Real-time vision pipeline

A real-time computer-vision pipeline that watches conveyor belts and detects objects with 97% accuracy — while never blowing its 50-millisecond frame budget at 20 FPS.

The problem

Object detection on a conveyor belt is unforgiving in a way offline ML is not: the belt does not pause while you infer. If a frame takes too long, you either fall behind the live feed or silently drop product. The system needed to handle two camera streams concurrently, on real hardware, with latency you can see.

The design

Detection. YOLOv8n — the nano variant — chosen deliberately: the smallest model that hit the accuracy bar, leaving headroom in the frame budget for everything that isn't inference.

Streaming. Camera feeds are ingested through MediaMTX, and results are pushed to clients over FastAPI WebSockets, keeping the browser view close to live rather than polling.

The StreamProfiler. The piece that makes it production software rather than a demo: a custom profiler that instruments every stage of the pipeline — capture, decode, inference, post-processing, publish — per stream, per frame. It tracks where the budget goes and flags any stage that threatens the 50ms ceiling, so degradation is visible and attributable instead of mysterious.

Results

• 2 concurrent streams processed in real time
• 97% detection accuracy on the production object classes
• <50ms frame budget held at a stable 20 FPS
• Per-stage latency visibility via the custom StreamProfiler