Skip to content

Privacy & ethics

People tracking is dual-use. This project is built so that the safe path is the default one.

Design choices

  • Session-local, anonymous IDs. A track ID is an integer valid within one video. Nothing links IDs across videos, cameras or sessions, and no identity (name, face, biometric template) is ever produced.
  • Appearance embeddings are working memory. ReID features exist only to keep identities straight within a clip, live in RAM, and are not persisted by the pipeline. (The benchmark caches embeddings of a public research dataset.)
  • The browser demo uploads nothing. Webcam and file inputs are processed locally; there is no server to receive them.
  • Analytics use geometry only (boxes, foot points, timing), never appearance.
  • Honest about what an ID is. The agent is instructed to say that "unique people" is an upper bound.

Out of scope on purpose

Cross-camera or cross-session re-identification, face recognition, attribute inference (age, gender, emotion), and watch-lists. Persistent re-identification turns a counting tool into a tracking-of-individuals tool and, in the EU, engages biometric-data rules (GDPR Art. 9) and, depending on the use, the AI Act's provisions on remote biometric identification and emotion recognition. This is general information, not legal advice -- deployments need their own legal review (purpose limitation, data minimisation, retention, signage, DPIA).

Using it responsibly

  • Prefer aggregate outputs (counts, flows, dwell distributions) over per-person trajectories; do not store trajectories longer than needed. Blur or drop frames at the edge and ship only counts.
  • Evaluate error rates before acting on outputs: HOTA ~0.4 on hard footage means identity errors are routine. Never use tracker output as the sole basis for a decision about an individual.
  • MOT17 contains real people filmed in public; its licence (CC BY-NC-SA 3.0) restricts commercial use.