What is Covariance Intersection?
Covariance Intersection is a conservative data-fusion method that combines uncertain estimates without requiring their unknown cross-covariance while preserving a consistent covariance bound.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Fuse in information space without inventing independence
For two Gaussian estimates (m_1, P_1) and (m_2, P_2), CI forms P_CI^-1 = w P_1^-1 + (1-w) P_2^-1 and m_CI = P_CI [w P_1^-1 m_1 + (1-w) P_2^-1 m_2], with w in [0,1]. The weight is chosen by an objective such as determinant or trace of the fused covariance.
Julier and Uhlmann's original paper introduced a non-divergent estimator for unknown correlations. CI does not estimate the missing cross-covariance and should not be described as the same posterior a centralized Kalman filter would compute.
Understand consistency and conservatism as separate properties
Consistency means the reported covariance does not understate the actual error covariance under the stated assumptions. It does not mean the fused mean is unbiased, outlier-resistant, optimally accurate, or calibrated under model misspecification. A biased local estimate can still corrupt the result, and contradictory inputs require validation rather than blind fusion.
The optimization criterion matters. Minimizing determinant targets uncertainty volume, while minimizing trace targets total marginal variance; scaling and units can change either choice. For more than two estimates, pairwise order can matter unless a joint optimization or a declared sequential policy is used.
Track provenance before accepting the conservative fallback
The distributed multisensor fusion review explains how common process noise and information recirculation create cross-correlation. If provenance permits exact cross-covariance tracking, decorrelation, or removal of common information, those approaches may retain more precision than CI.
Validate CI with Monte Carlo coverage and NEES across adversarial correlation values, not only independent samples. Record source identity, timestamps, frames, units, covariance semantics, weight objective, optimizer tolerance, fusion order, and rejected inputs. CI addresses unknown correlation; it does not solve time alignment, data association, sensor faults, or malicious data.
Key Characteristics
- Combines means and covariances in information space
- Requires no explicit cross-covariance between input estimates
- Uses convex nonnegative weights that sum to one
- Preserves a conservative covariance bound under its assumptions
- Trades statistical efficiency for protection against double counting
- Depends on a declared objective, scaling, and fusion order
Common Use Cases
- Decentralized multi-robot localization with shared information
- Track-to-track fusion when sensor histories overlap
- Combining estimates after communication paths lose provenance
- Fault-tolerant sensor networks without a central raw-data store
- Benchmarking safer fusion against naive independence assumptions
Example
Loading code...Frequently Asked Questions
When should Covariance Intersection be used?
Use CI when multiple estimates refer to the same state, their covariance bounds are meaningful, and their cross-correlation is unavailable or too expensive to track. If the exact correlation or common information is known, a less conservative fusion rule may be preferable.
Why is naive Kalman fusion unsafe with unknown correlation?
Treating correlated estimates as independent counts shared evidence more than once. The fused covariance can become smaller than the actual error covariance, making the system overconfident and potentially causing gates, planners, or later filters to reject valid information.
Does Covariance Intersection produce the optimal estimate?
It optimizes a selected covariance objective within the CI family while protecting consistency over unknown correlations. It is generally not the minimum-variance posterior that would be available if the true cross-covariance and joint model were known.
How is the CI weight selected?
Choose a weight in the unit interval by minimizing a declared scalar objective such as log determinant or trace of the fused covariance. State scaling, constraints, numerical tolerance, and multi-source fusion order must be recorded because they can change the result.
Can Covariance Intersection handle bad sensor data?
Not by itself. CI protects against unknown correlation, not bias, wrong frames, stale timestamps, false data association, outliers, or compromised sources. Validate and gate inputs first, then test fused consistency across relevant correlation and fault scenarios.