Axis-aligned Box Tracking

Use AABB tracking when object orientation is not part of the output.

boxmot track \
  --detector yolov8n \
  --geometry aabb \
  --tracker bytetrack \
  --source video.mp4 \
  --save

Canonical contract

import torch

from boxmot.structures import Boxes, Detections

detections = Detections(
    geometry=Boxes(xyxy_float32_cpu.contiguous()),
    scores=scores_float32_cpu.contiguous(),
    class_ids=class_ids_int64_cpu.contiguous(),
    sample_id="sequence-1:42",
)
tracks = tracker.update(detections)

print(tracks.geometry.values)       # float32[M,4] xyxy
print(tracks.track_ids)             # int64[M]
print(tracks.detection_indices)     # int64[M], -1 for coasting rows

Boxes validates finite float32[N,4] values with x2 > x1 and y2 > y1. Detections and Tracks keep geometry, scores, classes, IDs, masks, and embeddings aligned during immutable selection.

For a standalone box-only call, tracker.update(rows) also accepts an exact NumPy N x 6 matrix in (x1, y1, x2, y2, confidence, class_id) order. The return value is a C-contiguous float64 M x 8 matrix in (x1, y1, x2, y2, track_id, confidence, class_id, detection_index) order. Use Detections for Tracks, embeddings, masks, sample metadata, and pipeline composition.

All registered Python trackers support AABB geometry. Start with ByteTrack for a lightweight baseline, then choose an appearance-assisted tracker when occlusion-driven identity switches matter more than minimum latency.