How does this handle motion detection? Anyone tried it?

I have a bunch of Nest cameras and some cheaper Tapo cameras.

The "motion detection" is night and day different. The Tapo does some basic frame-diffing and it's awful. Shadows? DING DING DiNG MOTION DETECTED! gust of wind made some blades of grass move? Motion! Spider? Motion motion motion! Turn down the sensitivity a notch or two and it won't notice a human walk across the frame 2 meters away. The Tapo ones claim to be smart but twigs and leaves still trigger pet/person/motion alerts. It makes them essentially useless.

The Nest cameras are so much better and their "human" detected is usually zero-false-negatives at the cost of one or two false-positives perhaps once every 3 or 4 months, and their app is superior (tapo one frequently needs to be false-killed to load clips). Yes I am aware that the nest ones are streaming back to google 24/7.

Tldr: naive frame-diffing sucks for this sort of thing if used outside. An open source implementation that has accurate and reliable "human detection" would be amazing. Doesn't need to be "AI" - I would hope that there is some sort of computationally reasonable OpenCV way of doing person detection. Perhaps wait for frame-diffing to flag motion then feed it to a more expensive algorithm?

Seriously… how hard is it to setup an RPi with a camera and offload the detection to some open source software? Last time I checked there projects performing animal, face and even license plates reckon, extraction and db recording of “who”, “when”, “what” using cheap cameras and an RPi.

Try the motion Linux package, it can connect to tapos over SDP and then you can run Yolo or any other solution on the output