Industries · Manufacturing
Manufacturing video analytics: the full use case map
Manufacturing AI video analytics turns existing plant CCTV into a detection layer for safety, production and dispatch. IndoAI maps 45 use cases across six groups: safety and compliance, production, gate and yard, security, workforce, and housekeeping. Seven have dedicated pages today. All inference runs on the on-premise Edge Box.
This page lists every manufacturing use case IndoAI can run, not just the ones that photograph well. Seven have dedicated pages with worked numbers and deployment constraints. The rest are listed plainly, because a list you can trust is more useful than thirty-eight thin pages that all say the same thing.
The one thing worth knowing before you read the list
Safety use cases sell the meeting. Production use cases sell the purchase order. Most plants start in group A because EHS has the incident log, then stall because EHS has no capital budget. If you want the project funded, put one group B use case in the same pilot.
The 45 manufacturing AI video analytics use cases
A. Safety and statutory compliance
The group that gets the meeting. EHS owns it, incidents drive it, and every plant already has a paper process these models replace with evidence.
- PPE detection: helmet, vest, gloves, goggles, mask, shoes Live
- Fire and smoke detection Live
- Fall and man-down detection Live
- Restricted and hazardous zone intrusion Live
- Unsafe act and behaviour detection Live
- Forklift to pedestrian proximity and near-miss logging
- Working at height and harness compliance
- Machine guarding and interlock bypass
- Lockout-tagout verification
- Confined space entry monitoring
- Hot work and welding spark monitoring
- Emergency exit, aisle and hydrant obstruction
- Spill, leak and wet floor detection
- Smoking and open flame detection
- Muster point headcount during drills
B. Production and productivity
The group that closes the purchase order. The buyer here is a plant head with a budget, not an EHS officer without one.
- Cycle time and station time measurement
- Micro-stoppage and line idle detection
- Operator presence at workstation
- Assembly step verification, visual poka-yoke
- Missing or wrong component detection
- Label, barcode and date-code print verification
- Carton seal, count and packaging integrity
- Bin and kanban level monitoring
- Work in progress counting between stations
- Changeover and SMED time measurement
- Tool and fixture presence
C. Gate, yard and dispatch
Cheapest group to deploy because the cameras are already there and the disputes are already expensive.
- Gate ANPR with weighbridge integration Live
- Truck turnaround time
- Dock and bay occupancy
- Loading and unloading evidence for damage disputes
- Yard and parking slot occupancy
- Driver cab and PPE compliance at gate
D. Security and shrinkage
Traditional CCTV work, except the review happens automatically instead of after the loss is discovered.
- Perimeter line crossing and loitering
- After-hours movement detection
- Tailgating at turnstiles and man traps
- Scrap yard and dispatch gate pilferage
- Asset removal without gate pass
- Camera tamper, blinding and defocus detection
E. Workforce
The group with the heaviest privacy design load. Read the consent section before scoping any of these.
- Contractor headcount and attendance Live
- Shift-wise manpower against plan
- Canteen and washroom footfall for facility sizing
- Induction and training compliance at entry
F. Housekeeping and environment
Low urgency, high audit value. Usually added in year two once the first two groups have paid for themselves.
- 5S and housekeeping audit scoring
- Stack and chimney smoke opacity
- Effluent discolouration at outfall
What decides whether any of this works
Pixel density on the thing you want to detect
Not camera resolution. Pixels per metre on target is the number that matters. A helmet needs enough pixels across the head to separate it from hair and a cap; a plate needs enough across the characters. A 4K camera 25 metres away resolves less on target than a 2MP camera at 8 metres, which is why camera count often goes down after a survey rather than up.
Decode budget, not model count
People size appliances by counting models. The real constraint is video decode. Sixteen streams at 1080p and 25 frames per second consume a large share of an appliance before a single inference runs. Most of these use cases are perfectly served at 10 to 15 frames per second, which roughly halves the load. See Edge Box for the channel and tier breakdown, and the glossary entries on RTSP and ONVIF for how existing cameras are brought in.
Alert routing that someone can act on
A PPE alert to a control room nobody staffs is a logged non-event. The three-step rule we use on site: every alert type gets a named recipient, a target response time, and a defined action. If any of the three is blank, the model does not go live. This is the difference between a system that survives its first quarter and one that gets muted.
Tuning, budgeted honestly
Expect two weeks of tuning after go-live: zone geometry, schedules, minimum dwell, per-camera confidence. The first fortnight's alert log is calibration data, not results. Read the glossary entry on false positive rate for why this matters more than headline accuracy figures.
A worked example: PPE plus cycle time in one bay
Take a single assembly bay with four existing 4MP IP cameras, two of which have usable views. Add one 8-channel appliance. Activate helmet and vest detection on the entry view and station occupancy on the line view.
| Input | Assumption | Result |
|---|---|---|
| Shifts covered | 2 shifts, 26 days | 52 shift-days per month |
| Manual PPE rounds replaced | 3 rounds per shift, 12 minutes each | 31 supervisor hours per month |
| Station idle visible | 4 micro-stoppages per shift, 6 minutes each | 21 hours per month surfaced |
| Cameras added | None; 2 repositioned | Hardware limited to 1 appliance |
The point of the table is not the hours. It is that 31 supervisor hours is an EHS argument and 21 hours of surfaced idle time is a production argument, and you need both in the same room. Substitute your own shift pattern and round durations; the structure holds.
Where the privacy line sits
Thirty-nine of the 45 use cases on this page never identify a specific person. They detect a helmet, a flame, a forklift, an empty station, a truck. Personal data only enters when the system is asked who someone is: face-based attendance, contractor identification, visitor management.
For those, Rule 4 of the Digital Personal Data Protection Rules requires a clear consent notice from 13 November 2026, and the penalty provisions in Section 33 commence on 13 May 2027. Scope identity use cases separately, design the notice before deployment, and keep the retention window defensible. Our face recognition page for manufacturing argues against the product for most shop floor applications on accuracy grounds alone, before the compliance question even arises.
Common questions
Do I have to replace my factory cameras to run these?
No, in most cases. The Edge Box ingests RTSP from existing IP cameras and connects to analog and DVR systems through an encoder. The usual gap is not the camera but the placement: a camera hung for general surveillance often cannot resolve a helmet strap or a plate. A site survey identifies which existing views work and which need a reposition rather than a replacement.
How many of these 45 can run at once on one plant?
Multiple models can run against the same camera, and dozens can run across a site. The limit is decode and inference capacity on the appliance, not the licence. A 16-channel Edge Box sized correctly will typically carry three to five active models per stream at 12 to 15 frames per second, which is well above what most of these use cases need.
Which use case should a plant start with?
Start with one safety model and one production model in the same pilot area. Safety gives you the internal sponsor and a visible early win. Production gives you a number the plant head can put in a review. PPE plus cycle time in a single bay is the most common opening pair, and it needs two or three cameras, not twenty.
How accurate is PPE detection on a real shop floor?
Accuracy depends far more on the view than the model. A helmet at 30 pixels across in backlit conditions will be missed regardless of vendor. In a correctly placed view with adequate lighting, helmet and vest detection is reliable enough for exception reporting. Gloves and goggles are harder because of occlusion and scale, and should be scoped with lower expectations.
Does camera-based fire detection replace our fire alarm system?
No. Camera-based fire and smoke detection is supplemental to statutory fire detection and suppression systems and must never be specified as a replacement for them. Its value is early visual confirmation in large open areas, yards and outdoor storage where point detectors are impractical, and reducing the time between ignition and a human looking at it.
Do these use cases trigger DPDP obligations?
Most do not. PPE, fire, fall, forklift proximity, dock occupancy and cycle time involve no identification of a specific person. Face-based attendance, contractor identification and visitor management do process personal data and need a consent notice under Rule 4 of the DPDP Rules from 13 November 2026. Scope those separately and design the notice before deployment, not after.
Can we get numbers out of this, or only alerts?
Both. Every detection is an event with a timestamp, camera, zone and confidence. Counting use cases such as cycle time, TAT, dock occupancy and headcount produce time series you can trend and export. The value in production use cases is almost entirely in the trend, not the individual alert.
What about false alarms?
Nuisance alerts are the single most common reason a deployment gets switched off. The fix is zone geometry, schedule rules, minimum dwell thresholds and confidence tuning per camera, not a better model. Budget two weeks of tuning after go-live and treat the first fortnight's alert log as calibration data rather than as results.
Next step
Scope a pilot bay, not a plant
Describe one area of your plant in the AI Adviser: what you want detected, how many cameras are already there, and the lighting. It returns appliance sizing, which existing views are usable, and which models to activate, priced from the live catalogue.