Detecting an object in a single frame is one problem. Following that same object across hundreds of frames, keeping its identity stable as it moves, gets occluded, and changes direction, is a different one. That second problem is object tracking in video, and it is what stands between a detection model and a system that can actually count items on a line, trace a ball across a pitch, or measure how things move over time.
Until recently, object tracking in video on Roboflow meant working in the SDK or the CLI. That has changed. The tracking algorithms now ship as drop-in blocks inside Roboflow Workflows, so you can add tracking to a detection model the same way you add any other component.
In a recent webinar, Roboflow machine learning engineer Lee Clement builds a working tracking pipeline from a blank canvas and gets it running on live video. As he puts it while wiring up the first example, "It takes literally seconds to put this thing together."
This post is an overview of what he covers and why the change matters. The full walkthrough, including the live runs and the parameter tuning, is in the video.
Object Tracking in Video
What Roboflow Trackers actually is
Roboflow Trackers is a plug-and-play multi-object tracking library, designed to drop in on top of any detection model. If you are processing sequential video and you need to keep track of the identities of your detections over time, you need a tracker in the loop. The library gives you a clean, modular implementation of several popular tracking algorithms, and all of them are Apache 2.0 licensed, so you are free to use them in your own projects.
The headline of the webinar is that these algorithms have escaped the SDK and the CLI. They are now building blocks in Workflows, usable in the same way as the rest of the platform's components. Tracking becomes a post-processing step on top of a detection model you already trained or pulled off the shelf, not a separate codebase to maintain.
Three Algorithms on One Spectrum: SORT, ByteTrack, and OC-SORT
A useful reason to watch is the framing of how to choose an algorithm. Lee lays the three current options out on a spectrum that trades throughput against robustness, which makes the choice concrete instead of a guess.
SORT sits at the fast end. It is optimized for high throughput and maximum simplicity, and it leans heavily on having a strong detector. In a stable environment with a good detector, SORT can run hundreds of frames per second and will not be the bottleneck in your pipeline.
ByteTrack sits at the robust end. Its two-stage association algorithm handles high and low confidence detections together and does well with occlusions and clutter, at the cost of some throughput.
OC-SORT sits in the middle. It keeps the simplicity of a real-time online tracker but adds a better formulation of the problem, so it handles lost tracks, occlusions, and nonlinear motion, like the direction changes you see in sports or dancing, far better than vanilla SORT, without giving up much speed. More algorithms are in the pipeline.
Two Examples that Show the Range
The webinar grounds all of this in two builds rather than abstractions.
The first is an industrial counting case: butter moving down a production line. Lee trains a custom instance segmentation model on the platform to detect the barcodes on each package, then builds the pipeline live. He injects the model, adds the SORT tracker block, attaches bounding box and label visualizations set to display the tracker ID, and connects it to the output. That is the whole pipeline. He runs it against a virtual webcam streaming production footage on a loop, and each package picks up a unique, stable tracker ID as it enters the frame. It is a clear look at how little stands between a detection model and a working counter.
The second is a sports case: tracking a soccer ball through a youth game full of fast motion, direction changes, and kids' legs crossing in front of the ball. Here Lee uses an off-the-shelf RF-DETR model rather than training anything, since a ball falls under a common detection class, and adds a class filter so only the ball is tracked. OC-SORT is the natural fit for this kind of cluttered, nonlinear motion. Out of the box it holds the track for a while but switches identity four or five times across the clip, which sets up the most useful part of the demo.
Why the Tuning Matters
The soccer example is where the video earns a watch, because it shows that tracking is not all-or-nothing. With one change, lowering the minimum IOU threshold from its default of 0.3 down to 0.1, Lee makes a visible difference in tracking performance. That parameter controls how much a detection has to overlap an existing track to count as a match. Loosening it lets the tracker hold a single identity through far more of the clip, dropping the ID switches down to almost none until the very end, where the ball goes blurry and fast behind other players.
This is the practical lesson worth taking from the session. These trackers come with sensible defaults that work across the board, and a small number of tuning knobs let you adapt them to specific kinds of footage when the defaults are not quite right. You get strong results immediately and a clear path to better ones.
Watch the Webinar
If you are building anything that counts, traces, or measures movement in video, this session is a fast way to see what object tracking in video looks like when it is a building block instead of a project. It is most useful for engineers working on production counting and inspection, sports and broadcast analytics, logistics, and any setting where keeping a consistent identity on a moving object is the point.
The full webinar walks through both builds end to end, the live runs, and the IOU tuning in real time. Watch it on YouTube here.
Then try it yourself. Open a new workflow, drop in a detection model and a tracker block, and add tracking to a video pipeline of your own at roboflow.com.
Learn more: Multi-object tracking in video
Cite this Post
Use the following entry to cite this post in your research:
Contributing Writer. (May 12, 2026). Object Tracking in Video, Now a Building Block in Roboflow Workflow. Roboflow Blog: https://blog.roboflow.com/object-tracking-in-video-workflows/