What is Multiplicative Extended Kalman Filter?
Multiplicative Extended Kalman Filter is an attitude-estimation EKF that represents orientation with a unit quaternion while estimating a minimal local rotation error composed multiplicatively with that quaternion.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Keep global attitude and local uncertainty in different representations
Markley's comparison of multiplicative and additive filtering explains the core representation problem: rotations have no globally nonsingular three-parameter coordinates, while a four-component quaternion is redundant. MEKF uses the quaternion for the reference attitude and a three-vector for local deviations.
The true attitude may be written q_true = delta_q times q_hat or q_true = q_hat times delta_q. Both are legitimate, but they define different error frames. The propagation matrix, measurement Jacobian, correction order, and reset transform must all follow the chosen side.
Propagate, correct, inject, normalize, and reset
Gyroscope measurements propagate the reference quaternion and usually a gyro-bias estimate. Aiding vectors or an attitude observation produce an innovation in a compatible frame. The Kalman update estimates a small rotation vector; the implementation maps it through the exponential map, composes the resulting error quaternion with the reference, normalizes against numerical drift, and resets the local error.
Quaternion sign continuity matters because q and -q represent the same rotation but create different coordinate residuals. Before constructing a residual, implementations commonly select the shortest equivalent quaternion branch. Normalization alone does not repair a wrong convention or reset Jacobian.
Treat operational evidence as model-specific
The UVSQ-SAT in-orbit study demonstrates one MEKF deployment using gyroscope, magnetic, solar, and Earth-sensor information. Its reported accuracy belongs to that spacecraft, sensor availability, calibration, and eclipse conditions; it is not a universal MEKF guarantee.
Test quaternion norm and sign handling, left/right convention parity, stationary bias, constant-rate motion, vector-measurement geometry, large initial error, eclipses or dropouts, innovation consistency, NEES, and reset covariance. Compare with an additive EKF only under identical sensors, dynamics, initialization, and output conventions.
Key Characteristics
- Stores attitude globally as a normalized unit quaternion
- Represents local attitude uncertainty with three unconstrained coordinates
- Injects attitude corrections through quaternion multiplication
- Requires an explicit left- or right-error convention
- Can estimate gyro bias alongside attitude error
- Remains a local first-order approximation with finite convergence radius
Common Use Cases
- Spacecraft attitude determination from gyroscopes and star trackers
- Drone attitude fusion with inertial, magnetic, and navigation sensors
- Robot orientation estimation without Euler-angle singularities
- Gyroscope bias calibration during recursive attitude estimation
- Embedded orientation tracking with covariance output
Example
Loading code...Frequently Asked Questions
Why does MEKF use a multiplicative quaternion correction?
Quaternion multiplication composes a local three-dimensional rotation with the global unit-quaternion estimate. This preserves the rotation geometry and avoids treating four constrained quaternion components as an ordinary unconstrained vector.
What is the difference between MEKF and an additive quaternion EKF?
An additive filter updates quaternion components directly and must handle their unit constraint and singular covariance structure. MEKF estimates a minimal local attitude error and multiplies its quaternion into the reference attitude, then resets that local error.
Is MEKF the same as every Error-State Kalman Filter?
No. MEKF is an error-state construction specialized to attitude, usually with a quaternion reference and three-component rotation error. A broader ESKF may also include position, velocity, calibration, and bias errors with additive or manifold composition.
Does quaternion normalization make an MEKF correct?
No. Normalization controls numerical norm drift. Correctness also requires consistent multiplication side, active or passive rotation meaning, component order, sensor frames, Jacobians, bias model, residual branch, covariance injection, and reset.
Why can an MEKF fail with a large initial attitude error?
The local rotation error and measurement model are linearized around the current reference. A large or ambiguous initial mismatch can violate that approximation or select the wrong quaternion branch. Use a robust initializer, gating, and explicit convergence tests.