Use Case · Manufacturing · India
AI CCTV for Factory Safety
AI CCTV for manufacturing turns existing plant cameras into a safety and compliance system. Models detect missing PPE, entry into restricted zones, forklift hazards and visible fire or smoke, route each event to the supervisor on shift, and produce the compliance evidence an audit asks for.
Safety is measured after the fact, which is the problem
Most plants know their incident numbers precisely and their near-miss numbers barely at all. An incident generates paperwork, an investigation and a corrective action. A near miss generates a shrug, if anyone saw it, and no record.
That asymmetry means safety systems are built almost entirely on the events that already went wrong, while the far larger population of events that nearly went wrong stays invisible. A forklift and a pedestrian passing too close at shift change happens dozens of times before it happens badly once, and nothing in a conventional plant captures the dozens.
Cameras have been watching all of it. They just have not been telling anyone. The footage exists, and it is only useful if you already know what happened and when, which is exactly what a near miss denies you.
What the Factory Safety Agent does
An agent bundles the models a situation needs with the rules for when a detection matters, the workflow for who is told, and the reporting that makes patterns visible. The wider architecture is set out in the AI apps and agents catalogue.
PPE compliance
Helmets, high-visibility vests, gloves and footwear, checked per zone rather than plant-wide, since requirements differ by area.
Restricted zone entry
People entering machine areas, forklift lanes or isolated equipment during operation or maintenance.
Fire and smoke cues
Visible flame or smoke in line of sight, which can surface faster than a sensor in a large open bay.
Incident and compliance reporting
Evidence packs around each event and recurring reports showing where and when exposures cluster.
The reporting is the part that changes behaviour. A single PPE alert tells you one person was not wearing a helmet. Twelve weeks of alerts tells you that the loading bay at shift change is where your exposure actually sits, which is a different and far more actionable finding.
AI CCTV for manufacturing, zone by zone
Plant layouts vary far more than retail floors, but the risk categories are consistent, and the discipline is the same: every camera traced to a named risk, so no camera exists without a job.
| Zone | What it is watching for | What the camera must resolve |
|---|---|---|
| Loading bay and docks | Forklift and pedestrian proximity, PPE, vehicle movement | Full bay width, both traffic directions, high contrast range at the door |
| Shop floor walkways | PPE compliance, walkway discipline, congestion | Torso and head at working distance, helmet colour distinguishable |
| Machine and press areas | Entry during operation, guarding bypass, isolation breaches | Approach path and the guarded zone together |
| Restricted and isolation zones | Any entry, particularly outside permitted windows | Full boundary, capable in low light |
| Flammable storage | Visible smoke or flame, unauthorised access | Wide area coverage, clear line of sight |
| Gate and perimeter | Vehicle entry records, contractor movement, after-hours access | Number plates at ANPR quality, faces at the pedestrian gate |
The forklift problem
If a plant runs forklifts and people in shared space, that is usually where the serious risk concentrates, and it is the clearest case for analytics because the exposure is continuous rather than occasional.
A common approach specifies PPE detection across every camera in the plant. On a forty-camera site that generates alerts all day, most of them from areas where the risk is low, and within a month nobody opens them. A system that produces more alerts than a supervisor can act on is worse than no system, because it trains people to ignore the screen.
The alternative is to find where forklifts and pedestrians actually share space, and cover those zones properly: proximity and restricted-lane entry on the loading bay and the main cross-aisle, PPE checked in those zones, alerts to the shift supervisor's phone rather than a central mailbox, and a weekly exception report to the safety officer.
Six cameras carrying the analytics on a forty-camera site produces alerts few enough to be acted on, and a report the safety officer can take to a monthly review. The comprehensive version produces a larger invoice and a system that is off by quarter end.
What PPE detection can and cannot see
This is worth being precise about, because expectations set at the demo stage cause most of the disappointment later.
It sees presence, not correctness. A model can determine that a helmet is on a head. Whether the chinstrap is fastened, whether the helmet is within its service life, or whether it is the right class for the hazard are not things a camera resolves.
Occlusion breaks it. A worker behind a machine, a pallet stack or another worker is partially or wholly unobservable. Coverage geometry decides how often that happens, and it is a design problem rather than a model problem.
Dust, steam and glare degrade it. Indian plant conditions are harder than benchmark conditions. A model quoted at a headline accuracy in clean light behaves differently in a dusty bay at three in the afternoon with sun through the door.
Similar-looking items confuse it. A cap can read as a helmet in poor conditions, and a coloured shirt can read as a vest. The rate depends on the site, which is why realistic expectations come from the design stage rather than a datasheet.
None of this makes PPE detection unhelpful. It makes it a tool for surfacing patterns and prompting supervision, rather than an automated enforcement mechanism. Plants that deploy it as the former get value from it. Plants that deploy it as the latter end up arguing with their workforce about false positives.
Fire and smoke: a specific caution
Visual fire and smoke detection has a real advantage in large open spaces. A heat or smoke sensor in a high-ceilinged bay may need significant accumulation before it triggers, while a camera with line of sight can surface visible flame or smoke earlier.
Camera-based fire and smoke detection is a supplement to your statutory fire detection and suppression systems, not a replacement for them. It requires line of sight, it can be defeated by obstruction, and it is not a certified life-safety device. Any plant treating video analytics as a substitute for a properly specified and maintained fire system is making a serious mistake, and no responsible vendor should encourage it.
The compliance evidence, which is the underrated part
Manufacturing differs from retail in one commercially important way. A retailer buys loss prevention to protect margin. A plant buys safety analytics to protect people, and then discovers the second benefit: it can prove what it has been doing.
Statutory safety obligations in Indian factories, along with customer audits, insurer requirements and internal governance, all ask the same underlying question, which is whether your stated controls actually operate. A recurring report showing PPE compliance rates by zone over twelve weeks is a materially stronger answer than a policy document and a training register.
Customer audits
Export-facing plants face buyer audits with safety components. Evidence of continuous monitoring shortens those conversations.
Insurer conversations
Demonstrable controls and incident evidence support renewal discussions rather than relying on assertion.
Incident investigation
An evidence pack assembled at the moment of the event, rather than footage reconstructed days later from a rolling recorder.
Internal governance
Trend data by zone and shift, which turns the monthly safety meeting into an analysis rather than an anecdote exchange.
Your workers are data principals too
An employment relationship is not a blanket permission. Workers whose conduct is continuously analysed have rights over that processing, and DPDP Rule 4 commences on 13 November 2026, requiring organisations to describe what they collect, why, for how long, and on what lawful basis.
Manufacturing raises a distinct question that retail does not: proportionality. Monitoring a hazardous zone for safety compliance is straightforward to justify. Using the same cameras to measure individual productivity is a different purpose, with a different basis, and is where employers most often lose the trust of their workforce and the confidence of a regulator at the same time.
Tell people, properly. Notice to the workforce about what is monitored and why, at induction and on signage, not buried in a handbook nobody has read since joining.
Keep purposes separate. Safety monitoring and performance management should not share a lawful basis, a dataset or a retention period. Conflating them is the fastest route to an industrial relations problem.
Face recognition is a separate decision. Attendance by facial recognition is a heavier processing activity than PPE detection, warranting a documented impact assessment and its own retention and deletion policy for the enrolled set.
Retention has to be enforced. A stated period the system does not actually apply is a commitment you are visibly failing.
Running inference on site narrows the exposure substantially, because the hardest questions concern personal data replicated to infrastructure you do not control. Privacy and face masking can redact at the point of capture. None of that replaces notice, lawful basis and proportionality, which remain yours.
What a plant deployment includes
-
Design
Camera placement traced to named risks per zone, blind spots measured, and the occlusion problem assessed against real plant layout rather than a drawing.
-
Hardware
Cameras where the design needs them, the EdgeBox running models on site, storage sized on real bitrates, PoE switch, UPS, cabling and mounts rated for the environment.
-
Models
The Factory Safety Agent, configured per zone rather than plant-wide, since PPE requirements and restricted-zone rules differ by area and by shift.
-
Install and go live
Survey and marking, cabling, configuration and aiming, and an acceptance test that verifies detection actually works in the zone's real conditions rather than in ideal light.
-
Run it
Alerts to the shift supervisor, evidence packs for anything escalated, and recurring compliance reporting to the safety officer. Annual maintenance and one number to call.
Most plants keep their existing cameras. The EdgeBox ingests RTSP and ONVIF streams from any manufacturer, and analog estates connect through an encoder, so a site with cameras added across several expansion phases does not need to be rationalised first. The full sequence is in how IndoAI works.
Which EdgeBox a plant needs
Plants sit higher up the range than stores, for two reasons. Zones are more numerous, and safety models tend to run several at once on the same channel: PPE, restricted zone and proximity on a single loading bay camera.
| Site type | Typical appliance | Why |
|---|---|---|
| Small unit or single shed | EdgeBox 8 | Enough for one bay, a machine area and a gate where analytics stay targeted |
| Typical plant, multiple zones | EdgeBox 16 | Several models per channel across bay, walkways, machine areas and perimeter |
| Large or multi-shed site | Stacked appliances | Capacity added per shed or zone group, all feeding one dashboard |
| Group with several plants | One per plant, plus cloud dashboard | Inference local to each site, safety reporting aggregated for the group function |
The rugged specification matters more here than in retail. The appliance is fanless and rated across a wide industrial temperature range, which is why it survives in a plant room or a bay rather than needing a conditioned cabinet.
Which model tiers manufacturing draws on
A plant configuration typically spans all four tiers of the EdgeBox catalogue, and reaches further up the range than any other use case.
Basic
- Intrusion detection and virtual fence
- Zone and multi-zone counting
- Stopping detection
- Speed anomaly
- Privacy and face masking
Advanced
- Fallen person detection
- Hand and foot intrusion at machines
- Vehicle type detection and counting
- Thermal intrusion
- Crowd detection
Extra
- No PPE and no mask detection
- Forklift helmet and non-driver detection
- Hazard detection
- Tailgating at controlled doors
- Licence plate recognition at the gate
- Aggressive behaviour detection
Hybrid AI Boost
- Helmet not worn
- Out of uniform
- Spill detection
- Imminent threat detection
- Enhanced thermal intrusion
The Hybrid AI Boost tier blends on-device inference with cloud compute for the heaviest models, which means that for those specific models, and only when you enable them, some processing happens off site. It is opt-in model by model and the appliance default remains fully local. If your plant operates under a customer or contractual requirement that nothing leaves the premises, decide that at design stage rather than discovering it in a configuration screen later.
The forklift-specific models in the Extra tier are the ones worth noting, because forklift helmet and non-driver detection address the shared-space problem directly rather than approximating it with generic person detection.
The honest limits
It supplements supervision, it does not replace it. Analytics surface patterns and prompt attention. Somebody still has to own the response, and a plant without that ownership will not benefit regardless of what it installs.
Fire detection is not a life-safety substitute. Camera-based cues supplement statutory fire systems. They do not replace them.
Occlusion is a design constraint. Dense machinery, pallet stacks and moving equipment create blind spots that no model resolves. The design measures them and states them rather than pretending they do not exist.
Plant conditions degrade accuracy. Dust, steam, glare and low light all move the numbers away from benchmark figures. Realistic expectations come from assessing your zones, not from a datasheet.
Alert volume has a ceiling. Beyond what a shift supervisor can act on, additional detectors reduce effectiveness. Targeted coverage beats comprehensive coverage nobody reviews.
It will surface uncomfortable things. A plant that starts measuring near misses will find more of them than it expected. That is the system working, but it needs to be framed to the workforce before go-live rather than after.
Frequently asked questions
Do I need to replace my existing plant cameras?
Usually not. The EdgeBox ingests RTSP and ONVIF streams from cameras already installed regardless of manufacturer, and analog estates connect through an encoder. The common exceptions are zones where existing coverage has blind spots or where a camera was never positioned to resolve what the analytics need, which the design stage identifies.
How accurate is PPE detection?
It depends on your zones. Dust, steam, glare, low light and occlusion all move the numbers away from benchmark figures, and Indian plant conditions are harder than benchmark conditions. It also detects presence rather than correctness: a model sees that a helmet is on a head, not whether the chinstrap is fastened or the helmet is within its service life.
Can I use this to discipline workers?
Alerts are prompts for supervision, not determinations of misconduct, and acting on one without human verification is unsound both operationally and legally. Plants that deploy PPE detection to surface patterns get value from it. Plants that deploy it as an automated enforcement mechanism end up arguing about false positives with their workforce.
Does camera-based fire detection replace my fire alarm system?
No. It supplements statutory fire detection and suppression, and it is not a certified life-safety device. It requires line of sight and can be defeated by obstruction. Its advantage is in large open spaces where a ceiling sensor may need significant accumulation before triggering, but it is an addition to a properly specified system rather than a substitute.
How many cameras need to run analytics?
Far fewer than most plants assume. Coverage should follow where risk actually concentrates, typically shared forklift and pedestrian space, machine approaches and restricted zones, rather than being applied plant-wide. Running analytics on every camera generates alert volumes no supervisor can act on, which is how these systems get switched off.
What do I have to tell my workforce?
What is monitored, in which zones, for what purpose, and for how long footage is retained, communicated at induction and on signage rather than buried in a handbook. Safety monitoring and performance measurement should be kept as separate purposes with separate bases, because conflating them is the fastest route to an industrial relations problem.
Does the footage leave my plant?
Not by default. Inference runs on the EdgeBox on your premises and raw video stays there. What travels, if anything, are events and evidence packs you have chosen to route. Privacy and face masking can redact personal data at the point of capture, which is useful where zone monitoring does not require identifying anyone.
Can it handle different PPE rules in different areas?
Yes, and it should. Requirements typically differ by zone and sometimes by shift or task, so models are configured per zone rather than plant-wide. A blanket rule applied everywhere generates false alerts in areas where the requirement does not apply, which is a common reason early deployments lose credibility.
What reporting do I get for audits?
Recurring reports showing compliance rates and exception patterns by zone and shift over time, plus evidence packs assembled around individual events. Trend data over several weeks is generally more useful to an auditor than any single incident, because it demonstrates that a stated control actually operates.
Will it work in a dusty or high-temperature environment?
The appliance is fanless with a rugged industrial design rated for a wide temperature range. Camera selection and housing for the environment is a separate design question, and detection accuracy in dusty or steamy conditions is assessed per zone during design rather than assumed from benchmark figures.
Which EdgeBox do I need for my plant?
Most plants run an EdgeBox 16, because zones are numerous and safety models tend to run several at once on the same channel, such as PPE, restricted zone and proximity on one loading bay camera. Small single-shed units may suit an EdgeBox 8. Large or multi-shed sites stack appliances per zone group, all feeding one dashboard.
Can it cover multiple plants?
Yes. Each site runs its own appliance with inference local to that plant, and an optional cloud dashboard provides centralised monitoring across sites. Reporting can be viewed per plant or aggregated, which is how group safety functions generally want it.
Where should I start?
Identify where your exposure actually concentrates before selecting analytics, because a plant with a forklift and pedestrian problem needs a completely different configuration from one whose risk sits at the presses. Describing the site and the concern to the AI Advisor produces a design that works backwards from the risk rather than forwards from a feature list.
Describe your plant. Get the design back in one session.
Tell the AI Advisor your layout, your zones and where your near misses happen. You will get camera placement traced to each risk, blind spots measured, the model mix per zone, and an itemised bill of materials.
Open the AI AdvisorRelated reading: how IndoAI works, end to end, the AI app and agent catalogue, the IndoAI EdgeBox, and the AI vision glossary.
Co-founder and Chief Strategy Officer at IndoAI. He holds a PhD alongside an MBA and a B.Tech, and has spent more than two decades in business development and IT security across technology and non-technology sectors. He reviews IndoAI's published technical claims.