Detection finds objects in each frame but remembers nothing. Add re-identification to recover identity through occlusion, and put SORT, ByteTrack, OC-SORT, or BoT-SORT on any YOLO or RF-DETR pipeline, with zero tracking code.
Build a YOLO26 plus BoT-SORT tracking pipeline in Roboflow Workflows, from detection to tracked and visualized output, in under five minutes.
Add an Object Detection Model block. The Trackers library works with any model that returns supervision.Detections, including YOLO, YOLO26, and RF-DETR. Set a class filter to track only what you care about, like people or vehicles.
Place a tracker block (ByteTrack, SORT, OC-SORT, or BoT-SORT) after the detector. It connects detections across frames and assigns stable tracker_id values, with separate state per video stream.
When motion alone is not enough, re-identification matches appearance: the object is cropped, turned into an embedding, and compared against earlier tracks. A close match keeps the same ID, so an object that disappears and returns is not counted twice.
Add bounding box, label, and trace visualization, colored by track ID, then deploy via the Inference API on your camera or video file. The full pipeline takes under five minutes to build and requires zero tracking code.
The tracker predicts where each object is going with a Kalman filter and matches it by position overlap (IoU). It is fast and works well for simple scenes with predictable motion, but it breaks down when objects are occluded, cross paths, or briefly leave the frame, where it may assign a new ID.
When motion is not enough, re-identification compares appearance. The object is cropped and turned into an embedding, a visual fingerprint, and matched against stored embeddings from earlier tracks. A high similarity score keeps the original ID, re-linking an object even after tracking has lost it.
The strongest trackers use both. Roboflow trackers combine motion matching with appearance-based recovery, so identities stay stable through occlusions, crowds, and moving cameras.
Open source, detector-agnostic, and deployable without writing a line of tracking code.
The Roboflow Trackers library is compatible with any model that returns supervision.Detections, including YOLO, YOLO26, and RF-DETR. Pass detections in, get stable track IDs back, no need to swap your detector to add tracking.
SORT, ByteTrack, OC-SORT, and BoT-SORT, with clean implementations behind a single consistent interface. Switch algorithms to fit the scene, from simple real-time tracking to occlusions, crowds, and moving cameras, without rewriting your pipeline.
Native tracker blocks in Roboflow Workflows drop in after your detector and assign stable IDs automatically. Build a full YOLO26 plus BoT-SORT pipeline, from detection to visualized output, in under five minutes with no tracking code to maintain.
The Trackers library is open-source Python you can run anywhere. Pair it with RF-DETR under the permissive Apache 2.0 license to keep your detection-and-tracking stack commercial-safe, instead of the AGPL-3.0 copyleft terms that come with the YOLO family.
Half the Fortune 100 build computer vision with Roboflow, with tracking pipelines deployed for vehicle counting, retail foot traffic, sports analytics, and warehouse safety.
Trusted by teams at BNSF, Rivian, GE Vernova, Cummins, USG, Pella, and Peer Robotics.
YOLO Re-ID refers to re-identification, the process of recognizing the same specific object across different video frames or camera streams, especially after it disappears, becomes occluded, or reappears later. Object detection finds objects in each frame but remembers nothing between frames, so an object that disappears and returns can be treated as new. Re-ID recovers identity through appearance: the tracker crops the object, a deep neural network turns it into an embedding (a visual fingerprint), and the system compares it against stored embeddings to decide whether it is the same object and keep the same ID.
Three questions separate the tasks. Object detection answers what is this object. Object tracking answers where is this object moving across frames, usually by predicting motion with a Kalman filter and matching position overlap with IoU. Re-ID answers is this the same object I saw before, by comparing appearance embeddings. Re-ID is an optional but powerful component inside the tracking pipeline that takes over when motion alone is not enough, such as during long occlusions or when objects leave and re-enter the frame.
Roboflow Workflows includes native tracker blocks for ByteTrack, SORT, OC-SORT, and BoT-SORT. Place a tracker block after an object detection model and it connects detections across frames and assigns stable tracker_id values, outputting tracked detections, new instances, and already-seen instances. You can build a full YOLO26 plus BoT-SORT tracking workflow, from detection to tracked and visualized output, in under five minutes with zero tracking code. The open-source Roboflow Trackers Python library offers the same algorithms behind one interface for code-based pipelines.
SORT is the simplest and fastest, good for controlled scenes with reliable high-confidence detections and predictable motion, but it has no re-identification. ByteTrack matches both high- and low-confidence detections, which helps in crowded scenes, partial occlusions, and sports. OC-SORT extends SORT with observation-centric updates for frequent occlusions and non-linear, erratic motion like pedestrians or warehouse workers. BoT-SORT is the strongest for difficult scenes with occlusions, moving or shaking cameras, and similar-looking objects, and it can apply camera motion compensation. See the full benchmark comparison to choose.
Add Re-ID tracking to any YOLO or RF-DETR pipeline with the Trackers library or Roboflow Workflows. Keep a stable ID on every object.
Ask the Roboflow assistant about Re-ID, choosing a tracker, and building a tracking Workflow.
One open-source Python interface for SORT, ByteTrack, OC-SORT, and BoT-SORT.
Compare the four trackers head to head and find the right one for your scene.
A step-by-step walkthrough of building a tracking Workflow end to end.
