What we do
Detect on site. Confirm with a human.
Unmanned 24/7 studios trade staff cost for open hours nobody is watching. Obscala watches the camera feeds where they already terminate and escalates only what a person needs to see.
What we detect
Fall
Pose geometry flags a person down on the floor. The alert never waits on knowing who they are.
Camera tamper
A covered, moved, or blinded camera is an incident in an unstaffed room, not a maintenance ticket for Monday.
Intrusion
Presence in a zone that should be empty for that time of day, from schedule plus region-of-interest rules.
Tailgate
More bodies through the door than check-ins at the reader. Counted at the door, no recognition needed.
The loop
- 01
Edge
A GPU box in the studio decodes the camera feeds, blurs faces on the analytics path, and runs the detectors locally. Only event metadata and a short encrypted clip leave the building.
- 02
Cloud
The API records the event, raises an alert, and pushes it to connected responders over a live socket. Clips stay encrypted at rest.
- 03
Responder
A person opens the alert, plays the clip, and acknowledges, confirms, or marks it a false alarm. Adverse attribution is never automatic.
What we do not do
- No face recognition, gait, or body-shape matching. Session continuity uses external, non-biometric signals only.
- No member name attached to routine tracking. Identity comes from the access log, on a specific incident, and adverse attribution waits for a human.
- No analytics on identifiable individuals. Operators see aggregate occupancy and utilisation.