Continuous Extrinsic-Miscalibration Monitoring Above a Fleet-Visible EyeCAN Flag Baseline: A Mixed-Regime Study of NADIR Pulse, Meridian, Apex, and Echo with Noisy Cross-Modal Residuals and Bounded Software Compensation
Abstract
Forward-facing advanced driver-assistance cameras require a stable relationship between the camera coordinate frame, vehicle frame, and any radar or lidar frames used by downstream fusion. A shop calibration establishes that relationship at one time, while vibration, repair variation, mount disturbance, glass replacement, temperature cycling, and mechanical aging create a continuing monitoring problem. This paper studies whether a fleet software overlay can identify emerging extrinsic yaw disagreement before or in the absence of a useful fleet-visible binary warning. The comparison is intentionally narrow. The baseline is an observable EyeCAN-style flag monitor composed of lane-departure, tamper, maintenance, and vehicle-state fields. It is not a reconstruction of Mobileye's proprietary internal online-calibration logic, and the study does not claim that Mobileye lacks internal safeguards or correction mechanisms.
We evaluate four NADIR service tiers using a detector-independent mixture of micro gradual drift, moderate gradual drift, severe events, and harsh noisy episodes. Unlike an earlier stress-only fixture, the generator does not cap every drift below a chosen flag threshold. Flag probability rises smoothly with yaw, and the overlay residual channel includes heteroscedastic noise, persistent bias, missing observations, isolated outliers, weather effects, vibration, and variable transport cadence. The primary study contains 800 drifted episodes, 300 clean episodes, and 185,664 simulated frames, followed by factor sweeps. Pulse uses robust instantaneous residual evidence and persistence. Meridian adds motion-gated cumulative change detection. Apex adds a lightweight change-posterior confirmation path. Echo performs lower-cadence batch confirmation. A bounded advisory transform is evaluated against latent yaw so that estimation error can reduce or worsen correction quality.
Across the mixed drift set, the EyeCAN-Flag baseline reached 48.9% recall. Pulse reached 82.3%, Meridian 89.3%, Apex 89.5%, and Echo 87.3%. The largest separation appeared in micro gradual drift, where recall was 18.5% for the baseline, 57.0% for Pulse, 77.0% for Meridian, 78.0% for Apex, and 55.5% for Echo. In severe events the baseline reached 90.5%, while every NADIR arm reached 100%, showing that the baseline remains competitive when faults become large and explicit. Clean-episode false-alert rates were 8.3% for the probabilistic flag fixture, 0% for Pulse and Echo, and 7.0% for Meridian and Apex. These values characterize the simulation, not a field estimate of Mobileye false-warning behavior. Median algorithm compute remained below 0.17 ms per scored frame, but end-to-end alert time is governed mainly by ingest cadence. Bounded software compensation reduced latent yaw in 92.8% to 98.6% of detected episodes, depending on tier. The evidence supports a complementary overlay hypothesis, not a general superiority claim. Field traces with synchronized calibration ground truth remain necessary before safety, OEM, or production-effectiveness claims are justified.
Relation to current work
The observable-output method (treating a production system only through what it exports) is reused in the counterfactual engine.
Cite
@techreport{hegde202602,
title = {Continuous Extrinsic-Miscalibration Monitoring Above a Fleet-Visible EyeCAN Flag Baseline: A Mixed-Regime Study of NADIR Pulse, Meridian, Apex, and Echo with Noisy Cross-Modal Residuals and Bounded Software Compensation},
author = {Hegde, Dhruv and Balaji, Srivatsan},
institution = {NADIR Research},
type = {Technical Report},
number = {NADIR-TR-2026-02},
year = {2026},
month = jul,
url = {https://nadirai.net/research/nadir-tr-2026-02},
note = {Archived; synthetic evaluation}
}