Rejected events, for manual review
A spread of the events the filter suppressed, sampled across every category and every day. Each was decided automatically and no email was sent. The question for each one is simple: should it have been?
How to use this
Judge from the zoom first - at full-frame scale these subjects are a few dozen pixels, which is exactly what made an earlier review miss a real person. The full frame underneath is only for context.
Note the numbers of any case that looks wrong. A rejection is wrong if a person or vehicle did cross the line. Everything else - someone walking past, a parked car, an insect, a shadow - is the filter working as intended.
Camera saw something we did not — 23 of 85 shown
The category that hides false negatives. Our one confirmed miss came from here. Look at the zoom: is there a person in it?
Present, but never crossed the rule — 34 of 85 shown
Someone was there and we tracked them, but their path never intersected the tripwire. The largest rejection category, and the one that justifies analysing a window of frames instead of a snapshot.
Nothing of interest held across frames — 16 of 85 shown
No person or vehicle survived two frames. Expected to be shadows, headlights, rain and vegetation.
Present, but stationary — 12 of 85 shown
Detected, but it never travelled more than 15 px - parked vehicles, usually. Watch for anything that was actually moving slowly.
