Facial recognition in public safety is a technology that detects faces in camera video, converts each one into a numeric template and compares it against an authorized watch list (for example, people with arrest warrants or missing persons). When similarity passes a threshold, the system does not "identify" anyone: it raises an alert with a possible candidate that a human operator must review before any action is taken. Used well, it is a lead for police; treated as proof, it can end in a wrongful arrest.
How does facial recognition work, step by step?
A typical system in a monitoring center follows this chain:
- Detection: the software finds faces in every video frame.
- Quality and alignment: it discards or flags faces that are blurry, in profile, too small or covered.
- Template: it converts the face into a numeric vector. The photo itself is not compared; that mathematical representation is.
- 1:N comparison: it searches for that vector in a watch list. The US NIST defines "one-to-many" identification as the search that determines whether the person in the photo has any match in a database.
- Score and threshold: each comparison yields a similarity score; above the threshold, an alert is created.
- Human review: an operator compares the live image with the watch-list image and decides whether the match is plausible.
- Logging: the alert, the review and the decision are stored for audit.
How is it different from unlocking a phone?
Unlocking a phone is 1:1 verification: is this person who they claim to be? On the street it is 1:N identification against a list, with distant cameras, poor light and hard angles. That is why the margin of error is larger and human review is essential.
How accurate is it, and why does it make mistakes?
There are two kinds of error: a false positive (the system flags the wrong person) and a false negative (it misses someone who is on the list). In public safety the false positive is the more serious one, because it can lead to detaining an innocent person.
The reference study is NIST's work on demographic effects, which evaluated 189 algorithms from 99 developers and found demographic differentials in most of them. According to NIST testimony, false positive rates across demographic groups often vary by factors of 10 to beyond 100, are higher for women, the elderly and the young, and, because more accurate algorithms make fewer errors, they can be expected to show smaller demographic differentials.
In practice, accuracy also depends on:
- Image quality: lighting, resolution, angle, caps, face masks.
- The quality and legality of the watch list: old photos, or people who should not be on it.
- The chosen threshold: lower means more alerts and more false positives; higher means more missed cases.
How do you prevent wrongful arrests?
Two cases show what happens when controls fail. In Detroit, Robert Williams was wrongfully arrested in 2020 and jailed for more than 30 hours over a face match taken from blurry surveillance images; the city paid $300,000 and, under the settlement, police can no longer arrest people based solely on facial recognition results.
In Buenos Aires, a judge ruled in 2022 that the rollout of the city's Fugitive Facial Recognition System was unconstitutional; according to CELS, the system was used to search for more than 15,000 people who were not on the official fugitive list, with no prior impact assessment, no special legislative oversight commission in place and no reports from the city Ombudsman's Office. According to Perfil, the judge highlighted an Ombudsman report documenting several cases of people wrongly detained because of false positives from the system.
The lessons are clear:
- A match is a lead, not proof. Every alert must go through an operator and then through police verification.
- Watch lists with a legal basis. Only people who belong there, through the process the law defines.
- Every search and every alert, logged. Without logs there is no way to detect abuse.
- Legal framework and prior assessment. Facial data is especially delicate personal data, and in Buenos Aires the lack of a prior impact assessment was one of the failures identified. Each institution must comply with the applicable legal framework in its country.
- Video under state control. The fewer copies leave the institution, the lower the risk.
What options does a ministry or city have?
| Criterion | Axentra OmniSight | Manual monitoring | Generic cloud AI service |
|---|---|---|---|
| Camera coverage | ✓ All, 24/7 | ✗ Only those being watched | Depends on contract |
| Human review of every match | ✓ Always | ✓ Everything is human | Depends on vendor |
| Searches and alerts logged | ✓ Yes | Depends on procedure | Depends on vendor |
| Where video lives | ✓ Own servers | ✓ Local | ✗ Leaves for the cloud |
| Natural-language rules and search | ✓ Yes | ✗ No | Depends on vendor |
| Cross-camera tracking with route | ✓ Yes | ✗ Manual, slow | Depends on vendor |
| Plates, fleet numbers, lettering | ✓ Yes | ✗ By eye | Depends on vendor |
| Blocked, turned or offline cameras | ✓ Automatic alert | ✗ Noticed late | Depends on vendor |
| Existing cameras and VMS | ✓ Any brand, Milestone, Genetec | ✓ Current setup | Depends on integration |
Bottom line: manual monitoring keeps data local but leaves most cameras unwatched; the cloud can offer good analytics, but video leaves the institution. OmniSight combines full coverage with local control and human review.
Why choose Axentra OmniSight?
OmniSight is designed so facial recognition runs with the controls the cases above call for:
- Analyzes every camera in real time, 24/7, instead of relying on an operator watching the right screen, and alerts the operator.
- Face recognition against watch lists with human review of every match. People make the decisions; alerts and searches are logged.
- Reads license plates, fleet numbers and vehicle lettering, useful when a face is not visible.
- Tracks people and vehicles across cameras and generates their route.
- Natural-language rules ("alert me if you see an accident") and natural-language search across archived video.
- Detects blocked, turned or out-of-service cameras, so coverage is not lost silently.
- Works with what you already have: IP cameras via ONVIF and RTSP, brands such as Hikvision, Dahua, Axis, Hanwha, Uniview, Bosch, Avigilon, Verkada and Cisco, and VMS such as Milestone and Genetec, in one unified VMS.
- On-premise: installed on GPU servers in the institution's data center; video never leaves and there is no cloud dependency.
Results confirmed to date: more than 1 million vehicles detected, video searches in about 40 seconds and more than 1,000 alerts sent to operators. Licensing is per processed camera, per year, with floating licenses.
How can you start without risk?
OmniSight offers a free 10-day pilot with up to 10 of your institution's cameras: test human review and search logging on your own cameras before deciding.