Teach the Car to See: The Perception Validation Dossier

A pedestrian-detection feature ships over the air in two weeks and you hold the safety sign-off. Meet the night-road footage where the detection boxes flicker and drop, see the empty five-tab validation dossier you will build, accept the pass bar and the public-data norm, and bank a gut call: ship or hold, and the one condition you fear most. That prediction comes back at the end.

Learn the camera, lidar, and radar suite (each strong somewhere, blind somewhere) and pin down what catching something even means. In a Python cell you compute Intersection-over-Union on a worked pair and call it a hit, miss, or phantom, locking the IoU 0.5 threshold and the false-negative definition that the rest of the dossier is scored against. Produces Section 1 of your dossier.

Precision, recall, and the VRU false-negative rate, then run them yourself. In a Python lab over the real derived open-AV detections-vs-ground-truth table you get the confusion counts and read off your own precision, recall, and false-negative rate on pedestrians and cyclists. The aggregate recall looks reassuringly green, then the act bridges out on the warning that an average can hide a body. Produces Section 2.

Slice recall by condition and watch VRU recall collapse at night and under occlusion, then triage the failure modes so the rare-but-catastrophic miss does not sit in the green band. Write the ISO 21448 SOTIF safety case for each top mode (trigger, functional insufficiency, hazardous behaviour, concrete mitigation), then sign your GO / GO-WITH-CONDITIONS / NO-GO verdict citing your own numbers, with a live consistency check that challenges a self-contradictory call. Produces Sections 3, 4, and 5 and completes the dossier.

The camera pulls from your finished dossier to the real Uber Tempe 2018 fatal crash (NTSB), where the perception system flip-flopped the pedestrian's class and suppressed the brake at night, with your slide-one prediction surfaced beside the evidence. Write an evidence-based analysis of which of your own failure-mode rows the real system blew, anchored to a real public figure, teach back why recall on VRUs is the safety metric, then sign and submit the dossier.