What is Cardinalized Probability Hypothesis Density Filter?
Cardinalized Probability Hypothesis Density Filter is a Random Finite Set approximation that propagates both multi-object intensity and a cardinality distribution under an independent and identically distributed cluster model.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Propagate count and spatial information together
The predicted cardinality distribution combines surviving prior targets with spontaneous births, while the predicted intensity combines their spatial contributions. The two quantities are coupled through their shared expected count; an implementation must preserve normalization and avoid treating intensity as a single-object probability density.
Vo, Vo, and Cantoni's analytic CPHD implementation derives a Gaussian-mixture recursion under linear Gaussian dynamics and birth assumptions. Nonlinear extensions based on local approximations do not make the exact nonlinear recursion closed form.
Update an IID-cluster posterior from the measurement set
For the standard point-target model, the update accounts for detections, missed detections, and Poisson clutter over the whole measurement set. Elementary symmetric functions of measurement likelihood ratios enter the cardinality and intensity corrections; replacing them with independent per-measurement updates changes the represented posterior.
The CPHD family approximates the posterior by an IID-cluster RFS: the count distribution is arbitrary, but states are IID conditional on that count. This is richer than the Poisson count assumption commonly associated with PHD filtering and weaker than a posterior that retains object-specific existence and association hypotheses.
Evaluate cardinality without claiming identity
Trajectory PHD and CPHD research highlights that standard PHD and CPHD filters estimate the current target set but do not form trajectories from first principles. A Trajectory CPHD Filter instead defines the IID-cluster approximation on complete trajectories; tags attached to state-only components remain implementation aids.
Evaluate count RMSE and cardinality calibration together with GOSPA or OSPA decomposition, localization error, missed and false objects, covariance consistency, component count, and runtime. Compare PHD and CPHD under identical birth, survival, detection, clutter, pruning, and extraction settings; otherwise an apparent count improvement may come from different tuning.
Key Characteristics
- Propagates a full target-cardinality probability mass function
- Propagates first-order multi-object intensity alongside cardinality
- Uses an IID-cluster approximation conditional on target count
- Models missed detections and clutter through a set-valued likelihood
- Supports Gaussian-mixture and sequential Monte Carlo implementations
- Does not encode persistent target identity in its standard form
Common Use Cases
- Radar or sonar tracking when target count changes over time
- Improving count stability over a PHD baseline in dense clutter
- Sensor-control objectives that depend on target-number uncertainty
- Multi-object localization where persistent identity is not required
- Benchmarking IID-cluster filters against labeled or hypothesis-based trackers
Example
Loading code...Frequently Asked Questions
What does a CPHD filter add to a PHD filter?
A PHD filter propagates intensity, whose integral is only the expected target count. A CPHD filter also propagates a probability for each possible count. That extra distribution can make count estimates more stable, but it adds computation and still does not preserve target identity.
What probability model does a CPHD filter assume?
The standard CPHD approximation uses an IID-cluster Random Finite Set. It allows an arbitrary cardinality distribution, while target states are independent and identically distributed conditional on the count. Standard recursions also declare survival, detection, birth, and clutter models.
Does a CPHD filter perform explicit data association?
No explicit track-to-measurement assignment is emitted by the standard CPHD recursion. Association uncertainty is integrated into its set update. The calculation is still combinatorial in measurement-set statistics, and Gaussian-mixture or particle implementations need component management.
Can a CPHD filter produce persistent tracks?
Not in its standard current-target-set form. Extracted peaks can be connected with heuristics, but those links are not persistent identities represented by the CPHD posterior. Labeled RFS or trajectory-set filters are better fits when identity and track history are required.
How should CPHD and PHD filters be compared?
Use the same detections, motion and birth models, clutter intensity, survival and detection probabilities, pruning limits, and extraction rules. Report count calibration and RMSE, GOSPA or OSPA components, localization error, false and missed objects, runtime, and memory.