From what I can tell it looks like the pedestrian’s vector is heading away from the car (right to left) for a few frames before he pauses. Couldn’t a level 5 model infer intentionality in a direction based on the way the pedestrian was walking?
Stating the obvious here: It's pretty interesting to think about the tradeoffs here. As a user I understand the obvious privacy benefits of this move. As a developer, I found having access to that data quite useful sometimes.