VialVision : Edge-AI Computer Vision

Background

Manual inspection of test tubes and vials in laboratories and small-scale manufacturing environments is often slow, subjective, and difficult to scale. Cloud-based vision solutions introduce latency, infrastructure cost, and dependency on network availability.

VialVision is an ongoing project exploring whether a low-cost edge device—Raspberry Pi—can reliably run deep learning–based computer vision to detect, classify, and measure test tube outputs fully on-device.

The Problems

• Manual inspection lacks consistency and traceability
• Cloud inference adds latency and infrastructure complexity
• Edge devices have limited CPU, memory, and thermal headroom
• Camera pipelines introduce color and white-balance inconsistencies
• A usable system requires both UI and API, not just a trained model

Our Approach

VialVision is designed as an end-to-end edge AI system rather than a standalone model. The solution combines optimized YOLO inference, a FastAPI backend, and a lightweight web interface to support both image upload prediction and future live streaming.

Architecture

Camera / Image Upload
        │
        ▼
Raspberry Pi (Edge)
- YOLO Inference
- Pre/Post Processing
        │
        ▼
FastAPI Backend
- REST API (/predict)
- Future WebSocket Stream
        │
        ▼
Web Interface
- Image Upload
- Visualized Detections

The architecture cleanly separates hardware interaction, inference logic, API exposure, and UI rendering, enabling rapid experimentation without coupling changes across layers.

Results (Ongoing)

VialVision has reached a working end-to-end prototype. Image upload prediction produces annotated outputs, and the FastAPI service runs reliably on Raspberry Pi. Current work focuses on performance optimization, camera color stability, and preparing the system for live stream inference.