Facial Recognition System Data Collection

TL;DR

A global smart-home security provider needed training-ready data that looked like real video footage from consumers and security teams. This included fixed viewpoints, mixed lighting and people moving naturally through different zones of a property. They were already collecting data for home security cameras, but they needed support for more robust facial recognition training.

Defined.ai set up an on-device collection using the customer’s own camera equipment, apps and participant workflow. We then produced a structured dataset of surveillance video clips featuring varied expressions, appearances and activities. The project ramped quickly, with full onboarding and the first usable data deliveries on the first day. The final delivery totaled 119 participants, with each completing a complete station set within a single lighting track.

Our customer: scaling smart-home video for AI face recognition deployments

Our customer is a global provider of consumer home-security technology developing next-generation computer vision for doorbells and property cameras. They needed higher-quality, more diverse training data so their facial recognition system performed reliably across varied household conditions and behaviors.

They also needed a partner that could use their existing devices, apps, and internal participant software. The dataset had to match their AI-powered surveillance cameras workflow to reduce downstream integration tasks.

The challenge: improving behavior analysis from low-light, fixed-view footage

Training facial recognition technology for home-security scenarios is different from “studio” computer vision. The customer needed data showing what a fixed camera actually sees, such a partial faces, motion blur and blocked lines of sight. They also needed to capture uneven lighting, especially at dusk and at night.

Three constraints made this difficult:

Diversity targets and thresholds

The customer required a balanced dataset across demographic groups, while recording individuals to keep identity signals clean for biometric facial recognition.

Lighting realism without over-reliance on night mode

Exterior scenes needed low-light “looks” that still preserved useful RGB detail. This is similar to how an AI outdoor security camera must perform before switching into more aggressive night settings.

Security-relevant activity coverage

In addition to identity, the customer wanted data that supported behavior analysis. Capturing both everyday motion and suspicious behaviors would improve downstream alerts in artificial intelligence for video surveillance workflows.

Our capabilities: rapid, flexible, real-world AI security camera data collection

Defined.ai enabled the customer to extend capacity and broaden diversity without changing their stack:

Workflow flexibility

We can outsource data collection using customer-owned hardware, apps and platforms, rather than forcing a new pipeline. In this case, collection used only client tools for recording and participant processing and delivery.

Ramp-up

Our team moved from onboarding to delivery in just one hour, minimizing back-and-forth while still meeting strict procedural requirements.

On-site execution discipline

Our experience with real-world on-site dynamics—no-shows, reschedules and “day-of” friction—allowed us to build in practical mitigations (like controlled overbooking for night sessions).

Security-camera context expertise

We designed protocols to the client’s exact specifications to reflect how a surveillance camera system behaves in real homes, including fixed fields of view, realistic motion paths and various lighting conditions.

Carrying out all the recording at our Lisbon office came with other benefits. From an operational standpoint, using a single location meant that the changes in daylight could be predicted with greater accuracy. It also allowed us to have full control over the artificial lighting conditions for consistent results. In terms of flexibility, no shows could be covered by our colleagues, and the on-site IT technicians were on hand to promptly resolve any connectivity or hardware issues. Privacy-aware handling of sensitive inputs

We support data collection based on informed decisions and clear participant choice for sensitive biometric data projects, ensuring participants understand the data security conditions and can decline at any point.

The solution: protocol-driven time stamps and two-station, two-lighting-track capture

Defined.ai designed a surveillance-style, static-camera object detection protocol aligned to consumer deployment realities.

Capture design and tooling

• Static camera viewpoints: fixed positions representative of a facial recognition camera at a front door, garage and indoor room.
• Customer platform integration: participant information and recording boundaries were logged using precise start and stop time stamps. This enabled the customer to reliably locate and retrieve the exact video segments associated with each participant session.
• Recording format: footage was captured and stored through the customer’s camera app, consistent with typical surveillance video recording workflows (MP4-compatible downstream handling).

Two lighting tracks with two stations each

Participants completed a complete station set under a single track to avoid incomplete records:

• Daylight track: front door + garage station coverage (two stations).
• Low-light track: front door (outdoor low light) + indoor dark station coverage (two stations).

Low-light is where many security models struggle, so the capture plan treated it as a first-class requirement.

Outdoor low light was recorded under naturally dim conditions to match what an AI outdoor security camera sees around dusk and night. Indoors, darkness was simulated by controlling ambient light replicate realistic night-light camera scenarios, similar to an AI security camera.

This created a spectrum of conditions—usable RGB detail at the edge of low light, plus darker indoor scenes—without turning the dataset into an artificial “night mode only” collection. The approach supports evaluation of artificial intelligence for video surveillance pipelines that must handle shifting illumination, motion blur, and partial occlusions. It also produces footage that is directly comparable across conditions, making it easier to isolate what improves biometric facial recognition performance.

Scripted actions for identity and detection

To improve model robustness for AI face recognition and related security tasks, participants were prompted to produce controlled, repeatable variation:

• Expressions: neutral, smile (with and without teeth), and frown—useful for robustness checks and adjacent AI emotion recognition signals.
• Appearance changes and occlusion: hairstyle changes and face-occluding garments (for example, hoods) to stress-test facial recognition under partial visibility.
• Movement patterns: walking and running through and across frame to support motion-related evaluation and tracking physical activity signals in fixed views, including scenarios relevant to an AI human detection camera.
• Normal vs suspicious activity: everyday entry/exit behaviors as well as “suspicious” actions (for example, attempting to break in, access restricted areas or conceal identity) to strengthen downstream alerting in AI security systems and AI security camera deployments.

The results: usable biometric data at speed and scale

Defined.ai delivered a structured dataset tailored to the needs of consumer security AI technologies. We recorded 119 individual participants, each completing a two-station set with varied lighting.

A key success factor was minimizing disruption to the customer’s existing processes. The collection operated inside the customer’s participant-management environment, with Defined.ai adapting field operations to the client’s tooling and retrieval patterns rather than imposing a separate pipeline. This made it easier to ingest outputs into internal model development and analytics.

We generated fast time-to-value, delivering usable data on the first day, so the customer could begin training their facial recognition system straight away. The footage was AI-ready, achieving outdoor low-light conditions while preserving usable RGB detail. This improved evaluation fidelity for AI surveillance camera performance before more aggressive night settings take over.

Just as importantly, every session was bounded with start/stop time stamps, making it easy to identify clips for repeatable analysis and facial recognition search experiments across lighting, station, expression and occlusion variations.

Overall, the customer received training-ready, surveillance-style video built to strengthen both identification robustness and security-event understanding. Defined.ai helped their teams spend less time wrangling footage and more time improving performance in production-like home environments.