The problem is this: a human can make a mistake but a machine cannot. For example a human driver may need to choose between swerving left or swerving right and, in hindsight, either choice could be defended because we know he had to make a split-second judgment, the outcome wasn't premeditated, etc.

With a self-driving car though, some programmer sitting in an office in the cold light of day had to make a conscious decision to kill the driver, or to kill the pedestrian. The outcome is programmed, and there are no mitigating factors like instinct or accident that can excuse it. This poses a very different moral problem for the creator of the self-driving car.