Video analytics
Edge vs server analytics, common analytic types, accuracy and false alarms, privacy, and how to sell analytics honestly.
- Distinguish edge, server and cloud analytics
- Match analytic types to use cases
- Set realistic expectations for accuracy and false alarms
- Address privacy and policy considerations
Where analytics run
| Location | Pros | Cons |
|---|---|---|
| Edge (in the camera) | Scales with camera count, low bandwidth (sends metadata/events), works when the server is down | Limited by camera processing power; features vary by model |
| Server (on-prem GPU/CPU) | Powerful models, central management, works with any camera | Hardware cost, server sizing, needs the stream |
| Cloud | Elastic compute, rapid feature updates | Upstream bandwidth, latency, data residency and privacy questions |
Modern cameras with deep-learning processors (for example Axis cameras running AXIS Object Analytics) can classify humans and vehicles at the edge and send events and metadata to the VMS. The VMS can then use that metadata for alarms and forensic search.
Common analytic types
| Analytic | Typical use | Notes |
|---|---|---|
| Motion detection | Trigger recording or events | Basic; prone to false alarms from weather, lights, animals |
| Object classification (person / vehicle) | Reduce false alarms by ignoring non-human motion | Foundation of most modern analytics |
| Line crossing / area intrusion | Perimeters, restricted areas after hours | Combine with schedules and arming states |
| Loitering / dwell time | ATMs, back doors, retail | Tune dwell thresholds per scene |
| People counting / occupancy | Retail analytics, space utilisation, occupancy limits | Overhead mounting usually most accurate |
| Queue / crowd detection | Retail, transit, stadiums | Business value beyond security |
| License plate recognition (LPR) | Gates, parking, investigations | Needs dedicated LPR cameras (see Camera types) |
| Forensic / attribute search | Find "person in red shirt" or "white van" across cameras | Saves investigation hours; depends on metadata quality |
| Tailgating detection | Detect more than one person per access grant | Best done by combining access control events with video analytics or dedicated sensors |
| Face recognition | Watch lists, access verification | Heavily regulated or banned in some jurisdictions; requires legal review, consent and policy |
Accuracy and false alarms
- Every analytic has detection range, angle, lighting and pixel-density requirements published by the vendor. Design the camera placement for the analytic, not the other way round.
- Measure both missed detections and false alarms in a pilot, day and night, in bad weather.
- A monitored perimeter that generates hundreds of false alarms a night will be ignored or disabled. Operators' trust is the real KPI.
- Use schedules, arming states and multiple conditions (e.g. person + line crossing + after hours) to keep alarms meaningful.
Privacy
- Use privacy masking on areas that must not be recorded (neighbouring windows, keypads).
- Some platforms offer dynamic anonymisation that blurs people in live view while keeping the original recording available to authorised investigators (Genetec offers a privacy-protection module for this).
- Document purpose, retention and access for each analytic, especially biometric ones, and involve legal and privacy officers.
Selling analytics responsibly
Analytics are often over-promised. Tie each analytic to a measurable outcome (fewer false dispatches, faster investigations, occupancy reporting), set expectations with a pilot and agreed success criteria, and be explicit about conditions where it will not perform. This builds the trust that wins the next project (see Habit 4: build trust with honesty).
Common mistakes
- Promising 100% detection.
- Placing cameras for general coverage and then expecting analytics to work.
- Deploying face recognition without a legal and policy review.
- No pilot or success criteria, so nobody can say whether the analytic works.
- Ignoring night and weather performance.
Check your understanding
1.What is a key advantage of edge analytics?
2.Which analytic is most effective at cutting false alarms from swaying trees?
3.What should be agreed with the customer before deploying analytics?