When organizing detection results into events awaiting review, define valid inputs before persistence and ending conditions. Time rules can organize notifications. They do not independently prove that a real abnormality occurred or that a person confirmed it. This article describes a proposed method; it adds no event tests or device deployment.
Record the frame and time associated with each result
Keep input sources, frame identifiers, result validity, and model and rule versions. Distinguish image acquisition time, result-receipt time and screen-display time. Do not directly subtract times from different devices without establishing the clock relationship. A count of processed frames does not automatically represent elapsed time.
Specify the clock used for elapsed time
Python time documentation explains that a monotonic clock is unaffected by system-clock updates. Its reference point is undefined; differences between readings are meaningful. This is a mechanism reference for elapsed time within a program. It does not establish synchronized device clocks, known image acquisition times or use of that mechanism by a particular SDK.
Separate persistence, ending and invalid inputs
| Condition | Proposed handling | Evidence to retain |
|---|---|---|
| Valid results meet the start rule | Create an event awaiting review | Input segment and start evidence |
| Valid results continue to meet the condition | Update event persistence | Observation times and rule results |
| Valid results meet the end rule | Record event ending | End evidence and subsequent review status |
| Missing frames, stream loss or invalid results | Apply the invalid-input rule | Missing interval, cause and recovery |
Review state changes with recorded-video playback
Missing frames must not automatically mean target disappearance. An event ending does not mean that a person confirmed the absence of abnormalities. Design separate playback checks for brief interference, persistent targets, changed frame intervals, stream loss and program restarts. Record triggering, persistence, ending and repeated notifications. Set thresholds using sample videos from the intended physical installation site. Do not publish an alarm-delay value or a false-detection reduction percentage without corresponding measurements.
Related services: Vision algorithms and model engineering, project consultation.