From a study in Fairfax, Virginia:
The range in Fairfax is broadly consistent with other studies I’ve looked up (see page 33 of this report for summaries of a bunch of studies). The lowest rate I’ve seen is 2 violations per 10,000 vehicles or 0.02%.
This implies that autonomous vehicles need to stop for red lights 99.9% to 99.99% of the time to be ~2x better than the average for human drivers.
A quick search on Google Scholar turned up an example of a paper that achieved 86.6% mean average precision (mAP) on the Tsinghua-Tencent traffic light dataset and 91.1% mAP on the LISA traffic light dataset. Note that the training dataset used was small compared to what a major autonomous car company could compile. It was only 6,700 images.