> "Pedestrian outside crosswalk not assigned goal of crossing street", "Tracking history not considered when classification changes", "Predicted path depended on object’s goal"
To be fair, I could see myself implementing these sort of heuristics to get a working prototype. On the other hand, I deliberately avoid working on life-critical software because of how easily it can go wrong.
I work in AV eng/research, and reading the details of the Uber crash was utterly shocking to me in terms of the corners cut. I can get reflexively frustrated at the slowing of progress and pushing changes induced by our safety constraints and testing, but I've always been pretty happy with the choices we've made. Uber made multiple obscenely negligent choices that I don't think are remotely defensible. There are plenty of hacky heuristics in the industry, but it's pretty central to industry practice to avoid hacks that make you _less_ safe.
A human driver would also follow the same heuristic. That is why it was programmed that way. People don't assume people are going to randomly walk in front of their cars except at cross walks.
Idk about you, but if I see a human start to step into the road I slow down/stop depending on how close they are. Someone stopped and facing the road looking both ways obviously may be attempting to cross, even if "illegally"
I'm enough of a jerk that sometimes, if they are at the edge of the road, I will stop and honk. Even if they are engrossed in their phone. Because, in my jurisdiction, it is illegal to not give the pedestrian the right of way at a crosswalk. Pedestrians who aren't planning on entering the roadway shouldn't stand at a crosswalk and not enter.
