I’m no expert at all, but just from reading your post and the comments, I’ve got 2 ideas… they might be no good, but I’d like to share them in case they can help a little bit:
- For the late turning that fails to “cue human trust”, I’m inspired by Shauna Gordon’s comment to suggest some simple interface improvement that might avoid both pitfalls of terrifying the driver (so the feature might not be trusted and used as much as it can be) or inducing a exaggerated sense of security (so the driver might fall asleep at the wheel and a real accident might happen): Every time the car needs to turn, after the computer perceives the turn ahead, and about the time when the average driver starts to move her foot to the pedal (see Shauna’s comment), the computer should sound an alarm (the ringtone should be customizable :), and flash the screen with the words Turn Ahead / Preparing to Turn. This way the driver can both be reassured that the computer has “seen” the turn, and also be kept alert in case there is some unexpected occurrence that needs human intelligence to take care of.
- For the problem with classifying: it is great to know that the software is being improved and updated every 2 weeks; but I guess in the near future the classifying function would still be far from perfect. A easy way to remedy the situation, along the same line of thought as idea No. 1, would be to program the interface to sound a (louder) alarm and maybe change the color and/or brightness of the screen whenever the computer fails to classify an object within a certain distance set by research; the alarm should be louder when there are more unclassified objects within that distance, and loudest when most objects in that area are unclassified.