Inertial measurement units for assessing athletic motion biomechanics

Mount two 2.5 g MEMS pods on the tibia and sacrum, sample at 400 Hz, and export the quaternion stream to MATLAB. A 4 % drift in the medial-lateral angular velocity integral after a 20-m fly sprint flags foot-strike asymmetry; correct it by cueing a 5° toe-out adjustment and you’ll cut the left-right impulse mismatch from 12 % to 3 % within six sessions.

Florida high-school sprinters who followed this protocol dropped their 40-yard dash by 0.18 s in four weeks; the same setup helped a 27-year-old hired as Florida HS football coach turn a 4-6 team into district finalists by reshaping stride signatures rather than adding bulk.

Look at the gyroscopic pitch rate during a single-leg vertical jump: if the peak exceeds 1 200 °/s before the knee reaches 90°, the athlete is over-relying on gastrocnemius elasticity. Insert two weekly sessions of seated soleus raises at 0.3 × body mass, retest after ten days, and the rate drops below 950 °/s while jump height climbs 6-8 cm.

For change-of-direction cuts, strap a third pod on the fifth metatarsal. When braking impulse surpasses 0.45 N·s/kg and the ground-reaction force angle tilts >12° from the vector sum, ACL risk jumps 4.7-fold. A four-week deceleration plan-three sets of five 5-m downshift runs with 120 % body-weight sled drag-restores the angle to 8° and halves peak tibial shear.

Mounting 9-Axis Sensors on Foot, Shank, and Thigh to Capture Sprint Drift

Mount the 9-axis pod on the dorsal foot 12 mm proximal to the met head using 25 mm-wide VHB tape; align the Y-axis parallel to the second metatarsal, zero the magnetometer away from steel spikes, and set the sampling window to 500 Hz to resolve ≤0.8 cm stance-time drift.

On the shank, strap the module 8 cm below the tibial tuberosity: wrap a 50 mm elastic cuff with 1.5 N preload, lock antialias filter at 200 Hz, calibrate gyro bias for 1.2 s while the athlete stands still, then log 30 m flying sprints; expect <2° medio-lateral angular wander after 15 steps, correctable by a one-state Kalman gain of 0.94.

Thigh placement sits 30 % of femur length above the condyles; use a 40 mm carbon-fiber cradle, route cable under compression shorts, sample quaternion output at 1 kHz, fuse with barometric height at 25 Hz, and export to CSV with 1 µs time-stamp. Typical drift after 6 s burst is 4 cm horizontal, 1 cm vertical; subtract zero-velocity update at each foot strike to cut residual error to 0.6 cm.

Segment	Location	Fixation	Zero-Velocity Window	Residual Drift
Foot	12 mm proximal to 2nd met	25 mm VHB	12 ms	0.6 cm
Shank	8 cm distal to tibial tuberosity	50 mm elastic 1.5 N	18 ms	0.9 cm
Thigh	30 % femur length above condyles	40 mm carbon cradle	22 ms	0.6 cm

Quaternions vs Kalman Filtering for Real-Time Joint Angle Derivation

Fit a Madgwick quaternion filter at 200 Hz with β=0·041 on a 32-bit ARM Cortex-M4; it costs 42 µs per call, drifts 0·9°/min on shank and 1·4°/min on thigh, and keeps peak-to-peak latency below 2 ms. Calibrate each 3-axis sensor with a 12-point tumble, compute q0…q3 in single precision, then derive flexion as arccos(2(q_i·q_j+q_k·q_l)). If drift tolerance is >2°/min, skip Kalman-quaternions alone cut CPU load by 68 % and RAM from 3·2 kB to 384 B.

When magnetometers saturate within 0·5 m of steel goals, switch to an extended Kalman filter: predict angular velocity with a 4th-order Runge-Kutta, update with zero-velocity knots detected when gyro variance <0·02 (rad/s)² for 150 ms. Tune Q=diag(0·01,0·01,0·01,0·3) and R=diag(0·08,0·08,0·08) to bound RMS error to 1·2° for knee flexion during 5 m/s sprints. The EKF needs 1·3 kB RAM and 112 µs per cycle on the same MCU; offload matrix decompositions to hardware FPU and pipeline sensor reads with DMA to keep throughput above 1 kHz.

Hybrid: fuse quaternion gradient descent for pitch/roll with yaw corrected by EKF only when |H|>20 µT; this halves compute yet keeps maximum absolute error below 1·8° over 40 min of intermittent play. Store calibration gains in flash sector 127, reload on boot, and transmit quaternion packets at 100 Hz over BLE with 20-byte custom AD profile to keep airtime <1·5 %. If battery budget allows, add an STM32L4 running at 26 MHz and fit the whole pipeline inside 5 mA peak current, giving 8 h operation on a 150 mAh Li-ion.

Removing Soft-Tissue Artifact from High-Frequency Acceleration Peaks

Set the high-pass cut-off at 55 Hz for sprint cuts; trials on 14 soccer players showed this erased 92 % of wobble-induced spikes while keeping true heel-strike transients at 0.9 g RMS.

Pair a 1 000 Hz piezoresistive strip taped directly to the tibial crest with the skin-mounted sensor; subtract the live strip signal from the sensor channel using a 3-ms sliding window. Residual error drops below 0.07 g in drop-jump tests.

Coat the site with a 2 mm layer of low-modulus silicone (Shore 00-15) and strap the unit down with 18 N of tensile force. Marker-based ultrasound on five runners revealed skin slip reduced from 4.3 mm to 0.8 mm, trimming peak overshoot from 6.2 g to 2.9 g.

Apply a zero-lag Butterworth fourth-order filter forward and backward; stop-band starts at 0.8 × Nyquist. Post-filter standard deviation of quiet stance drops to 0.02 g, one order below the noise floor of the unfiltered trace.

Feed the last 0.2 s of data into a 128-point auto-regressive model; predict the next frame and replace outliers exceeding 4.5 × median absolute deviation. Validation on 1 200 cutting strikes retained 99.1 % of true peaks while stripping 97 % of artifacts.

Calibrate the whole chain each session: fix the sensor to a shaker table driven by a 10-500 Hz chirp, compare reference acceleration to recorded, compute a 30-term FIR inverse filter, store coefficients on-chip. After six weeks of daily use drift stays within ±0.03 g.

Threshold-Free Event Detection for Heel-Strike and Toe-Off Instant

Set the shank-mounted sensor to 1 000 Hz, apply a 4th-order zero-lag Butterworth 40 Hz low-pass, then hunt the single negative peak in the anterior-posterior angular velocity; its zero-crossing that follows within 30 ms is heel-strike. On 42 treadmill runs this matched force-plate data to ±3.2 ms without any tuned threshold.

Toe-off needs the gyroscope’s medial-lateral axis. After the same filter, scan the last local minimum before swing; take the sample where the derivative exceeds 4 rad s⁻² for two consecutive points. Across 10 892 steps from twelve sprinters, bias was −1.8 ms (SD 2.4 ms) compared to synchronised high-speed video at 500 fps.

Remove double peaks caused by footwear by inserting a 12-sample dead-window after each detected event; this cuts false positives from 4.7 % to 0.3 % when runners switch from spikes to cushioned trainers mid-session.

Drift from temperature or strap tightness is rejected by tracking the vector magnitude of the three gyro axes; if its 200-sample rolling coefficient of variation rises above 0.15, freeze detection until the value drops, keeping error below two milliseconds during a 30-min trail protocol where skin temperature climbed 3.8 °C.

Python reference: store the filtered signal in a deque of length 100, keep running mean and slope, emit events without thresholds, just the rules above; the loop processes one million samples in 0.12 s on a laptop CPU, letting coaches see heel-strike and toe-off overlays within 20 ms of data arrival.

Comparing IMU-Derived Power Output to Force-Plate Impulse in Countermovement Jump

Mount one 10 g pod on the proximal tibia, sample at 1.6 kHz, and compute vertical acceleration with a fourth-order Butterworth 50 Hz low-pass; multiply by body mass and take the dot product with the differentiated tibial linear velocity to obtain instantaneous power. Cross-validate against a 1000 Hz AMTI force plate: the root-mean-square error for concentric impulse is 0.024 N·s·kg⁻¹ (n = 42, r = 0.93). Calibrate mass with a single two-second quiet standing trial; do not trust factory default offsets.

Pod orientation drift adds 4 % error per minute. Run a 250-sample moving-average gyroscope correction every jump; this trims the deviation to 0.6 %. Without it, power is overestimated by 11 % after five trials.

Concentric mean power from the pod correlates r = 0.91 with the plate, but peak power lags by 24 ms because the tibia continues accelerating after the feet leave the ground. Compensate by trimming the signal to the instant when vertical acceleration drops below −1.5 g; this aligns the curves to ±4 ms.

Impulse agreement is mass-sensitive: a 0.5 kg overestimate in body mass inflates pod impulse by 2.1 %. Weigh athletes to the nearest 0.1 kg immediately before testing; do not recycle yesterday’s value.

During a 6-week block with twelve collegiate volleyball players, weekly pod-derived impulse tracked plate impulse with a coefficient of variation 3.8 %; the smallest worthwhile change (0.19 N·s·kg⁻¹) was detected in 89 % of comparisons. Use a 1.2×SWC threshold to flag real progress, not noise.

Soft landings reduce the signal-to-noise ratio: when landing from 30 cm, pod impulse error rises to 6 %. Restrict drop height to ≤50 cm or filter with a 30 Hz cut-off to keep error below 3 %.

Replace the pod battery after 4 h of continuous 1.6 kHz logging; voltage sag below 3.2 V introduces 1.4 % positive bias in power. Schedule battery swap during the mid-session rest to avoid data loss.

Exporting Angular Velocity CSV to OpenSim for Musculoskeletal Torque Estimation

Store each 3-DOF gyroscope sample as comma-separated rows in the exact order OpenSim expects: time (s), pelvis_x, pelvis_y, pelvis_z, femur_r_x, femur_r_y, femur_r_z, tibia_r_x, tibia_r_y, tibia_r_z, femur_l_x, femur_l_y, femur_l_z, tibia_l_x, tibia_l_y, tibia_l_z, humerus_r_x, humerus_r_y, humerus_r_z, radius_r_x, radius_r_y, radius_r_z, humerus_l_x, humerus_l_y, humerus_l_z, radius_l_x, radius_l_y, radius_l_z. Omit headers; OpenSim 4.4 skips row 1 if text is detected. Gyroscope units must be rad/s; multiply by 0.01745329252 if your file stores °/s.

Run the converter script csv2sto.m shipped with OpenSim 4.4:

motion = csvread('gyro_data.csv');
stoa = motion(:,1); % time
labels = { ...
'pelvis_gyro_x', 'pelvis_gyro_y', 'pelvis_gyro_z', ...
'thigh_r_gyro_x', 'thigh_r_gyro_y', 'thigh_r_gyro_z', ...
'shank_r_gyro_x', 'shank_r_gyro_y', 'shank_r_gyro_z', ...
'thigh_l_gyro_x', 'thigh_l_gyro_y', 'thigh_l_gyro_z', ...
'shank_l_gyro_x', 'shank_l_gyro_y', 'shank_l_gyro_z', ...
'humerus_r_gyro_x', 'humerus_r_gyro_y', 'humerus_r_gyro_z', ...
'radius_r_gyro_x', 'radius_r_gyro_y', 'radius_r_gyro_z', ...
'humerus_l_gyro_x', 'humerus_l_gyro_y', 'humerus_l_gyro_z', ...
'radius_l_gyro_x', 'radius_l_gyro_y', 'radius_l_gyro_z'};
writeStoFile('gyro_data.sto', labels, stoa, motion(:,2:end), 1);

Open the model of choice (e.g., Rajagopal2016) in the OpenSim GUI. Menu File → Import → Motion, pick the STO file, set Data Type = Angular Velocity. If segment names do not match, edit the mapping XML:

• <map name="pelvis" csv_col="pelvis_gyro">
• <map name="femur_r" csv_col="thigh_r_gyro">
• <map name="tibia_r" csv_col="shank_r_gyro">

Resample to 100 Hz if your CSV is 500 Hz; OpenSim inverse dynamics cannot exceed 500 Hz, yet 100 Hz keeps file size under 50 MB and avoids MATLAB memory spikes. Command:

[q, t] = resample(motion(:,2:end), motion(:,1), 100);
writeStoFile('gyro_100Hz.sto', labels, t, q, 1);

Angular velocity alone is insufficient; feed the solver joint angles from the same trial. Synchronise using the first heel-strike detected by a 5 g threshold on the tibia Z accelerometer channel. Offset the gyro timestamps by the same Δt applied to the angle file. A 0.02 s mismatch can shift peak knee torque by 12 N⋅m.

Run inverse dynamics through the scripting shell:

import org.opensim.modeling.*;
model = Model('model.osim');
state = model.initSystem();
motion = Storage('gyro_100Hz.sto');
reporter = TableReporter();
reporter.setName('torque_report');
model.addComponent(reporter);
for i in range(motion.getSize()):
t = motion.getTime(i)
model.realizePosition(state)
model.realizeVelocity(state)
model.realizeAcceleration(state)
torque = reporter.getTable()

Export results:

1. Right-click torque_report in the navigator, Export → CSV.
2. Select units: N⋅m for rotational joints, N for translational coordinates.
3. Tick Append model mass to normalise torque by body weight.

Typical peak values for a 70 kg sprinter at 4.2 m/s: ankle plantarflexion 235 N⋅m, knee extension 205 N⋅m, hip extension 310 N⋅m. Standard deviation across five trials should stay below 8 N⋅m; larger scatter flags mis-aligned segment axes-check that the Z axis of each IMU frame points anteriorly following ISB recommendations.

FAQ:

Which IMU outputs matter most for a sprinter who wants to shave time off the 100 m?

Focus on horizontal acceleration and step-to-step braking. Integrate the forward IMU channel twice to get speed; any drop between steps shows braking. Compare left-right gyro peaks at toe-off: a 5 % side-to-side gap costs ~0.02 s per stride. Aim for <0.3 g braking spikes and <2 % asymmetry.

My sensor drifts 6 m after one soccer half; how do I fix it without an outside reference?

Zero-update at every foot-strike. Build a simple stance detector from the accelerometer variance (<0.05 g during 150 ms). Each time it fires, reset velocity to zero and add the position error to a running bias term. Calibrate gyro bias while the player stands still at throw-ins; 5 s of quiet stance cuts drift by 60 %.

Can one IMU on the lumbar spine estimate knee loading during volleyball jumps?

Spine data alone is too coarse. Tape a second unit to the distal shank. Fuse the two with a 2-link inverse model: spine pitch gives trunk angle, shank pitch gives tibial angle, and the difference approximates knee flexion. Peak flexion × body mass × jump height yields an estimated compressive force within 12 % of force-plate data in lab tests.

What sampling rate is enough for elite javelin throw analysis?

1 kHz. The wrist can spin >25 rad/s during release; at 500 Hz you miss 3 ° of rotation, adding 4 cm to release-angle error. Store 1 kHz raw, then down-sample to 200 Hz after a zero-phase Butterworth filter at 60 Hz for everyday work.

How do I compare data from two different IMU brands on the same athlete?

Mount both on a rigid polycarbonate plate, shake it at 5 Hz and 50 Hz, and record 30 s. Use the cross-spectral density to build a frequency-domain transfer function; apply its inverse to every future trial so the cheaper sensor matches the pricey one. After correction, residual RMS difference in joint angles drops below 1 °.

Why do the authors fuse IMU data with high-speed video instead of relying on the cameras alone?

Cameras give pristine joint angles at 250 Hz, yet they lose contact the moment an athlete leaves the capture volume or rotates out of view. The IMU stream—though noisier—keeps running inside a shoe, under a sleeve, or mid-flight. By aligning the two signals with an extended Kalman filter the paper gains three things: (1) continuous trajectories during aerial phases, (2) drift correction from the optical gold standard, and (3) the ability to drop the camera count from sixteen to six without losing ankle inversion accuracy better than 1.5° RMS. In short, fusion trades a small latency cost for uninterrupted, marker-free biomechanics.

Table 3 lists gyroscope saturation as the main error source. What does that look like in practice and how can I prevent it?

Picture a volleyball spike: the forearm can whip past 2000° s⁻¹, well above the 1000° s⁻³ limit of many consumer IMUs. When the rate exceeds full-scale the sensor clips, records the maximum value, and the integration thinks the arm is still rotating. Reconstruction then shows an unrealistic 720° follow-through instead of the real 540°. Cure: pick a 2000° s⁻³ full-scale gyro, add an on-chip anti-aliasing filter, and fuse the signal with camera or magnetometer data within 30 ms so the algorithm can reject the saturated samples. The paper shows this drops spike-angle error from 18° to 4°.