When a production line needs low-latency decisions, sending every frame to the cloud may be unsuitable for network capacity, latency and data-transfer reasons. Edge inference near the camera can reduce those risks. A Hailo accelerator with Raspberry Pi 5 is one architecture option, subject to model compatibility, cycle-time, thermal and maintenance tests.
Why not the cloud?
High-frame-rate uncompressed video can consume substantial bandwidth. Calculate the required bitrate from the actual camera format, then measure WAN capacity, jitter and end-to-end round-trip time on site. If the actuator's decision deadline is shorter than the measured cloud path, use local inference or another deterministic control layer.
Model size and quantisation
INT8 quantisation can reduce model size and compute cost, but device memory, compiler support, latency, power and thermals require separate validation. Use a representative calibration set and compare precision, recall and error cost under line conditions. The lab-to-field gap is not a fixed percentage; report it alongside lighting, vibration, lot and camera placement.
Operator screen
The edge box should not emit only a score; it should show the decision image, confidence and an access-controlled set of recent examples. Set the number of records from retention and screen needs. Test gloved use on site, and keep data-collection or training mode off by default and permission-controlled.
Edge-AI performance is affected by mounting and environment as well as the model: include camera angle, LED flicker, dust and enclosure temperature in acceptance tests. Verify mechanical and lighting stability before interpreting model metrics.
