Retrofit Shelf Camera + Local Vision Model

## FridgeEye: Privacy-Preserving Shelf Vision ### Architecture 1. **Camera**: Raspberry Pi Zero 2W + wide-angle CSI camera, mounted inside upper shelf lip. €18 hardware. 2. **Inference**: runs locally using a MobileNetV3 model fine-tuned on grocery categories (~120 items). Runs at 2 FPS, uses <200MB RAM. 3. **Inventory model**: tracks items by visual delta (item disappears/appears = remove/add event). 4. **Interface**: local web app only — no cloud, no external API calls. ### Privacy model All processing on-device. The camera activates only when the door opens (magnetic reed switch). No video recording — only inference results stored. Image buffer: 3 seconds, then wiped. Compliant with GDPR Art. 5 data minimization principle. ### What it can't do Distinguish between two nearly identical items (two cans of similar beans). Requires good lighting — mount a small LED strip if needed. ### What it can do - Alert when milk is running low (95% accuracy on major brands) - Detect opened/closed packaging (via contour change) - Generate shopping list suggestions based on depletion patterns **BOM total**: €28. Payback in avoided spoilage: €8–15/month for average German household.