IMU Drift and Sensor Fusion: Why Your Data Wanders and How It's Fixed

Twenty minutes into a velocity-tracking session, the numbers stop making sense. The bar path readout shows the barbell sitting several centimetres higher than it did at the start, even though nothing about the setup has changed. The mean velocity on an easy warm-up set reads faster than a near-max single from earlier in the session. Coaches who work with inertial sensors run into this constantly, and the first instinct is usually to blame the app or swap the device. The real cause is almost always drift — a predictable, well-documented artifact of how inertial measurement units generate position and velocity data, and one that good sensor fusion design bounds rather than eliminates.

What IMU Drift Actually Is

An inertial measurement unit combines a three-axis accelerometer and a three-axis gyroscope — many sports units add a magnetometer — to estimate orientation, velocity, and position. None of these quantities is measured directly. Acceleration is measured; velocity is obtained by integrating acceleration over time; position is obtained by integrating velocity over time again. Integration is where the trouble starts. Every MEMS accelerometer and gyroscope carries a small bias error that shifts with temperature, supply voltage, and time since power-on, typically on the order of a few milli-g for consumer-grade accelerometers and a fraction of a degree per second for gyroscopes.

A bias that small sounds irrelevant until it gets integrated. A gyroscope bias of even 0.3–0.5°/s, left uncorrected, accumulates into tens of degrees of heading error within a minute of continuous integration. Because velocity is the single integral of acceleration and position is the double integral, a small constant accelerometer bias produces a velocity error that grows roughly linearly with time and a position error that grows roughly with the square of time (Woodman, 2007). That's the mechanism behind a tracker that reads accurately for the first three reps of a set and is visibly off by rep twelve: the error isn't noise averaging toward zero, it's a bias compounding on every sample.

Why the Error Compounds Instead of Averaging Out

Oliver Woodman's 2007 University of Cambridge technical report on strapdown inertial navigation is still the reference most sports-tech engineers point to when explaining why raw IMU output can't behave like GPS. In his illustrative calculation, a mid-grade MEMS accelerometer with roughly a 1 milli-g residual bias, integrated twice with no correction applied, produces a position error on the order of tens of metres within about a minute — enough that the paper's central conclusion is that an uncorrected low-cost inertial system is essentially unusable for standalone position tracking beyond a few seconds. Barbell and sprint trackers aren't running 60-second open-loop navigation, but the same growth curve applies to the shorter windows that matter in training: a residual bias too small to notice on a single rep is large enough to visibly skew displacement and velocity numbers by the end of a five-set block if nothing resets it.

Gyroscope drift makes the problem worse because orientation error changes how acceleration gets resolved into the correct reference frame. If the device's internal estimate of bar tilt is off by even a couple of degrees, part of the gravity vector gets misread as horizontal acceleration, and that misread quantity then gets integrated along with everything else. This is why single-sensor, gyro-only trackers tend to drift fastest during lifts with large orientation changes — a snatch or clean pull punishes uncorrected orientation error far more than a back squat, where the bar's orientation barely changes.

Sensor Fusion: How the Drift Gets Cancelled Out

Sensor fusion combines the gyroscope's short-term accuracy with a second, drift-free reference — the accelerometer's gravity vector, and on 9-axis units, the magnetometer's heading reference — so that the low-frequency error building up in one sensor gets corrected by information from the other. Gyroscopes are excellent over sub-second windows but drift over minutes; accelerometers are noisy from one sample to the next but their long-term average, gravity, never drifts. Every practical fusion algorithm is built around exploiting that trade-off rather than trusting either sensor alone.

Madgwick, Harrison, and Vaidyanathan (2011) introduced the gradient-descent orientation filter that carries Madgwick's name, now widely used in sport and rehabilitation IMUs. Validated against an optical motion-capture reference in the original paper, it held static orientation error under roughly 1° and stayed in the low single digits of degrees during dynamic movement, while running at a fraction of the computational cost of a full extended Kalman filter — a meaningful difference for a chip that has to run fusion math in real time on a coin-cell battery strapped to a barbell rather than on a lab workstation. None of these approaches make drift disappear. What they do is convert an error that would otherwise grow without bound into one that oscillates around a small steady-state value, provided the device gets a clean reference to correct against on a regular basis.

A Field Calibration Protocol That Actually Controls Drift

Most of the drift complaints we hear from coaches trace back to skipped or rushed calibration, not a faulty sensor. The sequence below takes under two minutes total across a session and materially changes how much error accumulates by the last set.

1. Static calibration, 15–20 seconds: before the first working set, set the device motionless on a flat, level surface — the rack shelf or the floor, not your hand. This gives the fusion filter a clean gravity reference to anchor to before any movement noise enters the picture.
2. Temperature settle, 2–3 minutes: in cold garages or unheated facilities, power the device on and let it sit for a couple of minutes before the first set. MEMS bias shifts measurably with temperature, and a sensor pulled straight from a cold gym bag hasn't reached thermal equilibrium yet.
3. Zero-velocity update between sets: rest the device motionless on the bar for 3–5 seconds during the inter-set rest, rather than picking the bar up immediately. This feeds the filter a known-zero velocity reference it can use to cancel accumulated integration error before the next set starts.
4. Re-anchor every 15–20 minutes: on longer sessions, build in a deliberate 10-second static pause roughly every 15–20 minutes, and always after any hard mechanical shock such as dropping the bar or slamming it into the rack.
5. Mount rigidly: a loose strap-on mount introduces a small relative rotation between the sensor and the bar that the filter reads as genuine orientation change. Snug the strap enough that the sensor can't be twisted by hand once it's on.

Camomilla, Bergamini, Fantozzi, and Vannozzi's 2018 systematic review of wearable inertial sensors in sport, published in Sensors, looked across a large body of in-field validation studies and flagged inconsistent or absent calibration protocols as one of the recurring factors limiting agreement between IMU-based measurements and gold-standard reference systems. The practical takeaway from that literature lines up with what shows up in the gym: a device isn't unreliable by nature, it's unreliable when it never gets a clean reference to recalibrate against.

Common Mistakes That Make Drift Worse

A handful of everyday habits reintroduce drift even on a device with good fusion software running underneath.

Loose mounting. A velcro strap that felt snug at set one is often noticeably looser by set four as the fabric stretches under repeated tension. That small amount of play lets the sensor rock independently of the bar, and the filter has no way to distinguish that rocking from real orientation change. Re-check strap tension every few sets rather than assuming the first fit holds.

Magnetic interference. Mounting a 9-axis sensor close to a steel rack upright, or wearing a wrist-based unit near stacked steel plates, can skew the magnetometer's heading reference noticeably — enough to matter for orientation-sensitive lifts. Most sport IMUs let you disable the magnetometer and fall back to gyro-plus-accelerometer fusion, which is often the more reliable default indoors around steel equipment anyway.

Cold-start bias. A sensor pulled from a car trunk or an unheated gym bag in winter and used immediately will show more drift in the first several minutes than the same sensor after it's warmed up, because MEMS bias is temperature-dependent and hasn't stabilized yet. This is the single most common cause behind the tracker-was-fine-yesterday-but-garbage-today complaint in outdoor or unheated facilities.

Fast, unfamiliar movement patterns. Fusion filters tuned for steady lifting cadences can lag briefly during very fast rotational movements — a striking combination in combat sports, or a golf swing — showing a short spike in apparent orientation error before the filter reconverges. That reconvergence window is normal filter behaviour, not a defect, and it's part of why sport-specific IMU firmware tunes its gain settings differently for cyclic lifting versus ballistic, high-angular-velocity movements.