What is Multiple Hypothesis Tracking?
Multiple Hypothesis Tracking is a multi-target tracking framework that keeps several competing measurement-to-track histories, scores them recursively, and postpones ambiguous association decisions until later observations provide evidence.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Branch on measurement origin and score complete histories
Reid's original MHT paper assigns each measurement to an existing target, a new target, or false alarm, then updates joint hypothesis probabilities recursively with detection, clutter, birth, and state likelihood information. This lets later scans influence which earlier association history survives.
Scores are commonly accumulated as log-likelihood ratios to avoid numerical underflow and to compare target-origin explanations with clutter. A score is valid only under the declared motion, measurement, detection, birth, and clutter models; adding arbitrary bonuses or comparing hypotheses with inconsistent constants can change the selected track set.
Bound the hypothesis forest without erasing ambiguity
Exact branch growth is exponential. Validation gates and independent clusters reduce candidate connectivity; ranked assignment generates likely children; probability or score pruning removes weak branches; merging combines sufficiently equivalent estimates; and N-scan pruning commits decisions older than a fixed horizon.
These controls change the approximation. Aggressive pruning can delete the branch that later evidence would support, while a long horizon can violate memory or latency limits. Track-oriented MHT stores compatible track hypotheses and solves a global selection problem; hypothesis-oriented MHT explicitly maintains global hypotheses. The names should not be treated as identical data structures.
Evaluate identity, latency, and retained alternatives
The SWTrack study maps multi-scan hypotheses to a sparse graph and sliding-window assignment for 3D multi-object tracking, illustrating how later frames can repair current-frame ambiguity. Its nuScenes results support that implementation under its detector and window, not universal superiority of MHT.
Report localization error, ID switches, fragmentation, false and missed tracks, track initiation delay, hypothesis count, pruned probability mass where available, peak memory, and deadline misses. Compare equal detections and track-management rules against GNN, JPDAF, and RFS trackers. Never evaluate an offline or long-window result as if it were zero-latency online output.
Key Characteristics
- Retains multiple competing association histories across scans
- Models measurements as existing targets, births, misses, or clutter
- Scores joint hypotheses recursively under explicit probabilistic models
- Uses later observations to resolve earlier ambiguous assignments
- Requires gating, clustering, pruning, merging, or bounded lookback
- Trades computation and memory for stronger identity continuity
Common Use Cases
- Radar and sonar tracking in dense clutter
- Autonomous-driving tracking through crossings and occlusions
- Air-traffic surveillance with missed and false reports
- Multi-camera identity maintenance over short temporal windows
- Benchmarking multi-scan association against single-scan trackers
Example
Loading code...Frequently Asked Questions
Why does Multiple Hypothesis Tracking delay decisions?
A measurement crossing or occlusion may be impossible to resolve from one scan. MHT keeps several plausible histories so later observations can favor the branch whose motion, detection, and clutter evidence is most coherent instead of making an irreversible early assignment.
What is the difference between track-oriented and hypothesis-oriented MHT?
Hypothesis-oriented MHT explicitly stores global association hypotheses. Track-oriented MHT stores local track hypotheses plus compatibility constraints and solves for compatible global sets. Both need pruning, but their storage, optimization, and implementation costs differ.
What is N-scan pruning?
N-scan pruning commits to the ancestry of the best current hypothesis at a decision point N scans in the past and deletes incompatible older branches. It bounds memory and delay, but a short horizon can lock in an association before enough evidence arrives.
How is MHT different from JPDAF?
JPDAF marginalizes current-scan joint events into one moment-matched estimate per track. MHT preserves alternative histories across scans. MHT can correct delayed ambiguity, while JPDAF generally has lower storage cost but may coalesce close tracks.
What makes an MHT implementation production-ready?
It needs explicit birth, detection, clutter, gating, scoring, compatibility, pruning, merging, and termination rules; bounded latency and memory; deterministic tie handling; and evaluation of ID switches, fragmentation, false tracks, state error, and deadline misses.