IMU-Guided Temporal Filter

• IMU-guided temporal filtering is a robust estimation technique that fuses high-frequency inertial data with auxiliary measurements to yield continuous, drift-corrected estimates of position, orientation, and velocity.
• It employs recursive filtering methods such as the EKF and manifold-based approaches to integrate state-space models, ensuring geometric consistency and mitigating inertial drift with corrective updates.
• Applications in robotics, VR/AR, and autonomous navigation benefit from improved accuracy and real-time performance, as evidenced by empirical benchmarks showing reduced RMSE and enhanced robustness.

An IMU-guided temporal filter is a class of estimation algorithms that fuse high-frequency inertial data from an Inertial Measurement Unit (IMU) with additional low- or moderate-rate measurements (e.g., visual, acoustic, wheel odometry, or pseudo-measurements) to provide continuous, drift-corrected estimates of dynamic state—typically position, orientation, and velocity. These filters utilize the IMU to drive rapid state propagation (temporal guidance), while corrective updates from other modalities anchor the solution and mitigate the intrinsically unbounded drift of inertial integration. The “temporal filter” aspect refers to recursive or sliding-window estimation frameworks (e.g., Kalman filters, observers on Lie groups, or their invariant/geometric or learned variants) that continuously update the state estimate as new IMU and auxiliary measurements arrive. IMU-guided temporal filtering is foundational in robotics, navigation, and autonomous systems, with architectures covering classic EKF/UKF, geometric observers, minimum-energy filters, and hybrid data-driven or structure-based approaches.

1. State-Space Formulation and IMU Propagation

The core of an IMU-guided temporal filter is a dynamical state-space model that integrates IMU signals to propagate estimates of pose, velocity, and often sensor biases. Formally, the state vector can encompass translational position $p \in \mathbb{R}^3$, velocity $v \in \mathbb{R}^3$, orientation $R \in \mathrm{SO}(3)$ or its parametrization (e.g., quaternion, DCM, or vectorized $c = \mathrm{vec}(C)$), and possibly IMU bias terms and other nuisance parameters. Discrete propagation equations typically take the form:

$\begin{bmatrix} p_{k+1} \[2pt] v_{k+1} \[2pt] c_{k+1} \end{bmatrix} = \begin{bmatrix} I_3 & \Delta t \cdot I_3 & 0 \ 0 & I_3 & 0 \ 0 & 0 & I_9 \end{bmatrix} \begin{bmatrix} p_k \ v_k \ c_k \end{bmatrix} + \begin{bmatrix} \frac{1}{2} \Delta t^2 a_k^b \ \Delta t a_k^b \ (\Omega(\omega_k^b) \otimes I_3) c_k \Delta t \end{bmatrix}$

where $a_k^b$ is the IMU-measured body acceleration, $\omega_k^b$ the measured turn rate, and $\Omega(\cdot)$ denotes the skew-symmetric matrix for angular velocity (AlSharif et al., 2024). Noise, biases, and process uncertainties are modeled via additive Gaussian noise to input signals, with covariance propagated using the linearized dynamics as in $Q_k = F_u Q_u F_u^\top$.

Variants exist for more general state spaces, e.g., $\mathrm{SE}_2(3)$ for minimum energy geometric filters (Henderson et al., 2020), group-extended representations with calibration and nuisance parameters (Zhu et al., 2022, Brossard et al., 2019), and multi-clone or window state for temporal displacement anchoring (Liu et al., 2020).

2. Measurement Models and Correction Modalities

To arrest drift, IMU-guided temporal filters are augmented with measurements from exteroceptive or auxiliary systems—visual features, acoustic signals, pseudo-odometry, or structureless keypoint tracks. The measurement function $y_k=h(x_k)+v_k$ typically extracts observable quantities (e.g., modulated positions, bearings, temporal displacements) related to the state. Representative models include:

• Acoustic Localization: Estimated triangle vertices are related to the centroid and attitude by $p_i = p_k + C_k d_i$, forming a linear mapping in the state (AlSharif et al., 2024).
• Visual Keypoints: Reprojection errors over tracked features are used as innovations; often processed via MSCKF-style structureless updates to reduce computational load (Huai et al., 2020).
• Learned Pseudo-Measurements: Deep networks predict displacements and uncertainties over windows of IMU data (e.g., $\hat\Delta p$, $\Sigma_\Delta$), providing pseudo-measurements within the EKF update step (Liu et al., 2020).
• Zero-Velocity or Kinematic Constraints: Pseudo-measurements enforcing physical constraints (e.g., wheel/foot non-holonomic motion, zero stance-foot velocity) are included with dynamically or adaptively tuned noise (Zhu et al., 2022, Brossard et al., 2019).

Measurement updates proceed by computing innovations, associated Jacobians, and the Kalman gain, leading to the recursive (EKF/UKF) correction of the state estimate and covariance.

3. Manifold and Geometric Filtering Frameworks

A major thread in IMU-guided temporal filtering is the formulation on nonlinear manifolds or matrix Lie groups, motivated by the inherent non-Euclidean geometry of orientation and pose. Examples include:

• SE(3) and Extensions: Filters on $\mathrm{SE}(3)$ or $\mathrm{SE}_2(3)$ employ group operations, exponentials, and adjoint actions to ensure geometric consistency and avoid singularities (Henderson et al., 2020, Hashim et al., 2021, Hashim et al., 2021).
• Quotient and Homogeneous Manifolds: State representations may exploit gauge-invariant variables by factoring out unobservable transformations (e.g., global yaw, translation), improving filter consistency and addressing the observability mismatch problem in naive EKF-VIO (Goor et al., 2021).
• Retractions and Projections: When fusing with Riemannian-based solutions, the output of the filter is “retracted” back onto a constraint manifold (e.g., isosceles-triangle manifold $M$) using specialized projection algorithms, often in the absence of a closed-form exponential or logarithmic map (AlSharif et al., 2024).

These frameworks underpin filters that are fully nonlinear, preserving the natural invariances and symmetries of the physical system.

4. Algorithms and Implementation Details

The temporal filter operates in a recursive predict-update fashion, often with the following structure:

1. Propagation (Prediction): Use IMU measurements to propagate the current state estimate and its covariance/pseudo-covariance through the process model. Manifold integration is performed with closed-form exponential maps or Euler integration, depending on the group structure.
2. Measurement Update: At measurement times, compute the innovation, the measurement Jacobian (linearized as needed), and perform an update using the Kalman gain or minimum energy correction. For stochastic cloning or keyframe-based filters, the measurement may link multiple time-indexed clones or windowed states (Liu et al., 2020, Huai et al., 2020).
3. Projection/Retraction: If the corrected state does not adhere to physical or geometric constraints (e.g., triangle geometry, group membership), apply an explicit projection algorithm to enforce invariants (AlSharif et al., 2024).
4. Sliding Window/Cloning and Marginalization: Many modern approaches rely on maintaining a window of prior states for batch updates, with marginalization policies to contain computational cost (Huai et al., 2020, Liu et al., 2020).

Typical pseudocode (as in (AlSharif et al., 2024)) is: initialize the state, iterate over sensor samples, propagate using IMU, correct on auxiliary measurements, optionally project, and repeat.

5. Performance Characteristics, Advantages, and Limitations

IMU-guided temporal filters exhibit several characteristic properties:

• High-Rate Propagation and Drift Correction: They leverage the high-frequency, locally accurate information from the IMU for temporal prediction, and rely on auxiliary sensing for global consistency.
• Geometric Consistency: Manifold-based approaches avoid local parameterization artifacts and inconsistency, yielding provable stability and convergence properties in many cases (Hashim et al., 2021, Hashim et al., 2021).
• Robustness to Sensor Quality: Filters incorporating adaptive or learned measurement covariances (e.g., via CNNs) can dynamically adjust to variable IMU quality, motion context, or application constraints (Brossard et al., 2019).
• Real-Time and Embedded Capability: Many algorithms achieve linear or near-linear computational complexity, suitable for real-time deployment on embedded systems (Hashim et al., 2021, Hashim et al., 2021).

Limitations may include dependence on observability from auxiliary modalities (e.g., line-of-sight for vision/acoustics, sufficient excitation for calibration), computational cost with large windows or numbers of landmarks, and sensitivity to model mismatch if geometric constraints or bias models are violated. Some methods require re-tuning or re-training for domain shifts (Brossard et al., 2019, Liu et al., 2020).

6. Application Domains and Empirical Benchmarks

IMU-guided temporal filters are ubiquitous in robotics, AR/VR tracking, mobile navigation, and autonomous vehicles. Empirical evaluations demonstrate:

• Indoor Position and Orientation Estimation: Fusing IMU with Riemannian acoustic localization, average RMSE in position $<$6 cm and in Euler angles $\sim1$–3° over walking experiments, outperforming GN approaches by up to 40% (AlSharif et al., 2024).
• Inertial Odometry: Neural network–augmented EKF systems yield 30–35% lower position/yaw drift than double integration or loose-coupling baselines (Liu et al., 2020).
• Visual-Inertial Odometry: Equivariant VIO filters achieve state-of-the-art accuracy on EuRoC sequences, with position RMSE $<$0.15 m per trajectory, rivaling more complex optimization-based pipelines (Goor et al., 2021, Huai et al., 2020).
• Navigation and SLAM: Nonlinear Lie-group observers converge to the true trajectory from large errors and maintain small bounded residuals in position and attitude, robust to IMU bias and noise (Hashim et al., 2021, Hashim et al., 2021).

A selection of performance metrics from (AlSharif et al., 2024) illustrates the empirical impact:

Scenario	Metric	Value
Simulated walking, $\sigma_a$	RMSE (cm) RSD/RTR	7–10 cm, GN ≈ 20 cm
Real exp. (MTi-1+acoustics)	RMSE (cm) UKF-RTR	5.00 (yaw 2.68°)
Semi-exp. (iPhone IMU)	RMSE <9 cm (80% quant)	GN <13 cm

7. Notable Variants and Research Directions

Recent progress expands IMU-guided temporal filtering into the following areas:

• Learning-Enhanced Filters: Tight integration of statistical priors (learned displacement and uncertainty) with traditional temporal filters for improved robustness and adaptability (Liu et al., 2020, Brossard et al., 2019).
• Invariant/Geometric Filtering: Design of invariant EKFs and observers that achieve autonomous, log-linear error dynamics, explicitly address sensor misplacement and group-affine process structure, and offer rapid convergence even with poor initialization (Zhu et al., 2022).
• Riemannian and Manifold-Constrained Filters: Incorporation of manifold constraints (e.g., isosceles-triangle geometry) via customized retraction/projection algorithms enables principled fusion of nonlinear geometric information (AlSharif et al., 2024).
• Minimum Energy and LTI/LTV Filters: Lie-group-based minimum energy filters accommodate asynchronous, multi-rate sensor fusion and offer deterministic stability properties, especially in high-noise or limited-update scenarios (Henderson et al., 2020).
• Keyframe-Based Structureless Filters: Efficient visual-inertial odometry pipelines that avoid explicit landmark state augmentation and support full online temporal calibration, scaling to multi-camera arrays (Huai et al., 2020).

Ongoing research focuses on improving filter consistency, cross-modal calibration, observability-aware design, and efficient implementation for large-scale, long-duration deployments.