A false positive is an alert that is wrong — the system says an event happened when it did not. Its opposite is a false negative, a real event the system missed. In surveillance these two errors trade against each other: dialling a system to catch everything (few false negatives) inevitably raises false positives, and tightening it to stop nuisance alerts (few false positives) lets some real events slip through. There is no setting that zeros both.
False positives are the dominant practical problem in deployed analytics because of the base-rate trap. Real incidents are rare, so even a low false-positive rate produces far more false alerts than true ones: a camera generating thousands of detections a day at a 1% false rate yields hundreds of false alarms against a handful of real events. The result is alert fatigue — operators stop trusting and stop checking, and the one true alert is missed in the noise. This is why analytics are best deployed as triage with human verification, not as autonomous alarms.
The way to manage false positives is layered, not magical. Use object classification instead of raw motion, restrict alerts to relevant zones, schedules, and dwell times, set the confidence threshold against the cost of a miss, and cascade filters so a noisy raw stream becomes a manageable trickle of credible alerts. Measure the false-positive rate honestly, accept that it is never zero, and design the operator's workflow around verifying a trickle rather than drowning in a flood.
