Median position deviation
Reported at 120 m altitude in the campus trial; this is a median, not a maximum-error bound.
POSITIONING RESEARCH Campus-tested prototype · Engineering migration
ZhiTu VisNav. Position from what you see.
Map-referenced visual navigation using a down-facing camera and edge compute. A reported 740-frame campus trial achieved a median position deviation of 9.8 m at 120 m altitude, with an unoptimized prototype time of 0.23 s/frame.
Discuss a civilian OEM evaluation ↗Technical highlights
The capability, in detail.
Reported capabilities and figures from the supplied technical profile, with their application context.
Consecutive localization frames
Profile-reported closed-loop sequence; not 740 independent flights.
Processing time per frame
Unoptimized prototype measurement; not a released embedded-platform latency specification.
Reference-map area
Campus environment used for the reported evaluation.
Source: Technical Capabilities — External Presentation · 2026-08 · pp. 13. Source and scope ↓
Engineering overview
ZhiTu VisNav map-referenced visual navigation
ZhiTu VisNav compares onboard camera imagery with a preloaded satellite or aerial reference map to produce an absolute visual position observation. The concept combines inertial information, learned reference-image matching and a quality gate. The profile describes a real campus trial and ongoing migration to an embedded platform.
Principle & workflow
ZhiTu VisNav: a visual position observation
InputProvide the references
The prototype concept uses a down-facing camera, a preloaded satellite or aerial map and inertial information. Map and image conditions define the evaluation context.
ProcessAssess the observation
At a system level, reference-image matching produces a candidate observation and quality checks assess its usability. The result can be reported as usable or unavailable depending on the observation quality.
OutputReview position + quality
Evaluate the output against independent reference information, including position error, timing, false matches and intervals without a usable observation. Embedded engineering remains in progress.
System concept
Camera and reference map
A down-facing camera and edge-computing board use a preloaded satellite or aerial map as a reference for absolute visual positioning.
System concept
Complementary information sources
The public architecture combines inertial information, reference-image matching and quality checks to assess whether a visual position observation is usable.
Trial evidence
Quality as part of the result
The profile reports rejecting confident but incorrect position matches. Evaluation therefore includes false matches and availability alongside typical position error.
Additional reported figures
| Measure | Reported value | Context |
|---|---|---|
| Reported reacquisition after map exit | 100% | Observed in the described sequence, with no loss or divergence reported; the number of reacquisition events is not given. |
| Largest cited false-position error | 430 m | The quality gate reportedly rejected high-confidence false matches up to this deviation. |
| Reported false-match rejection | 100% | Reported for the described test; the number of false-match events and broader operating-domain statistics are not provided. |
Reported campus evaluation
| Item | Condition / observation |
|---|---|
| Environment | Real campus scene with approximately 1300 × 1300 m reference map |
| Altitude | 120 m for the cited median-error result |
| Sequence | 740 consecutive localization frames |
| Position deviation | 9.8 m median |
| Prototype timing | 0.23 s/frame, unoptimized |
| GPS use | First-frame initialization; the profile states GPS was not used for subsequent position estimates |
| Reference comparison | The profile shows GPS ground truth compared with visual estimates |
| Reacquisition and false matches | 100% figures are sequence-specific; event counts are not disclosed |
Engineering maturity
| Stage | Status described in the profile |
|---|---|
| Campus validation | Completed prototype sequence reported |
| Rockchip RK3588 migration | In progress |
| NPU quantization and C++ pipeline engineering | In progress |
| Real-time platform interface integration | In progress |
Applications & value
Where the engineering matters.
Application context
Civilian mapping-platform research
Evaluate a map-referenced position observation in controlled conditions when conventional positioning is unavailable or unreliable.
Application context
OEM feasibility assessment
Review map dependence, camera conditions, latency, typical error and incorrect matches before defining an integration program.
Application context
Reference-map suitability studies
Compare the intended operating environment with the reported campus scene and identify the additional evidence needed.
Who it's for
ZhiTu VisNav map-referenced visual navigation
Civilian OEM navigation and autonomy teams with controlled engineering evaluation capabilities.
Bring the application, representative inputs and acceptance criteria. We will connect the relevant technology and platform to a defined engineering deliverable.
What we start with
- Civilian use case and operating domain
- Reference-map quality, age and image conditions
- Independent ground truth and integration constraints
What the scope can include
- A scoped research evaluation plan
- Error distributions and failure-case review requirements
- An engineering integration and maturity assessment
Before we begin
Common questions
What was measured in the campus trial?
The supplied profile reports 740 consecutive localization frames over a roughly 1300 × 1300 m campus reference area, 9.8 m median position deviation at 120 m altitude and 0.23 s/frame for the unoptimized prototype.
Was GPS used throughout the position estimate?
The profile states that GPS was used for the first frame and not for subsequent position estimates. It separately shows GPS ground truth for comparison with visual estimates.
What do the 100% figures refer to?
They refer to reacquisition after leaving the map and rejection of false matches in the described sequence. Event counts and wider operating-domain statistics are not supplied, so the figures are not universal guarantees.
Is the embedded version already released?
The profile describes migration to Rockchip RK3588, NPU quantization, C++ engineering and interface integration as ongoing. The offering is a prototype evaluation rather than a claimed released production SDK.
The next step