A Benchmark Protocol for Measuring Error in Video-Based ADAS Crash Reconstruction
Abstract
Courts increasingly hear disputes about whether a vehicle's driver-assistance system could have perceived a hazard and should have intervened, and the evidence available after a crash is usually a single dashcam or onboard video plus the event data recorder (EDR) pre-crash record. Reconstructions built from such evidence are only as useful as their known error.
This report pre-registers the protocol NADIR Research will use to measure the error of its video-based reconstruction pipeline before any accuracy figure is published. We specify the quantities estimated (ego and actor speed, position, relative gap, time to collision, event timing, and the probability that a modelled sensor suite would have detected a given actor), the ground-truth sources in three stages (public driving datasets with lidar and inertial ground truth, instrumented closed-track runs with RTK-GNSS on every vehicle, and blind comparison against credentialed reconstruction experts on closed cases), the metric definitions and interval coverage tests, the reporting rules, and the threats to validity. Acceptance thresholds are deliberately left as explicitly marked placeholders to be fixed by the team before data collection. No results are reported here.
Cite
@techreport{hegde202603,
title = {A Benchmark Protocol for Measuring Error in Video-Based ADAS Crash Reconstruction},
author = {Hegde, Dhruv and Balaji, Sri},
institution = {NADIR Research},
type = {Technical Report},
number = {NADIR-TR-2026-03},
year = {2026},
month = oct,
url = {https://nadirai.net/research/nadir-tr-2026-03},
note = {Draft}
}