Inertial Measurement Units: Principles & Applications

• Inertial Measurement Units (IMUs) are integrated sensor assemblies that measure linear acceleration, angular velocity, and sometimes magnetic fields using accelerometers, gyroscopes, and magnetometers.
• Modern IMUs leverage MEMS technology to achieve compact, low-power designs while contending with measurement noise, bias instability, and calibration challenges using filtering and optimization techniques.
• Applications span navigation, robotics, motion capture, and biomechanics, with sensor fusion methods like complementary and Kalman filters enhancing precision and mitigating drift.

Inertial Measurement Units (IMUs) are integrated sensor assemblies for measuring a body's linear acceleration and angular velocity in three dimensions, with some designs including magnetic field measurement. IMUs underpin a wide array of applications such as navigation, robotics, wearable motion capture, biomechanics, vehicular state estimation, and human-computer interaction. Modern IMUs exploit microelectromechanical system (MEMS) technology, resulting in miniaturized, low-power, and cost-effective devices, but face significant challenges due to measurement noise, bias instability, and intrinsic sensor imperfections (Conlin, 2017).

1. Fundamental Sensing Principles and Models

A typical IMU consists of three mutually orthogonal accelerometers and three orthogonal gyroscopes, providing six degrees of measurement (6D), with optional tri-axial magnetometers extending this to 9D. Their measurement models are:

• Accelerometers: The output for the x-axis is

$a_x = S_{ax} \cdot a_{x,true} + b_{ax} + n_{ax}$

where $S_{ax}$ is the scale factor (typically ≈1), $b_{ax}$ a slowly-varying bias, and $n_{ax}$ white Gaussian noise. At rest, the z-axis measures local gravity.

• Gyroscopes: The output model is

$$\omega_{i,meas} = \omega_{i,true} + b_{\omega_i} + n_{\omega_i},\quad i\in\{x,y,z\}$$

where $b_{\omega_i}$ represents drift, and $n_{\omega_i}$ is measurement noise. Integrating gyroscope output yields orientation, but bias causes drift that grows with time.

• Magnetometers (if present): The measurement equation is

$m_{i,meas} = S_{mi} m_{i,true} + b_{mi} + n_{mi}$

where $m_{i,true}$ is the true magnetic field projection.

The continuous-time kinematics for a rigid body, assuming $R_b^n(t)\in SO(3)$ (body-to-navigation-frame rotation), is:

$S_{ax}$0

Here $S_{ax}$1 denotes the skew-symmetric operator for vector cross product (Conlin, 2017). Magnetometer readings enhance heading observability by providing absolute yaw information.

2. Error Sources, Calibration, and Filtering

IMU accuracy is fundamentally limited by several error sources:

• Random (stochastic) noise: Thermal and quantization noise results in high-frequency errors, accumulating as random walks in orientation and position.
• Systematic (epistemic) errors: Scale-factor inaccuracies, axis non-orthogonality, constant biases, and temperature dependencies cause long-term drift and pose estimation errors (Faizullin et al., 2022, Faizullin et al., 2021).
• Bias Instability and Drift: Biases are modeled as random walks (Ornstein–Uhlenbeck processes), with orientation drift typically growing as $S_{ax}$2 or unbounded if unchecked (Faizullin et al., 2021, Conlin, 2017).

Calibration Procedures

Modern practice employs batch least-squares or nonlinear optimization to correct for biases, scale factors, and axis misalignments in both accelerometers and gyroscopes. The improved procedure in (Webering et al., 2022) models the measurement as $S_{ax}$3 and shows that with as few as 12 static orientations, sub-mg and sub-0.1° accuracy for bias and misalignment are achievable, substantially reducing calibration time.

Additionally, extrinsic (IMU–chassis) calibration, essential in robotics, is addressed using on-board sensors only, optimizing the 3D rigid-body transform to ensure accurate global pose estimation (Peng et al., 2020).

Sensor Fusion and Filtering

Robust state estimation demands fusing IMU outputs to absorb drift and reduce noise. Classical techniques include:

• Complementary Filter: Merges short-term gyroscope stability with long-term accelerometer reference. Discrete update:

$S_{ax}$4

where $S_{ax}$5 (Conlin, 2017).

• Kalman and Extended Kalman Filters (EKF/IEKF): Employed for nonlinear sensor fusion, explicitly modeling system and measurement covariances. The IEKF tightly couples IMU streams with multibody dynamic models, enforcing kinematic and kinetic consistency (e.g., for biological or robotic linkages) to reduce drift (Osman et al., 13 May 2025).
• Learning-based Correction and Uncertainty Propagation: AirIMU (Qiu et al., 2023) integrates data-driven per-sample correction of noise and bias with model-based pre-integration and covariance propagation, greatly improving odometry accuracy, especially under non-Gaussian or non-stationary disturbances.

3. Architectures: Single, Multiple, Redundant, and Hybrid IMUs

Single IMU and Classic Strapdown

The conventional form is a "strapdown" 6D or 9D IMU, fixed to a rigid body, providing real-time inertial data. These are the basis for both pedestrian- and vehicle-dead-reckoning (Wang et al., 2020, Khorrambakht et al., 2021, Qiu et al., 2023).

Multiple IMU Arrays

Combining the outputs of multiple co-located IMUs (MIMU) can increase accuracy and robustness:

• Probabilistic Fusion (AVE/Virtual IMU): Outputs from $S_{ax}$6 sensors are averaged, reducing random noise variance by $S_{ax}$7, but systematic errors are not eliminated (Zhang et al., 2019, Faizullin et al., 2022).
• Best Axes Composition (BAC): Selects, at each time, the three axes (from $S_{ax}$8 possible) with the lowest recent drift, constructing a virtual angular velocity vector more robust to systematic error than probabilistic averaging, yielding up to 20% orientation and position improvement for short dead-reckoning intervals (Faizullin et al., 2022, Faizullin et al., 2021).
• Gyro-Free and Distributed Redundant IMUs: Arrays of accelerometers alone can algebraically infer angular rates ("software gyros"), or distributed IMUs can be fused in a decentralized manner for fault-tolerance and improved observability (Conlin, 2017).

4. Drift, Magnetometer-Free Approaches, and Kinematic-Chain Correction

IMUs are inherently subject to gyroscopic drift. Classical correction exploits gravity (for roll/pitch) and the Earth's magnetic field (for yaw). In environments where these references are unavailable, innovative algorithms have emerged:

• Kinematic-Chain Drift Correction: In extended sensor networks (e.g., upper-body or full-body tracking), drift in a joint sensor can be compensated by enforcing constraints that the acceleration measured at a joint must be equivalent across parent and child segments after accounting for local rotational accelerations. Relative alignment is maintained even in microgravity or magnetically disturbed scenarios, with root joint yaw unobservable but relative orientations maintained within $S_{ax}$9 RMSE (Stretton et al., 2024).
• Magnetometer-Free, Constraint-Based Heading Correction: For bodies with range-of-motion-constrained joints, optimizing the heading offset to keep the sequence of measured orientations inside the physically feasible set, enables long-term drift-bounded tracking (<4° even without any absolute heading reference) (Lehmann et al., 2020).

5. Applications: Motion Capture, Robotics, Biomechanics, Interaction

IMU technology has pervaded a broad spectrum of domains:

• Motion Capture (MoCap): Both full-body and facial MoCap systems leverage IMUs for portable, robust tracking, benefiting from independence to lighting and occlusion (e.g., vision-free MoCap) (Wang et al., 2024).
• Human Biomechanics: Wearable IMUs support range-of-motion and functional assessments for clinical populations (e.g., frozen shoulder detection) (Lu et al., 2021), as well as gait characterization with large public datasets combining optical and inertial sensing (Santos et al., 2021).
• Robotics and Autonomous Vehicles: IMUs are integral to mobile robot state estimation, with tightly-coupled nonlinear optimization routines ensuring calibration and extrinsic alignment are maintained (Peng et al., 2020).
• Pedestrian and Handwriting Tracking: Dead-reckoning algorithms and domain-adapted neural filters enable long-term, infrastructure-free human trajectory estimation (Wang et al., 2020, Gupta et al., 2023).
• Human-Computer Interaction and Myoelectric Control: IMUs are embedded alongside EMG, clarifying the role of posture-induced separation vs. gesture-specific signal content in classification accuracy (Campbell et al., 2020).

6. Limitations, Open Issues, and Future Directions

• Systematic Error Mitigation: Systematic (non-Gaussian, non-stationary) errors remain a key limitation, especially in low-cost MEMS sensors. Adaptive axis selection, continuous calibration, and learned uncertainty propagation are active research areas (Faizullin et al., 2022, Qiu et al., 2023).
• Sensor Placement and Realignment: Wearable and mobile settings are prone to sensor misalignment and on-body slippage; online detection and correction strategies based on IMU-intrinsic invariants have demonstrated success with ~3 ms detection latency and immediate recovery of estimation accuracy (Yi et al., 2021).
• Model-Based Deep Learning Integration: Hybrid architectures (e.g., accurate preintegration with deep sequence models) show superior drift characteristics and computational efficiency, especially under variable motion dynamics (Khorrambakht et al., 2021, Qiu et al., 2023).
• Scalability and Cost-Efficiency: Distributed, redundant arrays (wireless, timing-based, or MEMS-driven) democratize high-fidelity motion sensing for consumer VR, spaceflight, and clinical monitoring, but raise algorithmic design complexity (Conlin, 2017, Stretton et al., 2024).

7. Summary Table: Key IMU Fusion Methods and Outcomes

Fusion Method	Error Improved	Best Performance Interval	Key Reference
Complementary Filter	Orientation drift	< 5 s	(Conlin, 2017)
Probabilistic Virtual IMU (AVE)	Random noise	Any, no systematic elim.	(Zhang et al., 2019)
Best Axes Composition (BAC)	Systematic drift	< 2.5 s orientation	(Faizullin et al., 2022)
Kinematic-Chain Drift Correction	Yaw drift (no mag)	Unlimited (rel. joints)	(Stretton et al., 2024)
IEKF + Dynamic Model	Drift + kinetics	Full trajectory	(Osman et al., 13 May 2025)
Learning-based Uncertainty (AirIMU)	Unmodeled error	All, sensor-agnostic	(Qiu et al., 2023)

The field continues to evolve rapidly along MEMS miniaturization, data-driven uncertainty modeling, multi-IMU fusion, and robust filtering architectures, targeting real-world conditions where environmental or mounting uncertainties challenge classical paradigm performance.