01Choose supported object classes
Use the available catalog, for example person, car, truck, couch, bed or suitcase. Car and truck are separate classes: select both when you want both. A bed label does not guarantee reliable mattress detection.
Selected classes describe objects, not intent. A couch in an elevator is not by itself proof of moving house; clothing color and helmet compliance are separate capabilities.
Choose objects and confidence in rules. Rules on the same camera share inference. Search labels come from observations meeting the corresponding rule object threshold, not a separate history setting.
02Set a confidence threshold
New configurations default to 70%; the editor accepts 10%–95%. Existing saved thresholds are not automatically changed. Higher thresholds generally admit fewer weak detections but can miss objects; lower thresholds admit more candidates and false positives.
A 70% score is the model’s confidence filter, not a guarantee of 70% accuracy. Small objects, low light, occlusion and compression can reduce reliability. Thumbnail badges show configuration, not current model health.
Scores belong to the event, not today’s settings. Current rules require the trigger-frame and recent tracking scores to meet their threshold. Scores are not accuracy guarantees.
03Connect detection to rules
Choose objects and confidence in rules. Rules on the same camera share inference. Search labels come from observations meeting the corresponding rule object threshold, not a separate history setting.
Searchable observations do not necessarily trigger an event: scope, activity, duration and notification cooldown are evaluated separately.
04Plan processing capacity
Each camera frame requires inference; multiple object rules can reuse that camera’s detections. Fewer selected classes do not reduce the main YOLO inference cost in proportion to their count, so there is no five-class hard limit.
More cameras, higher frame rates and PPE checks add work. Validate the actual processor’s throughput and image quality instead of treating the number of checked labels as a GPU budget.
Common questions
Do I have to draw a region?
No. A presence rule can use the whole image. Draw a zone only when you want to restrict where it applies.
Will adding a rule re-analyze every old image?
No. An administrator must explicitly start a history analysis job on retained images; it does not create live alerts.