Manufacturing / Perimeter
Perimeter intrusion detection for factory sites
Perimeter intrusion detection uses cameras to flag people or vehicles crossing a boundary they should not. On a large factory site the technical detection is the easy part. The difficulty is alert volume: a long perimeter with vegetation and animals generates enough false alerts to make the system unusable unless it is designed against that from the start.
Size the alert volume before you buy
Every intrusion deployment lives or dies on precision, and precision depends on how rare real intrusions are at your site. That is a number you know and a vendor does not, so the sizing calculation has to be done by you.
Worked example: an 800 metre perimeter
Monitored hours: 20:00 to 06:00, 10 hours nightly
Genuine intrusion attempts: about 1 per month
If each camera produces 2 false alerts per night:
12 × 2 = 24 false alerts nightly, roughly 720 per month
Against 1 genuine event, precision is about 0.1%
Two false alerts per camera per night sounds modest and is in fact a good result for an untuned system. It still produces 720 alerts a month against one real event. The lesson from the base rate arithmetic is that halving the model error changes almost nothing here. What changes the outcome is removing the sources of false alerts entirely.
What generates false alerts on a plant perimeter
- Animals. Dogs and cattle inside or along a boundary are the single largest source at most Indian industrial sites. Size filtering helps and does not solve it, because a large animal close to the camera occupies the same pixels as a person further away.
- Vegetation. Grass and branches moving in wind, particularly during monsoon.
- Lighting transitions. Vehicle headlights sweeping a fence, security patrol torches, and the dusk transition when cameras switch to night mode.
- Rain and insects. Rain streaks and insects attracted to IR illuminators, both close to the lens and therefore large in frame.
- Legitimate activity. The plant's own night patrol, which will cross the perimeter line repeatedly by design.
What actually reduces them
In rough order of effect:
- Restrict the schedule. Monitoring only during unmanned hours removes the largest block of legitimate activity, and it is free.
- Use direction, not presence. A tripwire crossed inward is an event. Movement parallel to the fence, which is what animals and patrols mostly do, is not. This alone removes a large share of animal alerts.
- Require persistence and a path. An object must be tracked across a minimum distance rather than appearing once. Insects, rain and lighting artefacts do not produce coherent tracks.
- Draw zones tightly. Exclude the sky, the road outside the fence and areas where movement is expected. Every excluded pixel is a false alert that cannot happen.
- Suppress the patrol. Either exclude patrol times or accept and acknowledge those alerts as a check that the patrol occurred, which turns a nuisance into a control.
Note what is absent from that list: a better model. Model quality matters, but at these base rates the structural controls dominate, and a vendor whose answer to false alarms is a model upgrade has misdiagnosed the problem.
Commissioning across a weather cycle
An intrusion system tuned on a clear week in February will behave differently in July. Tuning has to span at least one full lighting cycle and, in India, ideally includes monsoon conditions. Budget two to four weeks of live tuning after installation and before anyone is asked to rely on the alerts. A system handed over on defaults will be ignored within a month, and regaining a security team's trust after that is considerably harder than earning it the first time.
During tuning, log every alert with its cause. The resulting breakdown, by camera and by cause, tells you exactly which two or three cameras generate most of the noise, and the fix is usually a zone edit or a vegetation trim rather than anything technical.
Following an intruder across cameras
Once an intrusion is confirmed, the operational question is where the person went. Re-identification links sightings across perimeter cameras so a path can be reconstructed without an operator scrubbing twelve recordings. It works on appearance rather than identity, so it answers where they went without attempting to determine who they are.
Running detection and matching on site through edge processing matters here for a practical reason: perimeter cameras are frequently the furthest from the plant's network core, and an alert that depends on an internet link is an alert that fails at the moment a link is cut.
Frequently asked questions
How many false alerts should we expect from perimeter detection?
Untuned, expect one to several per camera per night, dominated by animals, vegetation and lighting transitions. After tuning with schedule restriction, directional rules, path persistence and tight zones, a well configured site can reach a small number per night across the whole perimeter. Size the expected volume against your real intrusion rate before committing.
Can the system tell a person from a dog?
Classification helps but does not fully solve it, because a large animal close to the camera occupies similar pixels to a person further away. Directional rules are more effective: animals typically move along or through a boundary while an intruder crosses inward toward the site. Combining classification with direction and path persistence works better than either alone.
Will our own security patrol trigger alerts?
Yes, unless handled. Either exclude patrol windows from monitoring, or keep the alerts and treat them as confirmation that the patrol happened on schedule. The second option turns an unavoidable nuisance into a supervisory control, and several sites prefer it for that reason.
How long does commissioning take?
Budget two to four weeks of live tuning after installation, spanning at least a full lighting cycle and ideally some adverse weather. A system handed over on default settings will produce enough false alerts in the first month for the security team to stop responding, and rebuilding that trust takes far longer than the tuning would have.
Can we use our existing perimeter cameras?
Frequently yes, provided they give adequate coverage of the boundary line and sufficient night-time illumination. The usual gaps are dark stretches between IR ranges and cameras aimed for general coverage rather than at the fence line. A site survey against the existing estate answers this faster than any specification exercise.
Perimeter alerts nobody answers?
IndoAI tunes intrusion rules against your own site footage across a full weather and lighting cycle before handover. Send us your perimeter length, camera positions and current nightly alert count.
Talk to an adviserReviewed by Dr. Vivek Gujar, Chief Strategy Officer at IndoAI Technologies Pvt. Ltd., a Pune-based edge AI camera platform founded in 2021. He reviews IndoAI's published technical claims for accuracy. Profile