But it is more complicated than that when you're talking about algorithms and complex systems. If you had three of the same exact system, they'd likely all make the same mistake and thus you'd gain no safety improvement (except actual malfunctions, not logical limitations).
I would like to see auto-braking taken out of autopilot completely, so if autopilot drives you into a wall at least the brakes slow you.
I'm not an expert...but it seems like for airplanes there are standards that manufacturers need to abide by for avoidance systems...which would mean testing and certification by an independent association before it's deployed. With manufacturers doing OTA updates...there's really no guarantee it's been tested...you'd have to trust your life to the QA process of each manufacturer. Not only the one of the car you're driving...but the car next to you!
Regardless, it should be trivial have a different behavior for a moving object that enters the road versus a rigid object that has not been observed moving.
On top of that, Kalman innately tracks the uncertainty of its combined estimate. So you can simply look at the Kalman uncertainty covariance and decide if it is too much for the speed you are going.
I really wonder if Tesla is doing that...
Basically it's a matter of having multiple redundant sensors and sounding the alarm if they don't all agree on what is going on, and also checking if what they think is going on is in the realm of reasonable possibility based on some a priori model (e.g. if suddenly all of your sensors tell you that you are 100 miles from where you were one second ago, that's probably a mistake even if all the sensors agree). That's a serious oversimplification, but that's the gist.