Published 15 November 2026 · Rathish Kumar, Invitty
A factory safety video analytics project in India should start with 2 or 3 clearly defined risks, a handful of cameras that genuinely see those risks, and a success measure agreed with the safety team before any software is bought. Run that as a pilot on your own footage for a few weeks. If the alerts are accurate enough for supervisors to act on, scale it; if not, the fix is usually camera placement, not a bigger licence.
Which safety problems video analytics handles well
Computer vision is good at repetitive visual checks that people get tired of doing. It is weaker at judging context. Match the use case to that strength.
| Use case | What it flags | What the camera must see |
|---|---|---|
| PPE compliance | Missing helmet, vest or mask at a line entry or work zone | People at a consistent distance, roughly head-height or slightly above |
| Restricted and forklift zones | A person entering a marked area or a vehicle path | The full zone boundary, in good light through the shift |
| Man-down and lone worker | A person lying still in an area where nobody else is present | An unobstructed view of the floor, not just the walkway |
| Fire and smoke | Early visual signs in storage or process areas | A wide view with few reflections or steam sources |
| Perimeter and after-hours intrusion | Line crossing or loitering at fences and gates | The fence line at night, with IR or lighting |
Camera requirements: pixels on the subject, not resolution
The single biggest cause of failed pilots is cameras that cannot see the detail the model needs. The IEC 62676-4 standard describes this as pixel density on the target. Axis, summarising the standard, gives these levels for a human face:
| Level | Pixels per metre | What it lets you do |
|---|---|---|
| Detection | 25 px/m | Tell whether a person is present |
| Observation | 63 px/m | See distinctive clothing and what the person is doing |
| Recognition | 125 px/m | Tell whether this is someone seen before |
| Identification | 250 px/m | Identify the person beyond reasonable doubt |
Most safety checks sit between detection and observation: you need to see a helmet and a vest, not a face. That matters for privacy as well as cost, because a system that only needs observation-level detail does not need identification-grade cameras. Angle, lighting through the day and frame rate matter just as much. A camera mounted very high suits counting but loses the side view a PPE check needs, and fast-moving vehicles need a higher frame rate than a corridor.
Edge, on-premises or cloud processing
For factories we recommend starting on site. Platforms such as NVIDIA's Metropolis framework and its Jetson embedded modules are designed to run video AI at the edge, close to the cameras. That keeps footage on your premises, keeps alerts fast, and means a broadband outage does not stop the safety checks. Cloud processing can make sense for multi-site reporting later, once you know which use cases earn their keep.
The DPDP Act points to settle before you switch it on
Footage of identifiable workers is personal data under the Digital Personal Data Protection Act, 2023. Section 7(i) lists processing "for the purposes of employment or those related to safeguarding the employer from loss or liability" as a legitimate use, which means safety monitoring of staff does not normally rest on consent. It does still need care:
- Notice: tell workers which areas are monitored, why, and who sees the alerts, in a language they read.
- Purpose: cameras installed for safety and later used to judge productivity are a second purpose. Decide that deliberately, not by drift.
- Retention: keep alert clips only as long as investigation and reporting need, and set the deletion schedule up front.
- Minimise: prefer detection and observation over face recognition unless there is a specific, justified need.
This is practical guidance, not legal advice. Involve your HR and legal owners, and see our DPDP Act compliance service if you need a structured review.
A pilot plan that produces a real decision
- Site walk and camera audit. Check position, angle, lighting at different shifts and frame rate for every camera covering a risk area.
- Choose 2 or 3 use cases tied to incidents or near-misses you have actually recorded.
- Agree the success measure. For example: of the alerts reviewed, what share were correct, and how many alerts per shift can supervisors realistically act on.
- Run on a small set of cameras with alerts going to the supervisors who will respond, not to a dashboard nobody watches.
- Review weekly with the safety team, then decide: extend, move cameras, or stop.
What drives the cost
We do not publish prices because they depend on the site, but the cost drivers are predictable: the number of camera channels analysed, any cameras that must be moved or replaced, edge hardware sized to those channels, software licensing per channel or per site, integration with your alarm or messaging systems, and the people time to review alerts. Ask every supplier to price the pilot and the full rollout separately.
Next step
Send us your site plan, camera list and the incidents you most want to prevent, and we will scope a pilot on your existing cameras. See AI video analytics for the full range of use cases we deploy, or our AI services overview.
Frequently asked questions
Can AI video analytics run on our existing CCTV cameras?
Does video analytics for staff safety need employee consent under the DPDP Act?
How many cameras should a factory safety pilot cover?
Should the analytics run in the cloud or on site?
Sources
Facts in this guide come from these pages, checked on 15 November 2026.