"However I think the feedback could be improved very much if each of the remote students would also stream their video (all at once, without audio)."
Raw streaming video isn't going to do you much good for even just a class of 20ish. But you do make me think of a crazy idea where each student has their camera on, and the client does facial analysis to do what would basically be sentiment analysis. Then just the sentiments get sent up to the streaming server (saving tons of bandwidth, which the clients may not much care about but the central server still cares about even in 2020), and the server composites the sentiments together and displays 3 or 4 synthesized faces representing the general sentiment. So you look at your 3 or 4 synthesized faces as you teach, and if suddenly 2 of them get a quizzical look on their face, you know that there's a non-trivial amount of the class you just lost. Or, alas, if 3 of the 4 of them are staring off into space, you know you've lost the class.
It would be a bit weird, but might be an interesting way to hack our human systems for facial reading into being able to read a virtual crowd.
Two possibilities come to mind; one is where the system reads sentiments, and then if it decides one of the faces is going to be confused, plays a pre-canned "confused" animation. Functional, easy, unless someone very skilled works at it it's gonna have a very 1990s animation feel to it, I suspect. The other is more complicated, but potentially more rewarding, where the system uses some more advanced AI and things like eigenfaces to synthesize a more continuous representation of what the crowd is expressing, which may do a better job of looking more human and may be a much more information-rich way of expressing the crowd's sentiments. In this case I imagine the system trying to maximize the number of real faces represented by the synthesized faces; I'd want to set it up so that a crowd of 90 neutral people, 10 bored people, and 5 confused people get the bored and confused a bit overrepresented, rather than just having 3 or 4 equally neutral faces.
You'll also see I'm pretty stuck on having more than one face, despite the fact I kinda expect any system that tried to do this would end up with just one, but I suspect that trying to represent an entire diverse crowd with just one face will wipe away too much information, both resulting in less information for the user, and also eventually the brain of the lecturer categorizing this as something other than a crowd of people. I think you need multiple faces for this to work, and you'd need to do something like what language analysts do to seek out the real information in a sentence and not be afraid to amplify some of the signals only a small minority are putting out.
Anyhow, the foundational tech for all this all basically exists now; this is definitely something a startup-class effort could do.