Fully Independent Aerotriangulation of a Vexcel UltraCam Dragon 4.2 Block
A GNSS/IMU-assisted bundle adjustment of a full Vexcel UltraCam Dragon 4.2 five-head block — 350 exposures, 1,750 frames over Graz — computed independently from the raw post-processed trajectory and the Applanix event log. The manufacturer's adjusted exterior orientation is not used as an input. On ten surveyed points held out of the adjustment entirely, the solution returns a check-point RMS of 0.035 / 0.069 / 0.069 m (X / Y / Z; 3D 0.103 m), with an overall image residual of 1.31 px (0.98 px median) and solver-vs-instrument attitude agreement of 1.06 σ. The result is returned in the client's native UltraMap formats — a drop-in substitution requiring no AG2I software downstream.
1. Data
The full f50 Graz demonstration block: 350 exposure events across five heads (four oblique, one nadir), 1,750 frames. Inputs are the raw post-processed GNSS/IMU trajectory (SBET, 200 Hz), the vendor's Applanix per-event log (trigger times, GPS seconds-of-week), and the sensor package calibration — per-head eccentricities, lever arms, interior orientation, and mounting conventions. The manufacturer's adjusted exterior orientation was not used as input to the solution. Of 57 candidate surveyed points, 43 were retained after a documented quality-control review and designated 33 control / 10 check; the check points were withheld from the adjustment entirely. The Graz demonstration block is publicly distributed.
2. Method — instrument-first, weighted not overridden
The adjustment is built instrument-first. The post-processed trajectory enters as per-exposure observations retained at their stated accuracies: positions at each recovered trigger instant, and attitudes from the 200 Hz inertial solution, constrained per exposure at 0.010° (ω, φ) and 0.020° (κ). Image tie observations are generated automatically — pyramid feature matching with sub-pixel refinement by least-squares correlation — then quality-gated and decimated to preserve the designed relative weighting between image information and the instrument constraints. The solved unknowns are deliberately few: per-exposure body states bound to the instrument observations; one self-calibrated boresight per oblique head; and a single, explicitly-solved trajectory-to-survey datum translation. The instrument is not overridden — it is weighted; the adjustment refines exactly the terms the instruments cannot observe themselves.
3. Independent check-point accuracy
Ten surveyed points held out of the adjustment entirely, evaluated against the solution (adjusted − surveyed, RMS):
| Component | RMS (10 check points) | — |
|---|---|---|
| X | 0.035m | — |
| Y | 0.069m | — |
| Z | 0.069m | — |
| 3D | 0.103m | — |
4. Aerotriangulation quality
| Quantity | Value | — |
|---|---|---|
| Image-space RMS (overall / median) | 1.31 / 0.98px | — |
| Image residual by track — nadir / oblique / mixed | 0.72 / 1.51 / 1.91px med | — |
| Solver-vs-instrument attitude agreement | 1.06σ | — |
| Solved datum translation (dE / dN / dH) | −0.033 / +0.069 / −0.050m | explicit |
| Tie network | 220,399pts | — |
| Stereo model consistency (349 nadir models) | 0.38 px median · 84% < 1 px | — |
The 1.31 px global figure is a weighted blend, not a uniform noise floor, and the split is instructive. Nadir-to-nadir matching is same-geometry template correlation and resolves sub-pixel — a median of 0.72 px on nadir-only tracks. Oblique rays, by contrast, are matched across a ~45° perspective change, where foreshortening and occlusion make 1.5–2.3 px the honest measurement noise of cross-head correlation. The mixed nadir–oblique tracks span the widest geometry and therefore carry the largest residual (1.91 px median) — yet they are precisely what braces height: withholding them was tested and degraded vertical accuracy by roughly 30 %. The global residual is thus a genuine average of near-sub-pixel nadir work and honest oblique cross-head noise — the signature of an instrument-held solution that refuses to bend the block to flatter the pixels, rather than a defect in the imagery.
The image residual reflects this instrument-weighted formulation: the adjustment does not minimize image residual at the expense of the instrument observations, and geometric quality is verified independently at the ground (§3). Relative residual between paired nadir frames — measured over 349 consecutive models and 103,892 shared tie observations — has a median of 0.38 px with 84 % of observations below 1 px, confirming the block supports stereoscopic compilation without operationally significant y-parallax.
5. Ground control — automated seeding, human-verified
The ground-control measurement is not fully automated, and no such claim is made. Candidate image measurements for all points were generated by an automated picking system; every measurement was then reviewed by a human operator in AG2I's GCP Point Picker — correcting minor auto-picker errors, re-measuring where a pick was unsatisfactory, and removing points that failed quality control. Fourteen of the 57 candidates were excluded — for poor or ambiguous ground marks, thin ray geometry, or residual outlier behavior — in a documented review. The final control comprises 43 points carrying 446 verified image rays across all five heads.
6. Delivery — native UltraMap substitution
AG2I returns a five-file package formatted natively for the client's production environment; each file replaces or augments its counterpart in the standard UltraMap project structure, and no AG2I software is required downstream. The revised trajectory carries the solved corrections at the full 200 Hz rate, so every downstream product — direct georeferencing, LiDAR georeferencing, orthorectification — inherits the same refinement from a single substitution. Because the trajectory-to-survey datum translation is stated as an explicit parameter rather than silently absorbed, delivery is available in either the trajectory frame or the survey frame at the client's preference.
7. Report
The full aerotriangulation report (per-point residual listing, control configuration, and the GCP removal reference) and the UltraMap method note are available on request.
Study data: Vexcel UltraCam Dragon 4.2 demonstration block — Graz, Austria (publicly distributed); all survey values in meters. Contact: info@ag2i.ai.