Research note 03
Why Sensor Placement Decides Your Data Quality
Departments spend a great deal of energy comparing sensor specifications and almost none on how the sensor is attached. The attachment is where most of the accuracy is lost, and it is the part you can actually fix.
A conversation we have often: a club is evaluating two telemetry systems and wants to discuss sampling rates, gyroscope ranges and filter design. Those things matter. They matter considerably less than whether the pod is held firmly against the athlete's upper back in the same orientation every session, which is the variable nobody puts in a tender document.
This note is about the mechanical interface between sensor and skeleton, why it dominates the error budget, and what a club can do about it.
Soft-tissue artefact
An inertial sensor measures its own motion. What you want to know is the motion of the underlying bone. Between the two sit skin, subcutaneous fat, muscle and the fabric of the harness, all of which deform under load. The resulting discrepancy is soft-tissue artefact, and in the biomechanics literature it is a long-standing and well-quantified problem for skin-mounted markers and sensors alike.
The artefact is worst exactly where the signal is most interesting: during high-magnitude, high-frequency events like foot-strike, landing and rapid deceleration. At those moments the tissue is loaded and the sensor is riding a moving surface rather than a rigid one. The consequence is a systematic distortion of peak acceleration — usually an over-estimate of magnitude with a low-pass-filtered shape — that no amount of downstream processing can undo, because the information was never captured.
The practical implication
Peak acceleration magnitudes from a body-worn sensor are not directly comparable to force-plate or accelerometer-on-bone values, and they are not comparable across athletes with different body composition at the mounting site. They are comparable to the same athlete's own values with the same harness fit. Design your questions accordingly.
Why the scapular site
We mount between the scapulae, over the upper thoracic spine, for three reasons. The tissue depth there is comparatively low and consistent across a wide range of body compositions. The site is close to the trunk's centre of mass in the anteroposterior direction, which simplifies interpretation of whole-body acceleration. And it is a location athletes tolerate: far enough from the neck to be comfortable, out of the way of contact in most sports, and not in a position where it interferes with equipment.
Alternatives exist and have their own trade-offs. A sacral mount is closer to the true centre of mass and better for some gait applications, but it is far less comfortable in contact sport and considerably harder to keep still under a waistband. Limb-mounted sensors give you limb-specific information at the cost of much larger soft-tissue artefact and much greater sensitivity to placement.
Harness tension is a measurement parameter
Treat harness tension as an instrument setting, not a comfort preference. Too loose and the pod translates relative to the trunk under acceleration, adding artefact that mimics real signal. Too tight and you restrict breathing, change the athlete's mechanics and guarantee the athlete will loosen it themselves before the next session.
The target is a fit that holds the pod pocket flat against the skin over the thoracic spine with no visible movement during a vigorous trunk shake, while allowing full inspiration. Our fitting protocol is a two-minute check: size selection, a standing shake test, a breathing check, then a five-step jog with a hand on the pocket. Run it once per athlete at pre-season and re-check after any significant body composition change.
Orientation and calibration
A pod rotated relative to its intended orientation contaminates every axis-specific metric. A 15° error in the transverse plane redistributes signal between the anteroposterior and mediolateral channels; a sagittal-plane error does the same between vertical and anteroposterior.
Some of this is recoverable. The magnetometer plus a short standing calibration — five seconds still, upright, arms at sides — lets the pipeline estimate a static mounting rotation and apply a correction matrix. What is not recoverable is a rotation that changes during the session, which is why the fit test matters more than the calibration routine.
Magnetometer-based heading has its own failure mode worth knowing about: ferrous structures and strong electrical fields distort local magnetic field. Indoor halls with steel roof structures, and the immediate vicinity of large electrical installations, both produce heading errors. Our pipeline detects magnetic disturbance and falls back to gyroscope-integrated heading with drift correction from the positional stream, flagging the affected period rather than silently degrading.
Things that quietly break your dataset
- Pods swapped between athletes without reassignment. Each pod carries individual factory calibration coefficients, and an unrecorded swap attributes one athlete's data to another. Our dock assignment step exists to prevent this and is the step most often skipped.
- A washed harness that has lost elasticity. Elastic degrades. A two-season-old harness does not hold the same tension as a new one, and nobody notices until the data does.
- Mixed harness types across a squad. If half your athletes wear the vest and half wear a strap arrangement, you have two measurement systems and one dataset.
- Pods worn over a base layer by some athletes and against skin by others. An extra fabric layer changes the mechanical coupling. Pick one convention and hold it.
- Undocumented mid-season equipment changes. Any change to harness, pod model or mounting convention should appear in the session record so that later comparisons know to stop at the boundary.
What we do about it in the product
The platform records harness size, fit-check date and pod assignment per athlete, and surfaces a warning when a fit check is more than a season old. Sessions where the mounting-rotation estimate differs materially from the athlete's established value are flagged for review rather than processed silently. And metrics that are known to be fit-sensitive — peak accelerations above all — carry that sensitivity in their documentation, so an analyst reading a spike knows the first question to ask.
None of this is glamorous, and it does not appear in a specification comparison. It is, in our experience, where most of the real difference between a trustworthy dataset and a suspect one is made. The related consequences for asymmetry work are covered in Detecting Sprint Asymmetry From IMU Data.
