---

name: tracking-and-trackers description: "Use BoxMOT for live tracking, tracker construction, tracker selection, and AABB/OBB output schema debugging." metadata: disco-role: operating disable-model-invocation: true license: AGPL 3.0

Tracking and Trackers

Use this sub-skill when the task is about boxmot track, direct tracker instantiation, tracker registry lookup, output schemas, or explaining how AABB and OBB detections flow through the tracker API.

Covers

• boxmot track and BoxMOT.track(...)
• boxmot.trackers.registry.create_tracker(...)
• direct tracker imports such as boxmot.trackers.OccluBoost
• tracker selection by name or class
• AABB and OBB detection layouts
• track result accessors such as xyxy, xywha, id, conf, cls, and det_ind
• per-class tracking and tracker output debugging

Does not cover

• benchmark cache generation or replay (generate, eval, tune, research)
• ReID training, evaluation, comparison, or export
• native C++ build steps and boxmot build

Use the sibling sub-skills for those routes.

Read first

• API reference
• Tracking workflows
• Detection and output formats
• Troubleshooting
• Smoke test script

Good prompts for this route

• "Track this video with ByteTrack and save the output."
• "Why does my tracker expect 7 columns instead of 6?"
• "How do I instantiate OccluBoost directly in Python?"
• "Why did my OBB track output lose the angle column?"
• "Which tracker objects support OBB input?"

Typical workflow

1. Identify the detector/tracker inputs and whether the user is working with AABB or OBB tensors.
2. Decide whether the user wants the CLI, the high-level BoxMOT facade, or a direct tracker class.
3. Use the tracker registry or the direct package export to construct the tracker.
4. Check the input tensor shape before asking the user to run a long command.
5. Confirm the returned TrackResults schema and the det_ind mapping when debugging downstream consumers.

Entry points

CLI tracking

boxmot track --detector yolov8n --reid osnet_x0_25_msmt17 --tracker botsort --source video.mp4 --save

Python facade

from boxmot import BoxMOT

model = BoxMOT(detector="yolov8n", reid="osnet_x0_25_msmt17", tracker="botsort")
run = model.track(source="video.mp4", save=True)

Direct tracker class

from boxmot.trackers import OccluBoost

tracker = OccluBoost(reid_model=None, with_reid=False)
tracks = tracker.update(dets, image=frame)

What to hand off to nearby references

• Column and shape details belong in references/data-formats.md.
• API signatures and registry details belong in references/api-reference.md.
• Concrete workflow examples belong in references/workflows.md.
• Shape errors, OBB support checks, and tracker misuse belong in references/troubleshooting.md.

When a user asks for a smoke test or contract check, use scripts/tracker_contract_smoke.py instead of hand-writing a one-off snippet.