Most manufacturing plants in India already record every shift on CCTV. The recordings are reviewed after an incident, if at all. AI video analytics changes that: software watches the same feeds continuously and flags the moments that need a person, such as a missing helmet, someone in a forklift lane, smoke in a store or a blocked exit. The first question every plant asks is whether it will work on the cameras they already have. Usually it does, with a few conditions.
What matters more than the camera brand
Hikvision, CP Plus, Dahua, Axis and Bosch cameras all provide standard RTSP or ONVIF streams, and that is all the software needs. What decides whether detection is reliable is how the camera sees the scene.
- Angle: a camera looking down at 30 to 60 degrees sees helmets and vests clearly; a camera looking straight along an aisle sees the backs of heads.
- Distance: a person should occupy a reasonable part of the frame. A single wide camera covering a whole hall is good for smoke and zone entry, weaker for gloves and glasses.
- Light: night shift, mixed lighting and glare from roof panels change what the model sees. A camera has to be validated across a full day and night, not one afternoon.
- Stream quality: heavy compression and low frame rates, common on older NVR setups, blur fast movement such as forklifts.
What existing cameras do well
- PPE checks at entries, workstations and walkways where people pass close to the camera.
- People entering marked zones: forklift lanes, machine cells, crane paths, electrical rooms.
- Visible smoke and flame in large open areas, where ceiling detectors react late.
- Blocked exits and material left in front of fire equipment.
- Headcount in a zone, and who is still inside during a drill.
What they do not do
They do not see what the camera does not cover, and they do not stop machines. AI video analytics alerts people; it is not a safety-rated system and should never replace guarding, interlocks or your procedures. It also cannot fix a camera pointed at a wall. Part of any rollout is moving a few cameras, not buying new ones.
Where the processing happens
Running detection on an edge unit inside the plant keeps footage on site, avoids sending video over the internet, and keeps alerts under a second. Blurring faces before anything is stored keeps the system about events rather than people, which is what makes it acceptable to employees and unions.
How to start
- Pick the two to four zones you worry about most, usually a loading dock, a press or welding area, and one place where night shift is thinly supervised.
- Check those cameras for angle, distance and night-time image quality.
- Run detection for a week and look at what it finds, by zone and by shift, before deciding anything else.
Written from deployment experience across multiple sites. Patterns are described without identifying any customer — we are under confidentiality with the operators involved.
Talk to an engineer about this