Manufacturing / Behaviour
Behaviour analytics on the factory floor
Behaviour analytics detects defined patterns of movement rather than objects: entering a machine danger zone, travelling the wrong way down an aisle, or leaving a load unattended. It works when each rule is written specifically. Generic anomaly detection, which promises to flag anything unusual, does not survive a real shop floor.
Specific rules work, generic anomaly detection does not
Behaviour analytics is sold in two very different forms, and the distinction determines whether a deployment succeeds.
Rule-based detection answers a question you wrote down: did a person enter this polygon while that machine was running? Did a forklift travel south in a northbound aisle? Was a load left in this zone for more than fifteen minutes? Each is checkable, tunable and explainable to a supervisor.
Generic anomaly detection promises to learn normal and flag deviation. On a factory floor, normal legitimately includes shift changeover, maintenance shutdowns, product changeovers, monsoon-season routing changes and audit days. The system flags all of them, an operator cannot tell why any given alert fired, and there is nothing to tune because the rule was never written down. This is the most common reason behaviour analytics projects are abandoned.
The practical test before buying: can the vendor state the rule in one sentence a supervisor would understand? If the answer is that the model learns what is unusual, expect an alert stream nobody can act on.
Rules that hold up in a plant
| Rule | What it detects | Depends on |
|---|---|---|
| Machine danger zone entry | A person inside a defined exclusion zone | Accurate zone geometry, and ideally a machine running signal to suppress alerts during shutdown. |
| Wrong way movement | Direction of travel against designated flow | A clear camera view along the aisle axis. |
| Person and vehicle proximity | Pedestrian close to a moving forklift | Ground plane calibration, otherwise image distance is not physical distance. |
| Unattended load | An object stationary in a zone beyond a threshold | A dwell threshold set to the real operational tolerance. |
| Restricted area occupancy | More or fewer people than permitted | Reliable counting at entry and exit points. |
| Loitering near a control panel | Prolonged presence where none is expected | Tight zones. Prone to noise if drawn generously. |
Danger zone rules and the machine state signal
Worked example: why the machine signal changes everything
Entries while the press is running: 2
Entries during setup, cleaning and maintenance: 38
Without a machine running signal: 40 alerts per day, 5% meaningful
With a machine running signal: 2 alerts per day, both meaningful
The camera cannot tell whether the press is energised. It sees a person in a polygon. Taking a running signal from the machine or the line PLC and using it to gate the rule is usually a small integration and the single largest precision improvement available on this class of rule. Where no signal is available, a schedule tied to production hours is a weaker approximation of the same idea.
This generalises. Most behavioural false alerts come from the system lacking context that the plant already has somewhere: production schedules, shift rosters, maintenance windows, work permits. Feeding that context into rule gating does more than any model improvement.
Ground plane calibration
Proximity rules deserve a specific warning. Two people who appear adjacent in an image may be metres apart in reality, and the error grows with camera angle and distance. A proximity rule that has not been calibrated to the ground plane is measuring pixels, not metres, and its output will be wrong in ways that vary across the frame.
A pedestrian and forklift proximity rule without calibration will generate alerts for people who were never at risk while missing people who were. Calibration is a commissioning step, not an optional refinement, and any proposal for proximity detection that omits it has been underscoped.
Deciding what to build
Start from your own incident and near-miss record rather than from a capability list. The rules worth writing are the ones that correspond to events that have actually happened at your plant, because those are the ones with an owner, a response and a reason to be reviewed.
Three questions before any rule goes live: what exactly triggers it, who receives it, and what they do about it. A rule that fails any of the three should not be enabled, however easy it is to configure. The precision arithmetic applies here as everywhere: rare events plus generously drawn zones produce alert streams that get ignored, and every ignored rule reduces trust in the rules that matter.
Monitoring workers
Behaviour analytics observes how employees move and work. Keep rules tied to specific safety or security purposes, report at zone and shift level rather than building individual behavioural profiles, state the purpose in your notice and posted signage, and consult worker representatives before deployment. Where a path needs reconstructing after an incident, appearance matching answers it without identifying anyone. Rule 4 commences on 13 November 2026. This is general information rather than legal advice.
Frequently asked questions
What is the difference between rule-based and anomaly-based behaviour detection?
A rule-based system checks a condition you wrote down, such as a person entering a defined zone while a machine runs. An anomaly system learns normal patterns and flags deviation. On a factory floor, normal legitimately includes changeovers, maintenance and audits, so anomaly systems flag all of them with no explanation and nothing to tune.
Why does a danger zone rule generate so many alerts?
Usually because it fires on every entry regardless of machine state, and most entries are setup, cleaning or maintenance rather than unsafe acts. Gating the rule with a machine running signal from the equipment or line PLC typically removes the large majority of alerts while keeping every meaningful one.
Can cameras measure distance between a person and a forklift?
Only with ground plane calibration. Without it the system measures pixel distance, which does not correspond to physical distance and varies across the frame with camera angle. An uncalibrated proximity rule alerts on people who were never at risk and misses people who were. Treat calibration as a required commissioning step.
Which behavioural rules should we start with?
The ones matching events that have actually occurred at your plant, taken from your incident and near-miss record. Those already have an owner and a response, which is what makes the alerts get reviewed. Starting from a vendor capability list produces rules nobody asked for and nobody closes.
Does behaviour analytics identify individual workers?
It does not need to. The rules are about movement and position, and the useful output is zone and shift level rather than individual. Where a path needs reconstructing after an incident, appearance matching links sightings across cameras without determining identity. Building individual behavioural profiles is a different purpose and should be treated as one.
Have a specific unsafe act to detect?
IndoAI writes behavioural rules against your own zones and floor layout rather than shipping a generic anomaly model. Tell us the specific behaviour you need flagged and we will tell you whether video can see it.
Talk to an adviserReviewed by Dr. Vivek Gujar, Chief Strategy Officer at IndoAI Technologies Pvt. Ltd., a Pune-based edge AI camera platform founded in 2021. He reviews IndoAI's published technical claims for accuracy. Profile