An office wants to know how its space is used: which desks sit empty, how long people are actually at them, when the room fills up. The usual answer is badge readers, which measure who entered the building, or camera systems with face recognition, which measure people.
This measures desks.
From an ordinary fixed camera it reports, per workstation: how long it was occupied, how long it stood empty, how many separate sessions there were, the longest one, and whether someone is sitting there right now.
No face recognition. No identity. No per-person tracking. A polygon is drawn around each desk and the only question ever asked is whether a person is inside it — which is the question the office actually had, and it happens to be the one that does not require surveilling anybody.
Setting it up takes a minute
The first frame of a recording opens and you draw around each desk with the mouse. Zones can be renamed afterwards — to a room name, a team, or a person's name if that is what the office wants.
Then one command turns a recording into a report and an annotated video:
ZONE | STATE | OCCUPIED | EMPTY | % | SESSIONS | LONGEST
-----------+----------+-----------+---------+-----+----------+---------
Ish-joyi-1 | OCCUPIED | 1h 23m 10s| 12m 04s | 87% | 3 | 52m 18s
Ish-joyi-2 | EMPTY | 24m 51s | 1h 10m | 26% | 5 | 9m 33s
The hard part is not detection
Detecting a person is solved. The difficulty is that a raw per-frame answer is wrong constantly — someone walks past the desk, someone bends down and vanishes for two frames, the model blinks.
Two waiting periods fix it. A desk only counts as occupied after someone has been there for several seconds, which ignores everyone walking by. It only counts as vacated after they have been gone for considerably longer, which absorbs the dropouts. Without both, a report full of one-second sessions is technically accurate and completely useless.
Measured, not claimed
Validated against the Edinburgh office dataset, which ships per-frame ground truth: 99.88% accuracy — four seconds of total disagreement across a 1 hour 52 minute recording.