These are important questions, but addressing them for each model built independently and optimizing for a low “racial bias” is the wrong approach.

In academia we have reference datasets that serve as standard benchmarks for data driven prediction models like pedestrian detection. The numbers obtained on these datasets are usually the referentials used when comparing different models. By building comprehensive datasets we get models that work well across a multitude of scenarios.

Those are all good questions, but need to be addressed when building such datasets. And whether model M performs X% better to detect people of that skin color is not relevant, as long as the error rate of any skin color is not out of an acceptable rate.