Queue head detection
QA and correction of head boxes on overhead fisheye CCTV: more than 50,000 boxes a month for a confidential US client.
Problem
Public-venue people counting from overhead fisheye CCTV. Counting models are only as accurate as the head labels they learn from, and fisheye distortion makes those labels hard to get right.
Data and task
QA and correction of head bounding boxes, rather than first-pass labelling only: reviewing existing labels frame by frame, adding what is missing, removing what is wrong and tightening what is loose.
Tools
Roboflow, used for review and correction of the existing labels.
Challenges
Fisheye views of crowded venues produce four recurring problems:
- Missing labels on heads the original pass did not catch.
- False positives on objects that are not heads.
- Overlapping boxes where people stand close together.
- A long-hair rule, so heads are boxed the same way on fisheye views whatever the hairstyle.
QA
Dedicated QA plus senior sign-off on corrected batches.
Delivery scale
More than 50,000 boxes a month for a confidential US client, in settings such as casinos, malls and hospitality.
Result and next step
Corrected, consistent head-detection data delivered month after month, in settings such as casinos, malls and hospitality.
Related industry: Gaming, Public venues. Related service: Computer Vision.