Reduce false alarms without hiding missed events
Use labeled events and exposure hours to measure an alarm rule.
Updated
Principle and limits
Separate true detections, false alarms and missed events. Precision asks what fraction of alarms were relevant; recall asks what fraction of real events were detected. Empty periods matter because operational burden accumulates with time and camera count. Overall accuracy can look excellent when almost all frames contain no event.
Worked example
If 12 false alarms occur across four cameras monitored for six hours each, the rate is 12/(4×6)=0.5 false alarms per camera-hour. That rate alone says nothing about missed intrusions. In a separate staged set of 20 intrusions, 18 detected gives 90% recall, but the small sample does not justify a universal 90% claim.
Installation and acceptance
Label test footage before adjusting thresholds. Tune zones, minimum duration and classes one change at a time; then test a different time period. Include shadows, foliage, rain, insects and partial occlusion. Record operating hours, event counts, firmware and settings. Review missed events alongside false positives before accepting a quieter alarm stream.