Sigma practical guide: missing-helmet event
The training design extends the previous graph. This is a project recipe that must be tested on your own dataset; it does not claim that the demonstration draft already contains every node listed.
Video Frame → Person and PPE Detector → Person Tracker
→ Add PPE attributes to the track
→ Production Zone Filter
→ Confirm the “Helmet-” attribute over time
→ Create Event → Send Event
↘ Local Visualization
| Stage | Setting | Check criterion |
|---|---|---|
| Tracker | Track the person using existing detections | One person has a stable ID; a neighboring person is not merged into it |
| PPE association | Associate the head/helmet with the correct person | Another person's helmet does not conceal the violation |
| Zone | Production zone type and intersection rule | Passing outside the boundary does not create an event |
| Confirmation | Confirmation count and window consistent with FPS | A brief error produces no event; a persistent violation does |
| Repetition | Interval and conditions for ending an episode | A series of adjacent frames does not turn into a stream of duplicates |
| Result | Violation type and required event fields | The log has the correct camera, time, category and image |
Calculation example, not a production recommendation: at 2 processed frames/s, a 10-frame window spans about 5 seconds. Requiring 6 positive observations means “at least 6 out of 10” if that is the mode implemented by the selected block; it does not mean “continuously for 3 seconds”. Explicitly compare the required business rule with the window implementation.
Test: helmet worn; helmet missing; head status unknown; a person crossing the boundary; two people overlapping; a person putting on a helmet; missing frames; stream recovery. After acceptance, publish and assign to one camera according to §4.11.
