Person Re-Identification

How It Works

A Re-ID pipeline produces a compact numeric signature for each observed person:

1. Person detection — a detector finds each person in each camera view.
2. Embedding — a specialized neural network produces a vector representation capturing clothing, shape, gait, and posture.
3. Matching — embeddings across cameras are compared; small distance means likely same person.
4. Tracking — the full trajectory of a subject across cameras is reconstructed.

Re-ID works even when faces are occluded, angled away, or too small to recognize — something face recognition alone cannot do.

Why It Matters

In large camera networks, a person of interest often appears briefly across dozens of cameras. Manually correlating these appearances is slow; face recognition requires clear frontal images. Re-ID:

• Links observations across cameras — reconstruct a full movement history.
• Works without face data — useful where face recognition isn't permitted or possible.
• Accelerates investigation — forensic search finds every camera a subject appeared on.
• Enables flow analytics — long-distance path tracking for planning.

IncoreSoft's Smart Tracking System / Person Re-ID module is used for investigation, visitor journey analysis, and operational security.

Use Cases

• Forensic investigation — tracing a suspect's full path through a facility
• Retail visitor journeys — understanding customer movement at mall scale
• Transit flow analysis — OD (origin-destination) studies without personal identification
• Lost person search — rapidly locating missing persons in a large facility
• Operational security — unusual movement pattern detection

Frequently Asked Questions

Is person re-identification the same as face recognition?

No. Face recognition identifies by name from a database. Re-ID only determines if two observations are the same person — without naming them. Re-ID works without any enrolled database.

Is Re-ID privacy-friendlier than face recognition?

Generally yes — no named identities are stored, and signatures are less identifying than face biometrics. However, some jurisdictions regulate any system that tracks individuals, regardless of identification.

What's the accuracy of Re-ID?

In research benchmarks, 80–95% rank-1 matching accuracy. In production deployments, accuracy is practical for investigation and analytics workflows, especially with human-in-the-loop review of top candidate matches.