A confidence threshold is the cutoff that decides whether a detection counts. Every AI detection comes with a confidence score — the model's estimate, say 0.42 or 0.88, of how sure it is — and the threshold is the line above which the system treats it as a real detection and below which it ignores it. It is the single most important dial an operator has for tuning analytics, because it directly sets the balance between catching events and raising false alarms.
Moving the dial moves precision and recall in opposite directions. Lower the threshold and the system accepts weaker detections: it catches more real events (higher recall) but also more false ones (lower precision). Raise it and it only acts on confident detections: fewer false alarms (higher precision) but more missed events (lower recall). There is no value that maximises both — the threshold is where you choose which error you can better tolerate for this camera and this job.
The pitfall is leaving it at the vendor default for every camera. The right threshold depends on the cost of a miss versus a false alarm: a perimeter where missing an intruder is unacceptable wants a lower threshold and more false alarms to chase; a retail watch-list where a wrong stop is harmful wants a higher one. Set it per camera and per use case against those costs, validate it against real footage, and remember that no threshold makes the underlying accuracy 100% — it only chooses how the inevitable errors are distributed.
