Re-identification (re-ID) is matching the same object or person across different cameras or across a gap in time — recognising that the person who left camera 3 is the one now entering camera 7. Single-camera tracking gives each object a local ID; re-ID is the harder step of carrying identity across views that never overlap, by comparing appearance features (clothing, shape, colour) rather than a face. It is what lets a system reconstruct a path through a whole site.
Re-ID is the engine behind cross-camera forensic search ("follow this person") and site-wide flow analysis. Crucially, ONVIF does not standardise cross-camera identity — a Profile M object ID is device-local — so stitching identities across cameras is the VMS's or analytics platform's job, not something the standard guarantees. That makes re-ID quality very implementation-dependent.
Two cautions define it. First, accuracy drops sharply going cross-camera: single-camera tracking around 80% on standard measures can fall to roughly 70–85% appearance re-ID, and stricter cross-camera measures lower still — it produces ranked candidates for a human to confirm, never certainty. Second, privacy: a movement trail singles a person out, so it is personal data even without a name (GDPR Art. 4(1)), though appearance-based re-ID is generally not Article 9 biometric data, whereas face or gait recognition is. The embedding-model internals belong to the AI for Video Engineering section.
