ByteTrack.NET 1.0.0

ByteTrack.NET

Unofficial C# port. This is a community port and is not affiliated with or endorsed by the original ByteTrack authors. The original work is MIT-licensed, © 2021 Yifu Zhang — see License & attribution.

A faithful C# (.NET 8) port of the tracking core of FoundationVision/ByteTrack (yolox/tracker/).

The YOLOX detector is out of scope — you feed the tracker detections that are already in image coordinates (bounding box + confidence), and it returns stable per-object track IDs across frames.

What's ported

Linear algebra sits behind ILinearAlgebra (default: dependency-free ManagedLinearAlgebra) so it can be swapped for a native/faster backend such as MathNet.Numerics without touching the filter. The linear assignment is a from- scratch LAPJV port with cost_limit support, matching lap.lapjv.

ReID-only helpers from the original matching.py (embedding_distance, fuse_motion, fuse_iou, gate_cost_matrix, v_iou_distance, merge_matches) are intentionally omitted — they are not on the BYTETracker code path.

Usage

using ByteTrack;

var config = new ByteTrackerConfig
{
    TrackThresh = 0.5f,  // high/low detection split
    TrackBuffer = 30,    // frames a lost track is kept (at 30 fps)
    MatchThresh = 0.8f,  // IoU match threshold (first association)
    Mot20 = false,       // skip score fusion for crowded scenes
};

var tracker = new ByteTracker(config, frameRate: 30);

foreach (var frameDetections in videoStream) // your detector's output per frame
{
    // Detection: (x1, y1, x2, y2, score) in image coordinates.
    var detections = frameDetections
        .Select(d => new Detection(d.X1, d.Y1, d.X2, d.Y2, d.Score))
        .ToList();

    IReadOnlyList<STrack> tracks = tracker.Update(detections);

    foreach (STrack t in tracks)
    {
        double[] tlbr = t.Tlbr; // [x1, y1, x2, y2]
        Console.WriteLine($"id={t.TrackId} box=[{string.Join(',', tlbr)}] score={t.Score}");
    }
}

Track IDs come from a process-global counter (as in the original). Call BaseTrack.ResetCount() between independent runs for deterministic IDs.

Build & test

dotnet build ByteTrack.sln
dotnet test  ByteTrack.sln

The test suite (18 tests) includes golden-value checks: the Kalman filter is compared against values generated from the original numpy/scipy code (tests/ByteTrack.Tests/Data/kf_golden.json), and LAPJV against an independent scipy.optimize.linear_sum_assignment oracle (lap_golden.json).

License & attribution

This project is licensed under the MIT License.

It is a derivative work (a port to C#) of FoundationVision/ByteTrack, which is also MIT-licensed. In accordance with the MIT License, the original copyright notice is retained:

• Original ByteTrack: Copyright (c) 2021 Yifu Zhang
• C# / .NET port: Copyright (c) 2026 Artur Buszka

See LICENSE for the full text and NOTICE for a summary of the provenance. This port is not an official release and is not affiliated with the original authors.

Citation

If you use this port in your research, please cite the original ByteTrack paper:

@article{zhang2022bytetrack,
  title   = {ByteTrack: Multi-Object Tracking by Associating Every Detection Box},
  author  = {Zhang, Yifu and Sun, Peize and Jiang, Yi and Yu, Dongdong and
             Weng, Fucheng and Yuan, Zehuan and Luo, Ping and Liu, Wenyu and
             Wang, Xinggang},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year    = {2022}
}