Repository navigation
v0.0.60 - `ultralytics-inference 0.0.60` Add object tracking with the six Ultralytics trackers (#635)
Latest🌟 Summary
v0.0.60 adds cross-frame object tracking to Ultralytics Inference, bringing six Ultralytics trackers to the Rust library, CLI, and browser package.
📊 Key Changes
- 🎯 Adds BoT-SORT, ByteTrack, OC-SORT, Deep OC-SORT, FastTrack, and TrackTrack—the default tracker. Trackers assign IDs that follow objects across video frames.
- 🧠 Adds optional camera motion compensation and ReID appearance matching. Native Rust can use the
yolo26n-reid.onnxmodel for ReID; browser tracking runs without ReID. - 🧰 Adds a
trackCLI command and Rust APIs for tracking frames, keeping tracker state between calls, and resetting it for a new video. - 🌐 Adds
model.track()andresetTracker()to the browser package for ONNX and LiteRT models. - 🖼️ Supports detection, segmentation, pose, and oriented bounding box models. TrackTrack also recovers detections kept by a looser NMS pass, matching Ultralytics behavior.
- ✅ Tracking is optional, and ordinary prediction results remain unchanged from v0.0.59.
🎯 Purpose & Impact
- 📹 Makes it possible to follow the same object over time, enabling applications such as video analytics and webcam tracking—not just per-frame detection.
- 🔄 Keeps the tracking workflow consistent across the CLI, Rust library, and browser, with tracker IDs validated against Python behavior.
- ⚙️ Uses a lower default confidence threshold for tracking so weaker detections can help keep existing tracks alive. Users can choose a tracker or provide a configuration file.
- 🧩 The new
trackfeature can be enabled separately; ReID is optional, so basic tracking requires no new dependencies.
What's Changed
ultralytics-inference 0.0.60Add object tracking with the six Ultralytics trackers by @onuralpszr in #635
Full Changelog: v0.0.59...v0.0.60