and end of each session; across 349 trigger sequences, a single trigger showed a mean error of 1.2 ms, and the evaluation averaged three consecutive triggers.

For active sessions, where the wearer moved, median alignment error fell as recordings lengthened: 0.442 s at 5 minutes, 0.074 s at 10 minutes, 0.044 s at 15 minutes, 0.041 s at 30 minutes, 0.028 s at 1 hour and 0.034 s at 6 hours or more. The headline 33.4 ms median across all synchronized recordings sits in that range. The paper also reports that BMAR reliably rejected non-overlapping traces, which is a separate claim from alignment accuracy: rejection concerns whether two recordings overlapped at all, not how precisely they were aligned.

Sleeping exposes the method's dependence on motion

Passive sessions isolate what happens when the accelerometer has nothing to work with. During sleep the experimenter stayed stationary, and the paper treats this as a worst-case baseline. Here the second stage could not exploit motion for refinement, so only the pressure-based pre-alignment remained. Median error was above 60 s at 5, 10 and 15 minutes, 20.2 s at 30 minutes, 5.42 s at 1 hour and 3.06 s at 6 hours or more.

The contrast is the paper's most decision-relevant comparison. The same method, on the same hardware, moves from tens of milliseconds to seconds depending on whether the wearer moved. Longer passive recordings help because the pressure trace has more time to constrain the alignment, but they do not restore the refinement step. The authors also state a limitation that follows directly: the accelerometer refinement depends on signal characteristics, no worst-case guarantee is offered, and if recordings contain no events or asynchronous patterns, refinement may worsen a pre-alignment result.

Recording length and wearer motion decide the fit

The practical split is between collections with body motion and collections without it. For activity monitoring, health tracking or motion capture where the wearer moves, the reported errors of roughly 28–44 ms at 30 minutes and beyond are the relevant figures. For sleep or other stationary monitoring, the relevant figures are the passive-session numbers, and a 3.06 s median at six or more hours is the realistic expectation rather than the 33.4 ms headline.

Two further conditions bound the claim. BMAR is designed for devices worn by a single person; the paper suggests pressure-based alignment could extend to multiple people or non-wearable devices if they share a location, but presents that as an expectation rather than a tested result. And because the method runs on sensors wearables already sample, it adds no runtime power cost and does not require devices to share a wireless network. The priority for a reader is therefore not the single median error but which session type their recordings resemble.

Sources used in this answer