🔍 Searching photos instead of video? Use the new Face Finder to find a person across a folder or ZIP of images. →

Last updated:

Video Person Tracker

Detect, track and re-identify every person in a video, entirely in your browser. Each person gets clickable timestamps for every appearance — plus optional face matching against reference photos and COCO object detection. No upload, no signup.

Load Video
🎬
Drop a video here or click to browse
The file never leaves your device — it's processed entirely in this tab.
MP4 · WEBM · MOV · MKV — best results under ~5 minutes
🔒
Nothing is uploaded. Detection, tracking and re-identification run with MediaPipe's WebAssembly models inside this browser tab. The video stays on your machine. Models are fetched once from a public CDN. Re-identification uses generic appearance embeddings — it's a strong investigative aid, not a biometric identity match, and results should always be human-verified.

Track every person in a video

The Video Person Tracker detects every person in a clip, follows each one frame to frame, and re-identifies them when they leave and come back — so one person stays one identity across the whole video. Every appearance lands on a per-person timeline with clickable timestamps, and the boxes are drawn back over the footage as you scrub.

Search a video for a specific person

Add reference photos of the people you are authorized to look for. Each is turned into an ArcFace face vector and matched against the faces in the video, with a confidence score on every likely hit. When a face is not usable, matching falls back to clothing and body appearance. It is an investigative aid to point your eye — not legal proof of identity — so verify matches yourself.

Private by design

Everything runs in your browser with MediaPipe and onnxruntime-web in WebAssembly. Your video is never uploaded; models are fetched once from public CDNs and then cached. Use it only on footage you have the right to analyze.

Frequently asked questions

How do you track a person across an entire video?
Upload a video and the tool steps through it frame by frame, detecting every person with MediaPipe and following each one with combined motion and appearance matching. When someone leaves and returns, an appearance embedding re-identifies them, so the same person keeps one identity. Each person gets a row on a timeline with clickable timestamps for every appearance and re-appearance.
Can you identify someone from a video?
You can search a video for a specific person by uploading one or more reference photos. The tool builds an ArcFace face vector from the photo and matches it against the faces it finds in the video, flagging likely matches with a confidence score. This is appearance-based similarity to help an investigator find someone they are authorized to look for — it is an aid, not legal proof of identity, and matches should always be verified by a human.
Does the Video Person Tracker do face recognition?
Yes. Faces are detected with BlazeFace, aligned to a canonical template using eye and nose keypoints, then embedded with a real ArcFace (MobileFaceNet) model running in WebAssembly through onnxruntime-web. Matching uses cosine similarity on those 512-dimension face vectors, with a fallback to clothing and body appearance when a usable face is not visible.
Can it detect objects in a video too?
Optionally, yes. A toggle enables MediaPipe's object detector across the 80-class COCO set — cars, bags, animals, laptops, and so on — and timestamps every class it recognizes, with its own cards and timeline rows alongside the people.
Is it free, and is my video uploaded anywhere?
It is free and runs entirely in your browser. The video is decoded and analyzed locally and is never uploaded to a server. The detection, face-recognition, and embedding models are fetched once from public CDNs and then cached; after that the analysis is fully client-side.
How accurate is video person re-identification?
Re-identification works well for clearly visible, reasonably frontal subjects and degrades with heavy occlusion, motion blur, tiny or far-away faces, and people dressed alike. Sensitivity is adjustable, identities can be renamed by hand, and a face-confirmation filter reduces false detections. Treat the output as an investigative aid to review, not a definitive identification.