Covariance Intersection

The defining property of CI is that it produces a consistent fused covariance matrix. This means the computed covariance is guaranteed not to be smaller than the true, unknown error covariance. By avoiding overconfident estimates, CI prevents filter divergence—a critical failure mode in standard Kalman filters where the estimator believes it is more accurate than it actually is. This guarantee holds regardless of the unknown cross-correlation structure.

CI fuses two estimates (a, A) and (b, B) by computing a convex combination parameterized by a scalar weight ω ∈ [0,1]:

• Fused covariance: C⁻¹ = ωA⁻¹ + (1-ω)B⁻¹
• Fused mean: c = C[ωA⁻¹a + (1-ω)B⁻¹b] The optimal ω is found by minimizing the trace or determinant of C, ensuring the tightest possible consistent bound. This optimization is computationally cheap, typically requiring a simple line search.

CI is inherently suited for decentralized sensor networks where nodes share processed estimates, not raw data. Key advantages include:

• No bookkeeping: Nodes do not need to track which information has been shared with whom, avoiding the complex pedigree management required by channel filters.
• Arbitrary topology: CI handles unknown correlations arising from loops in the communication graph, making it robust for ad-hoc and mesh networks.
• Scalable: Each node fuses incoming estimates independently without global knowledge of the network state.

A known limitation of standard CI is that it can be overly conservative, discarding known independent information. Split CI addresses this by partitioning each estimate into a correlated component and an independent component:

• The correlated parts are fused using the conservative CI rule.
• The independent parts are fused using the standard, optimal Kalman filter update. This decomposition yields a tighter, less conservative fused estimate when some independence structure is known, recovering the exact Kalman solution when estimates are fully independent.

Inverse Covariance Intersection provides an alternative formulation that addresses the conservatism of standard CI without requiring the explicit decomposition of SCI. ICI operates in the information space and guarantees a fused covariance that is bounded from above by both input covariances. It is algebraically equivalent to applying CI in the inverse domain, often yielding tighter estimates than standard CI while maintaining the same strong consistency guarantees and computational simplicity.

CI's primary use case is fusing estimates from unknown or unmodeled correlation sources, such as:

• Common process noise: Multiple trackers observing the same target are affected by the same unpredictable maneuvers.
• Shared prior information: Two local filters initialized with the same map or calibration data.
• Implicit data reuse: Information propagating through redundant paths in a sensor network. In all these cases, assuming independence leads to filter inconsistency; CI provides a principled, conservative alternative.

In collaborative Simultaneous Localization and Mapping (SLAM) for autonomous vehicle fleets, vehicles share map features and pose estimates. When two vehicles rendezvous and share information, their estimates become correlated through common prior observations. Covariance Intersection allows them to fuse shared data without double-counting information, producing a consistent joint map estimate that never overestimates precision. This is essential for heterogeneous fleet orchestration where maintaining a full cross-correlation matrix between all agents is computationally prohibitive.

In distributed defense systems, multiple radar stations independently track a target and transmit their local state estimates to a command center. These tracks are correlated because they observe the same target dynamics and may share a common process noise model. Track-to-Track Fusion using CI provides a consistent fused track that is guaranteed not to be overconfident, unlike a naive Kalman filter combination which can diverge. This is critical for Multiple Hypothesis Tracking (MHT) systems where correlation structure is complex.

A swarm of drones performing Visual-Inertial Odometry (VIO) can share relative pose measurements to improve individual localization. When drone A observes drone B, and drone B later observes drone A, their estimates become correlated through the bidirectional measurement chain. CI enables each drone to fuse inter-drone measurements into its own state estimate without requiring a centralized server to maintain the full swarm covariance matrix, enabling scalable, fault-tolerant swarm autonomy.

A manufacturing robot equipped with a LiDAR-Inertial Odometry (LIO) subsystem and a separate Radar-Camera Fusion module produces two independent 6-DOF pose estimates. Because both subsystems share the same physical motion, their errors are correlated through common IMU biases and vibration profiles. CI provides a principled method to fuse these modular outputs into a single, consistent pose estimate without requiring the vendor of each subsystem to expose their internal cross-correlation structure, preserving modularity.

When two robots independently explore adjacent areas and return with local occupancy grids, merging them requires knowing the correlation between their map estimates. This correlation arises from shared loop closures or common prior information. Grid-Based Fusion using CI allows the maps to be stitched together into a single consistent occupancy grid without double-counting the evidence from overlapping regions, preventing the fused map from becoming artificially certain about obstacle locations.

In safety-critical Fault Detection and Isolation (FDI) architectures, multiple redundant sensors measure the same physical quantity. If one sensor develops a subtle bias, its estimate diverges from the consensus. CI provides a robust fusion backbone that does not require an accurate model of the fault-induced correlation, allowing the system to maintain a conservative but safe state estimate even during transient fault conditions while the FDI system isolates the faulty unit.

A high-level fusion architecture where locally processed state estimates (tracks) from multiple independent sensor systems are combined. Covariance Intersection is the canonical solution for fusing these tracks when the cross-correlation between their estimation errors is unknown, preventing the double-counting of common information that would lead to overconfident, inconsistent global tracks.

The foundational recursive algorithm for optimal state estimation in linear dynamic systems. A standard Kalman filter requires explicit knowledge of the cross-covariance between estimates. Covariance Intersection provides a suboptimal but guaranteed consistent alternative when this cross-covariance is computationally intractable or impossible to maintain in a decentralized network.

A fundamental property where a state estimator's calculated covariance matrix does not underestimate the true error. Covariance Intersection mathematically guarantees this property by fusing only the unique information from each source. The fused covariance P_CI is always larger than or equal to the true error covariance, ensuring the system never becomes overconfident, which is critical for safety.

A network architecture where each node processes its own local sensor data and communicates only state estimates, not raw measurements. This topology inherently creates unknown cross-correlations due to common prior information and process noise. Covariance Intersection is the foundational algorithm enabling scalable, consistent fusion in these networks without a central fusion center.

The dual form of the Kalman filter that propagates the information matrix (inverse covariance) and information vector. The Covariance Intersection algorithm has an elegant closed-form solution in information space, where the fused information matrix is simply a convex combination of the local information matrices, weighted by the parameter omega (ω), which minimizes the determinant or trace of the result.

An advanced extension that separates the known independent component of an estimate's error from the unknown correlated component. By splitting the covariance into two terms, SCI allows the independent information to be fused with standard Kalman gain while only applying the conservative CI update to the correlated portion, resulting in a significantly less conservative and more accurate fused estimate.