Computer Vision Deployment

Run the economics before you buy, not after. Five sources of economic value: safety and incident cost, shrinkage and loss prevention, quality and rework, throughput, compliance and insurance. Cost structure: hardware, integration, licensing, annotation, retraining, support — with typical proportions, because customers almost always underestimate the last three.

The core idea of the module: CV does not create value, it creates a signal. Value is created by the changed response process. If nobody acts on the alert, ROI is zero at any model accuracy.

Artefact: ROI Calculator (Excel) with three industry scenarios.

A diagnosis: is your site physically ready? Cameras: resolution relative to distance and object size, fields of view, mounting height, FPS, IR and low light, thermal. What can be reused from existing CCTV and what has to be replaced — typically 30–60% of the fleet is unusable.

Network and bandwidth. Edge vs on-prem server vs cloud — with a selection matrix by latency, traffic cost and data localisation requirements. Separately — climate: dust, humidity, temperature, vibration, salt fog on the coast. For teams in the Gulf and Southeast Asia this is not an abstraction; it is why hardware fails in month three.

Artefact: Site Readiness Checklist, 40–50 points.

The domain gap in concrete terms: different workwear and PPE, different helmets and colour codes, different vehicles and equipment, different anthropometry, different lighting and weather, different site layouts. How much data you actually need, collection timelines, the cost and organisation of annotation, annotation quality and inter-annotator agreement, synthetic data — where it helps and where it hurts.

A dedicated block: who owns the data and the trained model. A contractual question that has to be settled before the pilot but is usually raised after — and it damages the vendor relationship.

Artefact: Data Collection Plan with volume, timeline and budget calculations.

Success criteria get written before the start. Metrics: false alarms per shift, missed events per week, operator response time — and how they relate, so that you can talk to a vendor as an equal. Baseline: how to measure what things were like before, without which the effect cannot be proven. Pilot duration, and why two weeks is not a pilot. Scope boundaries and protection against scope creep. Acceptance criteria in the contract.

Artefact: Pilot Charter, two pages.

How an alert turns into an action. Where the signal goes: SOC, control room, a supervisor’s mobile app, a turnstile, a line stop. Integration with MES, WMS, ERP, access control and EHS systems. Escalation matrix and SOP: who responds, within what time, what gets logged. Dashboards and reporting for management.

Alarm fatigue — how the sensitivity threshold kills the system when engineers set it instead of operations.

Artefact: Response SOP Template and escalation map.

A conversation with the workforce that does not end in resistance. The perception of total surveillance, working with frontline staff, the role of unions and works councils where they are strong. Positioning: the system protects the worker rather than punishing them — and how to prove that in practice, for example with anonymised analytics instead of individual violations in the first phase.

Operator training. Supervisor KPIs, and why they must not be tied to violation counts in the early months. Cultural adaptation: the line of acceptable monitoring falls very differently from country to country, and this has to be said out loud rather than assumed to be universal.

Artefact: Stakeholder Map and communication plan.

The difference between anonymised detection (a person in frame as an object) and biometric identification — legally these are two different worlds, and most of the risk disappears simply by not crossing into biometrics. Data residency and cross-border transfer. Retention policy. Notifying employees and visitors. Bias and model fairness.

An overview of regimes by region: GDPR as the reference point, the PDP Law in Indonesia, PDPA in Singapore, PDPL in Saudi Arabia and the UAE, sector-specific rules. This module gives you orientation, not a legal opinion. Regulation changes, and the final check is done by local counsel.

Artefact: Privacy Impact Pre-Screen, a 15-question screener.

What to do once the pilot has worked. Model drift, and why accuracy drops after six months. Quality monitoring in production. The retraining loop and what it costs. Standardising configurations across sites. TCO over three to five years. What to keep in-house and what to hand to the vendor. SLAs, support, vendor lock-in and exit strategy. Governance: who inside the company owns the CV platform.

Artefact: Scale-Up Roadmap across three horizons.

A full module, not an appendix at the back. Four to five post-mortems: a pilot with no baseline, cameras at the wrong angle, alerts with no response procedure, workforce resistance, a model that did not survive the change of season. Every post-mortem in the same format: what was planned → what happened → root cause → what should have been done differently.