Human pose keypoint annotation
Seventeen-keypoint skeletons for workplace-safety AI on industrial CCTV: more than 250,000 skeletons for a confidential EU client.
Problem
Workplace-safety AI needs reliable human pose estimation from industrial CCTV. To recognise posture, movement and activity, a model first needs consistent skeletons on real site footage.
Data and task
Seventeen-keypoint skeletons with visibility and occlusion rules, annotated on overhead industrial footage. Each keypoint carries a visibility state, so the model learns the difference between a joint it can see and one it has to infer.
Tools
Encord, used as the annotation and review platform for the programme.
Challenges
Busy industrial scenes make pose hard to label consistently:
- Occlusion: workers partly hidden by machinery, racking and other people.
- Left and right orientation: keeping limbs correctly assigned from overhead angles.
- Visibility tags: applying the same rule to every partly visible joint.
QA
Two-tier review, so skeletons are checked at two levels before delivery.
Delivery scale
More than 250,000 skeletons, delivered as one of three concurrent projects for the same confidential EU client.
Result and next step
A large, consistently annotated pose dataset for workplace-safety AI, delivered alongside the PPE compliance and workplace safety object detection projects for the same client.
Related industry: Industrial and workplace safety. Related service: Computer Vision.