Distribution-Free Fleet Monitoring of Post-Repair ADAS Extrinsic Drift Under Sparse Telemetry Proxies
Abstract
Post-repair advanced driver-assistance system (ADAS) calibration certifies extrinsic alignment at a single shop instant, yet latent drift δt ∈ ℝ³ continues to evolve under fleet operation when only controller-area network (CAN) and telematics proxies cross the fleet boundary. We formalize shadow monitoring as sequential inference on proxy innovations νt = h(δt) + εt, where νt collects z-scored camera, radar, and lidar residuals. The static OEM comparator fires when ‖νt‖₂ ≥ τstatic with τstatic = 2.6. Our Pulse stack computes a Huber-robust Mahalanobis score via iterative reweighted least squares on a calibration corpus, then applies Mondrian conformal gating: CAUTION is emitted only if the score exceeds the stratum quantile at nominal level α = 0.05. Meridian augments Pulse with per-modality CUSUM statistics and ego-speed motion gates; Apex fuses modality-wise Bayesian online changepoint detection (BOCPD) run-length posteriors. Per-frame scoring costs O(d²) after O(d²n) calibration with d = 3.
On an open synthetic benchmark (n = 500 Monte Carlo trials, seed 42), static thresholds detect drift in 100% of trials versus 63.4% for Pulse, with zero positive lead time when both alert (median Δ = t_static − t_Pulse = −21 frames). Conformal Pulse achieves a 4.2% drift-free CAUTION rate within a ±2 percentage-point band of the nominal target. Extrinsic-coupled simulation at 0.03° day⁻¹ yields 0% recall within a 20-frame post-onset window for eight detectors, including CUSUM, Page-Hinkley, generalized likelihood ratio, extended Kalman extrinsic tracking, and the full NADIR combined stack. All tables, graphs, and heatmaps regenerate from JSON artifacts. We release Algorithms 1–4 for calibration, online scoring, Meridian layering, and Monte Carlo evaluation. Claim boundary: pilot-ready API and SDK prototype for fleet shadow monitoring; not automotive safety certification.
Relation to current work
The BOCPD changepoint machinery described here is reused for braking-onset detection in the forensics pipeline.
Cite
@techreport{hegde202601,
title = {Distribution-Free Fleet Monitoring of Post-Repair ADAS Extrinsic Drift Under Sparse Telemetry Proxies},
author = {Hegde, Dhruv and Balaji, Srivatsan},
institution = {NADIR Research},
type = {Technical Report},
number = {NADIR-TR-2026-01},
year = {2026},
month = jun,
url = {https://nadirai.net/research/nadir-tr-2026-01},
note = {Archived; synthetic evaluation}
}