Glossary / Measurement
What is false positive rate?
False positive rate is the share of genuinely negative cases a detector wrongly flags as positive. In video analytics it matters more than accuracy, because when the event being detected is rare, even a very low false positive rate produces alerts that are mostly wrong. This is what causes operators to stop responding.
Four numbers describe a detector, and they answer different questions. Confusing them is the most expensive measurement mistake in video analytics.
| Measure | Formula | Question it answers |
|---|---|---|
| Sensitivity (recall) | TP ÷ (TP + FN) | Of real events, how many did we catch? |
| Specificity | TN ÷ (TN + FP) | Of non-events, how many did we correctly ignore? |
| False positive rate | FP ÷ (FP + TN) | Of non-events, how many did we wrongly flag? |
| Precision | TP ÷ (TP + FP) | Of the alerts I received, how many were real? |
Precision is the operator's experience of the system. It is also the number almost never quoted, because unlike the others it depends on how rare the event is, and a vendor cannot know that for your site.
Why 99 percent is not good enough
Worked example: perimeter intrusion at a single gate
Genuine intrusions per day: 2
Detector sensitivity: 99%
Detector false positive rate: 1%
True positives = 2 × 0.99 = 1.98
False positives = 9,998 × 0.01 = 99.98
Precision = 1.98 ÷ (1.98 + 99.98) = 1.9%
A detector that is 99% accurate on both measures produces alerts that are wrong 98% of the time. The operator receives around 100 alerts a day, of which two are real. Within a fortnight nobody is checking them, and the two real ones are missed along with everything else.
This is base rate neglect. The detector is not faulty. The rarity of the event is doing the damage.
What it takes to fix
Work the arithmetic backwards. To reach 50 percent precision, false positives must fall to roughly the number of true positives.
Required FPR = 1.98 ÷ 9,998 = 0.000198
Required false positive rate = 0.02%, not 1%
The false positive rate has to improve fifty-fold to make alerts a coin flip, and further still to make them trustworthy. This is why marginal improvements in model accuracy often make no practical difference to an alert stream, and why the productive interventions are usually structural rather than model work:
- Shrink the negative pool. Tighter zones, scheduled active hours and direction filters cut the 9,998 windows directly. A perimeter rule active only from 20:00 to 06:00 removes most of the denominator.
- Require persistence. Several consecutive detections rather than one, which multiplies the independent error probabilities together.
- Require corroboration. Two conditions together, such as a detection plus a direction of travel across a line, or a detection plus a second sensor.
- Accept lower sensitivity. Raising the threshold trades some recall for large precision gains. Missing one intrusion in fifty that gets acted on beats catching all of them in a stream nobody reads.
Reading a vendor accuracy claim
A single accuracy figure on an imbalanced problem is close to meaningless. A detector that flags nothing at all is 99.98% accurate on a problem with 2 events in 10,000 windows, and it is useless. Ask for sensitivity and false positive rate separately, measured on footage resembling your site, with the number of genuine events in the test sample stated. If the answer is one percentage figure with no denominator, it has not been measured in any way that predicts your experience.
Then do the precision arithmetic yourself using your own event rate. You are the only party who knows how often the thing actually happens at your site, which means you are the only one who can compute the number that will determine whether the system gets used.
Alarm fatigue is the real failure
The cost of a false positive is not the alert. It is the erosion of response. A team that has learned the system cries wolf responds slowly to everything, including the genuine events, which leaves the site in a worse position than before the system was installed while the organisation believes it is protected.
This is why persistence rules matter so much in safety analytics, and why tuning against real site footage during commissioning is not optional polish. A system tuned on generic defaults and handed over will be ignored within a month. Budget two to four weeks of tuning against live conditions before anyone is asked to rely on it.
Frequently asked questions
What is the difference between false positive rate and precision?
False positive rate is the share of non-events wrongly flagged, and it is a property of the detector. Precision is the share of alerts that turn out to be real, and it depends on both the detector and how often the event actually occurs. Precision is what an operator experiences, and it is usually far worse than the false positive rate suggests.
Why do I get so many false alarms from a 99 percent accurate system?
Because the event you are detecting is rare. If a gate has 10,000 evaluation windows a day and two genuine intrusions, a 1 percent false positive rate produces about 100 false alerts against two real ones. That is 98 percent of alerts being wrong, from a detector working exactly as specified.
How do I reduce false positives without missing real events?
The most effective changes are structural rather than model related: tighten the zone, restrict active hours, require the detection to persist across several frames, and require a second corroborating condition such as direction of travel. Each cuts the pool of opportunities for a false positive without much cost to genuine detections.
Is a single accuracy percentage useful when comparing vendors?
No. On a problem where events are rare, a detector that flags nothing scores extremely well on accuracy while being useless. Ask for sensitivity and false positive rate separately, measured on footage resembling your site, and ask how many genuine events were in the test sample.
What false positive rate should I aim for?
There is no universal figure, because it depends on how often the real event occurs. Work backwards from the alert volume your team can actually handle. If they can review ten alerts a day and you have two real events a day, you need false positives near eight per day, which fixes the rate you require.
Is it acceptable to miss real events to cut false alarms?
Frequently yes, and this is a judgement rather than a technical decision. A system catching 90 percent of events that operators trust and act on protects a site better than one catching 99 percent inside an alert stream nobody reads. The right balance depends on the consequence of a miss, so it should be an explicit decision rather than a default.
Drowning in false alarms?
IndoAI tunes detection thresholds and verification rules against your actual site conditions rather than shipping a default. Tell us your current alert volume and how many are genuine, and we will tell you what is achievable.
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