Self-Calibration of Vehicle-Mounted Sensor Extrinsics via the Nonholonomic Constraint and Turn-Segment Association
NH-Calib is a targetless extrinsic-calibration framework for sensors mounted on wheeled vehicles. It uses the metric ego-motion measured by each sensor together with locally planar vehicle motion and the nonholonomic constraint (NHC), which approximates the lateral velocity at the rear-axle center as zero.
- Each sensor independently estimates Roll, Pitch, Yaw, and X from its own linear and angular velocity, without cross-sensor observation correspondence.
- Only the relative lateral offset Y requires inter-sensor association. NH-Calib associates common turning segments, integrates forward motion over each segment, and differences the two sensor integrals to cancel the unknown vehicle-forward-speed component.
- Z is not observable from planar rotation and horizontal translation alone and is outside the estimation scope.
flowchart LR
A["Per-sensor metric ego-motion<br/>linear and angular velocity"] --> B["Planar-motion axis alignment<br/>Roll and Pitch"]
A --> C["NHC lever-arm estimation<br/>Yaw and X"]
A --> D["Turn-segment extraction"]
D --> E["Common-segment association<br/>integral and difference"]
E --> F["Relative Y"]
B --> G["Observable 5-DoF extrinsics"]
C --> G
F --> G
H["Z<br/>unobservable under planar motion"] -. "outside scope" .-> G
The method is being evaluated on public LiDAR and radar datasets:
- A2D2
- RadarScenes
- Ford Multi-AV Seasonal Dataset
The evaluation covers absolute and relative calibration errors, motion-based baselines, registration-based baselines, and controlled temporal-offset analyses.
This repository accompanies an ongoing research paper. Source code, configurations, and evaluation scripts are being prepared for public release.
Citation information will be added after publication.
Automotive Intelligence Lab, Hanyang University
https://autolab.hanyang.ac.kr