Cameras see well. RFID counts well. Each lies in its own way. The interesting work in industrial computer vision isn't about either sensor alone — it's about fusing them so the lies cancel out. Here's how we do it in production.
Why neither alone is enough
RFID gives you "this tag is in this zone" with near-perfect counts but no spatial information beneath the antenna footprint. Vision gives you precise positions but is blind to occluded objects, mistakes carbon-fiber for steel, and breaks under shift-changing light.
Fusion is where the magic happens. The camera tells you where; the RFID tells you what. Together you get item-level position tracking that beats either modality alone.
The pipeline
In production, our pipeline runs four stages — detection, association, fusion, and reconciliation. Each has its own failure modes.
Stage 1: Detection
A fine-tuned YOLO-class model identifies object candidates in each frame. We train per site because lighting, angles, and object classes are too varied for a generic model.
Stage 2: Association
RFID reads are timestamped. Vision detections are timestamped. We associate them using a Hungarian-algorithm-style assignment that considers spatial proximity, RFID signal strength, and time alignment.
Stage 3: Fusion
Once associated, an Kalman-filter tracks each object's position across frames. RFID updates anchor the track to identity. Vision updates refine the position.
Stage 4: Reconciliation
A nightly job reconciles the day's tracks against ground truth — manual cycle counts, ERP records, or the next-day reality. Drift gets fed back into retraining.
Calibration tip
The biggest source of error in fusion is RFID antenna placement. Spend a week walking the site with a calibration cart before you trust any model.
We see 99%+ match accuracy in production on yard operations. That number breaks if the antenna placement breaks, if the camera angles change, or if a new SKU enters without retraining. Maintenance discipline matters more than model architecture.