AI Video Analytics: Choosing Use Cases That Deliver Value

What video analytics can do

• Detect and classify objects such as people and vehicles.
• Raise alerts for intrusion into zones or line crossing after hours.
• Count people or vehicles and measure occupancy or queue length.
• Read licence plates at controlled entrances.
• Detect personal protective equipment in defined work areas.
• Speed up investigations by searching recorded video for attributes, such as a vehicle color.

Some capabilities, such as facial recognition, carry much higher privacy risk and deserve separate legal review. Video-based smoke detection can support, but not replace, certified fire detection systems.

Good and poor use cases

Define the use case before buying

• What exactly is the event, and where and when should the rule be active?
• Who receives the alert, and what will they do?
• How many false alarms per shift can operators handle?
• How will success be measured, such as incidents detected earlier or investigation time saved?

Check whether the scene is suitable

• Enough pixels on the target for the task: detection needs far fewer than recognition.
• Lighting at night, including infrared range and glare.
• Camera height and angle that avoid people and vehicles blocking each other.
• Few distractions such as trees, flags, reflections and moving shadows.
• Exposure to weather, dust and insects near infrared lights.

Where the analytics run

Analytics can run on the camera, on a server or in the cloud. Camera-based analytics reduce bandwidth and scale with the cameras. Server-based analytics allow more demanding models on existing cameras but need processing hardware. Cloud analytics reduce local hardware but depend on upload bandwidth and involve sending video off site.

Pilot and measure

1. Choose representative cameras, including difficult night and busy scenes.
2. Run the pilot long enough to cover day, night, weekends and busy periods.
3. Review every alert and classify it as correct or false, and check recordings or staged tests for missed events.
4. Calculate precision (the share of alerts that were real) and detection rate (the share of real events that raised an alert).
5. Tune zones, schedules and sensitivity with operators, then measure again.
6. Compare operator workload before and after.

Governance and privacy

• Document the purpose and why analytics are necessary and proportionate to it.
• Carry out a privacy impact assessment before deployment. NPC Circular No. 2024-02 applies its CCTV requirements to video analytics as well and expects privacy impact assessments to assess and minimize the risks.
• Update CCTV signage and privacy notices.
• Restrict who can view analytics results and metadata, and set retention for them.
• Require human review before any action that affects a person.
• Check vendor terms on whether your footage is used to train their models.

The NIST AI Risk Management Framework offers a structured way to identify, measure and manage risks such as inaccurate results and unfair outcomes.

Operations

• Re-tune rules when scenes change, for example after new racking, lighting or landscaping.
• Keep lenses and infrared illuminators clean.
• Track alert volumes and false alarm rates over time.
• Apply model and firmware updates through change control.