What camera accuracy numbers mean before you buy

By Gate & Shift Editorial Desk · Planned week of 2026-12-21 · Documentation reviewed 2026-11-15

If a proposal says “high accuracy,” make the salesperson finish the sentence: accuracy on which events, in which views, at which threshold? A loading bay, a dark service entrance and a wet shoulder are not one evaluation set.

The useful buying habit is to carry the counts with every percentage. The figures below are worked examples for a proposed pilot, not results for any product.

Define the event before counting it

A reference event is a human annotation naming the camera, target activity and start/end times. Decide what makes a system event a match before any results get evaluated. NIST's activity-evaluation work is a primary reference for activity-specific boundaries and missed detections alongside false alarms M01 M03.

For instance, decide whether a returning vehicle counts as another event and how much timing difference is allowed. Those are test-design choices. Change the rule after seeing the results and the comparison stops being worth much.

Two fractions that answer different questions

Recall and precision have different denominators M04. Our proposed burden and coverage measures are house definitions; ActEV's time-based false-alarm measure normalizes differently M02. Carry that distinction onto any chart that goes to procurement.

Why “accuracy” needs more explanation

Classification accuracy is correct classifications divided by all classified examples M04. For a camera event task, the examples have to be defined first, particularly the negative windows. Depending on that choice, long stretches without the target event can dominate a percentage.

Ask for positive and negative counts, thresholds, abstentions and excluded periods. Where an evaluation has no reference events, recall is undefined; where it has no scorable alerts, precision is undefined. Write “undefined” rather than turning either case into a perfect result.

Observation gaps are part of the buying decision

Report scheduled coverage separately. A camera that stops observing can produce fewer alerts without becoming a better detector. Note when the system was unavailable and whether a disputed event fell inside that gap M02.

Keep model performance during eligible monitored time visible, but show the whole operating watch too. The person answering the gate phone needs to know both whether the system detects the event and whether it was watching when it mattered.

Time the rule, then the recipient

Our proposed alert delay starts when the annotated rule first qualifies and ends when the chosen endpoint receives the alert M02. A rule that requires a sustained stop qualifies later than the vehicle's first appearance, so record both if the task needs them.

Keep each observation, not just the fastest result. Name the endpoint, clock source, rule version, view, lighting and interruptions. The pilot worksheet is where those requirements become a supplier demonstration.

FAQ

Can two vendor percentages be ranked?

Only if their evaluation units, samples, thresholds and conditions support the comparison. The source register for this publication does not establish a common vendor test set.

Which number should a supervisor choose?

Set acceptance conditions for the job, with both misses and unwanted alerts in view. A quiet phone and a complete watch are separate goals, and each needs its own evidence.