Insights

WA facial recognition: the systems behind the camera

On 22 June 2026, Western Australia Police launched a five-month mobile trial of live facial recognition (LFR) - the first operational deployment of its kind by any Australian law enforcement agency. In the first week, a single marked van equipped with NEC NeoFace cameras scanned 131,478 faces across crowded public spaces in Perth, generated 33 alerts, and contributed to 19 arrests. Two were false positives.

The public debate has focused on privacy and civil liberties. That debate is necessary. But it has largely missed the more consequential question for anyone working in law enforcement technology: what does the system architecture actually look like when you wire live biometric detection into existing police infrastructure?

From pixel to alert: the detection pipeline

NEC’s NeoFace system does not work like CCTV. A standard camera records footage for later review. LFR processes every face in the camera’s field of view in real time, whether or not the person is on any watchlist.

The pipeline has four stages. First, the camera detects individual faces and segments them from the background. Second, each detected face is aligned and normalised to compensate for lighting, head angle, and resolution - NEC’s documentation states NeoFace can work with as few as 24 pixels between the eyes. Third, the system measures geometric relationships between facial landmarks: eye spacing, nose width, cheekbone distance, jawline contour. These measurements are encoded into a mathematical feature vector - a biometric template that represents the unique geometry of that face. Fourth, the template is compared against the watchlist.

The WA trial’s watchlist contains approximately 4,000 entries: people with outstanding arrest warrants, registered sex offenders, missing persons, and individuals assessed as a risk to themselves or others. When the similarity score between a probe face and a watchlist entry exceeds a pre-set confidence threshold, the system generates an alert on the operator’s screen in the van.

NEC’s algorithm was ranked the world’s most accurate in 1:N identification by the US National Institute of Standards and Technology (NIST) in April 2025, with an authentication error rate of 0.07 percent against a database of 12 million. The WA trial’s 2 false alerts from 33 total alerts in its first week gives a false-alert rate of roughly 6 percent among actionable alerts - meaning officers were directed to approach an innocent person once for every 15 or 16 valid hits.

What happens after the alert

An alert is not an arrest. Commissioner Col Blanch has stated that algorithmic output alone cannot justify the exercise of any police power. Officers must physically observe the individual, conduct independent verification, and make their own assessment before any action is taken.

System-side integration is where the real architecture question lives. In a mature deployment, a facial recognition alert does not exist in isolation. It connects to the police agency’s Computer-Aided Dispatch (CAD) system, which logs the incident location, time, and responding officers. From there it links to the Records Management System (RMS), holding the individual’s criminal history, prior interactions, and current bail or warrant conditions. A further connection may reach intelligence databases that aggregate cross-jurisdictional data from the Australian Criminal Intelligence Commission (ACIC) or state-level intelligence holdings.

In the WA trial, the watchlist itself must be sourced and maintained from somewhere. Entries come from the RMS: outstanding warrants and missing persons reports, plus registered child sex offenders under the Community Protection (Offender Reporting) Act 2004. Each entry requires a source photograph with sufficient quality for template generation. Entries are added and removed as warrants are executed, offenders register or deregister, and missing persons are located.

Data flow is one-directional for non-matches. Faces that do not trigger an alert are pixelated on the operator’s screen and the biometric data is deleted immediately. No biometric templates are retained for non-matching faces, according to WA Police. Yet the OIC WA has confirmed that even this momentary processing constitutes “collection” under the PRIS Act, regardless of how briefly the data exists.

The integration question nobody is asking

A mobile van is one deployment model. But the technical architecture is not van-dependent. NEC NeoFace integrates with fixed CCTV networks, body-worn cameras, and fixed infrastructure cameras. The software layer that processes faces, generates templates, and compares against watchlists is the same whether the camera is mounted on a van, a lamp post, or a building facade.

What makes this interesting from a systems integration perspective is the automation layer that sits between the detection and the response. A manual process is not what we are talking about here. The system cross-references the face against the watchlist in milliseconds. But the downstream workflow - logging the alert, checking the individual’s history, dispatching officers, recording the outcome - is where the real integration work happens.

Picture a fully wired deployment. An LFR alert triggers automated enrichment: the CAD system receives the alert, pulls the individual’s last known address and associated vehicles from the RMS, and pushes a tactical summary to the responding officers’ mobile devices before they reach the scene. The intelligence database checks whether the individual is linked to any ongoing investigations, any associates, any prior encounters with specific units. Operators see a consolidated brief on screen, not just a face and a name.

This is not hypothetical. It is the architecture that London’s Metropolitan Police has been building toward since 2016. The Met’s LFR system is integrated with their Command and Control system, their custody databases, and their intelligence holdings. When a match fires, the system populates an alert package that includes the individual’s custody photograph, criminal history, and any active intelligence markers. The operator reviews the package and decides whether to dispatch.

The data management problem

The PRIS Act’s timing creates a specific challenge for system architects. The Act requires agencies to evaluate and demonstrate how they manage risks of bias, harm, and discrimination in automated systems. High-risk projects require a formal Privacy Impact Assessment (PIA). WA Police completed and published a PIA before the trial launched.

But the PIA covers the trial as designed - a mobile van with immediate deletion of non-match data. The harder question is what happens when the architecture scales. If LFR cameras are mounted on fixed infrastructure and integrated with the state’s broader CCTV network, the data management obligations multiply. Every face scanned becomes a collection event under the PRIS Act. Every template generated requires a lawful basis for retention. Every alert requires an audit trail.

The Office of the Information Commissioner WA confirmed on 9 July that it “was not invited to participate in any formal consultation process with WA Police and did not provide input into the design of the trial.” The OIC WA is now actively monitoring the trial and has stated that its observations will directly inform broader biometrics policy in Western Australia.

For system integrators, this means the governance layer cannot be an afterthought. The data retention policy, the audit logging, the template deletion schedules, the access controls on matched alerts - these are not compliance extras. They are architectural requirements that must be baked into the integration from day one.

Real-time insights and operational intelligence

The other side of the architecture coin is what happens with the aggregate data. Individual alerts are useful. Pattern analysis is more useful. When an LFR system scans 131,000 faces in a week and generates 33 alerts, that dataset contains information about crowd density, movement patterns, and repeat detections that has nothing to do with the watchlist.

In a mature deployment, this data feeds into a command dashboard that gives police commanders real-time situational awareness. Which locations are generating the most alerts? Are there individuals being detected repeatedly across different deployments? What does the detection timeline look like across a shift - are there predictable peaks? This is the kind of operational intelligence that changes how resources are allocated.

The UK experience is instructive. London’s Metropolitan Police deployed LFR 231 times in 2025, scanning approximately four million faces. The technology has been operational in the UK for nearly a decade. In 2020, the UK Court of Appeal ruled that South Wales Police’s 2017-18 trial violated human rights - not because the technology did not work, but because the legal framework around it was insufficient.

WA Police have stated the trial carries no commitment to permanent rollout, covert deployment, or integration with the state’s broader CCTV network. But the technical capability for that integration already exists. The question is whether the legal and governance framework will be built before the technology is expanded, or after.

With 45 percent of Australians now identifying facial recognition as their greatest privacy risk - up from 27 percent in 2023 - the public is paying attention. The PRIS Act’s first real test is happening on the streets of Perth right now.

Sources:

Some of the content on this site may have been generated by AI tools, with human oversight but without detailed human review. We are human and our oversight is not perfect, so there may be mistakes or inaccuracies. Please verify any critical information independently before relying on it.