What is Labeled Multi-Bernoulli Filter?
Labeled Multi-Bernoulli Filter is a multi-object tracker that represents each potential object by a unique label, an existence probability, and a conditional state density, then approximates the labeled Bayesian update as independent components.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Use labels as posterior state variables
A labeled state is a pair (x, label), and the distinct-label indicator prevents two objects in one realization from sharing a label. An LMB density is parameterized by {(r_label, p_label(x))}: r_label is the probability that the labeled object exists, and p_label is its conditional kinematic density.
Labels support track extraction, but they are bookkeeping variables under the model rather than observable identity truth. A poor birth model, unresolved crossing, or aggressive pruning can still initiate duplicate labels, lose a track, or switch a reported identity.
Approximate GLMB association mixtures
Reuter and colleagues' LMB paper interprets the LMB filter as an efficient approximation of the delta-GLMB filter. A standard implementation predicts LMB components, expands feasible association hypotheses for the update, and marginalizes the resulting GLMB mixture back into one existence probability and state density per label.
That collapse preserves useful first-order labeled information and cardinality but removes correlations between labels and mutually exclusive global hypotheses. LMB can therefore be much smaller than GLMB while becoming overconfident or ambiguous when several tracks compete for similar measurements.
Engineer birth, hypothesis generation, and extraction together
Runtime is governed by the number of predicted labels, measurements, and retained association hypotheses. Gating, ranked assignment or Gibbs sampling, pruning, merging, and existence thresholds define the practical algorithm. Report those settings rather than describing LMB complexity with one context-free asymptotic number.
Recent LMB implementation research continues to trade hypothesis-search cost against tracking fidelity. Evaluate GOSPA or OSPA, track switches, fragmentation, cardinality error, existence calibration, localization consistency, latency, and memory under crossings, births, deaths, missed detections, and clutter.
Key Characteristics
- Attaches a unique discrete label to every represented object
- Stores one existence probability and state density per label
- Produces labeled current-object estimates directly
- Approximates a GLMB update by marginalizing association hypotheses
- Reduces hypothesis storage relative to a full GLMB mixture
- Depends strongly on birth, pruning, and extraction policies
Common Use Cases
- Radar and sonar tracking that requires persistent track labels
- Tracking-by-detection for vehicles, robots, or pedestrians
- Distributed sensor fusion with explicit label matching
- Sensor management using per-track existence uncertainty
- Lower-cost baselines for comparison with GLMB or PMBM trackers
Example
Loading code...Frequently Asked Questions
What information does an LMB component contain?
Each component contains a unique label, an existence probability, and a probability density for the object's state conditional on existence. The collection induces a distribution over labeled object sets and a Poisson-binomial cardinality distribution.
How is an LMB filter different from an unlabeled multi-Bernoulli filter?
Both use independent Bernoulli components, but LMB includes a distinct label in each object state. That label supports track output and label-specific existence estimates. An unlabeled filter estimates the current set but does not represent persistent identity.
How is an LMB filter related to a GLMB filter?
A GLMB posterior is a mixture over label sets and association histories. An LMB filter commonly performs the measurement update in that richer form, then marginalizes the mixture into one Bernoulli component per label. This reduces cost while discarding cross-label dependence.
Does a unique label prevent identity switches?
No. Labels make identity explicit in the state, but incorrect birth, association, hypothesis truncation, or model mismatch can still attach the wrong label to observations. Track-switch and fragmentation metrics remain necessary.
How should an LMB filter be evaluated?
Measure set localization and cardinality with GOSPA or OSPA, then add trajectory metrics, ID switches, fragmentation, existence calibration, covariance consistency, and runtime. Stress crossings, occlusion, low detection probability, clutter, and uncertain births.