Data Association

The simplest and most computationally efficient data association method. GNN makes a single hard assignment per measurement by solving a global cost minimization problem, typically using the Hungarian algorithm or Munkres assignment.

• Mechanism: Constructs a cost matrix of all measurement-to-track distances and finds the unique assignment that minimizes total cost
• Key constraint: Each measurement can be assigned to at most one track, and each track receives at most one measurement
• Primary weakness: Makes irreversible decisions; a single incorrect association in a cluttered scene propagates error into subsequent frames
• Best for: Low-clutter environments with well-separated targets and high signal-to-noise ratios

A soft-assignment algorithm that avoids hard decisions by computing probabilistic weights for every measurement-to-track pairing. Instead of selecting a single winner, JPDA updates each track with a weighted composite of all measurements.

• Mechanism: Enumerates all feasible joint association hypotheses and computes the marginal probability that measurement j originated from target t
• Key innovation: The track state is updated using a combined innovation weighted by these association probabilities
• Limitation: Assumes a known, fixed number of targets; does not handle track initiation or deletion natively
• Best for: Dense target environments where measurements frequently fall within overlapping validation gates

A deferred-logic paradigm that postpones association decisions by maintaining multiple competing hypotheses about measurement origins. When ambiguity arises, MHT branches into alternative world states and lets future data resolve which was correct.

• Mechanism: Generates and propagates a tree of hypotheses, each representing a complete set of track-to-measurement assignments over time
• Key strength: Can recover from past ambiguities when clarifying measurements arrive later—something GNN and JPDA cannot do
• Pruning strategies: Hypothesis reduction via N-scan pruning, k-best enumeration, and track scoring using log-likelihood ratios
• Best for: High-stakes tracking where association errors are costly and computational resources permit deferred resolution

A mathematically rigorous framework that models the entire multi-target state—including the number of targets—as a random finite set. This elegantly unifies detection, tracking, and association into a single Bayesian filtering problem.

• Key algorithms: Probability Hypothesis Density (PHD) filter propagates only the first-order moment; Cardinalized PHD (CPHD) jointly propagates cardinality distribution
• Labeled RFS: The Generalized Labeled Multi-Bernoulli (GLMB) filter maintains target identities, enabling true track continuity
• Advantage: Eliminates explicit association as a separate step; measurement-to-track assignment emerges from the Bayesian recursion
• Best for: Scenarios with unknown and time-varying numbers of targets, low detection probabilities, and high false-alarm rates

An expectation-maximization (EM) approach that treats data association as a missing data problem. Unlike MHT's combinatorial hypothesis tree, PMHT iteratively refines association probabilities using batch processing.

• Mechanism: Assumes measurements are independent and each target can generate multiple measurements per scan, relaxing the one-measurement-per-target constraint
• Computational advantage: Linear complexity in the number of targets and measurements, making it tractable for large-scale problems
• Trade-off: The independence assumption can lead to track coalescence when targets maneuver in close proximity
• Best for: Offline batch tracking applications where computational efficiency is prioritized over strict association constraints

Emerging approaches that replace hand-crafted cost functions with learned similarity metrics from neural networks. These methods excel when appearance, motion, and context cues must be combined non-linearly.

• Siamese networks: Learn an embedding space where measurements from the same target are close and different targets are far apart
• Graph neural networks: Model the association problem as edge prediction on a bipartite graph of detections and tracks
• Transformer architectures: Use cross-attention mechanisms to jointly reason about all measurement-track pairs simultaneously
• Best for: Vision-based multi-object tracking where deep appearance features provide strong re-identification cues beyond kinematics alone

A deferred-logic tracking algorithm that maintains multiple competing data association hypotheses over time. When measurement origin is ambiguous, MHT propagates several possible interpretations simultaneously, pruning low-probability branches as new data arrives. Key characteristics:

• Propagates uncertainty rather than collapsing to a single estimate
• Computationally intensive due to combinatorial hypothesis growth
• Excels in high-clutter, high-maneuverability scenarios

A recursive mathematical algorithm that estimates the state of a dynamic system from noisy measurements. The filter operates in two steps: predict the next state using a motion model, then update that prediction using observed sensor data. It minimizes the mean squared error and provides a Gaussian uncertainty covariance. The standard Kalman filter assumes linear system dynamics and measurement models, making it the optimal estimator under these conditions.

A high-level fusion architecture where locally processed state estimates from multiple independent sensor systems are combined into a single global track. Unlike raw measurement fusion, this approach transmits only processed tracks, reducing bandwidth. The critical challenge is handling cross-correlation between tracks that share common process noise or prior information, requiring techniques like Covariance Intersection to avoid overconfident estimates.

A data fusion algorithm for combining state estimates when their cross-correlation is unknown or intractable. In decentralized sensor networks, tracks from different nodes often share correlated errors that cannot be tracked. Covariance Intersection computes a weighted average of the estimates that produces a consistent fused covariance, guaranteeing the result is never overconfident. This is essential for robust multi-sensor data association in distributed architectures.

A preprocessing step that defines an ellipsoidal region in measurement space centered on a predicted track position. Only measurements falling within this gate are considered candidates for association with that track. The gate size is derived from the innovation covariance and a chi-squared threshold. Key benefits:

• Dramatically reduces computational complexity
• Eliminates physically implausible associations
• Prevents distant clutter from corrupting state estimates