Over the past seven years, I have worked with computer vision for industrial safety across very different operating environments. One lesson has repeated itself: its biggest value is not simply seeing more than a person, but helping an organisation understand risk continuously rather than only when someone observes, reports or investigates it.
At one industrial site I visited, forklifts repeatedly crossed a busy corner of a loading bay throughout the shift. The corner was partially blind, yet a supervisor could observe only a fraction of those movements. Nothing happened, so nothing was recorded or discussed. But the risk was present every time a vehicle and pedestrian came close.
From sampling to continuous observation
A safety walk-through, however well conducted, is a sample. No supervisor can watch every interaction between a worker and vehicle, every restricted zone and every production area across every shift. CCTV increased visibility, but in most organisations it has remained largely passive: footage reviewed after an incident, or monitored by operators already responsible for several screens.
Computer vision changes that relationship. Existing cameras can become continuous operational sensors, identifying conditions such as missing PPE, entry into a hazardous zone, unsafe vehicle-pedestrian proximity, work at height or unsafe interaction with machinery.
In real deployments, this distinction matters. A camera may observe hundreds of normal movements before one interaction creates meaningful risk. The value is not simply detecting the event that becomes an incident. It is learning from the pattern of events that came before it.
Measurement matters more than detection
Serious incidents are relatively rare. Unsafe conditions and near misses occur far more often, yet they have historically been difficult to measure consistently. How many vehicle-pedestrian near misses occurred today? Which crossing creates the most exposure? Does PPE compliance deteriorate on a particular shift? Did an intervention actually reduce risky behaviour?
Safety teams have traditionally answered through observation, audits and employee reporting. All are valuable, but none provides continuous visibility. Computer vision can make observable events measurable over time, changing the discussion from “what happened?” to “what patterns are developing?”
After years of deployments, I have learned that the most useful output is rarely an individual alert. It is the trend behind the alerts: recurring events at one location, at a certain time, during a particular task or after a change in process. That is often where the real safety insight sits.
Not all near misses are equal
More detection does not automatically mean more safety. Too many alerts, without distinguishing what matters, quickly erode confidence.
A missing hard hat in a low-risk area should not carry the same weight as a pedestrian stepping into the path of heavy equipment, someone beneath a suspended load, or a person entering an area where hazardous energy could be released.
This is why Serious Injury and Fatality (SIF) potential matters. EHS teams need to distinguish frequent lower-severity deviations from conditions that could reasonably lead to life-changing injury or fatality.
Computer vision can help surface SIF precursors more consistently, including vehicle-pedestrian interactions, work at height, unsafe proximity to machinery, suspended loads and entry into high-energy zones.
The better question is not “how many violations did we detect?” but “where are we repeatedly seeing conditions with the greatest potential for serious harm?”
Trust is a precondition
Another lesson from deployment is that technical accuracy alone does not determine whether a programme succeeds. Workforce trust matters just as much.
Employees are entitled to know what is monitored, what happens to video, who can access it and how it will be used. These questions should be addressed before deployment. Organisations should decide retention, access, whether identity is required and where processing happens. Wherever possible, the focus should be hazardous conditions and patterns rather than profiles of individuals.
How findings are used matters even more. If repeated vehicle-pedestrian proximity events cluster at one intersection, the answer may not be discipline. The cause could be poor segregation, a blind corner, congestion, production pressure or an impractical workflow. Technology identifies the pattern; EHS professionals still have to understand the cause.
What seven years of deployment teaches you
Computer vision is not magic, and industrial environments are unforgiving. Lighting changes. Cameras move. Dust, weather and occlusion affect visibility. Different sites perform the same task differently. A detection that works well in a controlled pilot can behave differently when expanded across dozens of cameras and locations.
That is why I have become cautious about judging a deployment by a demo or headline accuracy number. The real questions are whether the system catches the risks that matter, whether false alerts remain manageable, whether performance stays reliable as conditions change and whether the site acts on the information.
A quiet camera does not necessarily mean a safe area, and a detected event is not automatically a near miss. Judgement remains essential.
From another dashboard to an operational system
The next stage for computer vision in EHS is not simply better detection. It is better integration.
Industrial organisations already operate EHS platforms, maintenance systems, manufacturing systems, access control and operational workflows. Computer vision should not become one more isolated dashboard.
Its value grows when what cameras observe becomes part of how the organisation works: a recurring proximity risk prompting traffic redesign, a machine-zone event feeding an engineering review, or trends in high-potential events becoming part of management safety discussions.
The organisations that gain the most will not necessarily be those with the most cameras. They will be those that connect what the cameras reveal to how safety decisions are made.
After seven years in this field, my view is simple: the technology is the easier part. The difficult part is selecting risks that matter, earning trust, maintaining performance, avoiding alert fatigue and building a process that turns observation into action.
The question worth asking is not “how many cameras can we connect to AI?” It is: “Which risks do we need to understand better, which could cause serious harm, and what will we do differently once we can finally see them?”
