AI Video Analytics: What Actually Works for Mines, Factories and Shops in Laos

From "review it afterwards" to "alert me now"

Most CCTV only gets watched after something has already happened. One guard looking at 32 tiles on a single screen through a night shift will not see everything. After the first hour, attention is gone. AI video analytics does the part people are bad at: it watches every camera at once and only raises the events worth looking at.

The AI can run in three places: in the camera (edge), in an AI-capable NVR, or on a GPU server for heavier work. Names you'll meet on datasheets include Hikvision AcuSense, Dahua WizSense and WizMind, Axis Object Analytics and Hanwha Vision's AI camera ranges. With a VMS such as Milestone XProtect, events from several camera brands can be collected in one place through ONVIF Profile M or the manufacturer's plugin.

Use cases that actually deliver

New in 2025: searching footage in plain words

The big manufacturers have started putting large AI models into NVRs and cameras. Hikvision, for example, launched its AcuSeek NVR range in mid-2025: you type a description such as "white pickup near the gate" and get results in seconds, without sending video to the cloud. Dahua has announced large models of its own. This is very useful for investigations after the fact. It doesn't change the basic rule, though: if a camera has too few pixels or the wrong angle, the AI won't find anything. These searches still tend to work best in English, so check whether the language your team uses is supported.

Example: a 2 km mine perimeter

• AI cameras with detection lines every 60–80 m depending on lens and terrain → roughly 25–35 cameras
• At night: thermal cameras where it's pitch dark or where morning fog sits
• Events go to the control room with a short clip. Operators check events instead of staring at screens all shift
• Fibre or point-to-point wireless back from remote sections, with surge protection and earthing at every pole

Do the alert arithmetic. 30 cameras with 5 false alarms each per night is 150 alerts. Two guards can't check them all, and they'll soon stop trying. Our target after tuning is fewer than one false alarm per camera per night. If you can't get there, shrink the detection zones before you add people.

Example: a PPE pilot on a construction site

Take a building site with one main gate and a hoist area. Put two AI cameras on the gate, framed at 2.5–3 m height so helmets and hi-vis vests are clearly visible, and one on the hoist landing. For the first two weeks, log alerts but don't act on them. Compare the AI's count of "no helmet" events with what a safety officer sees on spot checks. If the two numbers are close, start the morning briefing with yesterday's figures and let the foremen deal with it. If they aren't, the camera angle or light is usually the cause, not the AI. Workers in dark clothing at dusk and helmets in unusual colours are the cases we see trip up detection most often.

Shops and restaurants: counting worth having

People counting from an overhead camera at the door is accurate enough to compare days, branches and promotions. It is not accurate enough to audit cash. Put the counting camera straight above the entrance, not at an angle, and keep it away from glass that reflects headlights at night. Heatmaps are useful once, when you plan a layout. After that, most owners stop looking at them, so don't pay extra for them unless someone will actually use them.

Limits the brochures skip

Accuracy depends more on installation than on brand: camera angle, height, light and pixel density. No AI is 100% accurate. Expect some false alarms and some misses, and budget 2–4 weeks of tuning after installation.

Face and plate recognition need dedicated cameras at the right angle. A general wide-angle camera does both badly. The pixel and lens maths is in our CCTV resolution and lens guide.

Lao plates have their own layout and Lao script, and plenty of cross-border trucks carry Thai plates. Before you buy LPR, test it on real plates at your own gate rather than trusting the datasheet.

Site conditions in Laos matter too. Red dust on the lens in the dry season, heavy rain, and insects at night all push false alarms up. The maintenance plan needs lens cleaning and re-tuning when the season changes.

And privacy: face data is sensitive personal data. Write down what you collect, how long you keep it and who can see it, and tell staff before you switch it on.

Edge or server: how to choose

For provincial sites, edge usually wins. A GPU server pays off when you know exactly what you need that edge can't do.

How to start without wasting money

1. Pick one problem with a real cost: fuel going missing from trucks, site accidents, people entering restricted areas at night.
2. Decide how you'll measure it: incidents per month, response time.
3. Pilot 2–4 cameras for 30 days.
4. Count the false alarms, then tune zones, sensitivity and object size.
5. Decide who gets the alerts and what they must do. Alerts without a procedure are worthless.
6. Scale up once the numbers prove it.

Network and storage impact

Edge AI does the analysis in the camera and sends only events and metadata. That suits provincial sites whose link back to head office has little bandwidth.

Server-based AI needs every camera's full stream. 30 cameras × 4 Mbps = 120 Mbps, all the time, so plan the uplink and GPU capacity to match. Our bandwidth calculator helps with the numbers.

For evidence, store event clips separately and keep them longer than routine recordings.

What sinks AI projects

1. Every AI feature switched on for every camera. By week two the guards have muted the alerts because they never stop.
2. Expecting wide-angle cameras to recognise faces.
3. Nobody responsible for alerts at night.
4. Buying an expensive GPU server before testing whether in-camera edge AI is enough.
5. Using PPE detection to punish people. Workers learn where the cameras can't see instead of wearing the helmet.

Where we'd start

On remote sites we'd begin with edge AI intrusion detection: good value and light on bandwidth. LPR at mine and factory gates to control truck traffic comes next, after testing on both Lao and Thai plates. PPE detection on construction sites works as a safety prompt, not a disciplinary tool. And the response procedure gets written alongside the installation, not after. More detail is on our AI video analytics page.

Is there a spot that costs you money or keeps you up at night every month? Tell us about it and send a few photos of the area. We'll design a 30-day pilot with numbers you can measure before you commit to anything bigger.