Real-Time AI Object Detection for Conveyor Safety

Challenge

Recycling conveyor lines regularly encounter hazardous foreign objects mixed into incoming material — batteries, gas canisters, metal parts — that can cause fires, explosions, and equipment damage if they reach the wrong processing stage.

Manual detection wasn’t reliable enough. Operators couldn’t consistently catch these objects in time: the conveyor moves fast, visibility conditions vary, and the margin for error is narrow.

The client needed a system that could detect dangerous objects early enough to enable either an automatic line stop or an immediate operator alert — and one that could prove itself under real operating conditions before committing to full production deployment.

What We Set Out to Validate

This engagement was structured as a deliberate Pilot. Before building a production system, the client needed to answer a specific question: can an AI-enabled computer vision model detect hazardous objects in real time, using industrial cameras, under the conditions that actually exist on a working conveyor line?

Three validation criteria shaped the Pilot scope:

1. Real-time detection at production latency. The model had to perform inference on live video streams quickly enough to be actionable.
2. Resilience to operating conditions. The solution had to handle the variation that makes manual detection unreliable: changing lighting, partially hidden objects, and inconsistent frame quality.
3. A clear path to on-premise production. The Pilot was designed cloud-first but architected from the outset to support migration to on-premises infrastructure — so a successful validation directly translates into a production deployment plan.

Solution We Delivered

We built a real-time object detection pipeline that ingests live camera footage, runs GPU-powered inference using YOLO, and routes every detection event through an event-driven architecture that drives operator alerts and a live monitoring dashboard. The full Phase 1 scope — camera setup and data capture, annotation preparation, initial model training, deployment pipeline, and field validation — was delivered in three months on Google Cloud Platform.

We also evaluated RT-DETR as a challenger model alongside YOLO during the Pilot to establish a solid baseline for production model selection.

Alongside the live detection system, we built the data infrastructure for sustained model improvement: a structured data-collection process to capture footage of dangerous objects under real operating conditions, and an annotation workflow to label and prepare between 1,000 and 10,000 frames for training and validation.

Architecture and Workflow

The pipeline runs end-to-end on Google Cloud (GKE) and was designed so each component can be containerized and migrated to on-premise infrastructure for production deployment.

What the Pilot Established