Multi-object tracking
Stable temporary IDs, movement trails and class voting, so a car briefly read as a truck stays one object.
Object detection · tracking · counting
Upload footage, place a counting line, and get reviewable counts of vehicles, people and bicycles. The model runs on your device, directly in the browser.
How it works
YOLO26n (80 COCO classes) runs in the browser through ONNX Runtime, on the GPU via WebGPU when available and on the CPU via WebAssembly otherwise.
Confident detections start tracks; weaker ones may only continue them. Constant-velocity prediction and size-aware gating keep IDs through short gaps.
Each confirmed track is counted once per counting line, with a dead band against jitter. Zones measure occupancy; unique totals cover objects that already left.
Crossing events as CSV, a full JSON session report, or an annotated frame. Exports are the persistence: nothing is stored on a server.
Features
Stable temporary IDs, movement trails and class voting, so a car briefly read as a truck stays one object.
Horizontal or vertical counting line with A → B and B → A totals per class.
Draw a rectangle on the preview to count object centres inside it, frame by frame.
Load a YOLO26 or YOLOv8 ONNX export with your own class names. Same tracker, same counts.
Footage never leaves the browser. Model and runtime files download once and are cached.
Pause, step frame by frame, and check every count against the footage before exporting.
Evaluation
The tracker was checked on Intel's CC BY 4.0 sample videos by comparing track IDs with the objects visible in the footage. “Before” is the previous single-stage tracker. Detection itself is YOLO26n trained on COCO: top-down views and unusual objects need a custom model.
| Clip | In the footage | Before | Now |
|---|---|---|---|
| Parking lot, 54 s | 8 objects (4 people, 2 cars, 2 bicycles) | 14 IDs | 8 IDs |
| Car park, 30 s | 5 cars | — | 5 IDs (one car read as a bus from above) |
| Parking lot, 12 s sample | 2 people, 1 car, 1 bicycle | — | 4 IDs · 3 line crossings |
Tallycam counts objects with temporary, session-only IDs. It does not recognise faces, read number plates or link people across cameras. Obtain consent for recordings and follow local privacy rules.
Open the workspace and run the sample clip, a sample image, or your own footage.
Open Workspace