Seems a bit complicated and bandwidth heavy. I wonder if it really performs significantly better than just processing only the keyframes of the full HD stream

Seems a bit complicated and bandwidth heavy. I wonder if it really performs significantly better than just processing only the keyframes of the full HD stream? Most activities people want to monitor last longer than a keyframe interval, and object detection is pretty good these days. Plus, if you do your brief local cache with HLS segments, the HLS multiplexer makes sure that every segment is a valid independent block starting with a keyframe, so you can always reliably catch the activity that happens even a few seconds before the actual trigger event. But I suppose the 2 stream model makes sense if you want to record 24/7 no matter what, or if you need to catch brief faraway events that might not be clear enough to trip the motion/object detector.