The authors made an assumption that ad companies are using the same models as the open source project they tested this with. That's complete nonsense and not near reality at all due to the enormous difference in data used (source, I'm working with this tech).
Unless you test with the exact same models as are used in the cars this really proves nothing and really shows that the authors own misunderstanding and bias (eg lidar vs camera discussions, or overall size of children vs adults).
Does that include civilians, it sounds a bit on the low side for that for me, but I have no source or anything. I know that a lot of children was kidnapped and taken to Russia so that's quite more also.