There is a simple solution. In the analysis, they need to fix the width of the lines they use when comparing the position of the two players relative to human error. Basically, they should measure the variation produced by the operators when determining lines on a test set of images. The width of the line could then be expressed as a "confidence interval" rather than an arbitrary depiction.

Then when they perform the analysis during a live game, if the two lines are touching the ruling should be in favor of the offensive player (or the original call). Two overlapping lines would indicate that the system was not sensitive enough to truly determine if the attacker was in front of the defender.