Computer Vision in facilities management: sensors, safety, and space utilization

Facilities leaders are under sustained pressure: real-estate portfolios need to justify every square metre, safety teams face stricter inspection regimes, and building operations budgets are scrutinised like never before. Computer Vision—the branch of AI that interprets video and image feeds in real time—offers a practical path to addressing all three pressures without adding headcount. This listicle is for heads of facilities, workplace operations directors, and the IT leads who support them. It ranks eight proven use cases by operational maturity, explains what each requires to work, and ends with a vendor-evaluation checklist.

Why Computer Vision is gaining traction in facilities now

Three converging forces are accelerating adoption. First, the cost of IP cameras and edge inference hardware has fallen to the point where retrofit deployments are financially viable in mid-sized buildings, not just new-build flagship campuses. Second, hybrid work has made manual occupancy surveys obsolete: badge data tells you who entered the building but not where people actually sit, gather, or move. Third, regulators in multiple jurisdictions have sharpened requirements for documented safety monitoring—creating liability exposure for organisations that still rely on periodic human walkthroughs.

Unlike IoT seat sensors, Computer Vision derives multiple signals from a single camera feed. One device near a conference room can simultaneously report occupancy count, detect whether personal protective equipment (PPE) is worn in an adjacent equipment bay, flag an obstructed fire exit, and feed cleaning-dispatch logic. That sensor consolidation argument is proving persuasive to facilities procurement teams wrestling with integration complexity.

The ranking criteria

The eight use cases

Ranked comparison: use cases by maturity and deployment effort

Vendor categories to evaluate

• Workplace intelligence platforms: Software-first vendors that ingest camera, badge, and sensor data to produce occupancy and utilisation dashboards. Often offer plug-ins for major IWMS and desk-booking systems. Best for use cases 1 and 2.
• Industrial Computer Vision platforms: Purpose-built vision AI tools that support custom model training and deployment on edge hardware. Strong for PPE, anomaly detection, and safety use cases (3, 7, 8). Evaluate model management, retraining workflows, and edge runtime options.
• Video analytics and security intelligence vendors: Established players in physical security who have extended their platforms to include facilities-relevant models (tailgating, loitering, egress monitoring). Best for use cases 5 and 6. Assess whether their facilities modules are native or bolted on.
• Cleaning and facility services management software: CAFM and cleaning-route optimisation platforms that accept occupancy data as a trigger input. The CV layer typically comes from a partner or separate vendor; evaluate the API quality of the integration. Best for use case 4.
• Edge AI hardware and inference vendors: Providers of edge compute devices (GPU-enabled cameras, on-prem inference boxes) that run CV models locally, keeping video data on-site. Critical where privacy regulations or data-sovereignty requirements rule out cloud video streaming.
• Integrated facilities management (IFM) service providers with AI practices: Large IFM contractors who bundle Computer Vision tooling with managed-service delivery. Reduces integration burden but reduces flexibility. Evaluate whether the IP and data remain with the customer.

What to ask in vendor demos

1. Where does video processing happen? Ask the vendor to show the data flow: which frames leave the building, which are processed on-device, and where inference results are stored. This is a privacy and data-sovereignty question, not just a latency question.
2. How is the model retrained when conditions change? Building layouts, PPE standards, and equipment configurations change. Ask specifically how you submit new training images, how long retraining takes, and who controls the retraining pipeline.
3. What accuracy benchmarks apply to your specific environment? Vendor accuracy figures quoted in marketing materials are typically from benchmark datasets, not your building. Ask for evidence from a comparable deployment: similar lighting conditions, camera resolution, and density of people.
4. How does the system handle privacy and anonymisation? For occupancy and safety use cases, facial recognition is rarely required and frequently prohibited. Ask whether people are tracked as silhouettes, skeletal poses, or facial identities—and what controls exist to enforce anonymisation.
5. What does the alert-to-resolution workflow look like? A Computer Vision system that generates alerts with no structured routing to a responsible person produces alert fatigue quickly. Ask for a walkthrough of how an alert moves from camera detection to facilities-team action in their platform.
6. What integrations ship out of the box versus requiring custom development? Specifically ask about your IWMS, CMMS, access-control, and cleaning-management systems. 'We have an API' is not the same as 'we have a tested connector to your specific platform version'.
7. How is system performance monitored over time? Camera drift, lighting changes, and seasonal variation degrade model accuracy silently. Ask whether the platform monitors confidence scores over time and surfaces degradation before it becomes a false-negative problem.

Common pitfalls

• Treating existing CCTV as a free baseline. Legacy security cameras are positioned and configured for human review, not machine inference. Angles, resolution, frame rate, and lighting are often inadequate for occupancy counting or PPE detection without hardware remediation. Budget for a camera audit before committing to a solution.
• Deploying without an alert management process. Computer Vision systems can generate high volumes of alerts. Without a defined triage workflow—who receives the alert, within what SLA, and how resolution is logged—alert fatigue sets in and staff begin ignoring notifications within weeks.
• Conflating occupancy with utilisation. A room where one person sits for ten minutes registers as occupied. Utilisation analytics require sustained occupancy above a meaningful threshold. Buyers who skip this calibration step end up with dashboards that overstate space usage and support the wrong portfolio decisions.
• Underestimating model maintenance. Initial model accuracy reflects the conditions at deployment time. Seasonal lighting changes (winter daylight hours), building reconfiguration, new PPE suppliers, and camera hardware replacements all degrade model performance. Budget for ongoing model operations, not just initial deployment.
• Selecting a vendor without a data-ownership clause. Some platforms retain rights to use anonymised video data or inference outputs to improve their models. In regulated sectors, this may be unacceptable. Review the data processing agreement before procurement, not after go-live.

Buyer's evaluation checklist