Manual counting doesn't scale
Fish farmers were losing money on discrepancies between the agreed sale amount and the actual count — because seeds were counted by hand, one at a time. Sorting and counting 500 seeds (1–3 cm each) took about an hour; a single sale could run 1,000–3,000 seeds. Small errors at low volume became real losses at scale.
A camera rig that counts as seeds pass through
The physical setup channels seeds through four zones — camera, inlet, counting area, outlet — so each seed is measured exactly once as it crosses a tunable detection width.
Processing pipeline
- Take a picture (camera)
- Segmentation
- Median filter (noise cleanup)
- Morphological operations
- Object detection
- Determine centroid point
Counting logic
- Check if contour width & centroid meet threshold
- Frame the object, locate its centroid
- Is the centroid inside the virtual detector?
- If yes → count it; if no → loop
- Reset counter, repeat per frame
Why a "virtual detector," not a fixed line
Rather than counting every object seen, the system only counts a centroid the instant it crosses a defined width inside the counting area — this made accuracy tunable, and testing different widths was the whole optimization loop.
Where accuracy landed
Published research
Type: Consultancy Project · Level C2 · October 2016 · Read the paper on iopscience.iop.org ↗