Deep Learning to count manufacturing throughput using video analysis
The Challenge
Our client, a large towel manufacturing company needs to optimise throughput from their stitching tables. They had installed cameras to but needed a solution to use the video data to monitor employee breaks and automatically count the through put of each stitching table.
Hypothesis
- Advances in deep learning should allow us to build a pipeline to identify the folding of towels and counting the throughput from each of the tables captured by the installed cameras.
Execution
- As each workstation consisted of a table, a stitcher and a towel folder, in addition to an occasional supervisor and a loader, it was important to localise each workstation and automatically identify the roles of each person at the table.
- Pose estimation was used to extract key points of the folder. A sequence of key points was classified into a folding action of a single towel.
Outcomes
- Localisation of workspace delivered using object detection combined with human pose estimation and a mixture of heuristics.
- Folding of towels accurately predicted despite frames where the folder was obfuscated, using bespoke neural architecture designed by our team.
- Counts of towels output and time durations of breaks by workers output to SAP.
- Individual workstations successfully identified.
- Throughput and monitoring of employees accurately identified and integrated into SAP.
Project Highlights
92%
of employee line activity captured for reporting.