Real-time multi-camera vision for battery cell sorting

Problem

Cells arriving for recycling are not a uniform stream. They differ in format, in condition, and in what is printed on them — and the value of the downstream process depends on routing each one correctly. Doing that by hand does not scale, and doing it by sampling means accepting that some fraction of what goes downstream is wrong.

The goal was to inspect every cell, inline, without becoming the slowest step on the line.

Constraints

The constraints did more to shape this system than the algorithms did.

• Line rate is not negotiable. Inspection had to complete inside the time budget the line already had. A correct answer that arrives late is a stopped line.
• Cells are in motion. Anything that assumes a static subject is unusable.
• The environment is a factory, not a lab: vibration, ambient light that changes across a shift, and dust on optics.
• Two different questions — dimensional and visual — needed different sensing, and both answers had to arrive for the same physical cell.

Approach

Volume by laser triangulation. A projected laser line is imaged off-axis; the deflection of the line encodes height. Sweeping that across the moving cell yields a profile, and the profile yields volumetric measurement — without ever stopping the part.

Two camera cells, two jobs. Six monochrome cameras handle OCR, where contrast matters more than color. Six RGB cameras handle classification, where it does not. Both cells run pulsed lighting with short exposure, which is the piece that makes imaging a moving part work at all: the strobe is brief enough that the cell is effectively frozen, and bright enough to overwhelm whatever the ambient light is doing that day.

Capture and compute. I wrote the C++ GigE drivers handling frame capture and cross-camera synchronisation, feeding a rack-mounted multi-GPU compute module that ran inference for all twelve streams. Synchronisation is the subtle part — twelve cameras observing one moving object are only useful if you can say with confidence which frames belong to the same cell.

Outcome

The system ran inline on a live production line, inspecting every cell rather than a sample, and emitting a sort decision back to line control inside the cycle budget.

The lesson I keep from it: on a manufacturing line, the interesting engineering is rarely the model. It is synchronisation, lighting, and the discipline of making a system that degrades honestly when a camera drifts out of calibration at 3am.