Hazard Detection

Most site hazard detection today is computer vision applied to a specific catalogue of dangerous situations, plus a layer of sensors for things a camera cannot see. It helps to look at the common detection classes concretely, because each has its own strengths and its own ways of failing.

Missing protective equipment. The classic example: a model trained on many labelled images learns to find people and then check whether each is wearing a helmet, a high-visibility vest, sometimes a harness. It is genuinely useful for a basic, common lapse. It also struggles with the obvious: a helmet held rather than worn, an unusual colour or type, poor light, a distant or partly hidden worker.

Danger-zone incursion. A zone is marked in the camera's view - around an excavation, under a crane, near live plant - and the system flags when a person enters it. This turns a fixed camera into a tireless watcher of a boundary. It depends entirely on the zone being defined correctly and on the camera actually covering it.

Work near an unprotected edge or opening. The system watches for a person close to a fall risk - a slab edge, a shaft, an opening without a guardrail. Because falls from height are among the deadliest construction hazards, early warning here is valuable; because edges and openings change constantly on a live site, it is also hard to keep accurate.

Proximity to moving plant. Struck-by incidents - a person hit by a moving machine or load - are a major killer. Vision, and increasingly proximity and anti-collision sensors on the plant itself, can warn when a person and a machine get dangerously close. Sensors handle the blind spots a camera misses.

Falls, environment and structure via sensors. Beyond vision, wearables can detect a fall or a worker's location; environmental monitors watch gas, dust, heat and noise; instruments on scaffolding, formwork or an excavation watch for movement that can precede a collapse.

Across all of these, notice the pattern. Each detector converts a slice of physical reality into a flag: here, now, this may be dangerous. It measures and matches; it does not judge significance or decide what to do. That job - and the responsibility for it - belongs to a person, and the rest of this lesson is about why that division must never blur.

The core of visual hazard detection is object detection: a model, trained on large numbers of labelled images, learns to locate and classify things in a frame - a person, a helmet, a vehicle - and returns each with a bounding box and a confidence score between roughly nothing and near-certainty. On top of that sit simple spatial rules: if a person box has no helmet box on it, flag missing protective equipment; if a person box falls inside the exclusion-zone polygon, flag an incursion; if a person box and a machine box overlap or nearly touch, flag proximity. Sensors work similarly in spirit: a stream of readings, a threshold, an alarm when it is crossed. None of this involves understanding; it is pattern-matching and geometry, which is exactly why it is fast, tireless and consistent - and exactly why it breaks in predictable ways.

The first break is the confidence threshold. The system only flags when its confidence passes a set level. Set the threshold high and it stays quiet unless very sure - fewer false alarms, but more missed hazards (false negatives). Set it low and it catches more, but cries wolf - more false alarms (false positives) that train people to ignore it. There is no setting that removes both; every deployment is a trade-off, and on a safety system both errors are dangerous in different ways.

The second break is the world outside the training data. The model has learned from particular images; it degrades on what it has not seen. Rain, dust, glare, darkness, an unusual vantage, a hazard type or piece of equipment it was never trained on, a camera knocked out of position, a scene too cluttered to parse - any of these can produce a confident wrong answer or, worse, silence. And it can only ever flag what its cameras and sensors actually cover; the hazard forming in an unmonitored corner is invisible to it.

The third break is subtler: the system has no sense of consequence. To the model, a worker one step from a fatal edge and a worker one step from a harmless kerb can look almost identical; it has no idea which one matters. It cannot weigh severity, context or intent. That judgement is precisely what a human brings - and precisely why the flag must go to a human, not straight to a conclusion. Understanding these failure modes is not pessimism; it is the operating manual for using the tool safely.

The single most important habit in using hazard detection is to read every output as a prompt, not a verdict - because the system can be wrong in both directions, and each kind of wrong is dangerous in its own way. Lay the four outcomes out plainly. When there is a real hazard and the AI flags it, that is a true alarm - useful, provided a human verifies and acts. When there is no real hazard but the AI flags anyway, that is a false alarm - annoying, and dangerous in bulk, because a stream of false alarms teaches everyone to swipe the alerts away, so the system stops working precisely when a real one arrives. When there is no hazard and the AI stays quiet, that is a correct all-clear - but note carefully that silence is never, by itself, evidence of safety. And when there is a real hazard and the AI stays quiet, that is a missed hazard - the false negative, the outcome that on a construction site can be fatal.

That last box is why over-trust is so dangerous. If a team comes to believe the AI catches everything, they stop looking for themselves, and the one hazard the model never learned - the novel situation, the thing outside the camera's view, the confident error - passes unchecked straight to an incident. This is automation bias in its most lethal form: the confident, usually-right machine lulling humans out of the vigilance that is the last line of defence. The defence against it is structural, not attitudinal: keep human checks, physical controls and enforcement fully in place and independent of the AI, so that the AI's silence is never load-bearing.

So the discipline is simple to state and must be lived on site. Treat a flag as: "something here may be dangerous - go and check." A person verifies what is really happening; if it is real, they act - stop the work, remove the person, fix the control, protect the area, record it, follow up. If it is a false alarm, they dismiss it and, over time, tune the system so it stays credible enough to be trusted. Either way, a human closes the loop, and a human - never the model - carries the responsibility for the call. A flag that no one verifies and acts on has protected no one; the value of the whole system lives entirely in the human action it prompts.

Put the pieces together into a way of working. First, place detection where the watching gaps are, not everywhere for show: the blind corner, the busy edge, the crane radius, the night shift - the places a stretched team genuinely cannot watch continuously. That is where an extra tireless eye adds real value; scattering cameras where supervisors already look mostly adds noise.

Second, design the human loop before the technology. Decide who receives each flag, how fast they can verify it, what actions are available, and how it gets recorded and followed up. A flag with no fast, owned path to human action is theatre. The system's worth is measured not in alerts generated but in hazards a person removed because of them.

Third, tune for trust. Watch the false-alarm rate as carefully as the miss rate, because a system that cries wolf is a system that will be ignored, which is a safety failure dressed as a working deployment. Adjust thresholds, zones and rules with the site's real conditions in mind, and keep the people who receive the flags involved in that tuning.

Fourth, and above all, never let detection replace a human control or become load-bearing for safety. The guardrail, the exclusion zone, the toolbox talk, the competent supervisor, the enforced rule - these stay fully in place, independent of the AI, because the AI will sometimes be silent when it should not be. Its job is to catch more, earlier; its job is never to be the thing that would have caught it.

The honest summary, then: hazard detection is a real and valuable capability that turns a flood of site imagery and sensor data into timely prompts a human can act on, and for a dangerous, hard-to-watch industry that is worth having. But it is a prompt-generator, not a safety authority. It can be wrong both ways, it is blind outside its data and its sensors, and it has no sense of consequence. Every binding safety decision and duty - what is truly dangerous, whether to stop work, who is responsible - stays with the responsible site management, the qualified safety professionals and the governing law (NBC India, IS standards, construction-safety and labour law). The next lesson turns from safety to quality, where the same discipline applies to catching defects.