Optimise number of frame to create detections

[MateoLostanlen]

Pyro-Engine currently combines n frames (currently 6) by averaging the scores of overlapping bounding boxes. Based on the results from this summer, the detection seems to be overly sensitive: it often triggers on very short smoke events (mostly industrial emissions) that are not relevant for firefighters.

Once the summer data is fully annotated, we should revisit this algorithm — including the number of frames used and the method for combining detection scores — to reduce false positives and improve relevance.