False Alarms at a Windy Perimeter
A car park perimeter uses 8 thermal cameras for intrusion detection across a 200 m fence line. Since go-live, operators receive 60–90 alarms per night, almost all between 01:00 and 04:00 when wind gusts reach 40 km/h. The fence carries banners that flap, and small animals cross the lot. Analytics are enabled on the camera with default sensitivity. Operators have started ignoring alarms, and the security manager wants detection disabled at night. Recorded footage is required for 14 days.
1. Diagnose the likely causes and propose a tuning plan that keeps genuine detection capability.
Read the worked answer
The symptom pattern (alarm burst correlated with wind, plus animals) points to analytics triggering on non-human motion because sensitivity and object filters are at defaults.
Tuning plan:
- Use object classification (human/vehicle) if the thermal camera supports it, and set the rule to alarm on humans only. This removes most animal and banner triggers.
- Reduce sensitivity and raise the trigger threshold, then re-test with a person walking the fence at several distances.
- Reshape detection zones to exclude flapping banners and the ground strip where animals pass; use include/exclude zones rather than one large area.
- Add time-based confirmation (object must persist N frames/seconds) to reject brief gusts.
- Verify the scene - check whether the fence itself is in the detection zone and whether camera shake from wind is moving the field of view.
Never accept 'disable at night' as the fix: that removes detection exactly when intrusion risk is highest. Document each change and re-measure alarm counts over several nights, including a windy one, before declaring success.
2. How do you prove the tuned system still detects real intrusions?
Read the worked answer
Run controlled walk tests at multiple points along the 200 m line, at different distances and approach angles, day and night, and log detection vs miss. Include a crawling/low-profile approach if the threat model warrants it. Compare alarm counts on a comparable windy night before and after tuning. Acceptance criterion should be stated up front, e.g. detect a walking human at the fence within the zone with zero misses across N trials, while non-human alarms drop below an agreed threshold. Keep the test record as commissioning evidence.