This is not a good argument overall for several reasons. First, we should aim to greatly exceed human performance on safety. Second, whatever work goes into making the algorithms work well using vision can just as easily still inform a system that fuses that data with LIDAR for enhanced situational awareness and safety.
