Track Bounding Boxes (Hungarian Algorithm)

How it fits

Typical backends

Caveats

• I/O contractOutput is track-centric. Each frame emits the set of active CONFIRMED tracks, so the count is the number of live tracks, not the number of input detections.
• I/O contractThe [TrackedSegmentation] output arm is only selectable when the input is [Segmentation]; the carried mask comes from the input. With [BoundingBox] input it cannot be resolved and the worker errors at startup.
• I/O contractThe [Track] arm is the only one that fills each track's ``trajectory``; it holds the track's own past centre points, oldest first, capped at history_size.
• State lifetimeA new track is tentative until matched on min_hits consecutive frames, then confirmed and emitted. A confirmed track unmatched for up to max_age frames is kept and emitted with its predicted (occlusion-bridged) box; after that it is dropped and its id is never reused.
• AccuracyAssociation is geometric + confidence only, with no appearance model. Add reidentify_objects_torchreid downstream when ids must survive long occlusions or identical-looking objects crossing.
• Parameter interactionmax_match_distance is the key quality lever. Lower it to forbid loose links (fewer id swaps, more new tracks); raise it to tolerate fast motion / low overlap.
• Parameter interactionclass_agnostic false keeps each track on its own class id; cross-class detections never steal a track. Set true only when the upstream class id is unreliable.
• Parameter interactionocclusion_aspect is OFF by default. Enable it only for upright-person tracking; it forces every box to occlusion_aspect_ratio and distorts other classes.