You don’t need artificial intelligence or cameras to do that. Inductive loops can detect cars and bicycles and their velocity just fine and control nearby traffic lights, and have been doing that for decades.
I would think such local control is less prone to failures, too.
Interview with a traffic engineer (IMO one that’s a bit too focused on throughput): https://bicycledutch.wordpress.com/2016/06/21/traffic-lights...
Their systems aren’t completely local, given “every bus is tagged and the computer knows all the time-tables of every line. The system knows exactly when which bus needs to arrive at a certain intersection. The installation checks whether a bus is too early, on time, or running late and sets the lights accordingly”
The problem is you have to do this city by city and they are so many local regulations not to mention how cheap they are with budgets.
Hell no. You want something that's explainable and predictable, not something that will allow a truck to run over a bunch of kids, just because it's painted in a pattern engineered to make it be identified as a bunny.