Measurement protocol for camera events
Freeze the unit
A camera-hour is one camera under observation for one hour. Two cameras watching the same gate for half an hour contribute one camera-hour — but only if both actually monitored their feeds. Keep each camera's start, end and interruptions on record. Otherwise a missed shift can quietly lower a nuisance-alert rate.
Protocol CTR-1 runs in two stages. First, verify delivery, frame order and evidence capture with the replay pack. Second, collect permitted field sessions for the decision at hand. The pack holds six procedural controls, six excerpts from three camera recordings and three edited copies. A light pulse has an exact onset; a misclassified pulse therefore says little about recognition of a person. L01 · Manifest
Mark the field sessions
Choose the physical task and its rule first. For a vehicle entering a zone, save the polygon on an image and define where on the vehicle the contact point sits; draw it again if the camera moves. Obtain recording and evaluation permission. Record the viewpoint, lighting, traffic, time base, resolution and sampling rate. Include quiet periods, and enough separate sites and shifts to cover the intended deployment.
Two reviewers should mark eligible events independently before either sees an alert. Record class, camera, rule, start, qualifying instant and end. Settle disputes against the frames, and keep both marks plus the disposition. Mark an obscured or unobservable interval as exactly that. Reserve whole camera sessions for evaluation, and keep a camera's derivatives, adjoining cuts and tuning footage together. Release CTR-1 is public practice material.
Count events, time and windows
Use the one-to-one matching rule. TP is a matched pair, FN an eligible reference left over and FP an unmatched scorable alert. Recall is TP/(TP+FN); precision is TP/(TP+FP). Give the numerator and denominator beside each fraction, and call an empty denominator undefined. Activity definitions and false-alarm accounting are established concerns in evaluations such as ActEV. Our matching tolerance and hour-based measures are house rules. N01 · NIST evaluation plan
For the nuisance rate, divide FP by monitored camera-hours. For the negative-window false-positive rate, tile each session from its scheduled start into nonoverlapping 60-second windows, keep only full, continuously monitored windows with no reference event touching them, then divide windows containing at least one scorable alert onset by all eligible negative windows. Two alerts in one such window still add only one to its numerator.
Account for missing time
Coverage is monitored seconds divided by scheduled seconds. Take partially observed references out of the event-scoring set and report their count and reason, but keep that lost time in the coverage denominator. Measure receipt delay only for matched live notifications with synchronized clocks. An uploaded video can establish a source timestamp; its original recording clock cannot establish a live receipt delay.
Report counts by task, site, lighting and delivery mode. By way of illustration: one false alarm over two monitored hours gives 0.5 per hour, and two over ten hours give 0.2. Sort those rows by raw alarms and the ordering of the rates reverses. Neither calculation tells a buyer what happened during the excluded hours. Freeze the manifest, run evidence and scoring decisions together, so the next quarter can use the same definitions.
Record identifiers · L01 · L02 · N01