The oil and gas industry in the past decade has invested heavily in digital technologies to improve drilling efficiency, equipment reliability, and production performance. Automated drilling controls, predictive maintenance, and real-time operational dashboards have become increasingly common across modern assets.
Safety, however, has remained comparatively dependent on periodic human observation for a long time.
The latest safety performance data by the International Association of Oil and Gas Producers (IOGP), drawn from member companies across 87 countries, shows 946 lost-time injury cases reported in 2024. Incidents categorized as “caught in, under, or between” constituted 20% of the reports, and “slips and trips” accounted for another 22%.
On the fatality side, “struck by” incidents made up 13% of the 32 fatalities recorded. These reported incidents were not exotic failures but were traced back to the same handful of mechanisms where a worker is in the wrong place for a few seconds that repeated year over year.
There are always two sides to a rig safety story. The first is the one that everyone expects and plans on. For example, the toolbox talk, permit-to-work review, and red-zone briefing before the start of a trip.
Nearly every serious safety program on a modern rig is designed to handle the first one. Very few are built to intercept the next, which is the 40 seconds when the crew’s attention is on the string, not on the two people who have drifted into the swing radius of the tongs because the shortest path to the next task happens to run through it.
That gap between what health, safety, and environment (HSE) planning covers on paper and what actually happens on the live rig floor is where computer vision is starting to find a role. This article looks at the operational choices that determine whether that role gets filled well: what worked, what didn’t, and what other operators evaluating this category should do differently as a result.
Retrofitting Safety Intelligence Onto Existing Infrastructure
The first operational choice to make is what to build the system on, and the answer that holds up across sites is: mostly what’s already there.
Most rigs and offshore facilities run extensive closed-circuit television networks installed for security and remote operations monitoring long before any safety-automation project is proposed.
Retrofitting an artificial intelligence (AI) analysis layer onto those existing feeds, rather than installing new cameras, is what makes this category viable without a capital project attached to it, and it’s the default starting point for most deployments now.
The retrofit decision isn’t binary, though, and treating it as such is where several deployments lose time. Camera placement on most rigs is designed for security sightlines, not for the specific angles a safety model needs, and weather, spray, low light, and vibration degrade image quality in ways a facility built to information-technology-grade camera specifications was never designed to handle.
Successful deployments, therefore, begin with existing infrastructure, evaluate operational performance, and improve coverage selectively where evidence demonstrates genuine observational gaps.
This incremental approach reduces implementation complexity while ensuring that engineering decisions are guided by operational requirements rather than assumptions.
Calibrating the AI Systems for Operational Reality
A second operational decision to make is determining when the computer vision system is ready to scale.
The vision-based analysis layer separates analysis into two stages.
The first is object detection. Identifying that a person, a suspended load, or a piece of equipment is present in a given camera frame. This is a mature, well-solved computer vision task, typically handled by real-time object detection models like the YOLO family of architectures, which is a common industry baseline, running against each camera feed.
The second stage is scenario-based analysis: figuring out whether a detected person is in a defined red-zone boundary, whether a load is moving rather than just sitting there, and whether a hard hat is missing or a harness is not fastened or if a strap was just being adjusted for a few seconds.
The challenge in this stage is to store the spatial and temporal context from one frame to the next, with most of the false positives and missed detections that affect crew confidence in the system coming from this stage.
The rig operators who get better results treat the first several weeks of any deployment as a mandatory tuning period against that specific site’s lighting, camera angles, and crew movement patterns before expanding coverage to additional zones.
For instance, at an Abu Dhabi offshore site, the 80% reduction in safety violations didn’t arrive as a single step change. It was built over the deployment period, and the steepest decline came once supervisors started using the system’s own zone-by-zone violation data to identify which boundaries and shift patterns were generating repeat incursions, a pattern that incident reports alone never revealed.
Routing the Alert, Not Just Generating It
One of the most important deployment decisions is not the detection model itself but how alerts are incorporated into day-to-day operations.
Successful deployments avoid creating another dashboard that supervisors must monitor alongside existing control systems. Instead, alerts are integrated into the operational workflow already used to manage the shift.
Computer vision detects the hazard, while agentic AI validates the event, assigns a risk level, attaches the relevant video evidence, and routes a concise summary to the responsible supervisor or HSE team. This reduces the time spent reviewing footage and ensures that notifications contain enough context to support immediate action rather than further investigation.
Another lesson is to design the system around response time rather than detection accuracy alone. On offshore assets and remote facilities, edge/cloud architectures are increasingly preferred because video is analyzed locally on industrial edge devices, while only event metadata, short video clips, and reports are synchronized with central HSE platforms.
This allows hazard detection and alerting to continue even where satellite connectivity is intermittent while a complete digital record is maintained for compliance and incident review.
The objective is not to generate more alerts but to create a response workflow that enables supervisors to intervene while the unsafe condition still exists.
What Successful Deployments Have in Common
As computer vision deployments have matured across upstream operations, one observation has become increasingly consistent: Deployment methodology has a greater influence on outcomes than the underlying AI model.
The Incident Statistics Program (ISP) 2026 Q1 report, covering onshore and offshore operations across nine geographic locations, recorded 166 incidents. Set against the U.S. National Safety Council 2024 estimates of $48,000 for the average cost of a medically consulted workplace injury, that single quarter represents a cost large enough to make “detect it sooner” an engineering and HSE priority rather than an optional add-on.
Operators achieving the strongest results typically begin with a single high-risk workflow—such as pipe handling, lifting operations, or work at height—rather than attempting sitewide coverage from day one. Early deployment is treated as an operational calibration exercise, allowing supervisors to validate detections, refine exclusion zones, and establish response procedures before expanding to additional work areas.
Infrastructure decisions also matter. Offshore facilities increasingly favor edge or hybrid architectures to ensure hazard detection continues despite limited bandwidth or intermittent satellite connectivity. Processing video locally while synchronizing only alerts and event metadata reduces latency without sacrificing centralized reporting or compliance records.
Perhaps the most important lesson is organizational rather than technical. Computer vision delivers the greatest value when integrated into existing HSE processes instead of operating as a parallel system. AI-generated events should inform toolbox talks, permit-to-work reviews, shift handovers, and incident investigations using the same governance already established for conventional safety observations.
This architecture reduces bandwidth consumption, improves response time, and enhances operational resilience across oil fields.
Computer Vision as the Next Layer of HSE in Oil Fields
The strongest lesson from early deployments is that computer vision is most effective when it strengthens existing HSE systems rather than operating alongside them.
None of the computer vision deployments replaces the HSE management systems already built up over decades. Vision-based monitoring is an input into those systems, not a parallel one. A red-zone alert should feed the same logging and escalation path an observer’s report would, not create a second, disconnected record HSE has to reconcile after the fact. Where it earns its place is in providing visibility HSE teams didn’t have before, revealing near misses that would previously have gone entirely unrecorded.
It’s also worth being honest about what this doesn’t change. HSE judgment about how to design a job, sequence a task, or handle a genuinely ambiguous risk still sits with people who understand the operation. What changes is how much of the routine, moment-to-moment hazard exposure the people making that judgment are blind to.
