Camera-based perception is the hardest part of an ADAS stack to sign off on: the software behaves as designed, yet still misclassifies a backlit pedestrian or drops a target on a tunnel exit. Much of that exposure sits in SOTIF territory, not classic functional defects. This session shows how a [front-camera detection and classification] function was validated by shifting validation left – replaying every recorded drive as a regression asset against frame-accurate ground truth, then layering synthetic edge cases for glare, rain and occlusion. The presentation will cover ground truth challenges, the KPIs that correlate with field behavior.

• How to set up a shift-left, resimulation-based validation loop that replays recorded drives as regression assets and cuts road-test dependence
• Which perception KPIs actually correlate with field behavior, and how to build ground truth you can trust for camera detection and classification
• How to use synthetic edge cases to address SOTIF risks, and where simulation stops and real-world road testing still has to take over