From Passive Video Surveillance to Scalable HSE Intelligence: Field Lessons from AI-CCTV Validation in a Harsh Desert Oil Field

Abstract

Abstract At remote oil and gas facilities, closed-circuit television still relies on a person watching live views and searching stored footage after an event. As camera counts rise, that arrangement misses brief, distant, partly hidden, or visually ambiguous events more easily. This paper examines field validation of an artificial intelligence (AI)-enabled closed-circuit television (CCTV) system, referred to here as AI-CCTV, in a large onshore field exposed to intense sunlight, glare, dust, long viewing distances, low-light conditions, sparse infrastructure, and a mix of fixed optical and moving infrared or pan-tilt-zoom (PTZ) cameras. The workflow combined secure video ingestion, computer-vision detection, multi-object tracking, context-preserving evidence, structured vision-language assistance, a monitoring interface, and human alarm disposition. Representative scenarios covered personnel presence, personal protective equipment compliance, portable gas-monitor recognition, liquid-like anomalies, falls, selected unsafe behaviors, and restricted-area monitoring. The evidence shows functional feasibility and recurring failure mechanisms, but it does not justify a universal accuracy claim. In many cases, the limiting factor was whether the scene made the target observable, not which model was selected. Small or distant targets required changes to imaging, camera position, tracking, and evidence capture; detector tuning alone rarely solved the problem. Moving cameras required a separate analytical contract for camera state, dwell time, target reacquisition, and return-to-route behavior. Generic models were useful starting points, but field use still required local hard negatives, controlled configuration, shadow-mode testing, and a release process that could be audited. Nuisance alarms were treated as a system cost and managed with temporal persistence, context, semantic checks, risk rules, and human review. These observations are organized into six scale-up gates covering operational value, observability, local evidence, infrastructure, controlled use, and lifecycle ownership. A three-wave rollout and a minimum metric set are also defined. The result is a governed health, safety, and environment (HSE) decision-support service, not an autonomous protection layer.