Wearable gait analysis accuracy improves when multiple sensors and algorithms are combined to compensate for individual sensor weaknesses and to enforce biomechanical constraints. Causes of inaccuracy include inertial sensor drift, magnetometer disturbances, sensor misplacement and inter-subject gait variability; addressing these requires integrating complementary data sources and models.
Sensor-level fusion and orientation correction
Classic signal-level fusion uses Kalman filters and complementary filters to combine accelerometer, gyroscope and magnetometer readings into stable orientation estimates. Sebastian Madgwick University of Bristol developed an efficient orientation filter widely used in wearable systems to reduce gyroscope drift while remaining computationally light. Magnetometers can correct long-term heading but are vulnerable to indoor distortion, so practical systems often blend magnetometer corrections with zero-velocity or biomechanical resets.
Event-based corrections and biomechanical constraints
Detecting gait events with pressure insoles or footswitches lets systems apply zero-velocity updates (ZUPT) to reset velocity and reduce integration drift during stance. Constraining inter-segment kinematics with a biomechanical model or a multibody Kalman filter ties limb orientations together, improving step length and joint-angle estimates. Hugh Herr MIT Media Lab and colleagues have demonstrated how combining inertial sensing with mechanical models benefits prosthesis control and gait-rehabilitation devices by producing more physiologically plausible kinematics.
Combining different sensor modalities raises the reliability of stride and asymmetry metrics important for clinical use. Max Little University of Oxford used smartphone accelerometry to detect Parkinsonian gait signatures, showing that fused, contextualized signals can yield clinically relevant markers remotely. However, algorithm performance depends on consistent sensor placement and population-specific calibration, so validation on target demographic groups is essential for trustworthiness.
Machine learning and decision-level fusion
Data-driven fusion uses feature-level or decision-level combination of IMU, pressure, electromyography and optical data. Deep learning architectures such as convolutional and recurrent networks can learn complementary representations from raw sensor streams, often outperforming rule-based fusion for complex, noisy environments. Yet these models require large, representative datasets and careful guarding against overfitting to a particular terrain or cultural gait pattern.
Consequences of improved fusion include more accurate remote monitoring, earlier detection of pathology, and better control of assistive devices. Environmental factors such as urban magnetic interference, slippery rural paths, and cultural variations in footwear and gait mean systems must be validated across territories and user groups to be effective and equitable.