Well if it's using optical recognition then lower contrast or smaller object will always make the job more difficult then a bigger object with more contrast; then the reason why the model is better at recognizing white adult people. Even if the car is never going to strike a child or a black person the model recognition would still be better with them.