How a semi-automated labeling pipeline turns raw captured Basic Safety Message streams into annotated ground-truth events — near-misses, occluded-vehicle warnings, forward-collision precursors — used to score the vehicle's V2V software.
When two instrumented vehicles pass near each other, their independently logged RTK trajectories are time-aligned and differenced to compute a ground-truth closing speed and time-to-collision curve accurate to centimeters — a far stronger baseline than trusting either vehicle's own onboard estimate of the other's position derived from decoded BSM content.
Once an event corpus is labeled, the vehicle's V2V consuming software is run against each captured trace and scored on whether it raised the correct warning within the correct time window — precision and recall against the human-confirmed labels become the primary release-gating metric for every software build.