GPS-Denied Navigation Field Acceptance at 120 m | aervimo
GPS-Denied Navigation Field Acceptance: Testing Positioning Deviation at 120 m Altitude
Direct Answer: Map-referenced visual navigation is a GPS-denied navigation approach in which a down-facing camera on the aircraft matches live imagery against a preloaded satellite or aerial reference map, producing absolute position observations without any external radio-frequency signal. In a profile-reported campus closed-loop trial, the ZhiTu VisNav prototype held continuous localization across 740 frames at 120 m altitude with a 9.8 m median position deviation, processed at 0.23 s per frame on an unoptimized pipeline. For an OEM or research team, the practical takeaway is that a 120 m field acceptance test should verify median error, continuity, reacquisition, and false-fix rejection — not a single headline number.
What a 120 m Acceptance Test Must Actually Measure
A single "accuracy" figure hides almost everything that matters in GNSS-denied operation. A defensible field acceptance protocol at 120 m altitude separates at least four measurements:
• Typical error — median or 50th-percentile deviation against independent reference truth, reported in metres, over a defined flight segment.
• Continuity — the number of consecutive frames for which the system produces a usable position estimate without manual intervention.
• Reacquisition — whether the system recovers a valid position after leaving the reference-map coverage, and how that recovery is counted.
• False-fix rejection — whether the quality gate catches confidently wrong positions, and the magnitude of the largest rejected error in metres.
This structure mirrors how practitioners already think about instrument verification elsewhere in surveying: ISO 17123 defines field procedures for checking geodetic instruments against a stated measurand rather than accepting a brochure value, and the same discipline applies to a visual positioning prototype.
Reported Campus Trial Results for ZhiTu VisNav
The ZhiTu VisNav map-referenced visual navigation prototype combines a down-facing monocular camera with a Rockchip RK3588 edge compute board. A deep-learning matcher compares each frame against the preloaded reference map, inertial propagation bridges the gaps between matches, and a dual quality gate screens the output. GNSS is used only to initialize the first frame; every subsequent estimate is produced without it.
The profile-reported campus results are summarized below. All figures are brand self-reported, from a single sequence, without independent verification.
| Acceptance dimension | Metric (unit) | Reported value | How to replicate in your test |
|---|---|---|---|
| Typical error | Median deviation at 120 m altitude (m) | 9.8 | Compare each frame against surveyed truth, report the 50th percentile |
| Continuity | Consecutive localized frames, closed loop (frames) | 740 | Count frames with a valid estimate until first dropout |
| Map-exit recovery | Reacquisition rate after leaving map coverage (%) | 100 | Fly out of coverage deliberately; note no event count is disclosed |
| False-fix rejection | Largest rejected false position (m) | 430 | Log gate rejections and their deviation from truth |
| Compute | Prototype processing time (s/frame) | 0.23 | Time the full pipeline on your target board, not a workstation |
Two caveats belong in any honest reading of this table. The 100% reacquisition figure is sequence-specific — the number of map-exit events behind it is not disclosed, so it should be treated as a qualitative signal rather than a statistic. And 0.23 s per frame is an unoptimized prototype figure; the ongoing RK3588 NPU quantization and C++ pipeline migration are expected to change it, in either direction, until measured on the final build.
Why Median, Not Mean, Is the Right Statistic Here
Visual matching against satellite imagery produces a heavy-tailed error distribution: most frames match cleanly, while a small number degrade under shadows, repetitive texture, or seasonal change between the map and the live view. A mean would let a handful of bad frames dominate the result; the 9.8 m median (profile-reported) instead describes what a typical frame delivers. This is also why the dual quality gate matters more than the matcher itself — the reported rejection of a false fix as large as 430 m shows the gate, not the matcher, is what keeps tail errors out of the navigation solution. The 9.8 m median at 120 m is an observation-level figure suitable for OEM feasibility assessment and research evaluation; it is not a precision-landing or safety-of-flight number, and the system holds no airworthiness certification.
Field Acceptance Test Checklist
Before flying, fix the protocol so results cannot be renegotiated afterwards:
1. Independent truth. Establish reference positions from a source the system under test never sees, and time-sync it with the onboard log.
2. Defined segment. State altitude (120 m), speed, lighting window, and the exact map tiles loaded — seasonal imagery age changes matching difficulty.
3. Continuity rule. Define in advance what counts as a dropout and whether the 740-frame closed-loop criterion must be met in a single flight.
4. Deliberate map exit. Plan several out-of-coverage excursions so reacquisition is measured over more than one event, addressing the undisclosed event count in the reported 100% figure — and repeat the segment with GNSS degraded, as in the 740-frame field trial report.
Alternative Routes and When They Fit
Map-referenced matching is one of several GPS-denied navigation options, and a fair acceptance process benchmarks against them. Visual-inertial odometry (VIO) needs no reference map but accumulates unbounded drift over long flights. LiDAR SLAM builds its own map in real time at the cost of sensor mass and power. UWB and 5G positioning deliver metre-level fixes but only where dedicated infrastructure or network coverage already exists — precisely what a denied environment removes. The reported campus profile of ZhiTu VisNav fits OEM and research teams evaluating absolute, drift-bounded positioning where a reference map can be preloaded and no infrastructure can be assumed.
FAQ
What median positioning deviation did ZhiTu VisNav report at 120 m altitude?
The profile-reported campus closed-loop trial recorded a 9.8 m median position deviation at 120 m altitude, measured across 740 consecutive localized frames. This is a brand self-reported figure from a single campus sequence, intended for engineering evaluation, not an independently verified or certified performance value.
Does ZhiTu VisNav need GPS to operate?
GNSS is used only to initialize the first frame of a flight. All subsequent position estimates come from matching down-facing camera imagery against a preloaded satellite or aerial reference map, fused with inertial propagation, so the system does not depend on any external radio-frequency signal during operation.
What happens when the aircraft flies out of the reference map coverage?
In the profile-reported campus sequence, the system reacquired a valid position after every map exit — a 100% reacquisition rate. The number of exit events behind that figure is not disclosed, so it should be read as a qualitative signal, and your own acceptance test should include several deliberate out-of-coverage excursions.
Is the 0.23 s per frame processing time real-time?
0.23 s per frame is the profile-reported timing of the unoptimized prototype pipeline. Migration to the Rockchip RK3588 NPU with model quantization and a C++ pipeline is still in progress, so the production timing is not yet established and should be re-measured on your target hardware.
Plan Your Own 120 m Acceptance Flight
If you are evaluating map-referenced visual positioning for an OEM platform or a research program, we can share the reported trial dataset structure and walk through a field acceptance protocol matched to your operating area. Contact the aervimo engineering team to scope a 120 m evaluation flight and review the raw frame logs behind the figures above.