Temporal model on Raspberry Pi #2

During summer 2026, the temporal model ran on a server (see #1): the Pis send their detections, the server filters them with the temporal model, and only the kept ones are shown. The results are good, so we now want to run the model directly on the Pi. This removes most of the server load and cost, which matters for Chile.

How it will work

• The Pi runs the temporal model: it sends a detection only if the temporal score is above a low threshold (for example 0.1), together with the score. The API then shows only the ones above the real threshold, the optimal one found in Analyse the summer 2026 performance vision-rd#143 (epic Improve the temporal model #1). Detections between the two thresholds are stored for future analysis.
• The Pi does not run the temporal model (older Pi, service off): the server runs the model, as today.

Work already done

pyronear/pyro-engine#415 runs the ONNX version of the model on the Pi. Tested on a Pi 5: about 8 s per job and about 1 GB of RAM.

Done when

• Pis with the temporal model send their score, and the API uses it without calling the server
• Pis without the temporal model still work as today
• Detections between the low threshold and the real threshold are stored, not shown

Theme Engine · Priority P1 · Period T4-26

The breakdown is tracked in the sub-issues below.